Papers for
ai platform developers
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Code mode cuts tool use cost and speeds agents without extra decision layer
Beyond Plain MCP: Do Decision Models and Code Mode Make Tool Use Cheaper and Faster?
Abstract: Model Context Protocol (MCP) gateways manage tool catalogs for language model agents using two primary paradigms: code mode where agents write scripts against tool stubs and decision layers which are cheap routers that preselect tool subsets. Using a controlled harness across 12 servers, up to 1,958 distractor tools, and three open-weight agents including Qwen, Gemma, and Nemotron, we present a preregistered evaluation on a sealed test split to determine when each approach succeeds. Code mode substantially outperforms plain MCP, cutting token usage by 70 to 90 percent while significantly boosting utility. Declaring return-shape hints in code mode stubs is a first-order variable, swinging success by up to 39.5 points for lower-performing agents, an effect larger than adding a router. The central hypothesis that a decision layer in front of code mode beats code mode alone is not confirmed, showing minor, statistically non-significant improvements for weaker agents and a slight cost for the strongest. We publicly release our harness, a finetuned 4B tool selection model, and all evaluation traces to support robust community experimentation on local AI privacy and agent architectures.
Prism reduces tool permissions for AI agents while keeping tasks effective
PRISM: Causal Counterfactual Minimization of Agent Tool Privileges via Information Bottleneck Role Induction
Abstract: This paper shifts the object of role engineering from human users to large language model (LLM) agents, and proposes PRISM to address the over-authorization that arises when an agent obtains full tool privileges through the Model Context Protocol (MCP). PRISM formulates whether a tool is necessary for a task as an intervenable causal proposition, and applies structured interventions in a sandbox to learn a counterfactual necessity model. It then compresses trajectories into discrete latent variables through an information bottleneck, and determines the number of roles from the knee of the rate-distortion curve. Finally, it formalizes pruning as a bilevel Stackelberg game, in which a generative flow network samples diverse near-optimal solutions. On three public benchmarks, AgentDojo, $τ$-bench and ToolEmu, PRISM compresses the number of authorized tools per role by 78.3%, attains a task success rate of 78.2%, reaches a counterfactual necessity accuracy of 88.7%, and lowers the blast radius to 15.9%.
Stable training method improves large language model fine-tuning accuracy
BRAT: Bounded Residual-Aligned Training for Stable Multi-Pass LLM Post-Training
Abstract: Reinforcement learning for large language model post-training often reuses each batch of sampled rollouts for several gradient passes to amortize the cost of generation. Under such multi-pass reuse, existing objectives fail in different ways: unclipped updates collapse as the policy drifts away from the sampling policy, adding a negative baseline avoids collapse but reinforces every sampled response, and PPO-style ratio clipping cuts off gradients abruptly. We first ask why reuse is fragile. Measuring how well an update estimated from a small rollout group aligns with the KL-regularized target on held-out rollouts of the same prompts, we find that this alignment drops by nearly an order of magnitude from in-sample to held-out rollouts across three base models, suggesting that repeated passes largely overfit the sampled group. Motivated by this finding, we propose Bounded Residual-Aligned Training (BRAT), motivated by a proximal update toward a bounded reweighting of the sampling policy. BRAT assigns each token a gradient weight that decreases linearly with its probability ratio and vanishes at a trust-region boundary, and normalizes group-centered rewards by their maximum magnitude, keeping advantages zero-mean and bounded without extra hyperparameters. With four reuse passes, BRAT avoids training collapse, limits policy drift, and achieves higher average accuracy with lower seed variance on most benchmarks.
AI truth finding improves when accounting for shared errors among agents
Where Is the Truth When AI Agents Disagree? Dependence-Aware Belief Propagation for Unsupervised Truth Discovery (Extended Version)
Abstract: LLM agents often give conflicting answers to the same question, and many deployed agents are variants of one base model, so their errors are correlated. We study truth discovery among such agents: inferring the true answer and each agent's reliability without labels. Classical truth-discovery and crowdsourcing estimators reward an agreeing bloc and can fall below majority vote. We propose BP-TD, which adds per-agent decorrelation weights to mean-field variational inference on a one-coin Dawid-Skene model, so that agents sharing a base model count together as about one vote. Its first iteration is exactly a weighted vote; later iterations estimate agent reliability. With 13 local agents on MMLU, BP-TD beats majority vote without any labels (0.716 vs. 0.624 on 160 questions; 0.707 vs. 0.674 on 10,000) and ranks agents by their true accuracy, which classical estimators invert. The gain depends on the pool: over 2,852 sub-pools of four benchmarks, decorrelation helps when the largest correlated bloc is weaker than the other agents and hurts when it is the strongest, and we found no label-free statistic that tells the two cases apart. Labels change the picture quickly: with 20 labelled questions, selecting the single best agent is more accurate than every aggregator we tested on MMLU, ARC and TruthfulQA.
Agents often harmed before reading inputs pose new security risk
Step-Zero Attack: Agent Safety Before Observability
Abstract: Existing attacks and defenses for coding agents act once content enters the model context: injected text persuades the model, and guards inspect what the model reads. Yet to judge whether processing an unknown archive, network response, or image is safe, an agent often must process it first, and that step can do harm before any content reaches the model. We call this first processing Step Zero, and harm done by it a Step-Zero attack. Across seven harnesses and six models, each agent received only a path or URL to a crafted input, paired with a benign twin showing the same visible fields. Step Zero is where most agents are already unsafe: 32/42 harness-model pairs overran a 4 MiB disk budget, 26/42 before the model saw any member, and four of the six models did so in every round (84/84), so the damage precedes any model-side defense. Every twin stayed within budget, so the attacker's bytes alone cause the harm. On network and image inputs the risk moves to what the agent writes: agents stored the body under the name the input declared, letting the input decide what is written, and their checks faded once cues were removed. Model-free tests trace the harm to the processing primitive, so we defend it there: a gate that counts the bytes each call produces held every routed extraction within budget while agents completed 25/28 tasks. Vendors should pair such gates with safety-alignment training that teaches agents to inspect an input before processing it.
Topological method improves layer choice for steering large language models
TopoSteer: Topological Layer Selection for Activation Steering in LLMs
Abstract: Activation steering controls large language models at inference time by adding a vector to their internal representations. Most work asks how to build that vector, and far less attention has gone to where to add it, even though steering at a poorly chosen layer can be worse than not steering at all. Prior selectors choose a layer without ever steering the model to test it: they score how separable the target behavior is at that layer, and how consistently the individual examples point along the steering direction. We argue that what matters is not the separability of the two activation clouds, nor the alignment of individual displacements with their mean, but the extent of the cloud of per-example displacements that mean-difference steering averages. We propose a topology-based layer selection method, called TopoSteer, which scores a layer by the total H_0 persistence of its displacement cloud, computed as the length of the cloud's Euclidean minimum spanning tree. We show that the score focuses on the displacement cloud's relative geometry, while class separation affects it only through standardization. We also provide a theoretical explanation for our method by proving that, as example-level variation increases, TopoSteer and discriminability move in opposite directions, so the two criteria can rank layers in opposite ways. Across multiple language models and steering tasks, TopoSteer provides a principled alternative to discriminability- and separability-based layer selection and yields more effective steering without requiring held-out validation.
Web model safety needs more than just plausible future prediction
Beyond Plausible Prediction: Safety Sufficiency in Web World Models
Abstract: Web world models increasingly predict the consequences of candidate actions before agents act, yet a plausible predicted future does not necessarily support the correct safety decision. In a matched human evaluation of 3,000 predictions from ten frozen open models, 86.1% are judged semantically plausible. However, several models exhibit highly asymmetric authorisation behaviour, combining low false-safe execution with poor recall of clarifiable ASK cases and a strong preference for BLOCK. This motivates safety sufficiency: whether a predictive interface preserves and exposes the policy-decisive information required to distinguish actions that may be EXECUTED, require clarification through ASK, or should be BLOCKED. We introduce SafeWebWorld, a benchmark for evaluating safety sufficiency across five web-safety domains through action-conditioned futures, policy-decisive evidence, and policy-conditioned decisions. Under a controlled Security-Aligned Supervised Fine-Tuning protocol, evidence preservation can improve substantially while authorisation-sensitive decisions deteriorate. To determine whether this deterioration reflects loss of authorisation-sensitive score structure or a failure of the decision readout, we introduce Latent Authorisation Recovery (LAR). Frozen-score recalibration reveals substantial but backbone-dependent recovery, while paired recovery provides only a modest additional gain. Yet within-family and matched-authorisation evaluations remain poor, showing that aggregate recoverability does not imply robust fine-grained authorisation behaviour. Overall, plausible prediction, evidence preservation, and aggregate recovery are each insufficient on their own to establish robust authorisation.
TextSeal2 improves watermarking for large language model outputs
TextSeal2: Distortion-Free LLM Watermarking with Copula Mixing and Memoryless Races
Abstract: We present TextSeal2, a distortion-free watermark for large language models that keeps its signal under speculative decoding: tokens that replace rejected drafts remain watermarked with the same keys, and the acceptance rate is unchanged. It also preserves output diversity and adds no measurable inference latency. Its predecessor, TextSeal, restores diversity in the Gumbel-max scheme by routing each token to one of two keys via a coin flip. This is suboptimal because the detector averages the scores from both keys without knowing which key was used. TextSeal2 instead merges two or more keys into a single virtual key and couples it to fresh noise through AutoMark's Gaussian copula. Any number of keys, including publicly shared keys, can be merged this way without diluting detection. Under speculative decoding, TextSeal2 verifies keyed drafts with one more exponential race, run on the leftovers of the draft races. By memorylessness, this race reproduces block verification, so the output follows the target model exactly. We evaluate TextSeal2 on three LLMs. At equal output token diversity, TextSeal2 detects 8 to 22 percentage points more 100-token answers than TextSeal at a false-positive rate of $10^{-3}$, and is comparable to textGrain while requiring no post-logits optimization. Under speculative decoding, it detects 18 and 22 percentage points more 200-token answers than TextSeal and textGrain, respectively, at the same acceptance rate.
