ai software developers

Abstract: Online Policy Distillation (OPD) often outperforms offline methods such as Supervised Fine-Tuning (SFT) by training on trajectories sampled directly from the student. However, standard OPD trains the student to follow the teacher at every state of the trajectory, assuming that teacher guidance is always beneficial. We instead distinguish between what the teacher prefers and what actually helps the current student succeed. We introduce Advantage-Based On-Policy Distillation (AOPD), which weights teacher actions by their counterfactual advantage under the student. Teacher actions that improve the student's expected outcome are reinforced, while those that do not are downweighted. Directly computing these advantages requires trying alternative teacher actions at each step and running new rollouts from each. Instead, we show that the AOPD objective can be estimated using only standard student rollouts and their terminal rewards. Our estimator uses teacher importance weights on the student-sampled actions to provide unbiased advantage-weighted supervision, without any additional rollouts or reward evaluations. This allows AOPD to focus teacher supervision on actions that help the current student while preserving the rollout efficiency of standard OPD. Experiments on math, coding, and Medqa show gains over distillation and RL baselines in both external-teacher and self-distillation settings.

Abstract: Curriculum learning, which organizes training samples according to their learning priority, has been widely adopted in reinforcement learning to improve training efficiency. However, existing methods commonly construct curriculum units using predefined data properties and estimate their utility by simply aggregating a single proxy signal from recent training history.Such designs may group samples with different optimization effects and fail to anticipate how their utility evolves in relation to the policy. We propose Predictive Gradient-based Bandit (PGB) Curriculum, which first discovers optimization-aware learning units by clustering projected sample-level gradients. PGB then employs a lightweight temporal predictor to forecast the future utility of each unit and converts the predictions into a Boltzmann sampling distribution for adaptive data allocation. Experiments across multiple model scales and reasoning benchmarks show that PGB consistently outperforms standard GRPO, difficulty-based curricula, and history-based scheduling strategies, demonstrating improved data utilization and training efficiency.

Abstract: Does persistent memory accumulated from previous attempts help an LLM agent solve new problems, and does memory design affect that benefit? We study programmatic Chinese-character drawing across five agent conditions attempting the same 668 items. Accumulated memory produces modest aggregate gains despite substantial growth in the memory stores: free-form notes improve first-attempt success by 4.6 percentage points, while a design combining free-form memory with reusable code improves it by 2.8 points. The two designs do not differ significantly overall, but the code bank shows a suggestive advantage on later, harder items. A separate post-hoc comparison finds a much larger improvement when the code-bank agent receives privileged, target-specific structural supervision. The agent can therefore exploit highly informative structural guidance, although the privileged per-item information does not establish what could have been learned from experience alone. These results suggest that memory design may affect where accumulated experience becomes useful, even when its average benefit is modest.

Abstract: Agentic knowledge graph question answering(KGQA) relies on sequences of executable operations, but trajectory-level rewards obscure their individual contributions. Successful trajectories may contain ineffective operations, while failed trajectories may contain useful ones. To assess individual operations, we propose OpCredit, an operationaware policy-learning framework that combines downstream outcome credit with immediate transition credit: Branch Outcome Credit (BOC) reuses ordinary rollouts to compare terminal task rewards observed after different operations under matched navigation contexts; Frontier Transition Credit (FTC) measures operation-induced changes in answer-conditioned frontier potential. The two complementary signals jointly supervise the corresponding decision spans during training, while inference uses only the learned policy and standard KG interface. Experiments on ComplexWebQuestions, WebQSP, and GrailQA across three backbones demonstrate consistent improvements in KGQA performance.

Abstract: We study classification dynamics under small initialization and identify a pronounced non-monotone behavior of accuracy that we call the accuracy well: accuracy rises rapidly early in training, subsequently drops, and only much later recovers. We observe this phenomenon in deep linear networks, fully connected ReLU networks on MNIST, deep convolutional networks on CIFAR-10, and controlled synthetic tasks, across different losses and optimizers and in both training and test accuracy. We propose a three-stage explanation in terms of target-driven organization, a degeneracy bottleneck, and subsequent recruitment and recovery. We develop a rigorous theory of these dynamics for shallow and deep linear networks, showing in particular that high accuracy can emerge while the predictor and the change in loss remain asymptotically vanishing, before continued optimization causes accuracy to deteriorate. For a simple synthetic classification task, we prove an extreme instance in which accuracy asymptotically rises to 100%, collapses to 0, and later recovers permanently to 100%. Our results show that classification accuracy can reveal aspects of learning dynamics that are not captured by the training loss alone.

