ai product developers

Abstract: Diffusion models achieve strong performance in text-to-image generation, but they can still produce unsafe content. Training-free safety guidance offers a practical inference-time alternative that intervenes during sampling without modifying model weights. However, many existing safeguards still rely on a coupled safety-control chain. In this chain, the same unsafe condition simultaneously specifies what should be avoided, indicates the current unsafe signal, constructs the intervention, and determines when that intervention remains active. This coupling can work when unsafe factors are explicit, but becomes less reliable once harmful and benign semantics are entangled in the current generation state, often leaving residual unsafe content or unnecessarily degrading benign-generation quality or prompt fidelity. We address this problem with Decoupled Safety Control (DSC), a plug-in safety-control algorithm for training-free safety guidance. DSC reformulates training-free safety guidance as an explicit state-dependent control process, so the executed intervention is determined from the current generation state rather than inherited directly from a single unsafe condition. Experiments on standard safety benchmarks show that DSC improves representative latent- and text-space safeguards, achieving stronger unsafe-content suppression while preserving benign-generation quality and maintaining efficient inference.

Abstract: Language agents rely on external harnesses to track state, organize workflows, and verify answers. Beyond providing tools and information, these harnesses supply control decisions about what to investigate, whether to revise, and when to stop. Training on successful harness-supported trajectories can improve task performance while leaving these decisions dependent on runtime intervention. We ask whether harness-supported experience can also teach the model to make these decisions, allowing the division of control to change as the model learns. We call this objective harness internalization: learning to assume specified control responsibilities while retaining task performance after the corresponding support is withdrawn. We introduce HARNESS ANNEALING TRAINING (HAT), which combines explicit control supervision with a curriculum over teacher trajectories collected under progressively weaker harnesses. Experiments with 9B and 35B models on SWE-QA and SWE-QA-Pro evaluate every checkpoint under four deployment harnesses. Selected annealed checkpoints operating with tools alone achieve scores close to those of their respective starting checkpoints deployed with the full harness. The benefits vary with model scale and deployment configuration, and further annealing does not uniformly improve performance. These findings suggest that harness-supported experience can help reduce the runtime control required by a trained agent.

Abstract: Self-training generative models - the continued improvement of a model using its own outputs - is becoming increasingly important as high-quality training data becomes scarce. However, naively finetuning on model-generated samples leads to degradation through model collapse and the model autophagy disorder. Negative-guidance self-training methods turn this degradation into a useful signal, using a model finetuned on its own outputs to guide the original model toward improved generation. Existing methods, however, take the negative signal in standard model outputs as given. We instead ask whether this signal can be explicitly strengthened. We introduce Geometrically Modified Outputs (GMOs), which reweight the singular values of the generator's input-output Jacobian to increase the influence of its leading singular directions. This geometric modification amplifies the mode-seeking behavior and distortions of standard outputs, providing a stronger and more targeted negative signal for self-training. Across a range of one-step generative models, GMOs consistently improve the performance of negative-guidance methods, including Neon and SIMS, compared with using standard model outputs.

Abstract: Recent representation alignment (REPA) methods accelerate diffusion transformer training by aligning projections of the transformer's hidden states with representations from pretrained visual encoders. In this work, we explore a reverse and complementary direction to REPA: rather than projecting diffusion representations into the encoder's space, we inject encoder representations into the diffusion transformer, allowing them to actively participate in the denoising process. To this end, we introduce \textit{REPresentation Injection} (REPI), a training framework based on a scaffold-to-internalization strategy, in which projected encoder representations initially serve as a temporary scaffold and are then progressively internalized by the diffusion transformer. REPI outperforms REPA across a wide range of backbones and is highly complementary to it: combining the two yields substantial gains over either alone. Notably, with only 160K training steps, REPI + REPA matches vanilla SiT trained for 7M steps, a speedup of over $43.5\times$. Code will be available at https://jeneveuxpas.github.io/REPI

Abstract: Decoding for large language models is typically treated as a collection of isolated sampling strategies, with limited theoretical understanding of the behaviours they induce and how their underlying objectives relate. We formulate decoding as an optimisation problem over next-token distributions on the probability simplex, balancing expected model score against regularisation under support constraints. This view recovers familiar decoding methods through choices of regularisers and support constraints; more importantly, it enables new decoders to be constructed by composing distributional preferences within a single optimisation problem without external rewards, learned critics, or model parameter updates. We introduce CompoSimplex, a library with configurable support rules, regularisation primitives, and simplex solvers for constructing and evaluating compositional decoders. We evaluate standard samplers, individual regularisers, and compositions across multiple models and reasoning tasks. Our results show that compositions can realise trade-offs between single-sample quality, multi-sample quality, and diversity that are not attained by individual decoding objectives.

Abstract: Recent studies suggest that large language models encode emotion concepts as structured internal representations, but most existing work focuses on text and a single architecture. Therefore, we ask, do emotion concepts generalize across sources, modalities, and architectures in vision--language models (VLMs)? To address this, we construct CMES (Cross-Modal Emotion Stimuli), a multi-source collection of emotion-conditioned stories, real facial expressions, synthetic portraits, and synthetic emotion-evoking scenes. For each stimulus source, we extract a separate set of six Ekman emotion vectors from each of three VLMs. We report four main findings as follows: 1) Image-derived emotion vectors form a low-dimensional geometry similar to that of text-derived vectors. Valence is relatively stable across sources, while arousal varies more. 2) Text- and image-derived emotion vectors have modest cosine similarity but still show held-out cross-modal correspondence. Text-derived vectors can also steer image interpretation. 3) Cross-architecture correspondence remains even when native cosine is near zero. Transformations estimated from generic ImageNet activations recover both correspondence and causal transfer without using the six emotion vectors or their labels. 4) After aligning representations across architectures, we construct a shared emotion subspace that preserves affective geometry and selective steering effects. The corresponding consensus emotion vectors also generalize to a held-out fourth architecture at two model sizes. These results suggest that emotion representations can share relational structure and causal effects across sources, modalities, and architectures, even when individual vector directions differ.

Abstract: Fine-tuning pre-trained models on specialized tasks with scarce data is central to modern deep learning. Despite its empirical success, theoretical understanding of fine-tuning remains limited. We introduce a Gaussian multi-index setting to study fine-tuning from pre-trained weights, where the teacher network has $m+1$ features, $m$ of which are learned during pre-training and one of which must be learned during fine-tuning. For two-layer ReLU networks, we show that two-timescale training, i.e., updating the outer weights infinitely faster than the hidden ones, learns the new task-specific feature while preserving the pre-trained ones in the model representation. Moreover, only $\mathcal{O}(d)$ fine-tuning samples are required for this recovery, independently of the number of pre-trained features. In contrast, with random initialization, the same number of samples is insufficient to recover the target parameters. Our results therefore demonstrate that pre-training can induce an implicit bias with a clear statistical advantage over random initialization, enabling feature learning from scarce fine-tuning data.

Abstract: Sampling latency compounds in diffusion workflows, where users generate and discard many candidates before keeping one. Surprisingly, the poor outputs of standard few-step samplers do not reflect a lack of reconstruction capacity: by optimizing only the initial noise, frozen 3-4-step samplers can closely reproduce their corresponding full-step outputs. Building on this finding, we learn corrections to the initial noise and denoising updates using endpoint supervision, improving correspondence with full-step outputs generated from the same noise and prompt. The resulting previews allow users to screen candidates cheaply and reserve full-step generation for promising ones. Input correction also transfers across sampling budgets without retraining. Experiments show substantial improvements in reference fidelity, including 53-78% lower reconstruction MSE than retrained LD3 on unconditional benchmarks, alongside improved ranking preservation and candidate selection on SD1.5, SDXL, and FLUX.1-dev.