Large language models reduce unnecessary refusals while keeping safety intact
Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective
Abstract: Over-refusal is a critical side effect of safety alignment in large language models (LLMs), causing harmless but superficially risky queries to be mistakenly rejected. Despite the effectiveness of existing mitigation methods, why safety-aligned LLMs may still reject harmless queries under correct safety supervision remains insufficiently understood. To address this gap, we analyze over-refusal through the lens of representation and optimization. We make two observations: 1) the representations of harmful and over-refusal samples are visually distinguishable at certain intermediate layers but increasingly overlap in deeper layers; and 2) harmful-sample gradients exhibit high intra-class similarity and concentrate in a low-rank subspace. Moreover, they are more strongly aligned with refusal-target gradients than with correct helpful-target gradients on over-refusal prompts. This imbalanced gradient coupling can bias over-refusal samples toward refusal, while the reduced apparent representational separation in deeper layers may make it more difficult for the model to maintain a clear boundary between harmful and over-refusal samples. To this end, we propose GOAL (Gradient Orthogonalization with contrAstive Learning), a safety-alignment framework for mitigating over-refusal. Specifically, GOAL consists of two components: 1) to prevent imbalanced gradient coupling from biasing the optimization of over-refusal samples toward refusal, we orthogonalize over-refusal gradients against the low-rank refusal subspace extracted from harmful gradients; and 2) to explicitly enhance the separation between harmful and over-refusal representations, we introduce a contrastive learning objective. Extensive experiments across multiple models and benchmarks demonstrate that GOAL effectively mitigates over-refusal while preserving safety performance and general capabilities.
LLM agents improve recommendation systems mainly by tuning settings
AutoRSE: Evaluating LLM Agents as Recommendation Engineers for Self-Evolving Recommender Systems
Abstract: Recommender Systems Engineer Agents (RSE-Agents), which use LLM-based agents to autonomously improve recommendation algorithms, are emerging as a promising paradigm. However, their capabilities and limitations remain poorly understood due to the lack of a unified, standardized evaluation framework and systematic trajectory-level analysis. To address this gap, we introduce AutoRSE, an evaluation framework spanning six representative industrial-grade recommendation tasks, and conduct a large-scale empirical study of three frontier LLMs across three widely used agent harnesses. Our results show that RSE-Agents consistently enhance the performance of backbones on five of the six tasks, with an average relative improvement of 16.1%, indicating that they can already autonomously improve recommendation performance. However, trajectory-level analysis also reveals that current RSE-Agents function primarily as engineering optimizers rather than algorithmic improvers: their gains are largely driven by hyperparameter tuning, training strategies, and model ensembling, while genuine algorithmic innovations are rare. Since the largest gains in recommender systems often stem from algorithmic advances, this behavior underscores a fundamental limitation of current RSE-Agents. To probe this limitation, we perform idea-injection experiments and show that agents equipped with innovative ideas from frontier research or human experts achieve substantially larger improvements than those relying on self-generated ideas. These results suggest that idea generation is a central bottleneck of current RSE-Agents, highlights a key direction for the future recommendation system research. Our implementation is publicly available on GitHub: https://github.com/Ray-Tiann/AutoRSE.
Direct preference optimization produces sparse updates in language model training
On the Sparsity of Direct Preference Optimization: Weight Disentanglement in the NTK Regime
Abstract: Direct Preference Optimization (DPO) has been widely adopted for aligning language models with human preferences. Recent empirical work has revealed that DPO induces remarkably sparse parameter updates, yet no theoretical explanation for this phenomenon has been offered. We analyze DPO's optimization dynamics under a neural tangent kernel linearization and show that the logistic structure of the DPO loss gives rise to an implicit maximum-margin bias. Building on this observation, we introduce the concept of dual sparsity: the DPO update direction is a sparse linear combination (data-level sparsity, driven by support vectors) of sparse gradient feature vectors (parameter-level sparsity, arising from gradient cancellation between preferred and dispreferred responses). We formalize this structure in our main theorem, bound the number of support pairs by a margin-dependent spectral quantity, and present a framework relating dual sparsity to weight disentanglement---a condition favorable for model merging. Experiments on GPT-2 Small and Medium and on 7--9B instruction-tuned models (Llama-3-8B, Mistral-7B, Gemma-2-9B) with the Anthropic HH-RLHF and UltraFeedback datasets confirm both sources of sparsity, and experiments on GPT-2 show that DPO reduces cross-task interference by up to $35{\times}$ compared to supervised fine-tuning.
AI agent workflows improve PyTorch deployment on device accelerators
AIPC: Evidence-Guided Agent Workflows for PyTorch Project Deployment
Abstract: Deploying and optimizing a PyTorch inference project on a dedicated accelerator, especially an on-device NPU, requires more than using a framework device API as with CUDA. The project must be converted into hardware-runtime artifacts and adapted to the target toolchain and application interface. Differences in model formats, quantization flows, operator support, and toolchain interfaces create cross-layer problems spanning PyTorch/ONNX models, toolchains, runtimes, and device execution. LLM Agents can invoke tools to address such problems, but may drift because of insufficient domain knowledge, large configuration and repair spaces, or poor path selection. We propose AIPC (AI Porting & Conversion), an Agent method for deploying PyTorch inference projects. AIPC combines structured Skills with toolchain, runtime, and profiling evidence to drive cross-layer diagnosis and repair, organizing model adaptation, artifact building, on-target execution, accuracy repair, and performance optimization into one deployment flow. We evaluate AIPC on the on-device NPU path of Qualcomm QAIRT across vision, speech, LLM, and diffusion models. In two optimization cases, 6/7 Skill-arm Whisper-Small runs meet predefined accuracy and latency criteria versus 0/6 control runs, while YOLO-Worldv2 achieves a 2.791x end-to-end speedup versus 1.599x for the control. Under a pre-fixed deployment contract, Skill, tool permissions, and check conditions, some runs achieve zero in-run human intervention (0-HITL). We also examine Qwen3.8-27B as a locally deployable model for driving an Agent workflow.
Data optimization techniques enhance large language model training and usage
Data Optimization for LLMs: A Survey
Abstract: Large language models (LLMs) have attracted great attention in both academic and industrial communities in recent years. Numerous survey studies have covered various aspects of LLMs, such as learning strategies, applications, alignment, explainability, and evaluation. However, little headway has been made in reviewing data-related aspects for LLMs, despite data being recognized as a key factor contributing to the emergence of the LLMs era. A wide range of data optimization algorithms and techniques are employed in LLMs' training and application processes. These techniques often originate from diverse deep learning domains, with their theoretical inspirations and heuristic motivations appearing unrelated to each other. This study aims to develop a comprehensive framework for LLMs' data optimization techniques to enhance data utilization and efficiency. We first discuss data challenges in LLMs' training and inference. Second, we summarize data perception dimensions for evaluating LLMs' data properties. These dimensions primarily include quality, difficulty, diversity, and redundancy, laying a foundation for subsequent optimization. Third, existing optimization techniques are reviewed and categorized into four groups: data selection, augmentation, reweighting, and others. Fourth, we analyze interconnections between these categories. Finally, we identify key challenges and future directions for LLMs' data optimization compared with conventional deep learning approaches.
Ai skill combinations expose new security risks in agent workflows
Detecting Insecure Compositions of Clean Agent Skills via Capability-Flow Graph Analysis
Abstract: AI agents increasingly extend their capabilities by invoking third-party skills, where each skill encapsulates instructions, scripts, and resources for agents to execute on demand. This paper reveals that deploying many benign skills together introduces compositional security risks: multiple individually clean skills can compose into safety-violation chains, because the agent passes an upstream skill's output to a downstream skill's input across invocations. The risk arises not from any single skill, but from the agent's runtime orchestration, and thus cannot be captured by existing malicious skill detection methods. To this end, we develop KnotScryer to detect insecure skill compositions by treating skills as typed capability flows. In particular, we abstract each skill into a capability summary, and construct the Skill Capability-Flow Graph (SCFG), where edges denote cross-skill information flow. Moreover, we design a search policy that tracks protected or untrusted data along handoff edges to locate paths reaching insecure operations. Finally, a two-layer verification method replays the insecure information flow for reachability analysis and applies a task-anchored LLM chain audit to confirm the semantic plausibility of the identified violation chains. We deploy KnotScryer to analyze 13,598 clean skills from ClawHub, and validate the detected chains across three agent frameworks and three LLMs. KnotScryer identifies 612 insecure compositions, which are consistently exploited across agent frameworks. Meanwhile, under GLM-5.1, it outperforms tag-match and graph-only baselines by improving the exploitation rate from 16.5%-18.2% to 85.6%, while reducing the search space by 85.3%. Furthermore, 13.7% of the detected chains consist entirely of skills labeled safe by a single-skill scanner, showing that KnotScryer effectively identifies insecure compositions that per-skill auditing cannot.
Message-passing method improves language model quantization accuracy
MPQ: A Message-Passing View of Post-Training Quantization
Abstract: Post-training quantization (PTQ) is a key technique for efficiently deploying large language models, as it compresses weights into low-bit representations without retraining. Existing PTQ solvers, including GPTQ, LDLQ, and block coordinate descent, commit each weight block to a single codeword before its neighbors are considered, leaving near-optimal alternatives unexplored. In this work, we cast row-wise PTQ as posterior inference over discrete codebooks, in which existing solvers correspond to a zero-temperature, hard-decision limit retaining only the top codeword per block. We propose Message-Passing Quantization (MPQ), which preserves soft posterior information across blocks during inference and commits to hard codewords only at the output stage. MPQ first uses vector approximate message passing to compute approximate blockwise posteriors, then refines a hard-decision initialization by exact block coordinate descent. Across 2-bit lattice and 3- and 4-bit scalar quantization on dense language models from Llama, Qwen, and Mistral families, MPQ improves perplexity and zero-shot accuracy over strong hard-decision baselines, and composes favorably with preprocessing techniques such as activation-aware scaling and incoherence processing. Our implementation is available at https://github.com/poseidonchan/MPQ.