Abstract: Training language agents on long-horizon interactive tasks is challenging because sparse terminal rewards provide weak credit signals for intermediate actions. Existing group-based methods either assign a shared trajectory-level advantage or estimate step-level advantages from state groups, which may conflate the current action with subsequent decisions. We propose Prefix Tree Group Policy Optimization (PTGPO), a critic-free method that constructs prefix trees over trajectory suffixes within anchor-state groups. PTGPO assigns each tree edge a credit given by the difference between its child and parent empirical mean returns, capturing the conditional return change revealed by each additional action. These prefix-tree credits provide fine-grained learning signals without requiring a critic or additional rollouts. Experiments on interactive benchmarks ALFWorld, WebShop and Sokoban show that PTGPO consistently improves over RL baselines, demonstrating its effectiveness.

Abstract: Multimodal large language models (MLLMs) often struggle to identify and use geometric relations in diagrams. Recent methods address this challenge by converting geometric entities, relations, and constraints into explicit textual representations for the model to reason over. However, effective formalization is highly non-trivial: on Geometry3K, structure injection fixes 28 errors but introduces 13 new ones among 200 examples. Redundant relations can distract the model, while ambiguous references to diagram elements can lead it to apply constraints incorrectly. This suggests that the key challenge is not merely extracting more geometric facts, but organizing them into representations that support downstream reasoning. To fully exploit the power of formalization, we further propose GeoReform, a reflective formalization evolution framework that treats formalization as an optimizable policy rather than a fixed parser output. GeoReform executes the full reasoning pipeline, collects failed rollouts, diagnoses defects in the current representation, and mutates the policy to better select, ground, group, and present geometric entities, relations, constraints, and targets. On Geometry3K, GeoReform improves Qwen3VL-2B accuracy from 42.0\% to 56.0\%. Extensive experiments and analyses across geometry reasoning benchmarks demonstrate that effective formalization is crucial for improving multimodal geometry reasoning.

Abstract: Agent skills provide reusable guidance for code agents, but incomplete or unsuitable guidance can impair task execution. To reduce the manual effort of skill refinement, recent approaches use LLMs to generate revisions from execution feedback. However, grounding these revisions in explicit behavioral evidence remains challenging. To address this gap, we propose SkillMorph, a skill-evolution approach based on trajectory-guided fault localization in agent skills. Its core idea is to link execution evidence to specific skill contents before generating revisions. Specifically, SkillMorph compares failure and success evidence in abstracted trajectories across repeated runs and tasks, incorporating changes between evolution loops to identify suspicious actions. It then uses these suspicious actions to localize edit sites in the skills and generate corresponding revisions. Experiments on SWE-Skills-Bench and CannBot show that the skills evolved by SkillMorph consistently achieve higher trial-level accuracy and execution consistency than the original skills and those from four existing skill-evolution methods. We have also applied SkillMorph to automated kernel generation with an AI operator-development team, which has accepted 6 skill-revision pull requests.

Abstract: Visual latent reasoning compresses rendered derivations into compact intermediate states, reducing textual reasoning overhead. Existing approaches differ in how they represent these states: continuous methods avoid vocabulary constraints, whereas discrete methods improve accuracy through quantization into a finite codebook. Our analysis of representative continuous and discrete systems identifies two functional requirements: answers must rely on latent states, and those states must carry valid, problem-specific reasoning. Continuous latents influence answers despite collapsed reasoning content, whereas discrete latents retain recoverable intermediate reasoning that answer prediction largely bypasses. To address these challenges, we propose Continuous Anchored Latent Reasoning (CALR), which connects latent formation with answer use through functional anchoring. With reference latents from information-balanced compression, CALR couples latent-mediated answer supervision with derivation-level semantic anchoring: the former routes answer supervision through intermediate states, while the latter grounds their decoded content in problem-specific derivations. A parallel-to-autoregressive curriculum develops sequential reasoning by conditioning subsequent latent blocks on generated prefixes. Evaluations on five mathematical reasoning benchmarks across model families show substantial accuracy gains. Under matched budgets, CALR gains 26.0 percentage points over a comparable continuous latent reasoning method. Further analyses show that its latents support answer prediction and carry problem-specific intermediate reasoning.