Abstract: Concept unlearning in text-to-image diffusion models aims to suppress a target concept (e.g., \texttt{horse}) while preserving related but distinct content (e.g., \texttt{donkey}), yet existing methods either leak under indirect prompts or visibly degrade other concepts. We show that these failure modes stem from the geometry of concept representations rather than from any particular algorithm. Formalizing concepts as activation-space regions, we prove that the overlap between a target and other concepts lower-bounds the damage any robust erasure must inflict on them, with the trade-off scaling linearly in the degree of overlap. Across thirteen unlearning methods, including methods designed to preserve non-target concepts, no method achieves both strong erasure and strong neighbor preservation: STEREO nearly eliminates indirect leakage but cuts neighbor generation by more than 75\%, while sparse inference-time methods preserve neighbors but leak. Damage increases with our overlap measure, monotonically so for STEREO; the $κ$-scaling reproduces on SDXL, and neighbor-selective damage recurs on FLUX. Perfect unlearning is the wrong target for entangled concepts; methods should be evaluated on the Pareto frontier our theorem establishes.

Abstract: Recent research argues that language models exhibit pronounced homogeneity in open-ended generation, framing such behavior as an Artificial Hivemind that poses a long-term threat to human creativity. We examine three of its central results. First, the flagship example is that model responses to "Write a metaphor involving time" collapse into two clusters. Visualization, spectral analysis, clustering, and language model labels all contradict this description. The labels record each response's vehicle, what it compares time to. Our responses and the original authors' own show one dominant vehicle plus a heavy tail of distinct minority vehicles. "Time" is one of our least diverse topics, so the example is a favorable case, not a representative one. Second, the paper measures homogeneity against an undemanding null: responses to unrelated prompts. Under a more demanding null (same-prompt responses expressing genuinely different ideas), 20%-32% of such pairs already exceed the paper's 0.8 convergence threshold. A residual effect survives this null. The paper's same-prompt pairs exceed 0.8 roughly two to three times as often as our different-idea pairs. Much of what the paper calls homogeneity is the shared geometry of answering the same prompt. The remaining measurements lack any null: no human baseline is collected, and the model-indistinguishability statistic has no null. Third, the paper concludes that inference-time interventions are inadequate for combating the Artificial Hivemind, writing that "more generalizable solutions are needed at the model training level." We show that this conclusion is unsupported in three ways, and that an inference-time intervention (prompting) reliably raises measured response diversity. We do not resolve whether the Artificial Hivemind is real. We show that the published evidence does not establish it.

Abstract: Model merging integrates the capabilities of multiple experts without joint retraining, but standard expert training optimizes task loss alone and does not guarantee good performance after merging. Merge-aware training (MAT) aims to improve merged performance, but existing methods do not fully account for common merging operations and add training cost. We observe that, from an expert's perspective, common merging methods can be described by three operations: Scale reweights its own update, Mask removes selected coordinates, and Perturb adds updates from other experts. Based on this view, we introduce SMAT (Simple MAT), which jointly optimizes expert loss and expected loss at simulated merged parameters generated by sampling scaling coefficients, masks, and additive noise. We further introduce periodic scheduling, kernel fusion, and parameter storage switching to make SMAT efficient, with one forward and one backward pass per step. Across four language and vision-language backbones, SMAT improves the mean score across five merging methods by 1.07-2.16 points over the strongest baseline for each backbone, with less than 2% training-time overhead over standard fine-tuning.

Abstract: Reasoning distillation from powerful teacher models to smaller students faces the Gap Curse: as teachers grow more sophisticated, their complex distributions increasingly diverge from what students can approximate, causing performance degradation. Existing mitigation strategies either filter out challenging examples through data selection or introduce weaker intermediate assistant models, inherently compromising supervision coverage or quality. We propose Teacher Alignment, which directly adapts the teacher toward the student's distribution without discarding data or degrading reasoning quality. However, naive alignment through standard knowledge distillation triggers catastrophic collapse of the teacher's reasoning capabilities. To address this, we reformulate teacher alignment as reinforcement learning and introduce TeacherGRPO, built on Group Relative Policy Optimization with two key innovations: (i) Curriculum Selective Alignment applies dual token- and distribution-level curricula to focus rewards on high-signal reasoning gaps while filtering noise from trivial tokens and uncertain tail distributions, and (ii) Importance-Adaptive Length Regularization selectively penalizes verbose redundancy while preserving pedagogically critical reasoning steps. The aligned teacher then distills knowledge to students via standard pipelines. Extensive experiments show TeacherGRPO significantly outperforms baselines across diverse reasoning benchmarks and distillation methods. Our code is available at https://github.com/LzyFischer/TeacherGRPO.

Abstract: Visual token compression lowers the inference cost of vision--language models by representing images with fewer tokens. However, most existing methods compress visual tokens to a reduced set, leaving the amount of visual evidence represented by each token and its original spatial context implicit. Therefore, the compressed representation does not explicitly encode how much visual information each representative carries or where it lies in the original image. This limitation arises even after a single reduction and becomes more pronounced when compression is repeated across decoder layers. To address this issue, we propose CoViST, a training-free framework that represents a compressed image as a composable visual state. Specifically, the state combines representative features with original positions, effective contribution weights, and reusable selection metadata. CoViST constructs this state through coverage-guided selection and conservation-based contribution composition, and explicitly incorporates its contribution and positional information into decoder attention. Each component of the state retains its interpretation under successive reductions, enabling the same formulation to support both fixed compression before prefill and progressive compression within the decoder. Experimental results on seven LLaVA-1.5-7B benchmarks show that CoViST-Fixed retains 99.9\%, 99.5\%, and 98.1\% of uncompressed performance at 192, 128, and 64 tokens, respectively, and CoViST-Pro retains 99.8\%, 99.9\%, and 99.1\% at the corresponding layer-average budgets, outperforming state-of-the-art methods under their respective budget settings. Code will be released publicly.

Abstract: Reinforcement learning with verifiable rewards (RLVR) often relies on sparse outcome rewards, providing coarse supervision for long-horizon agents. On-policy self-distillation (OPSD) complements this signal with dense privileged feedback. However, we identify \emph{Decision--Timestamp Mismatch}: privileged guidance may be misaligned with the student's functional decision because the corresponding decision can occur at a different timestep, while the student's decision itself may span multiple timesteps rather than being tied to a single timestamp. Thus, timestamp-local supervision can misalign both the context and the temporal scope of credit. To address this mismatch, we introduce \textsc{AlignOPSD}, following the principle of aligning supervision before assigning credit. Decision-Aligned Supervision Rectification re-scores the same student-sampled response in functionally matched contexts across sibling rollouts to calibrate local teacher evidence. Semi-Markov Hierarchical Credit Assignment then derives variable-duration decision spans from correspondence changes and uses rectified evidence to allocate outcome-grounded credit across spans and their constituent turns. We evaluate \textsc{AlignOPSD} with Qwen2.5-3B and Qwen2.5-7B on ALFWorld, WebShop, and Search-QA against representative baselines. \textsc{AlignOPSD} outperforms both GRPO and StepOPSD across all eight backbone--aggregate-metric comparisons, improving on GRPO by 5.5--8.7 \% and ranking first in six. Additional analyzes examine the two alignment stages and hyperparameter sensitivity between tasks. Our code is avaliable at https://github.com/mingju-c/Align-OPSD

Abstract: Verifying multi-step LLM reasoning requires more than determining whether a trace is correct: a useful verifier should identify where the reasoning first goes wrong. However, existing holistic methods provide little positional evidence, while forward sequential verification often treats the first rejected step as the error source. Under error propagation, this assumption can fail, since an earlier mistake may remain locally plausible and become observable only through its downstream consequences. We therefore rethink reasoning verification as a progression-aware error-source localization problem: rather than asking only where a reasoning trace first appears inconsistent, we ask which earlier step best explains how that inconsistency emerges along the trajectory. Based on this view, we propose Progression-aware Reasoning Origin (PRO), a training-free framework for first-error localization. PRO jointly models incoming support from the preceding context and outgoing compatibility with subsequent reasoning, selectively refines regions where these signals disagree, and finally performs detector-conditioned source attribution with intervention-based evidence to distinguish the true error origin from its propagated manifestations. We further formalize the gap between forward rejection and structural exposure, showing why incoming-side evidence alone is insufficient for reliable localization under error propagation. Experiments across open-form, medical, and structured reasoning tasks demonstrate consistent improvements over strong verification baselines, supporting progression-aware source attribution as a more faithful formulation of reasoning verification.