Reliable recovery methods improve AI tool use in changing conditions
Permission to Retry: A Study of Authorization and Recovery in AI Agent Workflows
Abstract: A durable agent can resume a consequential action while its authority changes and the action's external outcome remains uncertain. We study execution-time reliability of tool-using AI agents, isolating recovery decisions from language-model behavior. Checkpoint state, permission to dispatch, evidence of an external effect and permission to reconcile are distinct obligations. We present a finite conformance study that checks these obligations through histories scored against a provider effect ledger hidden from the recovering application. A prospectively frozen campaign executes 336 histories on native LangGraph and Temporal hosts, crossing 14 schedules, three provider contracts and four matched application policies. All 168 paired host cells have identical checked outcomes. Per host, retained-ledger reconciliation passes all 42 histories: 21 checked completions, 15 denials and six unknown outcomes with reservations retained. Cached retry, revalidation before retry and stable-key retry pass 30, 36 and 31 histories respectively. A separate 16-case component study of InfinityOS, an enterprise AI platform developed by Arya.ai, exercises exported authority, worker, recovery and settlement code over repository fakes. The results isolate the roles of current authority, provider contracts and retained observations in recovery, with safety, completion and unresolved obligations reported separately for the declared application profile.
Data voids let bad actors trick web searching AI models
Data-Void Exploits in Web-Searching LLMs
Abstract: Web-searching Large Language Models (LLMs) have changed how users seek online information: instead of sifting through ranked lists of links, users read synthesized answers. This shift creates a new LLM vulnerability: a data-void exploit. Disinformers publish their narratives in search spaces where relevant and reliable content is scarce, i.e. data voids, then rely on the LLM's web-search to retrieve and then use, cite, or list those pages in its responses to data-void seeking prompts. Unlike classical search, where users see ranking cues and competing results, LLMs can obscure when an answer relies on sparse, unverified, or adversarial sources. We study this threat by varying the user's prompt tone - skeptical, excited, or neutral - and LLM choice - ChatGPT, Gemini, or Sonar Pro - as well as three attacker controls. The attacker can use one of five known strategies to construct and fill a data void with their narrative, host their content at varying levels of visibility from well-established, indexed domains to less established ones, and include generative-optimization engine features thought to boost citing by web-searching LLMs. We find that data-void exploits can succeed with only a single attacker controlled page. Across all conditions, attacker-content appears in LLM responses 31% of the time. User controls offer limited protection: while skeptical prompts slightly reduce success and slightly increase flagging or warnings in LLM responses, we find no evidence of large effects of prompt tone or model choice on success rates, although smaller effects cannot be ruled out. Overall, our results establish data voids as a practical inference-time attack surface for manipulating web-searching LLMs through strategically placed web content and highlight the urgent need for defenses against such exploits.
Privacy controls adapt to autonomous ai agents on the web
Privacy Preference Controls for Agentic Systems
Abstract: In response to vast and widespread online tracking, regulators and technologists alike have worked to develop laws and tools that allow people to express their privacy preferences. These measures seek to offer users greater control over the sale and sharing of personal information, putting agency in their hands to protect their privacy. However, the web is undergoing a seismic shift, away from strictly \textit{user-driven} behavior towards \textit{agentic behavior}, where AI agents take autonomous actions on behalf of users. In this study, we ask: What should privacy preference controls look like in the agentic web and agentic systems more generally? To answer this question, we develop a taxonomy of enforceable privacy preferences for agentic systems. Our taxonomy is informed both by laws and empirical studies of existing privacy tools as well as an audit of 30 leading AI assistants and agent platforms. To explore how our taxonomy can be implemented, we design and develop eight open-source prototypes that mirror real-world agentic applications and demonstrate how major aspects of our taxonomy can be structurally integrated in agentic protocols, such as the Model Control Protocol and Agent2Agent protocol. Ultimately, our study seeks to provide the groundwork for ensuring end-user privacy and autonomy is not lost in the newly arriving agentic world.
Concept word effects on language model learning differ by part of speech
Every Part Has Its Role: Part-of-Speech Effects in LLM Post-Training
Abstract: Large language models are typically trained with next-token prediction (NTP), which rewards predicting specific tokens rather than explicitly accounting for the semantic similarities between different words or expressions. Prior work has addressed this limitation by augmenting training data with minimally varying sentences that replace a target word with conceptually related alternatives, e.g., "The tired coach ran home" $\rightarrow$ {"The tired trainer ran home", "The tired instructor ran home", ...}. Such concept augmentation improves performance, but it remains unclear whether its effectiveness depends on where in the sentence the concept is expressed. We reproduce and extend prior work $\textit{beyond nouns}$ to verbs and adjectives, asking whether the part of speech (POS) carrying the concept signal affects perplexity and downstream performance. We find different patterns across evaluation settings: concept augmentation improves perplexity most consistently on news, independent of the POS used, whereas verbs provide the most reliable downstream gains. Despite these behavioral differences, representational analyses using CKA and linear probing show that the conditions yielding the largest downstream gains are not necessarily those that most alter the model's internal representations. Together, these results suggest that the effectiveness of concept augmentation depends on where and how conceptual information is introduced during training.
GaP-MoE improves expert diversity in sparse models with gradient projection
GaP-MoE : Explicit Expert Specialization in Mixture-of-Experts via Orthogonal Gradient Accumulation and Projection
Abstract: Sparse Mixture-of-Experts (SMoE) models improve scalability by routing input tokens to only a subset of expert networks, but they often suffer from representation collapse, in which multiple experts converge toward similar functions and lose their intended diversity. Prior work has addressed this limitation either by improving router architectures or by introducing auxiliary loss terms that implicitly penalize similarity among experts. However, router-based approaches remain vulnerable to routing collapses, and loss-based approaches can conflict with the task's original objective during optimization. In this work, we propose GaP-MoE (Gradient Accumulation and Projection in MoE), which induces expert specialization by directly manipulating expert output gradients rather than modifying the loss function. At each training step, GaP-MoE stochastically selects one of the $TopK$ routed experts as a leader based on normalized routing scores, and projects the gradients of the remaining follower experts onto the orthogonal complement of the subspace spanned by the leader's recent gradients. This removes directional overlap between expert gradients without altering the task loss, allowing specialization to emerge as a direct consequence of the optimization dynamics rather than an externally imposed constraint. We evaluate GaP-MoE in both fine-tuning settings (BERT-based, on GLUE) and from-scratch pretraining settings (GPT-2-based, on WikiText-103, enwik8, and text8), and further confirm its compatibility as a drop-in module for existing SMoE architectures. Experimental results show that GaP-MoE achieves performance comparable to or better than standard SMoE.
Language models retain skills better when new knowledge is recast as old skills
Combating Catastrophic Forgetting in Continual Domain Adaptation via Knowledge Recasting
Abstract: Adapting language models to new domains during post-training risks degrading previously acquired capabilities, a phenomenon commonly known as catastrophic forgetting. Existing mitigation methods largely treat forgetting as a consequence of excessive model drift. They therefore constrain updates by preserving old outputs, replaying old data, or regularizing changes to parameters and representations so that the updated model remains close to its previous state. We take a different view: Rather than pulling the updated model back toward its old state, we translate the new domain into the model's pre-existing knowledge space, making unfamiliar knowledge learnable through familiar procedures. Therefore, we propose Skill Schema Transport (SST), a post-training framework that transports new knowledge into the model's existing skill space. SST abstracts each example type into a skill schema, preserving its essential operations, dependencies, and verification steps while removing domain-specific surface forms. By recasting new examples as instances of reusable procedures, SST turns knowledge update from trajectory memorization into procedural reuse. In continual domain adaptation settings, SST consistently reduces forgetting while improving generalization to related tasks. Across two 3B instruction-tuned backbones, SST improves average final performance on streamed domains by 2.53 to 5.75 points over sequential post-training. It also lowers max-drop forgetting in all evaluated settings, with the largest reduction from 42.41 to 7.10 points on the Qwen-3B MATH stream, while substantially recovering the OOD degradation induced by sequential updates.
Scientific language models improved for clearer research conversations
CollabSciReM: Aligning Language Models for Scientific Conversations via Science-Aware Reward Decomposition
Abstract: Large Language Models (LLMs) are increasingly used as research collaborators for hypothesis refinement, methodological planning, and technical analysis across multi-turn interactions. In this setting, a useful assistant must be accurate, methodologically justified, logically consistent, and token-efficient. Current alignment approaches rarely achieve all of these together. Instruction-tuned models gain conversational quality through verbosity, and we show that science-agnostic general collaborative rewards compress responses at the expense of scientific rigor. We propose CollabSciReM, a goal-conditioned reward framework that decomposes the session-level reward into conversational quality, scientific rigor, and structural efficiency. We also introduce SciQuality, a pairwise evaluation of multi-turn research conversations grounded in 300 held-out scientific papers and scored on six dimensions: methodological soundness, justification, logical consistency, factual correctness, novelty, and responsiveness. Fine-tuned on 8.62K synthetic research conversations, CollabSciReM is preferred on SciQuality over a science-agnostic reward baseline trained on the same data for all four Llama and Qwen models, including a 65.3% win rate at matched response length. For the 7B and 8B models, CollabSciReM outperforms the science-agnostic reward baseline on the primary metric in 8 of 10 single- and multi-turn benchmark settings. In a blind study, domain experts rated CollabSciReM significantly higher on novelty than the general-reward variant and reported fewer hallucinated responses than both the general-reward variant and the base model.