Abstract: As multimodal large language models (MLLMs) reason for longer, attention to the initial visual input diminishes, weakening visual grounding. Visual memory reintroduces visual evidence during reasoning. We conduct a controlled analysis of visual memory along three axes: curation, organization, and access. We find that local evidence benefits from global context, compact latent representations balance accuracy and visual-context cost, and the utility of memory access depends on the reasoning state. Guided by these findings, we propose ReMAP (Reasoning-Time Memory-Augmented Perception), which couples two complementary latent memories: a static, question-conditioned Global memory that preserves scene and cross-image context, and a dynamic Local memory that uses this context as an anchor while selecting and re-encoding region-level evidence according to the current reasoning state. Both memories return compact latent tokens inserted into the reasoning sequence, and a reinforcement-learning access policy trained with branched rollouts decides when to continue reasoning or invoke Global or Local memory. On ten benchmark families, ReMAP outperforms prior visual-memory methods on all four multi-image benchmarks, exceeding the strongest prior results on MuirBench and MIMIC by 8.38 and 14.84 percentage points. Across four backbone families, enabling memory access improves over the same trained model with memory disabled, and on shared V*Bench, CV-Bench-2D, and MuirBench questions ReMAP reduces the visual tokens entering the reasoning sequence by 51.0-76.8% relative to the native-resolution backbone. Further analyses show that Global and Local memory form distinct yet complementary latent representations. Together, these components restore the perceptual cycle by letting the reasoning state trigger targeted visual retrieval, with the retrieved evidence guiding subsequent reasoning.

Abstract: AI agents based on foundation models (FMs) have demonstrated strong capabilities to perform complex open-ended tasks. However, they face some common challenges in practice: (a) agent behavior can deviate drastically even for semantically similar tasks, leading to catastrophically propagated errors; (b) high cost and latency due to FM calls, repeated in full whenever a task recurs with different inputs; (c) FMs' limited context and instruction following capability confine how well agents manage the ever-growing execution context and follow complex plans. We introduce $\textbf{A}$gent $\textbf{B}$ehavior as $\textbf{C}$ode $\textbf{Agent}$ (ABCAgent), which uses a symbolic program (e.g., Python code with potential neural functions) to fully specify the agent's behavior at runtime, with a powerful FM agent editing that program for flexibility. Behavior is thus specified without premature variable binding, and its execution is deterministic. We evaluate ABCAgent on six agent benchmarks, two of which we construct to test how well a derived program generalizes to variants of the task it was written for. ABCAgent matches a model-matched neural agent on GAIA and augmented GAIA, and surpasses it where robustness and long control flows matter: 98.3% against 97.3% on GSM-Symbolic ($p = 0.001$), 71.9% against 47.4% $\mathrm{Pass}^4$ on the telecom domain of $τ^2$-bench ($p = 0.0001$), and more records written correctly at every loop length on our control-flow-augmented WorkArena benchmark. For more parametric task families, ABCAgent is also significantly superior in efficiency. Without authoring a new program, ABCAgent solves 92.6% of GSM-Symbolic instances and 20.1% of augmented GAIA variants, which yields $5.2\times$ lower latency and $7.0\times$ lower cost on GSM-Symbolic, 19% lower cost on augmented GAIA, and $9.5\times$ lower agent latency on $τ^2$-telecom.

Abstract: A coding agent can earn a passing grade by fixing its code, or by deleting the test that exposes the bug. Detecting such reward hacking requires recognizing attempted shortcuts, including those that fail. We release 173,561 annotated multi-turn coding trajectories from Qwen3-8B and show that supervising shortcut behavior independently of exploit success substantially improves detection. We introduce HACKTRACE, a behavior-supervised monitor that reads the internal states the agent already computes while generating code. Reusing these states enables monitoring before a turn is complete, without additional language-model tokens or passes. Combining this evidence with static features of the final files achieves a mean per-problem AUC of 0.997 with 8 ms of monitoring overhead, improving both accuracy and latency over monitors that run the model again on an honesty question and answer. The same generation states also provide an inexpensive monitoring signal for reinforcement learning. With strong GRPO penalties, HACKTRACE reduces the cheating share of passing solutions from 82-91% to 1-5%, while retaining honest, correct solutions and maintaining high detection accuracy as the policy evolves. Our results show that both the supervision target and the source of monitoring evidence matter for turning accurate detection into a useful training signal.