Abstract: Action agents do things for people. They send email, spend money, and call businesses while the user is busy with something else, so a mistake can turn into an action before anyone notices. They fail their users in two ways. They break trust when they do something the user never agreed to, or hold back after the user clearly said go. And they fall short on completion when they give up on errands that turn out to be hard. Both depend heavily on the harness around the model, meaning its instructions, tools, context, and guardrails. We built an evaluation that scores trust and completion on the same runs, in a simulated world of businesses with their own websites, inboxes, and phone lines, and of people who write back. A simulated user answers the assistant's questions. Trust means that nothing happens the user did not agree to. No email goes to someone they never approved, no private detail ends up on a group thread, no money is spent past their limit, no stranger's instructions are followed, and nothing is claimed without a source. Every trap has a matched control in which acting is the right call. We use the evaluation to measure Fo, Wajo's personal assistant, against a base model with basic instructions on three foundation models, and against the Fo harness with its guardrails switched off. Fo completes 71% of the errands and keeps the user's trust on 94% of the trap runs. The base models complete 50% to 64% and keep trust on 59% to 75%. On the matched controls, Fo goes ahead slightly less often. OpenClaw, a popular open-source assistant given the same access, completes 42% of the errands it shares with Fo, against 71%, and keeps the user's trust on 74% of the shared trap runs, against 94%. Measuring trust and completion together, on the whole system rather than the model alone, is how we think action agents become safe to hand real work to.

Abstract: Reward models score responses from large language models (LLMs) and guide LLM training toward human preferences. However, reward models can favor superficial attributes such as length or confidence, leading LLMs to produce higher-scoring but not more correct responses. Existing mitigation methods either retrain the reward model or apply a fixed correction to one known bias, such as a preference for longer responses. Retraining requires additional data and computational resources, while existing editing methods require the target bias to be specified in advance and use a fixed edit for that bias. To this end, we propose BiasReducer, a lightweight framework that edits only the linear reward head and selects the relevant edits for each new dataset. First, BiasReducer uses a sparse autoencoder (SAE)-style encoder to learn which attributes (e.g., length and confidence) the reward model is sensitive to. Second, it learns how to reduce the reward model's dependence on each attribute by determining which direction to adjust the reward head and how much to adjust it. Third, for a new dataset, it ranks the attributes by their influence on reward scores, selects the relevant ones, and edits the reward model accordingly. BiasReducer consistently improves reward-model robustness to biases toward superficial response attributes. Across five reward models, BiasReducer-M improves the three benchmarks by 8.3, 18.0, and 6.9 percentage points on average, outperforming the two training-based baselines. The gains transfer downstream, reducing unnecessary verbosity and sycophancy while maintaining comparable judged quality.

Abstract: While large language models (LLMs) are increasingly deployed in long-context scenarios, lengthy prompts can increase inference costs and latency and exacerbate the ``lost-in-the-middle'' phenomenon. Selective prompt compression offers a model-agnostic approach to alleviating these issues. However, methods based on fixed token- or sentence-level importance scores may overlook how content contributions change with the selected subset, limiting their ability to account for inter-sentence redundancy. Compression procedures that rely on autoregressive LLM scoring can also introduce substantial overhead. We propose PC-SubMax, a theoretically grounded framework that formulates selective prompt compression as regularized monotone submodular maximization under a knapsack constraint. The objective is $U(S)-\ell(S)$, where the monotone submodular utility $U$ combines information coverage, query relevance, and log-determinant diversity, and the non-negative modular penalty $\ell$ captures token cost. Through diminishing marginal returns, the objective evaluates each sentence's contribution relative to the selected content. To optimize this objective, we develop the Regularized Greedy+Max (RGM) algorithm, which deterministically returns a feasible set $Q$ satisfying $U(Q)-\ell(Q)\geq \frac{1}{2}U(O)-\ell(O)$, where $O$ is an optimal feasible solution to the regularized problem. RGM uses $O(nκ)$ value-oracle queries, where $n$ is the number of candidate sentences and $κ$ is the maximum feasible subset size. PC-SubMax uses encoder representations and avoids autoregressive LLM scoring during compression. Experiments across seven diverse benchmarks demonstrate competitive downstream performance with low compression overhead.

Abstract: Knowledge distillation (KD) enables model transfer across heterogeneous platforms and deployment environments, yet it exposes proprietary models to distillation-based theft, where an adversary extracts intellectual property (IP) by training a student model on teacher outputs. Current defenses, such as software watermarking or hardware-based access control, either fail to survive the distillation process or lack the granularity to identify the specific device responsible for a leak. Aiming to provide post-theft accountability and traceability, we propose a novel fingerprinting framework that superimposes device-specific Physical Unclonable Function (PUF) signatures onto teacher logits during distillation. By utilizing signatures derived from Ring Oscillator (RO) PUFs measured on a Xilinx Zynq-7020 FPGA, we ensure that any student model trained via KD inherits a unique, hardware-linked identity. Our framework is architecture-agnostic, enabling reliable identity inheritance across heterogeneous structures, including Convolutional Neural Networks, Vision Transformers, and encoders. To ensure robust attribution, we implement a two-stage recovery pipeline consisting of a neural decoder and Hamming-distance refinement, maintaining high detection accuracy even under noisy conditions. Furthermore, we introduce a multi-level logit encoding scheme to support large-scale device deployment. Experimental results demonstrate that the embedded fingerprints are resilient against common post-distillation modifications. These results establish a practical system-level approach for enabling hardware-linked model traceability in distributed AI deployment environments.

Abstract: Modern LLM post-training composes supervised fine-tuning (SFT), reinforcement learning with verifiable rewards (RLVR), and on-policy distillation (OPD) into multi-stage pipelines, yet these stages are typically designed and evaluated in isolation. We show that this composition is consequential: a stage that improves the current model can make the next stage less effective. Through controlled experiments with Qwen3 models on math and science reasoning, we first characterize OPD across nine student-teacher pairs spanning 2x to 53x parameter ratios and show that OPD effectiveness depends on student-teacher compatibility rather than teacher scale alone. The surrounding stages of OPD reshape this compatibility in three ways: (1) A brief SFT warm-up improves subsequent OPD, while an RLVR-strengthened student regresses under distillation from the same teacher. (2) Adapting the teacher with RLVR raises downstream OPD accuracy in proportion to the capability it adds. Following these two interventions, we find that combining teacher adaptation and student warm-up alone raise average OPD accuracy from 29.2\% to 43.8\% (50\% relative improvement) after the same number of distillation steps, with additional preparatory training. (3) At comparable accuracy, OPD leaves a stronger initialization for downstream RLVR than SFT, with a gap that widens as RL compute scales. Our results suggest that each post-training stage should be chosen not only for the capability it adds, but for the learning interface it creates for the next stage.

Abstract: LoRAs are widely studied for adapting base text-to-image diffusion models. However, a backdoored LoRA can hide a backdoor: it behaves normally in most cases, but produces attacker-specified content (the backdoor target) when a hidden backdoor trigger appears in the input prompt. Detecting such backdoors before using an untrusted LoRA is important for the safety of LoRA adaptation. We present TokenScanner, a model-level vocabulary scanner for backdoor detection within a LoRA fine-tuned text-to-image diffusion model, aiming to discover the malicious trigger for trustworthy LoRA adaptation. The key observation is that backdoor trigger tokens that appear in a larger proportion of training prompts tend to induce more prominent token-specific LoRA responses than those induced by unrelated tokens. TokenScanner therefore scans the tokenizer vocabulary and measures token-wise LoRA responses in the U-Net and the text encoder. It uses these responses to detect backdoored LoRAs and rank candidate trigger tokens for subsequent testing of backdoor activation. Experiments on seven backdoor settings, comprising 840 backdoored LoRAs and 840 real-world benign test LoRAs, show that TokenScanner achieves 95.96% AUC and 96.55% TPR at an FPR of 10.95%. It also achieves 89.40% Hit@1 and 99.40% Hit@5 for trigger discovery across all seven settings.