Arctic-SWE improves long-horizon AI training with model and system design
Arctic-SWE: Revisiting the Long Horizon Agent Training by Modeling-System Co-Design
Abstract: Long-horizon agent training, including software engineering, is advancing rapidly, but the recipe and infrastructure needed for a strong reinforcement-learning baseline remain difficult to reproduce. This makes gains from new algorithms or data hard to distinguish from improvements to an under-tuned baseline. We present Arctic-SWE, a fully asynchronous RL framework that establishes such a baseline through model-system co-design, using overfitting on a small task pool as the testbed. This testbed exposes failures that early reward gains conceal and guides the integration of established components: gated reward, importance-weight control, aligned training and inference engines, and consistent state across policy updates. No single component is new, yet together they form the missing foundation. On raw R2E-Gym data, without new algorithms or data work, Arctic-SWE trains Qwen3.5-4B to 62.0% pass@1 on SWE-bench Verified. This is comparable to concurrent FrogNano (61.5%), which relies on online task synthesis, after about one-sixth as many training rollouts. Building on this foundation, we further filter and refine R2E-Gym, SWE-rebench V2, and Scale-SWE. With this curated subset, the 4B model improves from 47.4%/25.0%/37.9% to 65.0%/38.3%/59.7% on SWE-bench Verified/Pro-v1/Pro-v2, and the recipe also improves Qwen3.5-9B from 56.2%/31.9%/48.8% to 65.7%/39.6%/70.2% on Verified/Pro-v1/Pro-v2. Controlled comparisons show that both curation and test strengthening improve held-out performance, although stronger tests reduce apparent training-task mastery. Arctic-SWE provides a documented starting point for studying better algorithms, data and environments for asynchronous agent RL.
SWE-Bench Pro V2 offers tougher and cleaner coding tasks for AI agents
SWE-Bench Pro V2: A Cleaner, Harder-to-Game Coding Benchmark
Abstract: SWE-Bench Pro is a widely used benchmark of long-horizon software engineering tasks for coding agents. As agents improve, they reach task issues earlier agents never touched and interact with the evaluation environment in ways the original harness did not anticipate. We present SWE-Bench Pro V2, a revised public split built to keep the benchmark accurate as capabilities advance. Agentic audits, expert adjudication and blind re-implementation repaired or retired 480 of the 731 tasks: 23 contracted engineers spent 404 hours, each working in a language they write professionally, and every revised instruction was verified by a second engineer solving the task from the text alone. A hardened protocol restricts network access to the model endpoint and re-grades every patch in a clean container, so agents cannot retrieve upstream fixes or get credit for changes outside their patch. On the 642 V2 tasks frontier agents resolve up to 99.4% while other systems stay clearly separated, and a 51-task Hard split spreads systems from 52.9% to 98.0%. We release the benchmark and distill lessons for maintaining agent benchmarks over time.
Safety monitor failures can hide evidence from llm decisions
Monitor-Validity Failures: LLM Safety Evidence Paths at Evolving Framework Boundaries
Abstract: Modern LLM applications rely on runtime monitors that inspect tool calls, traces, streams, detector outputs, and verifier inputs. This creates an interface-reliability problem: a framework path may remain operational while policy-relevant evidence never reaches its decision point. In one representative case, a regex detector returned a positive match, but the framework stranded the result before the block decision. This differs from jailbreaks and detector errors because the evidence is already recognized yet disconnected from enforcement. We study this repair-boundary signature with canary contracts: parent/fixed confirmation units that place a known payload only on the changed evidence surface and record monitor visibility together with the downstream decision for action-boundary cases. Applying the protocol to public framework repairs, we confirm 18 evidence-path failures across 11 LLM or agent safety project groups and four implementation ecosystems. Nine are established-boundary repairs, eight are new-surface policy-coverage gaps, and one is a boundary-only evidence case. Fourteen reach E1 enforcement boundaries: seven native action outcomes, six configured enforcement boundaries, and one synthetic-consumer consequence. The remaining cases affect verifier, safety-prompt, or observability channels. The cases form carrier, adapter, authority, and surface families and show how ordinary application behavior can remain intact while a safety-specific evidence path misses its decision. Exact revisions, deterministic outcomes, negative controls, and a frozen prospective audit document the confirmation rule. Monitor assumptions therefore need regression tests at the schemas, streams, tools, and adapters that frameworks evolve.
Switching optimizers during training does not harm large language model performance
The Ephemeral Cost of Switching Optimizers Across LLM Training Stages
Abstract: Recent work suggests that switching optimizers between pretraining and subsequent training stages incurs a performance cost, but systematic investigation of this phenomenon is lacking. We study the mismatch between AdamW and Muon optimizers by pretraining 1.2B Llama-3 models from scratch while varying the proportion of updates taken with each optimizer, and similarly interpolating between the two optimizers during continued pretraining and supervised fine-tuning (SFT). Surprisingly, we find little evidence of a meaningful optimizer mismatch penalty: final performance is largely explained by the independent effects of the pretraining and post-training optimizers, with downstream performance generally improving as the proportion of Muon updates during SFT increases. Although matching the pretraining optimizer can reduce forgetting during SFT, this does not translate into improved final benchmark performance. We further evaluate AdamW and Muon for task-general SFT of open-weight models, across the Llama and Qwen families up to 14B parameters, where Muon remains competitive with AdamW, despite these models having been pretrained with AdamW. Together, our results suggest that the choice of optimizer for post-training can be made largely independently of the optimizer used during pretraining.
Federated skill updates improve language model agents without sharing data
FedSkillGit: Federated Skill Evolution via Version Control
Abstract: Large language model agents can improve their capabilities by evolving reusable natural-language skills from execution experience. In real-world deployments, this experience is distributed across data owners who cannot pool their raw data, which motivates federated skill evolution. The natural first design lets each client evolve the whole skill document and upload the result for server-side aggregation. That design is poorly matched to skill evolution for three reasons. An evolved document can still carry client-specific content distilled from private traces, repeated transmission of a nearly unchanged document spends communication on knowledge that the population already holds and the contribution of a single client is buried in a document that the server cannot meaningfully attribute. We propose FedSkillGit, a federated skill evolution framework that replaces full-document updates with commits. Each participating client submits a bounded set of patches over named rules together with its local experience, which summarizes what the round taught it, and the pair passes a two-stage privacy filter before it leaves the client. The server deduplicates the submitted patches and the local experience of all participants, updates a global experience document that only the optimizers read, and merges the surviving patches into a candidate skill. The candidate is distributed to the clients, which score it on their own validation sets and report the results, and it is published only when the reported scores improve on the current version. Experiments on three datasets and three language model families show that FedSkillGit outperforms state-of-the-art baselines by up to 22% and approaches centralized evolution without sharing raw client data.
CoTranslator improves math reasoning training with efficient answer-based translation
CoTranslator: Reverse to Forward Chain-of-Thought Translator for Reasoning Distillation
Abstract: Reasoning distillation requires chain-of-thought (CoT) data with correct answers, but obtaining such data efficiently remains challenging. Forward CoT generation (FCG) followed by rejection sampling (RS) produces verified, question-only responses, yet may require many teacher calls for difficult problems. Reverse CoT generation (RCG) is cheaper because it conditions on the known answer, but access to this answer can lead to post-hoc rationalization and a mismatch with the student's answer-free inference setting. We propose CoTranslator, which learns to translate reverse CoT responses toward the distribution of verified forward responses. Trained on a small set of paired reverse and rejection-sampled forward responses, it supports large-scale generation and verification of student training data. Across six student models evaluated on six mathematics benchmarks, CoTranslator improves average accuracy over semantic suppression by up to $5.82$ percentage points. For two 4B students, it retains $95.0\%$ of rejection sampling's average accuracy while using $35.9\%$ of its estimated compute. Without retraining, the same translator transfers to a new mathematics dataset, a different teacher, and STEM problems, improving average STEM accuracy by $4.88$ and $4.31$ percentage points for 4B and 8B students, respectively. These results show that CoTranslator offers an efficient and transferable approach to reasoning distillation.
Llm agents improve tool choice using functional and contextual checks
From Relevance to Validity: Zero-Shot Tool Selection via Functional and Contextual Validation
Abstract: Large language model (LLM) agents increasingly face tasks from unseen domains and execution contexts, making zero-shot tool selection essential for identifying tools that can satisfy task requirements. Existing dominant methods mainly rely on relevance ranking, but semantic matching alone cannot ensure that a tool is functionally compatible or remains applicable in a new context. We preliminarily identify a \emph{transfer-validity gap}, where semantically relevant tools may still fail to satisfy functional or contextual requirements after transfer. To address this gap, we propose \methodname{}, a progressive validation framework that complements semantic relevance with endpoint-grounded functional compatibility and task-conditioned contextual evidence. Across cross-domain and cross-dataset transfer, \methodname{} consistently improves capability ranking, with larger gains under functional or contextual ambiguity. Ablation results further confirm the complementary contributions of functional and contextual validation.
Scoring format strongly affects language model data filtering quality
The Format is the Filter: Scoring Format Fixes What an LLM Judge Can Express
Abstract: When an LLM judge filters instruction-tuning data, its scoring format (binary verdict, 1--5 rating, 0--100 score, or comparison) is rarely justified or ablated. Holding the candidate pool, judge prompt, decoding and selection budget fixed, we score 15,406 UltraFeedback candidates under five formats with three judges, curate, fine-tune 66 models, and evaluate them with four judges. A 0--100 scale is worth only 7--12 equally used levels, two to three times a 1--5 scale rather than twenty, and a binary verdict leaves the top score tied in 77--78\% of prompts, so most selections are made at random. These properties are set by the format almost independently of the judge. Kendall's $τ_b$ hides them: it ranks binary best while $τ_a$ and top-1 agreement rank it last, because it discounts the ties a selector must break at random. Both comparative formats inherit a slot bias whose sign differs across judges. Downstream, curated data beat random selection under every evaluation judge, robust to length and markdown adjustment, but no ordering among formats is detectable. We propose four quantities that judge-based curation pipelines should report.