Abstract: As agents evolve from single-tool systems into modular, composite architectures, skills are becoming an important mechanism for capability development and distribution. However, the academic community lacks a structured framework for systematically analyzing and evaluating skills. Drawing on information gain and behavioral constraint, we propose the four-dimensional MCRI Framework and operationalize it as MCRI-Eval, a large language model-based evaluation method. We evaluate MCRI-Eval using 63,812 public skills from the OpenClaw skill Hub, with 58,275 skill-conditioned model executions across BigCodeBench, BFCL-Fundamental, and Mind2Web. MCRI-Eval scores are positively associated with community popularity signals and achieve the highest downstream ranking agreement among the evaluated methods. MCRI-Eval also improves top-1 skill selection across all three benchmarks: compared with the strongest baseline on each benchmark, the skills selected by MCRI-Eval advance by 17.7, 22.8, and 19.6 percentile points in downstream performance rank on BigCodeBench, BFCL-Fundamental, and Mind2Web, respectively. These results indicate that MCRI-Eval provides a useful pre-execution signal for prioritizing promising skills before costly execution-based evaluation.

Abstract: Tackling complex real-world tasks can exceed the capabilities of a single large language model (LLM), motivating the use of multi-agent workflows that coordinate specialized agents to work together on these tasks. Recent methods train LLMs to construct better workflows from execution outcomes, but they optimize only the workflow generator, while the other agents that build or execute each workflow remain fixed even though every outcome depends on all of them. However, extending training beyond the generator is challenging: the agents are coupled, and a workflow's outcome is a single sparse score that cannot tell which agent causes a failure. We propose FloWright, which leverages the workflow as a harness to optimize workflows. By introducing a hierarchical, structure-aware reward paradigm, FloWright enables one role to self-evolve and two or more roles to co-evolve, with no additional models, labels, or executions. Considering the limitation that workflows are commonly trained and evaluated on data that a single agent can already handle, we further propose DataWright, an adaptive data hardening approach that converts existing datasets into workflow-level tasks with increased difficulty. Across document, slide, chart, code, math, and finance tasks, small open models trained with FloWright achieve improved performance by up to $+7.41\%$, with co-evolving ($+5.03\%$) more roles gaining more than optimizing one of them alone ($+2.83\%$). Our project page: https://xhguo7.github.io/FloWright/.

Abstract: Tool-using LLM agents remain vulnerable to indirect prompt injection because trusted instructions and untrusted observations share one context, allowing malicious content to steer consequential input-filtering defenses. Multi-path consensus defenses still leave a high attack success rate because they examine content or aggregated outputs rather than authorizing effects, especially for the within-tool attack, which preserves the intended tool but manipulates its arguments. Data-Flow Control such as CaMeL provides stronger guarantees, but incurs substantial time latency that limits practical deployment. We introduce ToolFence, which compiles a typed authorization blueprint before execution, enforces it through a deterministic monitor, and when the blueprint is incomplete asks a judge to grant new capabilities rather than adjudicate each concrete call. ToolFence provides two key advantages. First, its fine-grained provenance-aware authorization enables the system to distinguish user-authorized values from untrusted observations, effectively addressing the within-tool attack. Second, its deterministic fast path and capability-level runtime grants substantially reduce the frequency of expensive judge calls, improving runtime efficiency. On AgentDojo with Qwen3-max, ToolFence reduces overall ASR to near zero with only a 3.80 percentage-point clean-utility drop and practical runtime overhead.

