# Research Papers — 2026-09 · bundle page 13 of 266

> Full-text markdown of papers #601–650 of 13298 published in 2026-09, newest first. AI-readable mirrors of the full 13298-paper 2026-09 archive (50 papers per file).

Canonical page: https://aipod.dev/papers/archive/2026-09/page/13
[← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/12.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-09/page/14.md)

### Mitigating the Length-Scaling Tax with Online Distillation

Authors: Xu Wan, Wenyue Xu, Shengjie Zhao, Mingyang Sun
Published: 2026-09-30
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.38854v1

Length scaling during reinforcement-learning (RL) post-training is often viewed as a sign of improved reasoning ability, especially on difficult problems, but may also make responses to already-solved problems unnecessarily verbose. We quantify this side effect as the length-scaling tax (LST): excess response length on already-solved queries without a commensurate accuracy gain. To mitigate LST, we propose Length Self-Distillation (LSD), which routes solved prompts to on-policy distillation and retains the original RL objective for unsolved prompts. LSD uses an exponential moving average of the online policy as its teacher, requiring no external model. We find that LSD achieves comparable or better performance than RL across multiple variants, while substantially curbing response-length growth on easy queries. LSD reduces LST from 19.0% to -3.7% on single-turn reasoning and from 31.4% to 13.7% on multi-turn agentic tasks, demonstrating that LSD effectively preserves concise response patterns on easy queries while supporting efficient exploration on difficult queries during RL post-training.

---

### Visualizing Distribution Coverage in Generative Diffusion Models

Authors: Yifei Wang, Xiaoyu Wu, Tsu-Jui Fu, Chen Chen, Liang-Chieh Chen, Zhe Gan, Chen Wei
Published: 2026-09-30
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.38853v1

Diffusion distillation is widely adopted to accelerate sampling, and the resulting few-step models are broadly believed to match or even surpass their multi-step teachers in generation. However, standard evaluations such as GenEval2 typically draw only one sample per prompt, so improved scores may fail to reveal losses in distribution coverage. We therefore revisit whether distilled models truly match their teachers beyond single-draw performance using \textbf{pass@$\mathbf{k}$}, which measures the probability that at least one of $k$ independent samples satisfies a quality criterion. At $k{=}1$, pass@$k$ reduces to standard single-draw evaluation. As $k$ grows, the curve reveals whether additional draws find genuinely different successes or merely revisit the same modes, directly exposing how broadly a model covers the space of valid outputs. We first show that classifier-free guidance (CFG), whose quality--coverage tradeoff is well established, is the clearest case: higher guidance improves pass@$1$, but its advantage shrinks and reverses at larger $k$. Applying pass@$k$ to few-step distilled models, we find the same tradeoff splits along training objectives: distribution-matching objectives concentrate the student's output distribution, boosting early-hit rates while eroding large-budget coverage, whereas consistency and trajectory-based objectives better preserve the teacher's coverage even at large $k$. We further show that this tradeoff extends to few-step causal video generation. Our findings reveal a previously overlooked cost of diffusion distillation: across both image and video generation, the choice of training objective fundamentally determines whether a few-step model inherits its teacher's distribution coverage or trades it away for single-draw quality.

---

### Where MLLMs Fail and Why: Causal Task Decomposition for Capability Failure Diagnosis

Authors: Xia Hu, Brian Potetz, Chun-Ta Lu, Huanfen Yao, Leonidas Guibas, Zhicheng Wang, Howard Zhou, Pengfei Xing, Andrew Gallagher
Published: 2026-09-30
Categories: cs.CL, cs.AI, cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2609.38851v1

End-to-end accuracy on compositional tasks records how often MLLMs fail, but cannot distinguish whether a failure reflects an intrinsic deficit in the targeted capability or a cascading error from an upstream prerequisite. We propose a causal decomposition framework that isolates these two failure modes through controlled interventions on the prerequisite dependencies of each task. Our capability metrics (NC, IC, RC) score each task under unassisted, correct, or incorrect prerequisites to diagnose where failures arise; contribution metrics (N-Score, S-Score), adapted from probabilities of causation, quantify each prerequisite's necessity and sufficiency to determine why. We instantiate the framework in CADET, a diagnostic benchmark of 10 composite tasks decomposed into 46 unit tasks with over 33,000 human-annotated questions spanning perception, spatial, temporal, and cognitive categories. Diagnosing frontier MLLMs with our framework uncovers systematic patterns that end-to-end accuracy obscures. Capability-wise, supplying correct prerequisites eliminates 54\% of errors on cognitive tasks, lifting them from weakest to above spatial and temporal. Prerequisite-wise, causal contributions are concentrated in a few critical prerequisites, and supplying the single most important one alone captures 84\% of the gain from supplying all prerequisites.