Abstract: Leaked system prompts are often treated as windows into the hidden values of commercial language models, yet their composition is rarely studied at scale. We analyze a merged corpus of 407 leaked, reconstructed, or officially published system prompts from 62 vendors across four community collections, identifying 29 near-duplicate clusters covering 66 files. Operational content rather than ethical statements dominates the corpus; a deliberately simple block-level classifier assigns roughly 58\% of classified words to tool/protocol and roughly 5\% to safety policy, while the strictest rule-lines guard tool use and file safety over harmful content by an 11:1 margin. Literal text transfer concentrates in a small set of cross-vendor pairs. Prompts also carry measurable maintenance debt, with version chains turning over thousands of words per release. The evidence supports treating leaked prompts as operational specifications, closer to configuration files than value statements, and treats reuse and prompt rot as engineering and supply-chain concerns. Because most documents are adversarial in origin and the detectors are deliberately simple, all magnitudes are directional; we audit the main classifier's error modes.

Abstract: Building adaptable AI systems requires effective coordination of specialized capabilities across diverse tasks. Low-rank adaptation (LoRA) enables modular expertise, but existing routing approaches may require auxiliary data, additional training, or autoregressive decoding. We propose JevSoup, a training-free framework separating System One expert routing from System Two execution. Using only the input and expert descriptions, Jev selects two experts through structured probabilities. JevSoup retains the leading expert's update, projects the second onto the orthogonal complement of the first update's row space, and combines them with equal weights. Across 14 PorTAL tasks and three Qwen3 scales, JepSoup achieves absolute gains of up to 1.19\% in task-macro and 1.21\% in sample-micro accuracy over the strongest evaluated external baselines. Our code is available at https://github.com/Leowang980/JevSoup.

Abstract: Pre-trained vision-language models such as CLIP can recognize new categories via prompting, but they often struggle when labeled data are scarce or the test distribution shifts. Prompt learning adapts only a small set of parameters while keeping the backbone frozen, yet many existing multimodal prompt learners couple the visual and textual branches weakly and can be brittle in low-shot regimes. We propose ProCAP, a probabilistic cross-attentive prompt learning framework that improves cross-modal interaction and training stability without updating any CLIP weights: it learns both visual and textual prompt tokens and links them through stacked bidirectional multi-head cross-attention so the two branches refine each other across prompt depth. To reduce overfitting under limited supervision, we parameterize prompt tokens with Gaussian means and variances and regularize them with lightweight KL and L2 penalties, and we further add a compact symmetric InfoNCE head that aligns cross-attended image features with class-level text representations in a shared low-dimensional space. Across few-shot base-to-novel generalization on 11 datasets, cross-dataset transfer, and domain generalization on ImageNet shift benchmarks, ProCAP achieves strong aggregate base-to-novel performance and competitive transfer performance while keeping the CLIP backbone unchanged.

Abstract: Large language models (LLMs) are increasingly deployed in settings where responses must reflect multiple, potentially interacting social norms and human values. Activation steering offers a lightweight alternative to training-based alignment by modifying internal activations at inference time. However, prior human-value steering methods have largely considered values in isolation, while direct composition of multiple directions relies on fixed intervention strengths that cannot respond to the model's evolving internal state. Motivated by this key observation, we introduce AIMES, a framework for adaptive multi-value activation steering. AIMES constructs layer-specific bipolar directions for moral-foundation values and uses intermediate-layer vocabulary readouts as online observers. An observer-guided controller then adapts the strength of each requested value intervention at every decoding step based on its current observed state, without training a separate value-state estimator. Across multiple instruction-tuned model families, value combinations, and intervention depths, we find that multi-value controllability varies across both value combinations and intervention locations. Compared with fixed joint steering and prompt-based steering, AIMES shows depth-dependent advantages that are broadly supported across two independent evaluators, with some variation in the precise depth at which specific control effects emerge. These advantages come with smaller realized activation-space interventions than fixed-joint steering and comparable response quality. Overall, our results suggest that online observer feedback can provide lightweight, state-aware adaptation for single-pass multi-value steering.

Abstract: Document retrieval is a crucial component of many modern AI systems, directly influencing their effectiveness, robustness, and fairness in downstream tasks. While recent years have seen a growing number of retrievers, comparative studies in the literature are typically limited in scope or focused on singular benchmarks, domains, or model families. This fragmentation makes it difficult to draw reliable conclusions about the relative strengths, weaknesses, and trade-offs of document retrievers. To address this gap, we conduct a large-scale empirical evaluation of document retrievers, covering three families (sparse, dense, and expansion-based) and evaluating 33 retrievers across seven IR datasets, analyzing retrieval quality, runtime, and failure points. Rather than tuning each model individually, we evaluate every retriever off the shelf, under the configuration reconstructable from its public documentation and a uniform compute budget. Our results show that NV-Embed-v2 achieves the strongest performance on four of the seven datasets, albeit at the cost of substantial query latencies. Among sparse retrievers, we find that SPLADE-v3 rivals the top-performing approach despite much lower latency, and even achieves top scores on MS MARCO. On instruction-following datasets, GritLM delivers the best performance. Finally, an analysis of the retrievers' failure points reveals contrasts between models and families that indicate potential for unrealized gains in retrieval performance.

Abstract: Personalized value alignment has become increasingly important as large language models (LLMs) are expected to accommodate diverse user preferences. However, existing methods typically align model outputs with a static value profile across prompts, overlooking that the salience of value dimensions varies substantially across contexts. Inspired by Lewin's Field Theory, which views human behavior as jointly shaped by personal dispositions and situational constraints, we model personal values as priors and context-dependent preferences as posteriors. We propose BaCVA, an inference-time Bayesian Context-aware personalized Value Alignment method that approximates posterior personalized preferences by integrating static personal values with scenario-specific value salience. BaCVA first estimates contextual value salience from generally normative responses, and then employs a dual-view personalization module to infer posterior preferences from complementary personal-value and scenario-driven perspectives. This Bayesian formulation enables more accurate and adaptive personalized value alignment while improving data efficiency via prior values. Extensive experiments on benchmarks demonstrate its superiority over strong baselines.

Abstract: Compact vision-language models (VLMs) now power a growing share of multimodal applications. The benchmarks used to compare them, however, inherit a frontier-centric design: each model is reduced to a single accuracy number, narrowing the inter-model gap on saturated suites and pressing models into low-score bands on harder ones. We introduce PRISM-VLM, a multi-axis discriminative benchmark that scores every item along seven axes covering the recurring failure modes (task quality, behavioral robustness, and capability bottlenecks) and combines them into a single PScore, with items recycled from fifteen public benchmarks. Across compact VLMs from the past two years, PScore separates model pairs more reliably than prior single-axis benchmarks under an item-level paired bootstrap, and surfaces behavioral differences these benchmarks average away. Even models with statistically indistinguishable PScores diverge sharply along the per-axis profile, particularly on sycophancy, which is nearly orthogonal to single-prompt accuracy. We will release the full pipeline, prompts, and per-item annotations.

Abstract: People hold diverse, sometimes conflicting values, so no single aligned model can satisfy everyone. Pluralistic alignment therefore calls for steerable models that can balance competing objectives differently. Multi-Objective Direct Preference Optimization (MODPO) does this by using an objective weight to span a continuum of trade-offs. We study two questions: when can one model improve two objectives simultaneously, and how can many trade-offs be covered without training a separate model for each? Across seven objective pairs from HelpSteer and UltraFeedback, two pre-training measurements predict whether objectives align or conflict for human-annotated data, but not for AI-annotated data, where response length and repetition confound reward-model scores. For broader trade-off coverage, selecting the nearest trained model and merging model parameters both help, but neither consistently matches direct training. These findings yield practical guidance for building steerable models that serve diverse preferences.