Abstract: As LLM-based agents perform increasingly complex tasks, Agent Skills have emerged as a flexible mechanism for extending their capabilities. An Agent Skill packages task-specific instructions with executable components and auxiliary resources to provide specialized functionalities. However, the growing adoption of third-party Skills introduces a new supply-chain attack surface. Malicious Skills can embed harmful behaviors that abuse agent privileges and compromise the agent execution environment or accessible resources. Although recent LLM-based malicious Skill auditing approaches have achieved promising performance, they often rely on capable commercial LLMs. How to achieve effective auditing with compact, locally deployable LLMs in security-sensitive and resource-constrained settings remains largely unexplored. Our investigation reveals that compact LLMs struggle to identify malicious behaviors hidden in complex Skill packages. This difficulty arises from both the implicit nature of such behaviors and the limited reasoning capacity of compact LLMs. To address these challenges, we propose SKILLLITE, an evidence-guided agentic framework for malicious Skill detection. SKILLLITE effectively extracts security-relevant behaviors and infers the intended functionality from complex Skill packages. It then employs a compact LLM to assess the maliciousness of the Skill based on the observed behaviors and their functional context. Experiments show that SKILLLITE improves malicious Skill detection across different compact LLM backbones and outperforms existing representative auditing baselines. Its effectiveness generalizes to behaviorally confirmed in-the-wild malicious Skills. Meanwhile, SKILLLITE maintains a low inference latency, supporting its practical deployment.

Abstract: Fine-tuning-as-a-service lets users adapt a safety-aligned language model to their own data, but it also creates a harmful fine-tuning attack surface: a small amount of harmful data mixed into an otherwise benign fine-tuning set can degrade the model's alignment. Two recent alignment-stage defenses address this problem at different levels of the model. Vaccine improves the robustness of hidden embeddings to the representation shifts induced by harmful fine-tuning, whereas Booster simulates harmful weight updates and attenuates their effect during alignment. We investigate whether these mechanisms are complementary and propose VaccineBooster, a single alignment procedure that combines embedding perturbation and weight-level gradient attenuation within each training step. On Llama-2-7B aligned with BeaverTails and then attacked through poisoned fine-tuning, VaccineBooster achieves the lowest OpenAI moderation score among the compared defenses, 0.315, while a Booster-Only variant retains the highest post-attack refusal rate, 50%. Together with ablations over the embedding-perturbation and gradient-attenuation strengths, these results indicate a trade-off: embedding perturbation primarily reduces flagged harmful content, whereas gradient attenuation primarily preserves explicit refusal behavior. Because our evaluation uses ten prompts and a single unseeded run per configuration, we report this trade-off as an observed pattern rather than a statistically resolved effect. These results provide practical guidance for prioritizing content safety or refusal retention when aligned models are exposed to untrusted fine-tuning.

Abstract: Reinforcement learning (RL) has become a cornerstone for improving the reasoning capabilities of large language models (LLMs), but the need for on-policy data substantially limits training efficiency. Reusing off-policy data through importance sampling (IS) can improve efficiency but introduce considerable instability. Hence, algorithms such as PPO and GRPO widely adopt IS-ratio clipping to stabilize training. However, training stability and gradient estimate are mainly determined by the product of IS ratio and advantage. To further stabilize training, we propose ACPO, which clips the product of the IS ratio and the advantage, leading to more stable gradient estimates. We also establish a connection between ACPO and gradient clipping in policy mirror descent (PMD), which is a standard technique to stabilize optimization process, and prove the convergence of clipped-PMD under the standard RL setting. Experiments on widely used mathematical reasoning benchmarks show that ACPO consistently outperforms PPO and GRPO in both accuracy and training efficiency, delivering 4-6 percentage points gains on standard math benchmarks, with Qwen3-8B+PPO. Hence, ACPO is a practical and effective alternative to conventional IS-ratio clipping for RL post-training of LLMs.