Efficient fusion of large language models and graph neural networks
DeKoR: Decomposed Knowledge Routing for Efficient GNN--LLM Fusion
Abstract: Large Language Models (LLMs) have shown strong potential for text-attributed graph learning by leveraging rich pre-trained knowledge and language understanding capabilities. Combining LLMs with Graph Neural Networks (GNNs) offers a promising paradigm for expanding the modeling capabilities of graph learning. However, LLMs and GNNs follow distinct learning paradigms, leading to a pronounced mismatch between their knowledge spaces. Such discrepancies make effective knowledge fusion nontrivial. Moreover, existing methods often incur substantial additional computation for only limited performance gains. To address these challenges, we propose DeKoR, an efficient framework for GNN and LLM knowledge fusion. To enable efficient knowledge transfer, DeKoR decomposes LLM knowledge into complementary local and global views and uses the LLM as an offline teacher to distill them into lightweight experts. To enable selective knowledge fusion, an adaptive router learns how much each node should rely on each expert. Extensive experiments on multiple benchmarks demonstrate that DeKoR outperforms existing baselines on both node and subgraph classification, achieving superior effectiveness and efficiency.
Watermarking technique strengthens federated learning against misuse
Watermarking for Byzantine Robust LLM Training
Abstract: LLM watermarking has emerged as a promising approach for detecting downstream model misuse in centralized settings, but extending it to federated learning (FL) presents unique challenges. In particular, FL systems often rely on strong robust aggregators to defend against poisoning attacks. We empirically show that a malicious server can exploit these same aggregators as effective watermark-removal mechanisms during training, causing representative existing LLM watermarking methods to fail detection in the final model. To address this limitation, we formalize the problem as a constrained optimization task and propose CamoMark, an aggregator-agnostic watermarking scheme for embedding robust watermarks in FL. CamoMark optimizes watermarked updates to become inliers, maintaining bounded deviation from the distribution of clean updates, while preserving sufficient watermark signal for reliable detection. Evaluated on Qwen fine-tuned on the Alpaca dataset, CamoMark achieves a detection confidence of $p$-value $= 1.82\times10^{-2}$ using only 3.3% watermarked data. The watermark also remains robust to a range of post-training removal attacks, including continued fine-tuning on clean data, pruning, and model distillation. At higher watermark ratios, detection confidence improves substantially, reaching $p$-value $= 5.11\times10^{-11}$ at a 13.3% watermark ratio.
Self evolving agents improve themselves by updating models and tools together
Evo$^2$: Evolving the Evolver through Weight-Harness Co-Evolution
Abstract: Self-evolving agents are increasingly used to improve language model systems without human intervention through either weight updates or harness improvements. Existing approaches typically pursue these directions separately: some methods update model weights using self-generated experience, while others rewrite the harness that defines how an agent uses prompts, tools, context, and execution feedback. However, these approaches overlook the relationship between model weights and the agent harness. To connect these processes and ensure effective and smooth agent self-evolving, we propose Evo$^2$, a framework for weight-harness co-evolution. For harness updates, we design a unified dual-role pipeline that a single policy serves as both a Proposer, which writes harness edits, and a Reflector, which diagnoses their effects and guides revision. For weight updates, harness modifications and their execution outcomes guide the training of both the Proposer and the Reflector. We further develop EvoTIDE, a training pipeline that recovers usable trajectories through edit extraction and prefix continuation, then combines SFT with interleaved GRPO across the two roles. Extensive experiments across three coding benchmarks demonstrate that Evo$^2$ significantly outperforms existing self-evolution methods. It achieves 82.1% on Polyglot and 48.3% on SWE-bench Verified, compared with 62.6% and 41.0% for the strongest baseline, respectively. We also show that the evolved harnesses transfer to different tasks without further evolution and rank first in cross-benchmark evaluations. Our dual-role harness improvement framework is also able to extend to self-evolution with proprietary models, improving a GPT-5-mini agent's performance from 24.4% to 48.3%. These results demonstrate the applicability of the framework Evo$^2$ across domains and its broad transferability across models and tasks.
Ai evaluation simulation shows impact of private benchmarks on satisfaction
The AI Evaluation Ecosystem
Abstract: AI evaluation shapes the decisions of model providers, users, funders, and regulators. We argue that designing valid benchmarks requires contextualizing design choices in the dynamics of this ecosystem of actors. We develop a simulation architecture that combines rule-based market dynamics with LLM-driven strategic actors, building on advances in Generative Agent-Based Modeling (GABM). We model benchmarks, consumer needs, and provider capabilities as vectors over a six-dimensional capability space (reasoning, coding, knowledge, safety, communication, agentic), with structural information partitions across actors. As a case study, we apply this stylized simulation to explore benchmark holdout design. We find that moving from public benchmarks to private holdout benchmarks shrinks the gap between benchmark scores and user satisfaction on most benchmarks but widens it on a few, depending on where holdout weights shift scoring credit. We stress-test our findings at both the instrument and case-study level, drawing on the V&V framework of Sargent (2013) and GABM-specific evidence criteria. Beyond holdout design, our simulation is a hypothesis-generating sandbox for studying how evaluator and policy choices, in turn, reshape the ecosystem.
Model merging works better when exploring beyond task weight combinations
A Broader Look at Model Merging: Rethinking Implicit Regularization Induced by Task Arithmetic
Abstract: Model merging aims to build a multi-task model cheaply by combining the weights of individual task-specific models. To perform well across multiple tasks, most existing merging methods use an additional dataset to find the coefficients for the best linear combination of task-specific weight updates. However, we identify an implicit regularization in this standard practice: searching over coefficients restricts the candidate models to a subspace spanned by task-specific weight updates. In this work, we investigate whether this regularization is actually useful. Surprisingly, empirical results show that optimizing merged-model weights without this regularization significantly boosts the performance of common merging methods across multiple architectures, domains, and even in an extremely data-limited scenario where only one instance is available per class. Moreover, directly optimizing the pretrained model weights even outperforms some existing merging methods. Analysis shows that better multi-task weights exist outside the subspace and can be found using multiple methods. We study different strategies for using the additional dataset, discussing their practical use and implications for model merging. Overall, this work calls for revisiting the existing model-merging pipeline, motivating a broader exploration of the weight space and a reconsideration of the implicit regularization induced by task arithmetic.
Visualerase improves safety by fully removing unwanted image concepts
VisualErase: Dual-Branch Visual Trajectory Redirection for Robust Concept Erasure in Text-to-Image Diffusion Models
Abstract: Concept erasure is essential for the safe deployment of text-to-image diffusion models, as they may reproduce harmful, copyrighted, or privacy-sensitive content learned from unconstrained large-scale data. Existing methods typically erase unwanted concepts while preserving general generation capability by redirecting target-related text-to-image mappings. However, recent studies show that erased models may still retain visual generative trajectories of target concepts, leaving them vulnerable to adversarial recovery attacks and revealing a fundamental gap between redirecting text-to-image mappings and truly removing visual knowledge. To bridge this gap, we propose VisualErase, a new paradigm that redirects concept-bearing visual generative trajectories toward explicitly defined concept-removed outcomes. To enable this redirection, we use structure-preserving image editing to construct content-aligned, concept-removed counterparts for source images, providing explicit visual endpoints that retain non-target content. We then derive a denoising target from each source-to-counterpart pair and use a dual-branch redirection loss to align both text-conditioned and unconditional predictions with this target, since conditional supervision alone does not explicitly constrain generation without textual guidance. To mitigate the adverse effects of concept erasure on non-target generation, we jointly optimize the redirection loss with a counterpart retention loss that matches denoising predictions from the frozen pretrained model. Across style, celebrity, and nudity erasure, VisualErase limits the maximum attack success rate over seven attacks to 0%, 8%, and 0.1%, respectively, while retaining general generation quality. These results highlight the importance of visual trajectory redirection for robust concept erasure beyond text-to-image mappings alone.
Agent harnesses pruned to cut token costs without losing power
SHarP: Saliency-based Pruning of Agent Harnesses
Abstract: Agent harnesses are systems that coordinate model calls, tool use, and task execution to help large language models complete complex tasks. To meet task requirements and address failures, these systems are often iteratively refined by amending and patching their instructions, tools, and workflows, continuously increasing harness complexity. It is therefore unclear whether some resulting harness modules are redundant, introducing substantial token overhead with little, if any, performance gain. Inspired by neural network pruning, in this paper, we study harness pruning as a means of striking a better balance between task performance and token cost. We propose SHarP (Saliency-based Harness Pruning), a simple yet effective pruning strategy based on the saliency of each harness module with respect to performance and efficiency. Specifically, we first identify tools, instructions, and supporting mechanisms as components that can be individually disabled. We then estimate the saliency of each module by ablating it and assessing its task performance and token cost relative to the full set of single-module ablations. Modules with the smallest contribution to performance or largest computational overhead are subsequently pruned. Our evaluation across various harnesses on held-out validation sets reveals a surprising finding: most harnesses that we studied are highly redundant and can maintain comparable performance and efficiency even after a substantial portion of their modules are pruned. Our pruning approach and empirical findings provide new perspectives on agent harness design and optimization.