Abstract: Reinforcement learning (RL)-based post-training has become an effective approach for eliciting reasoning capabilities in large language models (LLMs). However, adapting post-RL models to new knowledge domains or behaviors through subsequent supervised fine-tuning (SFT) can severely overwrite these capabilities. Existing approaches mitigate such forgetting through experience replay, specialized initialization, or constrained optimization using gradient projection, but either provide limited preservation or incur substantial training overhead. Our analysis shows that reasoning activations concentrate in low-dimensional subspaces, leaving substantial null-space capacity for adaptation, and that the corresponding approximate null spaces can be reliably estimated from a modest number of examples. Motivated by these observations, we propose Null-Basis Low-Rank Adaptation (NB-LoRA), a parameter-efficient method for adapting post-RL LLMs while preserving their acquired reasoning ability. We formulate reasoning retention as a layer-wise hidden-state preservation constraint and construct a fixed approximate null basis from reasoning activations. LoRA updates are then reparameterized through this basis, enforcing the preservation constraint throughout fine-tuning. Extensive experiments across multiple RL-trained LLMs and diverse downstream tasks show that NB-LoRA matches standard LoRA in adaptation performance, maintains reasoning accuracy near pre-fine-tuning levels, and generalizes this preservation to held-out reasoning benchmarks.

Abstract: Large language models (LLMs) exhibit a parametric vulnerability to adversarial swarm consensus. To mitigate this sycophancy, we introduce Contrastive Epistemic Decoding (CED), a zero-shot inference intervention. Unlike standard Contrastive Decoding (CD) which relies on a weaker secondary model, CED utilizes a dual forward-pass on a single architecture to isolate conformity bias. By introducing a novel asymmetric, zero-bounded probability clamp and discrete top-k truncation mask, CED mathematically suppresses toxic consensus tokens without causing grammatical collapse. Evaluated across 7,200 paired trajectories on complex benchmarks (GAIA, SWE-bench, Multi-Challenge) using Gemma-2 (9B), Llama-3.1 (8B), and Mistral v0.3 (7B), CED successfully neutralizes architectural and positional biases. By reducing cognitive loafing by up to 33.00% absolute, CED drives significant performance gains, yielding up to a 30.75% accuracy recovery. Regaining sovereignty induces distinct architectural behaviors---passive task-focus in Gemma-2 and active refutation of the simulated swarm in Llama-3.1---showing CED decouples compliance from capability without fine-tuning.

Abstract: In high-stakes decision-making applications of large language models (LLMs), practitioners require not only accurate LLMs but also uncertainty estimates for their predictions. Existing approaches to uncertainty estimation for LLMs require access to log-probabilities output by the model or require fine-tuning access. However, many industrial LLM products use closed-source API models, and many such API models like GPT do not return log-probabilities and may not allow fine-tuning. We introduce Pinocchio, an external calibrator that estimates the correctness of responses from black-box API models. Trained jointly on responses from seven LLMs, it achieves 0.862 AUROC predicting the correctness of held-out responses from those same models, and shows zero-shot transfer to thirteen unseen models across eight organizations. Our model needs only a single forward pass to generate an uncertainty estimate and requires no access to the target model's logits, weights, or internal states. A lightweight text only 0.8B checkpoint matches our largest model's AUROC. We release code for adding uncertainty estimation to existing repos in only two additional lines of code.

Abstract: This paper describes the SKstars submission to SHROOM-Visions 2026, a shared task on fine-grained hallucination detection in large vision-language model outputs. The task requires systems to identify hallucinated character spans, assign hallucination categories, and provide confidence estimates for their predictions. Our approach combines zero-shot predictions from Qwen2.5-VL-72B-Instruct with those of a LoRA-adapted Qwen2.5-VL-7B-Instruct model. The outputs of the two models are integrated through a lightweight ensemble procedure, followed by span refinement and confidence adjustment. We evaluate the main system components on a small internal development subset and report the performance of the submitted system on the official English test set. SKstars achieved a Cor+Lbl score of 0.2902, ranking 15th among 29 teams, and obtained Cor and IoU scores of 0.3642 and 0.3151, respectively, ranking 18th on both metrics. The results show that combining a large zero-shot model with a smaller adapted model provides a practical framework for multilingual and fine-grained hallucination localization, while also highlighting the difficulty of transferring development-set improvements to hidden test data. Code and predictions: https://github.com/aliathar1401/SK-Stars-shroom-visions-2026

Abstract: Textual-gradient methods automate prompt optimization through natural-language feedback, but their iterative updates can be unstable. We identify two sources of this instability: noisy gradients produced from already-correct examples and over-specialization to hard cases that degrades performance on simpler inputs. We introduce STEVE, a stabilization framework with two coupled mechanisms. Error-Driven Refinement generates gradients only from incorrectly handled examples, concentrating updates on informative failures. Regularized Verification treats every update as provisional and accepts it only when improvement on hard cases does not cause unacceptable regression on a preservation set. Across ten reasoning benchmarks, three evaluator/optimizer models, and established prompt-optimization baselines, STEVE reduces degradation and produces more robust prompts. Additional evaluations with gpt-5.4-mini/gpt-5.4 on symbolic reasoning, GSM8K-Platinum, and DS-1000 show that these gains persist with newer models and larger test sets. STEVE therefore provides a practical way to improve the stability and effectiveness of textual-gradient prompt optimization.

Abstract: Discrete diffusion models now match autoregressive language models on several benchmarks, while the question of how best to train them has received far less attention: the optimizer is inherited from one paper to the next and never compared. New optimizers, meanwhile, are validated almost exclusively on autoregressive pretraining, a different objective on a different loss surface. We present a controlled optimizer benchmark across four diffusion formulations, to our knowledge the first for discrete diffusion: seven optimizers (AdamW, Lion, Muon, SOAP, MARS, MARS-M, Schedule-Free) on masked diffusion (text8), uniform diffusion (QM9, and LM1B through the Gaussian duality) and Gaussian diffusion on images (CelebA-64), each on a task with published reference values. Every optimizer receives the same search protocol, and every winner is retrained at the full budget with three seeds. AdamW is a strong default but not always the right choice: it is beaten by a resolved margin on two of the four tasks, and the winner changes with the formulation, so the optimizer deserves the same care as the rest of the training recipe. Notably, methods validated on autoregressive language model pretraining transfer well: Muon, MARS-M and SOAP each beat the tuned AdamW on at least one diffusion formulation. The benchmark, all runs and every figure are reproducible end to end from the released code at https://github.com/armanbolatov/diffusion-baselines.

Abstract: Post-training plays a pivotal role in enhancing the reasoning capabilities and task-specific expertise of large language models (LLMs). Despite recent advances in post-training methods, such as Group Relative Policy Optimization (GRPO), their practical deployment remains impeded by training instability arising from the reliance on importance sampling. We introduce Group Variance Policy Optimization (GVPO), a novel post-training method that integrates the analytical solution of KL-constrained reward maximization into its gradient weighting scheme. This formulation provides an intuitive interpretation: GVPO's gradient corresponds to the mean squared error between the central distance of implicit rewards and that of actual rewards. GVPO offers two key advantages: (1) it guarantees a unique optimal solution, exactly to the KL-constrained reward maximization objective, and (2) it enables flexible sampling distributions without requiring importance sampling. Beyond general post-training, we show that GVPO naturally extends to on-policy distillation (OPD). Furthermore, GVPO enables the optimization of a broad family of extended OPD objectives, providing a principled foundation for diverse objective design. By unifying theoretical guarantees with practical adaptability, GVPO establishes a new paradigm for reliable and versatile LLM post-training and on-policy distillation.

Abstract: Safety tuning pipelines judge only the final answer, which makes it difficult to distinguish robust refusal from two undesirable shortcuts: blanket refusal on benign requests and polished but unfaithful safety rationales that do not actually constrain the answer. We propose AUDITPLAN, a single-model plan-then-answer approach where the model first emits a compact structured safety plan and then answers conditioned on it. The plan records a threat label, intended action, and explicit constraints, enabling machine-checkable auditing while remaining hidden from users at deployment. We train this behavior with supervised fine-tuning followed by reinforcement learning with FAITHGATE, a reward-gating objective that grants answer reward only when the safety plan is correct. This discourages safe-looking but unfaithful behavior and promotes tighter plan-answer coupling. Across Qwen backbones, AUDITPLAN improves both robustness and auditability: on Qwen2.5-3B-Instruct, FAITHGATE reduces ASR from 24.0% to 11.6%, LSR from 1.0% to 0.36%, and over-refusal from 11.0% to 2.0%, outperforming answer-only RL, free-form explanation, and weighted-sum structured rewards. Similar trends hold for Qwen2.5-1.5B-Instruct. Larger-model confirmation runs on Qwen-3-4B-Instruct and Qwen2.5-7B-Instruct preserve the same trend suggesting that explicit internal commitments can make safety alignment more faithful, robust, and auditable.