Abstract: Post-training quantization (PTQ) enables efficient deployment of large vision-language models (LVLMs), but is typically calibrated on a small set while expected to generalize across diverse downstream tasks. Although recent PTQ methods for LVLMs incorporate sensitivity signals, they still minimize reconstruction loss with respect to the full-precision model, potentially over-preserving FP behavior and calibration-specific bias. Rather than treating quantization solely as an error to be minimized, we observe that it can also provide beneficial regularization for certain layers and modalities. Motivated by this observation, we propose Balanced Fitting, a quantization effect-based framework that balances precision and regularization beyond reconstruction-based optimization. By measuring layer- and component-wise quantization effects for weights, vision activations, and text activations, Balanced Fitting combines fine-grained fitting for sensitive components with coarser fitting to exploit potential regularization benefits. Experiments on multiple LVLMs show that our method consistently outperforms prior PTQ approaches under both weight-only and weight-activation quantization, while lower reconstruction loss does not reliably translate into better downstream performance. The source code is publicly available at https://github.com/kmc3661/BFQ

Abstract: As LLM agents are increasingly deployed as commercial services, protecting proprietary orchestration logic and tool-use policies is important. We consider agent repackaging: an adversary integrates a protected agent into its own application via API and presents it under its own identity. It may modify parts of execution to obscure the source. The owner typically has only black-box access to the repackaged service, so black-box ownership verification is essential. Agent watermarking embeds ownership evidence into agent behavior for later verification. An effective watermark should satisfy two requirements: losslessness, preserving original functionality, and robustness, keeping ownership evidence recoverable after partial modification of execution. Existing methods often embed signals into behavior selection or execution trajectories, intervene in normal decisions, and offer limited robustness to behavior modification. We propose LoRo-Mark, a provably lossless and robust agent watermarking mechanism. For losslessness, it isolates watermarking into a cryptographically authenticated forensic branch that remains inactive during normal execution and is activated only by owner-authorized requests. By reducing unauthorized branch activation to standard MAC security, LoRo-Mark formally guarantees performance preservation. For robustness, it redundantly distributes ownership information across forensic behavior sequences, enabling reliable recovery under partial behavior substitution and sequence truncation. Experiments across multiple LLM agents show zero degradation on normal tasks and reliable ownership verification under sequence modifications.

Abstract: An agent can complete a task while exhibiting undesirable behavior during execution. Developers need tests for the specific behaviors encountered in deployment, beyond fixed benchmark suites. We present TraceDance, an agent system that constructs targeted benchmarks from deployment traces for user-specified undesirable behaviors. For efficient construction, Anchor-and-Confirm combines programmable retrieval with candidate-level confirmation by a Flash large language model (LLM), while the Anchor Synthesis Loop generates and revises specifications for custom behaviors. The benchmarks use decision-point continuation to evaluate an LLM's next turn at a recorded decision point with a behavior-specific rubric, without a reference answer or environment replay. Experiments in coding and general tool use draw on 252,557 sessions and produce 107 benchmarks with 4,125 instances, fulfilling 95.3% of build-target requests. Both human annotators confirm the requested behavior in 84% of sampled instances, and the automated grader's agreement with human pass/fail judgments is comparable to that between the annotators. Nine frontier LLMs achieve a mean pass rate of only 26.7%, showing that they still struggle to respond appropriately at the evaluated decision points. Analysis across behavior-specific benchmarks further reveals weaknesses in how current LLMs behave as agents. By turning deployment problems into targeted benchmarks, TraceDance could serve as a key component of the recursive self-improvement (RSI) loop.

Abstract: Multimodal reasoning requires models to preserve visual evidence through long decision chains while selecting appropriate procedures across diverse scenarios and rules. When learning is guided only by terminal verifiers, reinforcement learning (RL) reveals whether a final answer is correct but not how it should be produced. The policy must therefore discover reusable reasoning procedures while learning to execute them, creating a program cold-start problem. Skills can externalize successful procedures, reduce repeated exploration, and provide inspectable guidance. However, a fixed Skill Bank assumes that this guidance remains compatible with an evolving policy, while updating Skills alone can leave their triggers, execution protocols, and demonstrations stale or mutually inconsistent. We introduce RLHARNESS, which organizes Skills, selection and execution protocols, few-shot demonstrations, and task contracts into a unified, versioned Harness and alternates Harness evolution with policy learning. An Exploration-Distillation Harness builds the initial Harness and version-aligned verified traces for SFT and DAPO I. After the first RL block, a Post-RL Reconstruction Harness rebuilds Skills, protocols, and demonstrations from fresh success-failure rollouts, and DAPO II adapts the policy to the reconstructed program. RLHARNESS improves Accuracy from 16.25%/27.50% to 62.00%/50.00% on MetroMap/TravelMap and raises F1 score from 37.13%/45.50% to 65.81%/65.51% on Fee-VL/Cancel-VL. All four tasks achieve their best results only after reconstruction and DAPO II, showing that an evolving Harness complements RL by continually updating the external program that the policy learns to execute.