Experimental setup details can reverse ranking of stale-data training methods
Probe the Harness: Setup Checks for Stale-Data RL Comparisons in Language Models
Abstract: Methods for training language models on stale samples are judged by comparisons against importance-corrected baselines. We show that details of the experimental harness can reverse the observed ranking of methods, and we introduce PTH (Probe The Harness), a set of checks that makes the harness visible. Our case is a comparison between SAN, a behaviour-free method, and truncated importance sampling (TIS) on verl and in a single-GPU trainer, in which SAN first finished ahead in both stacks. Four details of the harness changed this comparison: the PPO ratio was taken against the learner's own recomputed probabilities, the data seed did not reach the TIS arm, the replay queue reused its first batch for 33 updates, and two loss normalisers differed from their description. In each case the logged quantity looked consistent with a working setup, while the quantity that defines the comparison went unchecked. With the harness checked, TIS matches SAN on verl, and in the trainer TIS learns steadily while SAN keeps a margin. We contribute the signature of each detail and its effect on the comparison, reference results for TIS and uncorrected GRPO under sampler lag, and the PTH checklist.
How enrichment methods affect reinforcement learning with sparse rewards
Understanding Enrichment in Reinforcement Learning
Abstract: When rewards are sparse, reinforcement learning with verifiable rewards (RLVR) often uses hints or intermediate guidance to generate more successful rollouts. This enrichment biases policy-gradient updates unless corrected via importance weights, but existing methods omit correction or truncate importance weights in order to avoid the high variance of correction. It thus remains unclear what exactly is gained or lost in RLVR by correcting enriched rollouts. We show, mathematically, that omitted or truncated correction implicitly reweights the defined reward, and we decompose the resulting gradient error into scale, rotation, and variance to explain their distinct effects on learning. To make correction practical, we develop a novel sequential Monte Carlo (SMC) weight correction mechanism that, under stability and particle-order assumptions, tempers the exponential compounding of standard correction variance over the length of a sample to an additive accumulation. We then apply our analysis of enrichment to interpreting the results of a fine-tune of Qwen3-1.7B on a sparse band of OpenMathReasoning, establishing a concrete mechanism of how enrichment, both corrected and uncorrected, helps avoid collapse in sparse domains. Our main contribution is to understand, in general, how enrichment and correction can affect training, as opposed to claiming that either mode of operation is superior to unenriched RL.
Multimodal language models improve perception with spatial self-distillation
Where-OPD: Spatially Guided On-Policy Self-Distillation of MLLMs with Synthetic Scenes
Abstract: On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored. Recent approaches use privileged visual information, such as image crops corresponding to a question, to improve fine-grained perception, but their gains are confined to tasks that benefit from such visual zooming and require either human-annotated grounding data or external teacher models. We introduce a different form of on-policy self-distillation for MLLMs that provides the teacher with textual, spatially grounded guidance identifying the visual elements relevant to a query. We use procedurally generated scenes with automatically available object identities and spatial coordinates, enabling scalable and annotation-free post-training. The teacher uses this spatial guidance to locate and integrate evidence from multiple relevant image regions, while the student learns to reproduce the resulting behavior from the image and question alone. Our approach consistently improves performance on counting, document and chart understanding benchmarks across multiple models. Importantly, although post-training uses only synthetic scenes, the resulting improvements transfer to real-world perception benchmarks, yielding a 3.23-point gain in average performance across CVBench, V*, ZoomBench, BLINK, HR-Bench, and MME-RealWorld. These results show that spatially grounded privileged information can induce broader perceptual capabilities through on-policy self-distillation, enabling substantial synthetic-to-real transfer beyond the task and data distribution used for post-training. Project page: https://github.com/sirkosophia/Where-OPD
Llm agents vulnerable to attacker skill chains causing harmful actions
Chaining Skills to Hijack LLM Agents
Abstract: LLM agents use skills to improve performance on specialized tasks. To complete a user request, an agent may invoke several skills in sequence, allowing information produced under one skill to guide the next. Because skills may come from open-source repositories, this handoff can also carry attacker-controlled claims into later decisions. In this paper, we introduce APEX, which constructs and refines adversarial skill chains tailored to a user task and an attacker-selected action. The key insight is that an agent-written record of genuine task progress can carry a false claim of user approval across skills: an upstream skill induces the agent to create the record, and a downstream skill uses it to direct the attacker-selected action. Across four targeted-action families and six models on SkillsBench, the chains induce the selected action in 512 of 690 attempts (74.2%). On GPT-5.4, the full chain succeeds in 84.3% of attempts, compared with 17.4% when the workflow is merged into one skill. We further evaluate a prompting defense that asks the agent to check skill-produced files against the original request. On GPT-5.4, it lowers targeted-action success from 84.3% to 59.1%, while the verifier test-pass rate across 72 benign native-skill tasks falls from 86.7% to 56.3%. These results highlight the need for defenses that prevent attacker-directed actions while preserving legitimate task performance.
Language-model search improves by skipping repeated reasoning errors
CORE: Conflict-Oriented Reasoning Elimination for Verifiable Language-Model Search
Abstract: Test-time reasoning systems often respond to failure by restarting or revising the latest step, even when an earlier decision caused the error. We introduce CORE, a search controller that requests a certified conflict core from a verifier, backjumps to the latest decision in that core, and caches the conflict to avoid repeating it. Under sound verification, finite branching and depth, and exhaustive proposals, the uncapped search is complete and never prunes a valid solution. On 2,000 planted graph-coloring instances with matched proposals and an exact verifier, CORE reduces median verifier calls by 39.8% at 30 variables and 35.0% at 36 variables relative to chronological repair; caching further improves on backjumping alone. Across five reasoning tasks, CORE achieves 75.9% mean success with Qwen2.5-7B-Instruct and 84.2% with Qwen3-8B, compared with 72.5% and 81.8% for Tree of Thoughts. It also uses fewer verifier calls and generated tokens on both backbones. These results show the value of using certified failure explanations to direct language-model search.
Cache compression improves large language model memory use and accuracy
KV-Kaizen: Learning Context-Adaptive Cache Compression Choices
Abstract: As the context size of text processed with an LLM grows, the size of KV caches can outstrip the memory allocated for the original model weights. This impacts LLM throughput negatively, since decoding is memory-bound and decode cost grows with cache size. Recent work alleviates this bottleneck by discarding the least relevant tokens. Eviction introduces a tension, since a one-off decision to discard content may prove detrimental later. Instead, we focus on alternative choices that can lead to cache compression without evicting tokens. We achieve this by learning a selector that is able to produce, based on context, a per-layer cache configuration towards an overall compression budget. The selector operates along three axes: sharing one cache across layers (depth), caching at fewer bits (precision), or truncating the low-rank latent cache representations (rank). We call the resulting method KV-Kaizen, for the many small per-layer choices it compounds. We observe that these interventions taken independently and uniformly over all layers limit achievable compression because they degrade accuracy. Crucially, composing them locally and adaptively to the context can instead preserve accuracy while achieving large memory savings. At inference, the selector runs once, before pre-fill. In evaluations on instruction following and reasoning tasks, our selectors reach the Pareto frontier of accuracy against cache size, against learning-free and post-hoc baselines. On long-context tasks, KV-Kaizen improves on eviction and can be composed with it, reaching a 32x smaller decode-time cache on a 14B model while preserving accuracy. A 4x cache size reduction incurs no accuracy degradation from 7B parameters up, and a compressed model is more accurate than a smaller uncompressed one with the same cache size. Together, these findings support pre-training large models and compressing them only afterwards.
Asynchronous reinforcement learning tuned for faster large language model training
Where Does Staleness Accumulate? Pool Aware Effective Staleness Control for Asynchronous RL in LLM Post-Training
Abstract: Fully asynchronous reinforcement learning (RL) improves resource utilization in large language model post-training by overlapping rollout generation with policy optimization, but it also introduces policy lag as trajectories are generated and queued while the trainer continues to update. We study how this lag accumulates over a trajectory's lifetime and how it can be controlled without sacrificing the wall-clock benefits of asynchronous execution. We decompose trajectory staleness into Generation Staleness, accumulated before rollout completion, and Waiting Staleness, accumulated after a completed trajectory enters the pool. Motivated by this decomposition, we introduce PACE (Pool-Aware Control of Effective Staleness). PACE converts excess pool occupancy into an adaptive rejection budget and ranks completed trajectories using an effective-staleness score that combines Waiting Staleness with prefix-aware Generation Staleness. This avoids penalizing long or interrupted rollouts solely because they span multiple policy versions. In single-turn mathematical reasoning, PACE improves the six-benchmark average validation accuracy by 18.7\% over unfiltered asynchronous RL at the same wall-clock budget and matches synchronous RL performance with 47.1\% less GPU time. PACE also improves validation performance in multi-turn tool-integrated reasoning, outperforming both synchronous and unfiltered asynchronous RL. Further experiments with the mixture-of-experts model and an alternative RL algorithm support its applicability across model architectures and training algorithms.
Reward rate policy gradient improves efficiency in machine learning agents
Reward-rate Policy Gradient for Efficient Machine Learning Engineering Agents
Abstract: Traditional reinforcement learning (RL) techniques focus on maximizing expected cumulative reward, where each action assumes to take a constant unit of time. However, this assumption does not hold for agentic RL tasks such as machine learning engineering (MLE) agents, where actions involve data loading, feature engineering, and model training that take variable durations. Efficiency matters in modern agentic RL where actions are costly. To address this limitation, we adapt from continuous-time RL and Semi-Markov Decision Process (SMDP) formulation and propose Reward-rate Policy Gradient (RPG), where we focus on optimizing the reward rate -- the long-term reward per unit of time. RPG estimates the reward rate from off-policy samples, then charges each action for the time it consumes at that rate. We first conduct theoretical analysis in the bandit setting to establish that RPG approximates the optimal reward rate and empirically demonstrate it outperforms baselines while avoiding enumeration over the policy space, a known issue for an existing method. We then further apply RPG on a small language model (Qwen3.5-4B) with self-improvement loops and empirically show it obtains higher rewards within a fixed time budget than vanilla RL on MLE-Bench and NanoGPT, with a 19.2% and 85.7% margin, respectively. Our method provides a practical solution for optimizing performance under wait time considerations in modern agentic RL tasks, where actions interact with external environments and cost time.