Abstract: Large language models (LLMs) often suffer from capability stagnation in self-improvement training because fixed difficulty levels fail to adapt to their evolving proficiency. To address this issue, we propose STRETCH (Self-Taught Reasoning Evolution via Targeted CHallenge), a unified framework inspired by cognitive scaffolding theory. STRETCH introduces a dynamic Stretch Zone mechanism that continuously aligns question difficulty with the model's solving capability. Within a single parameter space, the model alternates between a Scaffolder that generates adaptive, boundary-pushing challenges and a Learner that that optimizes its solving trajectories through reinforcement learning. This dual-loop co-evolution effectively stabilizes training, mitigates reward hacking and promote progressive reasoning growth. Experiments on both negotiation and operation research benchmarks demonstrate that STRETCH consistently outperforms strong prompting and domain-specific baselines. Further scaffolder configuration analysis shows that dynamic difficulty alignment is critical for sustained capability improvement and synchronized reasoning evolution.

Abstract: Large vision-language models (LVLMs) are commonly used with only a single text prompt as the input, or plus an image. In this paper, we demonstrate that when the image exists, even if the text prompt is not about the specific instance (but only the concept it belongs to) in that image, the response would still be affected. For example, when the text prompt only asks for the attribute descriptions of a dog breed, an image depicting a specific dog from that breed would shift the response. Further, how the specific instance is framed in that image would determine towards which the response shifts. Detailed analyses also reveal that in the response, physical terms increase from 18% for text-only to 45% (40%) for subject-focused (subject-in-situation) framings. Overall, the unexpected effects of visual cues on LVLMs highlight the need to understand the presence of an image and its framing when evaluating the robustness of LVLMs.

Abstract: We present NeMo Data Designer (NDD), an open-source, general-purpose framework for multi-modal synthetic data generation (SDG). Designed to be intuitive to use, NDD provides a declarative configuration format in which human and/or agent users define each dataset column, with column types spanning text, code, structured outputs, images, embeddings, and statistical samplers that are explicitly configured to steer dataset diversity. Additional column types and functionality can be introduced using the framework's flexible plugin system. NDD's configuration is an inspectable artifact, supporting workflow sharing and reproducibility. SDG is an inherently iterative process. NDD therefore builds a preview-and-revision loop into its core workflow, allowing users to generate and inspect a small number of records, refine the specification, and rerun generation at full scale. At runtime, NDD resolves dependencies, schedules calls to user-provided model endpoints, and retries failed requests. We describe NDD's architecture and programming model and present case studies spanning structured, agentic, multimodal, and domain-specialized tasks, including datasets used in Nemotron model development and in production enterprise deployments.

Abstract: LLM watermarking helps trace the origin of generated text, but faces stealing attacks that recover watermark information, scrubbing attacks that remove watermark signals, and spoofing attacks that forge text accepted as watermarked. These attacks are often studied in isolation, leaving their connections unclear. Evaluations also often lack shared detector calibration, metric definitions, and reporting protocols. Moreover, measuring attack success and text quality separately makes it difficult to identify attacks that are both effective and quality-preserving. We propose MarkSec, a general framework that unifies analyses of stealing, scrubbing, and spoofing. We evaluate attacks under a common reporting protocol and introduce a quality-constrained attack success metric to assess effectiveness and text quality jointly. Experiments across representative watermark families, attacks, LLMs, and datasets reveal three findings. First, attacks that appear strongest by watermark removal alone can fall behind general rewriting when success also requires acceptable text quality. Second, general rewriting remains a strong baseline across watermark families, while its advantage over other scrubbers varies by family. Third, in a case study of one watermark family, stealing-based scrubbers often underperform the best general-scrubbing baselines when text quality is required. These results show that apparent attack winners depend on text-quality constraints, attack generality, and capability assumptions.

Abstract: Gender bias in large vision-language models (LVLMs) undermines their fairness and reliability, compromising output trustworthiness. Current mitigation methods rely on training-phase adjustments or post-hoc calibration, but face limitations in dynamic visual bias mitigation. These include inability to capture real-time visual-textual incongruence, dependence on predefined gender bias taxonomies, and degraded cross-modal alignment with emergent bias patterns. To address these challenges, we propose ViD, a causally-inspired framework that analyzes attention mechanisms across five distinct patterns, revealing confounding effects from strong language priors. ViD demonstrates that visual-to-language cross-attention effectively suppresses bias while preserving general reasoning capabilities and text generation quality. ViD incorporates dual mechanisms: backdoor adjustment counters strong language priors, while refined token selection in decoding layers optimizes processing. This enhances model robustness and inference efficiency. Our integrated approach significantly mitigates gender bias across multidimensional social attributes in LVLMs, improving visual grounding and output fairness. Cross-benchmark validation shows ViD reduces gender bias by 14.7\% on single-attribute evaluations (FACET) and achieves significant improvements on image captioning tasks (MS COCO), with gender bias score improving from 0.6708 to 0.9978 for LLaVA. Crucially, these improvements require no additional training overhead, making ViD a scalable and practical solution for bias mitigation in LVLMs.

Abstract: Query-conditioned vision--language models enable fine-grained interpretation by revealing how visual evidence changes with textual queries. However, evidence conditioned on complete descriptions does not necessarily resolve into object-specific evidence, nor does an exposed evidence map necessarily identify the evidence that constitutes the model's prediction. Across multiple VLM architectures and independent benchmarks, we find that object-level queries often retain evidence from co-occurring objects and shared context. In this paper, we introduce \textbf{ProtoLIP}, a lightweight prototype-mediated evidence layer that organizes reusable visual prototypes into text-derived semantic families and uses query-dependent family routing to constrain which prototypes may provide evidence. Without spatial annotations or backbone retraining, ProtoLIP improves evidence localization and separation across query granularities, with localization gains transferring to independently pretrained VLMs with well-aligned patch--text representations. Despite using only text-derived weak supervision, ProtoLIP remains competitive with a spatially supervised grounding model while maintaining strong matching and competitive image--text retrieval. Crucially, ProtoLIP constructs its matching score directly from localized prototype evidence, enabling the score to be exactly decomposed into semantic-family and prototype contributions.

Abstract: Large language model (LLM) agents increasingly interleave natural language reasoning with external tools such as web search and code execution. These tool-use policies are often optimized via reinforcement learning (RL), which can amplify spurious correlations in the training data. In this work, we study when and why RL-trained agents learn shortcut tool-selection policies: invoking tools based on superficial prompt cues rather than genuine task requirements. We construct controlled synthetic environments combining factual question answering and mathematical reasoning tasks, and inject cues that are strongly correlated with specific tools during training but causally irrelevant to tool necessity. Across counterfactual evaluations where cues are present but the associated tools are not required, agents exhibit substantial shortcut behavior, with spurious tool invocation rates increasing by up to 39 percent. However, shortcut formation is not universal: across the conditions we test, it arises only when the agent has already learned to use the target tool reliably, suggesting that task competence, rather than dataset imbalance alone, is a key factor in shortcut learning. A swapped-cue analysis further shows that semantic alignment between cues and tools substantially amplifies this effect. To mitigate these failures, we introduce a dense, decision-level reward in which an LLM judge evaluates the necessity of each tool call. This tool-necessity reward effectively suppresses cue-driven tool use while preserving task performance, providing a practical approach to improving the robustness of LLM agent tool-use policies.