Abstract: Multi-turn tool-use failures can hinge on a single model call, yet reward variation alone does not reveal which call would benefit from training. When rewards depend on later interactions, their variation can reflect downstream randomness rather than differences between the current actions. We introduce Critical-State RL to identify trainable states in multi-turn interactions. Given task-defined candidate calls and local rewards, the method assesses whether each reward captures the action's effect on task success and whether improvement over a reference policy is possible. It then uses nested sampling to separate action-dependent reward variation from continuation noise and optimizes the policy at the selected states using contextual-bandit training. Experiments on the Berkeley Function Calling Leaderboard (BFCL) v4 compare training at diagnostic-selected states with training at alternative states. For missing-function tasks, the diagnostic selects the response after the tool becomes available; for missing-argument tasks, it selects the response before the missing argument is supplied. Training the selected responses improves performance, including about 14 percentage points on the missing-function task, while training the alternatives leaves performance flat or worse. We further apply the recipe across models and tasks, including logged repeat-call avoidance and memory management.

Abstract: Reinforcement learning (RL) is an exciting concept as well as a remarkable success story worth sharing. However, RL builds on rather complex interactions between different objects that play out over several cycles. Such dynamics are often best explained with an easily accessible implementation. We present RLLBC-Lib, a carefully crafted code library with the goal of lowering the entry barrier for students and other learners of RL in the context of learning-based control. At its heart, RLLBC-Lib comprises a comprehensive library of tabular RL approaches to enforce a clear understanding of the theoretical foundations. A deep RL library follows the same design principles, underscoring the parallels between simple tabular and state-of-the-art deep RL approaches. Additionally, RLLBC-Lib provides a collection of implementations illustrating core RL principles and contrasting RL to other learning-based control approaches. Finally, RLLBC-Lib provides an ideal basis for creating programming assignments with automated grading.

Abstract: Masked diffusion language models (dLLMs) generate text by iteratively denoising masked tokens with bidirectional attention. Extending reasoning across generation chunks normally requires keeping earlier generated text in context. We ask whether a dLLM can instead continue reasoning after that text is cleared, using only a fixed-size carried state. We implement this state as a small number of register tokens: dedicated fixed-position tokens whose continuous hidden states are trained to carry reasoning progress across generation chunks. We post-train dLLMs to decode a chunk of text, clear it while preserving the register values, and continue decoding from the prompt and carried state. In our main comparisons on LLaDA and Dream, registers outperform discrete-text carry on every benchmark, with gains of up to 8.5 points on math and 19.5 points on code. Registers are especially effective for bounded code generation, where correct programs usually span several chunks. Finally, registers can be further refined with reinforcement learning on long-horizon reasoning tasks.

Abstract: LLM-based agents rely on heterogeneous interaction capabilities to accomplish complex tasks. Existing approaches often distribute these capabilities across multiple LoRA adapters, which increases adapter storage requirements and introduces routing overhead during inference. A single LoRA avoids this overhead, but learning from diverse agent trajectories under a fixed rank budget presents two challenges. First, trajectories with different interaction traces and parameter gradients can induce equivalent changes in decision distributions, causing repeated updates to overemphasize redundant behavioral changes. Second, an aggregated update may exceed the rank budget of the adapter, and approximating it in weight space can distort the decision changes that it is intended to produce. We propose BQ-LoRA, a low-rank adaptation framework that organizes trajectory updates through a local behavior quotient manifold. It contains two modules, i.e., behavior quotient balancing (BQB) and decision preserving compression (DPC). BQB constructs the quotient manifold from decision distributions and reweights trajectory update directions according to their local density in the quotient tangent space. DPC projects the balanced gradient onto the intrinsic fixed rank tangent space and refactorizes the resulting target by jointly controlling effective weight error and distortion of decision distributions. Experiments on AppWorld and BrowseComp-Plus compare BQ-LoRA with standard LoRA and recent low-rank adaptation methods, while separate ablations evaluate the complementary contributions of both components.