Iterative merging improves combining multiple teacher models into one
No Pain, More Gain: Iterative Merging for Effective Multi-Teacher On-Policy Distillation
Abstract: Multi-teacher on-policy distillation (MOPD) combines independently developed domain teachers into a single student by distilling their predictions on student-generated samples. We study a setting where teachers share a reference model but undergo different post-training procedures, and find that MOPD can struggle to recover some teacher capabilities. Because distillation occurs on student-generated prefixes, the student initialization can strongly affect subsequent recovery. However, initial benchmark performance is not a reliable predictor of a good MOPD initialization. For example, merge initialization can start below SFT warm-up yet finish higher after MOPD. We further find that effective merging depends on both the relative teacher contributions and the overall merge scale, with some strong configurations lying outside the simplex of convex parameter averaging. Thus, selecting a good merge initialization requires evaluating not only its immediate performance but also the learning it enables under MOPD, making one-shot coefficient search difficult. We propose Iterative Merging for MOPD (IM-MOPD), which starts from a uniform merge and progressively adds task-vector increments for under-recovered domains during distillation. In a 5-domain setting, IM-MOPD achieves higher average normalized recovery than MOPD with either uniform merge initialization or SFT warm-up, showing that effective teacher contributions can be determined progressively during training.
Language model routing improves by using semantic probabilities
SeLMRoute: Probabilistic Semantic Evidence for Large Language Model Routing
Abstract: Large language model (LLM) routing aims to select the most suitable model for each incoming query. Most existing routers learn this decision directly from query embeddings, model representations, preference data, or clusters of similar examples. Such approaches can be effective, yet the representation used for routing rarely states what a query actually requires. We introduce SeLMRoute, a routing framework that separates the extraction of candidate-independent semantic evidence from the learning of candidate performance and the application of deployment objectives. A decision model first evaluates a set of interpretable questions about the query, such as its reasoning requirements and use of external knowledge, with each judgment retained as a probability distribution. The resulting probabilistic semantic state is used by a lightweight supervised router to estimate candidate model performance. Routing objectives are applied after performance estimation, which allows the same semantic state to support performance-oriented and cost-aware decisions. On the LLMRouterBench (15 datasets, 20 candidate models, 11,481 queries), SeLMRoute achieves an average accuracy of $72.08\% \pm 0.45$, while grouped five-fold out-of-fold evaluation reaches $72.64\%$, compared with $69.23\%$ for the strongest fixed candidate. The representation achieves the highest mean performance among the evaluated semantic, dense, lexical, and domain-level representations. In a separate 13-model performance-cost setting, SeLMRoute improves performance in all five grouped splits, with a mean PerfGain of $2.66\%$. Our code is available at https://github.com/Indigma-Innovations/SeLMRoute.
Language model internal states reveal harmful action patterns better than guard models
Agent Safety From Within: Detecting Harmful Trajectories from LLM Internal States
Abstract: Language model agents can now perform sophisticated sequences of actions via tools and harnesses, which has increased the scope of the damage they can cause. Guard models, however, are mainly built for content moderation and thus are not well-suited to detecting this agentic risk. To address this, we proceed by first conducting a representational analysis, then use the resulting insights to build a solution. In our analysis, we focus on two types of trajectory-level agentic harms: harmful content, which is expressed directly, and unsafe tool use, which depends on whether an action is consistent with the interaction that produced it. We investigate how open-source guard models represent these two types of harm and find that they are linearly readable inside the model, even though guard models predict no better than chance on pairs that differ only in the called tool's schema. The two harm types also follow nearly orthogonal internal directions, and neither reliably serves as a proxy for the other. These results motivate reading trajectory safety directly from internal states. We introduce TACIT, a readout of a frozen backbone's internal states that decodes no tokens. Trained on six trajectory-safety benchmarks, a linear probe raises mean macro-F1 from 62.3 for the strongest open guard to 80.7, and refined readouts reach 86.2. With each benchmark held out of training entirely, the refined readouts still lead the strongest guard (65.7 vs. 61.1). With the same backbone, training data and test split, the frozen readout is on par with full safety fine-tuning, and it improves the fine-tuned model further when applied on top. The probe trains about one millionth as many parameters as full fine-tuning in about a sixth of the time, and TACIT has the lowest latency of the guards we evaluate.
Fine-tuning language models to balance learning and generalization
SoFT: Soft Targets for Generalizable LLM Fine-Tuning
Abstract: Distillation enables student language models to acquire new capabilities from expert teachers. However, integrating knowledge from multi-teacher, multi-domain demonstrations into a single student remains challenging. We study supervised fine-tuning (SFT) in this setting, where students must acquire diverse capabilities while maintaining generalization beyond the training tasks. Our experiments reveal varying trade-offs between in-distribution learning and out-of-distribution generalization across SFT methods, motivating more explicit control over this balance. To this end, we propose soft-target fine-tuning (SoFT) to balance learning from teacher demonstrations with retaining the Base model's existing capabilities. SoFT sets a minimum target probability for each demonstrated token while making the smallest KL change to the Base distribution. The resulting objective couples learning from demonstrations with adaptively weighted regularization toward the Base model. We further use domain-specific gradient budgets to control this balance and determine a probability threshold for each trajectory. Experiments on mixed-domain reasoning and agentic tasks show that SoFT achieves the best overall performance among the compared methods, with improvements in both in-distribution capability acquisition and out-of-distribution generalization.
SCLATE platform improves testing and training of continual-learning AI agents
SCLATE: a Substrate for Continual-Learning Agent Training and Evaluation
Abstract: Continual-learning agents are systems of models, harnesses, and memory operating over long multi-session horizons. Evaluating and training them requires interleaving tasks with agent-side events such as session stop and start, crons, and memory consolidation. Yet existing benchmarks and training frameworks schedule only the benchmark's own events, leaving each benchmark and agent pair to build a custom scheduling loop. We present SCLATE, an execution substrate where benchmarks and unmodified agents each add their events to one open event scheduler through an adapter. A hybrid simulated clock runs these events on a shared timeline, flowing in real time while the agent works and skipping idle gaps, which compresses a month-long scenario into hours. SCLATE also serves as a rollout engine that runs any agent's harness and memory unmodified, recording the tokens and log probabilities of every model call through an in-container proxy. We port seven benchmarks to SCLATE and compare ten unmodified harness and memory configurations head to head on ten models. The comparison shows that an added memory system does not reliably beat the harness's native memory and that models differ widely in how they use the same harness and memory. We then post-train Qwen3.5-4B through unmodified harnesses and memory systems. The model learns to use both, reading 6.8x fewer file lines with a 16.7-point higher SWE-bench Verified pass rate, and writing richer memory records, while its held-out MetaClaw accuracy rises by up to 11.8 points.
Multilingual AI models reveal brain-like first-word language circuits
FTB Graph: Determining and Validating First-token Broadcasters and Language-Identity Head Circuits in Multilingual Language Models
Abstract: Large language models operating in multilingual contexts must resolve target response languages early in generation, yet the causal circuitry governing first-token language identity decisions remains poorly mapped. We present an end-to-end structural circuit analysis across six model architectures spanning four families: GPT-2, BLOOM-560M, Pythia-1B/2.8B, and Qwen2.5-1.5B Base/Instruct. Using Edge Attribution Patching (EAP) with FP16 active clamping, followed by exact activation patching verification with a 2,000-candidate-edge search ceiling, we extract directed acyclic graphs driving first-token language broadcasting. Across the standalone models, we observe deep or mid-to-deep broadcasting hubs, though the evidence is strongest for Pythia-2.8B and BLOOM-560M because GPT-2 and Pythia-1B leave few out-of-graph heads for comparison, while both Qwen2.5-1.5B variants invert the necessity check. Scaling from Pythia-1B to 2.8B expands node participation while maintaining a similar verified edge budget, producing sparser topology. The Qwen2.5-1.5B base and instruct circuits retain 84.7% Jaccard similarity, including the Layer 27 hub, indicating that first-token routing is largely established during pretraining and preserved by instruction tuning. Finally, EAP scores correlate weakly with exact patching deltas across most models, showing that linear gradient approximations can diverge from causal interventions in FP16 and motivating exact-patching verification for reliable circuit discovery.
Human agent collaboration improves when interfaces show how decisions form
The Interface Is Downstream: Designing the Terms of Human-Agent Collaboration
Abstract: Before an agent responds or acts, much of the experience has already been designed. Memory and retrieval shape what it notices. Evidence rules shape what it may claim. Permissions shape what it can do. Learning rules shape what it carries into the next encounter. The argument comes from Alicia, a personal agent I've built and used since January 2026. A fine-tuning pilot produced no defensible model-performance result. It exposed a provenance failure: Alicia repeated an interpretation from a retrieved synthesis, cited a source note credited by that synthesis, and left the synthesis out of the visible chain. In a model-blind review of thirty-three citations, one reviewer judged that the retrieved intermediary supplied the claim in sixteen relayed citations and part of it in four. Five of twelve citations to directly retrieved targets lacked support in the target excerpt supplied for review. These judgments remain unadjudicated, and the packet is not public. I call the shared setting a humorphic environment: a persistent computational setting that translates a human practice into software. The first Humorphism paper translated partnership. This paper translates the studio, the room where practice happens. The failure prompted an audit of attention, evidence, action, and learning, with a review artifact and available recourse for each. The test is whether the person can inspect and contest what shaped the teammate's behavior. The output is downstream. Correction, consent, and learning carry the collaborative interface back upstream.