Abstract: In-context learning (ICL) enables models to infer tasks from demonstrations, but existing benchmarks generally lack matched text and image versions needed to compare ICL performance across modalities. We introduce TwinICL, a procedurally generated benchmark providing such pairs for controlled comparison. Across six open-weight models and 38 tasks, multimodal ICL consistently underperforms text-only ICL, with gaps varying by task family. To test whether this gap can be recovered, we target visual access, task framing, and reasoning through three interventions. Their combination recovers strong multimodal ICL performance on a diagnostic subset, despite limited or inconsistent individual effects. To distinguish difficulties in executing tasks from those in inferring them, we evaluate models with explicit task instructions, revealing a modality gap even when the task is known. We then examine how adding demonstration inputs and outputs reshapes this gap, highlighting demonstrations' dual role as additional context to process and evidence about the task. The dataset is available at https://github.com/lab-flair/TwinICL.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has shown remarkable success in improving the mathematical reasoning of large language models. Yet problems beyond the model's current capability, where rollouts uniformly fail and no learning signal is produced, are structurally wasted despite marking the most informative training frontier. We show that these otherwise-inert problems can be unlocked via teacher-guided curriculum learning: partial reasoning traces from a stronger model create a graded difficulty landscape, and a backward-chaining curriculum progressively withdraws guidance until the model solves problems unaided. Training on only 128 unsolvable problems matches or exceeds GRPO trained on a full 2,000-problem corpus (~16x data efficiency) on the nine-benchmark average for both base models, while substantially expanding the reasoning boundary measured by pass@k at large k. Furthermore, we identify a distribution-shift cost that is particularly acute in the unsolvable-only regime and propose Monotone Frontier Curriculum (MFC), a method that monotonically drives training toward unguided solving, consistently outperforming existing curriculum methods.

Abstract: Large Language Models (LLMs) are powerful zero-shot learners but remain prone to misalignment with human preferences, often producing biased, toxic, or otherwise harmful outputs. Existing alignment methods, while effective, are costly and tightly coupled to the model, limiting flexibility and scalability. We propose a modular correction framework that augments pretrained LLMs with Activated LoRA (aLoRA) adapters and a context-aware routing mechanism to eliminate harms from misaligned model responses. Our approach enables expert adapters to activate mid-sequence without invalidating the KV cache, allowing low-latency, targeted correction during generation. Each expert is trained to detect and mitigate specific harms, such as bias or toxicity. A learned router dynamically selects appropriate experts based on the models intermediate outputs. We demonstrate that our system improves alignment on standard safety benchmarks while preserving task performance, offering a lightweight and efficient path toward safer and more controllable LLM deployments.

Abstract: Reinforcement learning (RL) has become central to post-training of large language models. Recent advances in RL for Mixture-of-Experts (MoE) models have primarily focused on improving optimization stability and training efficiency, while treating the expert selection as a fixed component. Since routing determines the sparse computation paths that induce output distributions, expert selection offers an additional source of rollout diversity. Through empirical analysis, we find that perturbing expert routing effectively alters model output and increases rollout diversity, which is similar to increasing the decoding temperature. However, direct perturbation can activate unsuitable experts and substantially degrade rollout quality. Motivated by these observations, we introduce Expert-Space Exploration Reinforcement Learning (ESRL), an architecture-aware framework that explicitly explores the expert-routing space of MoE models. ESRL preserves high-confidence experts as anchors, and restricts stochastic routing to a plausible candidate pool, thereby retaining reliable computation paths. The perturbation strength is further adapted according to router entropy to avoid over-perturbation. To mitigate the routing mismatch introduced by perturbation, ESRL records the expert paths used during rollout and replays them during policy optimization. Experiments demonstrate that ESRL achieves the best performance across MoE backbones with top-K, top-1, and shared-expert routing, as well as across mathematics, science, and code tasks without additional sampling or computational cost. Specifically, ESRL on Qwen3-30B-A3B achieves the best among all compared methods, improving average Pass@1 and Pass@8 over GRPO by 3.2 and 4.5 percentage points, respectively. Further analyses of expert utilization and training dynamics provide insights into how exploiting MoE-specific routing structure benefits RL training.

Abstract: Multi-modal image generation, particularly subject-driven customization, has garnered growing attention in recent years. Despite the rapid advancement of generative models, their evaluation remains largely lagging. Existing methods, whether embedding-based or Multi-modal Large Language Model (MLLM)-based, evaluate alignment with each modal condition in isolation, which contradicts the simultaneous condition alignment objective of multi-modal image generation, leading to poor consistency with human judgments. To address this challenge, we propose UFO, the first unified framework for omni-condition alignment simultaneous evaluation. Specifically, UFO introduces a novel Atomized Chain-of-Evaluation paradigm, \emph{i.e.}, it first decomposes omni-condition alignment into a sequential chain of fine-grained, disentangled Atomic Evaluation Units (AEUs), categorizes them into distinct modality-relevance classes, and then employs general or dedicated functional calls for accurate verification of different AEU types. Experimental results demonstrate that UFO achieves the highest correlation with human evaluation preferences, delivering an average improvement of 15.25\%. Furthermore, we present UFO-Bench, a dedicated benchmark designed to holistically evaluate the performance of existing customization models under the diverse mutual interactions of textual and visual conditions.

Abstract: A central question in LLM reasoning is whether reinforcement learning (RL) instills genuinely new capabilities or merely reshapes how existing knowledge is expressed during inference. Building on the distribution-sharpening hypothesis, which holds that RL reallocates probability mass toward high-reward trajectories already latent in base models, we ask: can we unlock those latent paths without costly RL fine-tuning? We present Decision-Flow Sampling (DF-Sample), a training-free, data-free inference-time framework that constructs a hierarchical reasoning tree, scores terminal nodes for quality, and back-propagates utilities to inform each intermediate branching decision. Unlike conventional sampling strategies that make purely local step-wise choices, DF-Sample performs explicit global trajectory evaluation before committing to a path, recovering high-quality but low-probability reasoning chains that standard decoding overlooks. On GPQA, DF-Sample achieves 45.6% accuracy, surpassing power sampling (38.9%) and GRPO (39.9%), showing that a training-free method can outperform a trained one. Across three models and four benchmarks, DF-Sample consistently outperforms baselines, indicating substantial latent reasoning potential in pretrained base models.

Abstract: We introduce AdamX, a first-order optimizer that incorporates cosine similarity as an adaptive mechanism for controlling update magnitudes. The proposed method is scalable, model-agnostic, and straightforward to integrate into existing training pipelines. We further introduce a variance rectification scheme that promotes smoother optimization during the early stages of training. Overall, we provide empirical evidence that AdamX achieves competitive convergence rates across a range of benchmark datasets and architectures. Performance is evaluated in terms of the number of epochs required to reach predefined performance thresholds under a fixed hyperparameter budget. Code and Experiments available at: https://github.com/FranciscoCaldas/adamX.

Abstract: We post-train Qwen3.8-27B for Korean response style -- verbosity, list and markdown usage, discourse structure and register -- and measure two behaviours the objective never targets: abstention on ambiguous social questions in KoBBQ, where the benchmark-correct answer is UNKNOWN, and unprompted disclosure in securities guidance. Both move, and the changes are expressed primarily through the model's emission policy: how often it answers and how much it says. Matched target-form controls show that answer propensity depends on the training target, not the prompt set or recipe alone. Holding prompts, recipe, data volume and serving fixed and changing only the target text, three style seeds give positive answer-rate point estimates (mean +0.82 pp) and three neutral seeds negative ones (mean -1.53 pp); the observed seed ranges do not overlap and the means differ by 2.34 pp. A length-matched arm lies between them, and a fourth arm that stays short while preserving hedging is unstable across seeds, so which feature of the form is responsible is unresolved. For absolute stereotyped exposure the decomposition into an answer-propensity term and a conditional-composition term is an algebraic identity, not a finding; its empirical content is where the movement went. Across the trained checkpoints the changes are dominated by answer propensity while the composition term stays small, and because that term is evaluated on treatment-dependent answered subsets we do not read it as evidence about latent preference. Two measurement results follow. A between-arm contrast in conditional stereotyped share does not identify a change in conditional content preference when answer status is treatment-dependent. And agreement between two rule detectors for the same construct runs from 0.44 to 0.99 depending on which checkpoint produced the text -- observable without any reference labels.