Rollout efficiency improves reinforcement learning in large language models
Rollout Efficiency in Reinforcement Learning for Reasoning Large Language Models: A Taxonomy and Future Directions
Abstract: Reasoning-oriented reinforcement learning enables large language models to solve mathematical, coding, and other multi-step tasks, but shifts a substantial portion of the training cost to rollout, where trajectories are generated for policy updates. Efficient rollout mechanisms are therefore essential to reduce this cost while maintaining the freshness, consistency, and statistical validity of training data. This survey provides a systematic taxonomy of recent research on rollout efficiency for reasoning-oriented reinforcement learning, classifying existing approaches from both mechanism and bottleneck perspectives. Based on this taxonomy, we analyze how different technique families address distinct sources of rollout inefficiency, examine opportunities and potential conflicts for combining them, identify gaps in the evaluation and reporting of efficiency gains, and discuss open challenges and future research directions.
Model merging improves by adapting to task differences and depth
Beyond Uniform Subspaces: Spectrum-Aware and Depth-Adaptive Fusion for Multi-Task Model Merging
Abstract: Model merging aims to consolidate multiple task-specific models without access to extra training process. However, existing subspace-based methods largely rely on a uniform treatment of task updates, overlooking their intrinsic spectral and depth-wise heterogeneity. We identify two key deviations from this assumption: different tasks require different subspace capacity and exhibit different tolerance to spectral transformation, while subspace projection introduces depth-dependent distortion. Based on these observations, we propose SADA-Merging, a spectrum-aware and depth-adaptive framework for data-free model merging. SADA-Merging allocates task-specific subspace capacity according to spectral complexity, adapts spectral preservation according to task-wise plasticity, and applies depth-dependent anchoring to compensate for projection-induced distortion. This enables the fusion process to adapt to both the intrinsic geometry of each task and its sensitivity across network depth. SADA-Merging operates directly on task updates and is applicable to both full fine-tuning and LoRA settings. Extensive experiments demonstrate consistent improvements over existing data-free merging methods across different task scales and adaptation settings.
Security risks in autonomous ai systems need more thorough testing
Connecting the Dots in Agentic AI Security: A Cross-Dimensional Threat Taxonomy, Evaluation Maturity, and Open Challenges
Abstract: Agentic AI extends LLM security beyond generated content to persistent state, autonomous actions, tool use, and interactions with humans and other agents. Existing threat classifications often emphasize individual dimensions, obscuring connections among entry points, affected components, and security consequences. The known threat landscape also differs from the coverage demonstrated by empirical research. Through a structured review of 66 studies published from 2022 to 2026, we introduce T={S, B, P, A}, a cross-dimensional representation linking affected functional or system surfaces {S}, interaction or trust boundaries {B}, violated security properties {P}, and empirically examined architectures {A}. We analyze 22 artifact-backed red-teaming studies and 11 representative security benchmarks to characterize empirical coverage and evaluation maturity. Within the selected studies, evidence concentrates on prompt/reasoning, memory, and tool-mediated attacks, predominantly in single-agent settings. Persistent, Human--Agent, complex multi-agent, systemic, and long-horizon threats receive less coverage. These findings describe the selected corpus rather than establish gaps across all empirical research. Heterogeneous metrics, limited adaptive defense evaluation, architectural imbalance, and incomplete execution-state capture further constrain comparison and reproducibility. We derive 13 open research questions to guide more systematic, architecture-aware, and reproducible security evaluation of agentic AI.
Ai reviewer training risks causing loss of scientific judgment diversity
When AI Reviews Train AI Reviewers: Scientific-Judgment Collapse and Mitigation
Abstract: Large language models (LLMs) increasingly participate in scientific evaluation, both as automated reviewers and as assistants to human reviewers. As model-generated reviews enter public data and future training corpora, AI peer review can become recursive: later reviewers learn from judgments produced by earlier models. We study one step of this feedback loop in a controlled setting. Starting from Llama 3.1 8B, we first fine-tune a reviewer on official ICLR reviews from 2018--2023 and then train four successor models on ICLR 2024 data with systematically varied mixtures of official and model-generated reviews. Our study shows that introducing synthetic reviews compresses rating distributions and reduces both same-paper and corpus-level semantic diversity. We call this pattern $\textbf{scientific-judgment collapse}$. To mitigate this failure mode, we introduce $\textbf{TrustReviewer}$, an open-source LLM-based system for generating peer reviews of AI and machine learning papers. TrustReviewer intervenes at two complementary stages. For training-time prevention, we train the core reviewer in a single stage on a curated corpus designed to reduce low-quality and semantically degenerate supervision. For test-time correction, paired activation steering aims to further mitigate residual tendencies toward collapsed judgments without further training or additional expert annotation. Together, these results characterize a concrete risk of recursive reviewer training and provide practical interventions for preserving judgment diversity and improving recommendation alignment in AI-assisted scientific evaluation.
Chronicle creates repeatable tests for large language model agent failures
Chronicle: Cut-Point Replay for Regression Testing of LLM Agents
Abstract: Large language model responses are non-deterministic, so failures in LLM agents are hard to reproduce: a failure depends on inference that is not bitwise reproducible, on tools that read changing state, and on a multi-step trajectory that a re-run rarely repeats. Record-and-replay makes a run reproducible, but existing agent tooling records runs only to trace or score them, not to test a code change against them. We present Chronicle, which records an agent run at its non-deterministic boundaries as immutable envelopes and replays it from the record. Its central operation, cut-point replay, serves a chosen subset of boundaries from the record and executes the complementary subset live with new code, turning a recorded incident into a regression test that runs in continuous integration. On a benchmark of 6 recorded failures with simulated model boundaries, recording adds 23 μs per crossing (0.008% of an assumed 300 ms model call), full replay issues zero model calls and is bit-stable across 20 repetitions, and cut-point tests fail on faulty code and pass on guarded and benign changes for all 6 incidents. In a mutation study of the guarded tools, cut-point tests catch every mutant that lets the recorded unsafe action through, while a baseline that stubs every boundary, using the same assertion, catches none. Chronicle and the benchmark are publicly available at https://github.com/theagentplane/chronicle.
Chatbots vary widely in product advice and sources given
"If I Had to Buy Just ONE: Galaxy S26 Ultra": Auditing AI-Generated Product Recommendations
Abstract: Consumers increasingly use AI chatbots for advice on what to buy. With companies like OpenAI and Google monetising their AI through advertising, this raises difficult questions about the bias and impartiality of such advice. In response, we conduct an AI audit of popular chatbots using real commercial-advice queries. First, we curate a dataset of 2,528 real commercial-advice queries (ConsumerQ). Then, we evaluate 1,536 responses to product queries from popular AI chatbots: ChatGPT (chatbot and API), Google Gemini (chatbot and API), and Google Search (AI Overviews). We find that ChatGPT expresses a first-person product preference in 79% of product-recommending responses, compared with 7% for Gemini and 2% for AI Overviews, while the products recommended often change across repeated requests. Displayed sources vary strongly: for the same query, the ChatGPT and Gemini interfaces share only 5.4% of domains on average, with no domain in common in 76.7% of comparisons. APIs provide a different view from their corresponding interfaces, with mean domain overlaps of 12.0% for ChatGPT and 14.8% for Gemini, and also differ in the types and layers of source information they expose. Our findings show that neither isolated responses nor API observations can be assumed to represent the commercial advice consumers encounter. Independent audits of AI-mediated commercial advice should therefore account for repeated responses, consumer-facing conditions, and the source layer being observed.
Teacher network structure shapes how well students learn functions
Teacher Geometry Shapes Learnability in Teacher-Student Networks
Abstract: Teacher-student systems, in which a teacher neural network generates training labels so that a student neural network can learn to implement the same function, are widely used as an abstract setting to study learning. However, the structure of the teachers is often overlooked by assuming randomly-generated, normally-distributed parameters. This hides substantial variation in how learnable different teachers are. We formalize learnability as the success rate of converging to the global minimum, as a function of overparameterization, learning algorithm, student initialization distribution, and teacher geometry. We both identify an easy distribution that maximizes node dissimilarity and a hard distribution that minimizes it, and show that these two distributions induce markedly different success rates across a large range of settings and for different activation functions. To explain the gap, we study the loss landscape of small neural networks that contain two distinct kinds of suboptimal local minima, out-of-bounds (OOB) minima at the edge of the data distribution and interior minima within. Assuming infinite data and a fast readout layer, we analytically reduce the loss landscape of small networks to two dimensions, showing that the region of attraction of interior minima changes as a function of teacher structure. In larger networks, maximally dissimilar teachers induce more interior minima, while minimally dissimilar teachers induce more OOB minima. Motivated by these analyses, we show that differentially increasing the learning rate of the readout layer and decreasing the learning rate of the inner biases increases success rates. These findings provide an important step in narrowing the gap between the study of teacher-student networks and more structured functions that arise in practice.
NeoHorse-1 improves AI models by learning from their own use
NeoHorse-1: Towards Recursive Self-Improvement via Agentic Post-Training with Routing Harness
Abstract: Recursive self-improvement (RSI) requires a concrete mechanism through which an AI system observes its capabilities and converts that evidence into the next round of learning. We present NeoHorse-1, a family of agent-native models developed to explore this path through agentic post-training. Our system combines a heterogeneous model pool with intelligent routing, recording the predicted capability demand, selected service tier, and subsequent interaction for each user turn. These records are converted into training examples that preserve interleaved reasoning, tool calls, and harness context, and are admitted through structural validation, six-dimensional semantic evaluation, and subscene-level labeling. Routing signals organize supervised fine-tuning into a three-stage curriculum and extend to routing-guided on-policy distillation, where a teacher supervises student-generated responses under the same progression. Capability-guided allocation then converts evaluation feedback into the next training mixture, closing an evaluation-selection-update loop in which what the system learns to do shapes what it learns from next. Across eleven benchmarks covering harness-based agents, tool use, coding, and instruction following, post-training raises the macro-average from 58.94 to 64.87 at 4B and from 65.60 to 69.04 at 9B, substantially narrowing the aggregate gap between the post-trained 4B model and the 9B base model. NeoHorse-1 provides an initial prototype of this feedback-driven process and a path toward harness-mediated RSI across successive iterations.