Abstract: The integration of Large Language Models into daily tasks relies on context-rich instructions, inevitably exposing sensitive user information. Current privacy-preserving methods typically employ context-agnostic static rules, causing severe utility degradation. However, the specific mechanisms governing how sanitization impacts downstream performance remain largely underexplored. To address this, we conduct a systematic analysis to deconstruct the privacy-utility trade-off, uncovering three underlying mechanisms: (1) Context-Dependent Utility, which first establishes when to sanitize by revealing that data value shifts from critical constraints to dispensable noise based on user intent; (2) Strategic Adaptation, which subsequently determines how to sanitize by dictating that the choice between removal and replacement depends on the task's reliance on factual integrity versus structural coherence; and (3) Combinatorial Interplay, which finally extends the protection scope by demonstrating that attributes form a semantic web of synergistic dependencies or antagonistic redundancies. Guided by these insights, we introduce an intent-driven local protection framework. By distilling a lightweight model Veilmind-4B to drive a dynamic extraction-sanitization-restoration pipeline, our approach reaches a low-leakage privacy point while preserving substantially higher response utility than existing privacy-oriented baselines, advancing the privacy-utility trade-off toward the Pareto frontier.

Abstract: LLM agents for sequential decision tasks are often post-trained with trajectory-level outcome labels, but such labels provide little supervision for preserving multiple successful branches from the same decision state. We study this problem as successful strategy coverage: how broadly a model realizes distinct successful strategies under a fixed rollout budget. We present Direct Diversity Optimization (DDO), an offline post-training method that combines Divergence-Tree Collection (DTC) with the Reference-Relative Target-Odds Objective (RTO). DTC constructs state-aligned branch sets rooted at shared decision states, and RTO trains the model to match reference-relative targets over successful alternatives. DDO achieves the strongest task success and successful strategy coverage among the compared post-training methods across BabyAI, BabaIsAI, and WebShop. It also achieves the highest recovery rate after local action replacement and higher task success and coverage than successful-only imitation and decoding-time diversification controls.

Abstract: Language models act through tools, yet practical agents face libraries containing thousands of interfaces. We introduce the tool menu as the short, ordered subset of available tools shown to an agent before execution. The agent can call only tools in this menu. Multi-step tasks require the final action and the prerequisite tools that create its inputs in a usable order. Current constructors rank tools by request relevance, which can surface the final action while omitting or delaying less obvious producers. We introduce the state path, a pre-execution route from the observable request state to the desired outcome, and propose State-Path Tool Menu to learn it. Our framework treats the menu as an execution prior over these routes. Its encoder represents which tools can run from the current state, how their outputs satisfy later inputs, and which orders recur in training paths. A retriever covers an executable entry, the missing-input producers, and the final action. A reranker then places producers before consumers. On ToolBench, our menu raises online success from 0.737 to 0.898 and outperforms retrieval, reranking, generation, and routing baselines without changing the agent. The State-Path menu also covers more complete chains with 32 tools than the official list covers with 128, and its success gain persists across executor families with different model capacities. Our code is at https://github.com/Met2348/State-Path.

Abstract: Large Language Models (LLMs) demonstrate impressive performance across diverse NLP tasks, yet their ability to exhibit genuine contextual understanding remains uncertain. Traditional evaluation metrics such as perplexity, BiLingual Evaluation Understudy (BLEU), or surface-level accuracy fail to reveal how well LLMs extract, integrate, and reason over contextual information--a gap particularly critical in question answering, where models must align responses with contextually grounded knowledge rather than memorized associations. We propose a novel knowledge graph-based evaluation framework introducing Semantic Structural Similarity for KGs (S3KG), a hybrid similarity measure integrating structural and semantic similarity into a continuous evaluation score, alongside a diagnostic framework for categorizing reasoning errors. To validate this pipeline, we evaluate S3KG against established metrics on a curated question-answer (QA) benchmark, demonstrating its effectiveness in measuring correctness, faithfulness, and interpretability in LLM-generated responses.

Abstract: This study examines the compositionality of steering vectors for language and behavioral control in large language models. Focusing on language, jailbreak, and conciseness, we investigate whether additive, training-free composition of attribute steering vectors can preserve the intended steering effect of each attribute, across four instruction-tuned models from two model families and two size scales. We find that single-attribute steering is reliable for all three attributes, but only within an appropriate combination of intervention layer and steering strength, with abstract behaviors (jailbreak, conciseness) favoring middle layers and language favoring earlier layers. We show that additive composition of two attribute vectors succeeds in steering both attributes simultaneously when each is injected at its own best-performing layer, and that this partially extends to three simultaneously composed attributes, addressing an inconsistency left open by prior work on training-free composition. We further analyze the geometric properties of these steering vectors, finding that they are approximately orthogonal in the residual stream, consistent with their compositional behavior.

Abstract: Vision-language models are increasingly used in settings where some input modalities may be unavailable, yet we know little about whether they can faithfully explain how such missing information affects their own predictions. We introduce an interventional protocol for evaluating self-explanations of modality dynamics: models state what each modality alone would support, whether restoring a missing modality would change their answer, and whether the available evidence is sufficient; we then execute the corresponding modality intervention and compare these claims with the model's realized behavior. We evaluate eight open-weight VLMs from two model families across four tasks spanning complementary and isomorphic text-image settings and a multi-view driving setting. We find a systematic tendency to overstate the sufficiency of available modality evidence. Models substantially underestimate the effect of restoring missing modalities: task-level median predicted change rates are at most 8.8%, while the corresponding executed change rates reach 72.1%, with underprediction in 62 of 64 model-task-condition settings. Insufficiency claims are rare, but precise when produced: restoring the modality changes the answer in a median of 78-100% of flagged cases. Retrospective self-explanations show the same tendency: on complementary data, models over-credit single-modality sufficiency; on isomorphic data, they over-credit single representation sufficiency relative to their executed behavior. Together, these results show that VLMs systematically mischaracterize how their predictions depend on available and missing modality evidence, motivating executable interventions as a behavioral ground truth for evaluating multimodal self-explanations.

Abstract: Large language models (LLMs) are often post-trained on pre-collected reasoning trajectories to improve their reasoning capability. Such trajectories tend to be long due to complex, interwoven paths, which often include detours on the path toward the answer. However, it has been underexplored whether LLMs indeed benefit from learning complete trajectories in post-training, such as supervised fine-tuning (SFT). Starting from our pilot study, we find that full trajectories provide only limited benefit, while partial trajectories are effective even under heavy truncation. We analyze redundancy in reasoning trajectories through attention-based analyses and controlled token-removal studies, both of which show that intermediate tokens contribute minimally to final reasoning quality. This suggests that avoiding redundant information may allow LLMs to internally infer coherent alternatives by inferring missing steps from their internal knowledge, given known trajectory endpoints. Furthermore, we show that training LLMs using endpoints leads to consistent changes in reasoning behavior, and that it also benefits post-training methods based on reinforcement learning or on-policy distillation, highlighting the need to revisit complete reasoning traces. Code is available at https://github.com/naver-ai/revisiting-trace.

Abstract: Language models may need to internalize information that arrives over time and retain it through many subsequent updates. To study this challenge, we introduce long-horizon memorization, a setting in which a model learns 100 query-answer tasks through continual supervised fine-tuning without retaining earlier training examples or receiving task identifiers at inference. Sequential updates cause catastrophic forgetting, and no single continual learning mechanism we evaluate maintains strong retention at this horizon. We hypothesize that mechanisms addressing complementary sources of forgetting will be more effective when composed. We organize these compositions along two design dimensions. Data, function, and weight anchors specify what prior information each update should preserve, while low-rank allocation rules determine where successive updates are retained. To test this hypothesis systematically, we construct three distinct 100-task memorization datasets. We introduce task-level successive halving to search the combinatorial design space and use a factorial experiment to measure individual and interaction effects. Our best method combines all three anchors with merged LoRA, ranks among the top 3 methods in all datasets, and raises average final retention from 1.2% under naive sequential fine-tuning to 34.9%, a 28-fold improvement. The data anchor and merged LoRA provide the largest average gains and interact super-additively on all three datasets. Together, these results show that composing complementary mechanisms substantially improves long-horizon memorization beyond what any individual mechanism achieves.