---

### OpenJev-RLCD: A Working RLCD Implementation

Authors: Zhimin Gao, Pichao Wang
Published: 2026-09-30
Categories: cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2609.38850v1

Decision models such as Jev answer questions with probabilities, which are only useful if they are calibrated. Open-source reproductions rely on supervised fine-tuning plus temperature scaling, while reinforcement learning from verifiable rewards (RLVR) makes reasoning models overconfident. We present a working implementation of reinforcement learning for calibrated decisions (RLCD) for reasoning models: the model samples a rationale, and we score the answer distribution it commits to afterwards with a strictly proper scoring rule. A variance identity shows that scoring the mixture of several samples rewards disagreeing rationales, and that RLVR is exactly this mixture objective without its diversity term. Optimized naively, the per-rationale objective either switches reasoning off or is drowned out by policy-gradient noise, which leads to a two-stage recipe: calibrate, then reinforce. With Qwen3-1.7B on two reasoning tasks (3 seeds, paired tests), RLCD matches or beats SFT, RFT/STaR and GRPO (each temperature-scaled) in accuracy and beats all of them in selective prediction; on GSM8K answer verification a single query decides \gvTwoCovFive\% of the items at $\le$5\% error, versus \gvGrpoCovFive\% for GRPO. When uncertainty comes from annotator disagreement, RLCD provably cannot beat cross-entropy. Code and results: https://github.com/ZimmyGao/openjev-rlcd.

---

### Scoring Higher, Answering Worse: Mitigating Reward Hacking in Rubric-Based RL via Protocol-Level Rubrics

Authors: Maoqi Liu, Junwei He, Bowen Zhang, Feiran Li, Wentao Ma, Rongyi Lin, Shuhan Zhong, Quan Fang
Published: 2026-09-30
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.38847v1

Rubric-based reinforcement learning (Rubric-RL) trains language models where no verifier exists. A judge checks each criterion of a rubric, and the verdicts are aggregated into a reward, most often by a weighted sum. We show that this additive aggregation is the weak point. Under a sum, criteria compensate for one another: a policy that misses the one decision that matters can buy the points back with advice nobody asked for. On clinical consultation, such a policy scores higher and answers worse. Rubric coverage rises while appropriateness on held-out physician criteria falls below the untrained model. The medical criteria are not to blame. Grouped so that they must hold together, the same criteria, unchanged to the word, recover a third of the loss; shorter answers recover almost none. We therefore propose Protocol-level Rubrics (ProRubric), which keeps what the criteria ask for and changes how they are aggregated. It groups a checklist into a few protocol-level dimensions. A dimension counts only when all of its criteria hold and its failure clause does not fire. The grouping is done once, offline, and leaves the optimizer unchanged. ProRubric raises appropriateness by 10.8 points without losing coverage and has the best seven-benchmark average at both scales. Reward validity is set not only by what a rubric verifies, but by how it aggregates. Code is available at https://github.com/Estrellajer/ProRubric

---

### Learn-Then-Differentiate Gradient Estimation

Authors: Nifei Lin, Qingkai Zhang, L. Jeff Hong
Published: 2026-09-30
Categories: stat.CO, cs.LG
arXiv: https://arxiv.org/abs/2609.38842v1

Learn-then-differentiate (LTD) estimates gradients by fitting a model to simulation outputs and differentiating it. We develop a unified framework explaining what LTD differentiates and how accurately it estimates gradients. For models with a weighted representation, LTD differentiates a learned representation of the underlying probability measure. We then show how accuracy guarantees for fitted models translate into guarantees for gradients and higher-order derivatives, with rates approaching the standard Monte Carlo rate under suitable smoothness conditions. The framework recovers established results for kernel regression, local polynomial regression, and kernel ridge regression, and yields further guarantees for multiple kernel learning and smooth neural networks. These results provide a common foundation for understanding and analyzing LTD across learning methods.

---

### scTrilemma: Balancing Identity, Invariance, and Fidelity in Single-Cell Representation Learning

Authors: Yunhak Oh, Yoonho Lee, Junseok Lee, Namkyeong Lee, Sang-Yeon Hwang, Yinhua Piao, Hyomin Kim, Seonghwan Kim, Jaechang Lim, Woo Youn Kim, Sungsoo Ahn, Chanyoung Park
Published: 2026-09-30
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.38840v1

Single-cell RNA-seq representation learning is fundamentally label-free: cell identities, states, and contexts are not fixed training targets, so what constitutes signal or nuisance is analysis-dependent. A single representation must therefore preserve biological identity and state, remain robust to nuisance context, and retain the gene-level variation needed for expression analysis, three demands we call the representation trilemma. To tackle this problem, we introduce scTrilemma, a latent-bottleneck VAE that routes expression-derived variation to the embedding, the decoder, or the prior rather than forcing all of it through one embedding. It gates gene tokens by expression, routes the cell representation through the decoder, and conditions the prior on unlabeled pseudo-bulk context, under a single reconstruction objective and without target annotations or auxiliary representation losses. In release-based zero-shot evaluation on successive CZ CELLxGENE Census releases, scTrilemma leads all three demands at once and preserves biological-state, differential-expression, and pathway structure across multiple disease settings. Latent interventions further show that context can be removed at almost no cost to the other demands, leaving identity against fidelity as the remaining tension. Code is publicly available at https://github.com/yunhak0/scTrilemma.

---

### FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation

Authors: Bo Yin, Xiaobin Hu, Jiaqi Zhao, Shuicheng Yan
Published: 2026-09-30
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.38839v1

Long-horizon video generation requires models to effectively leverage an increasingly long generation history. As the generated history grows, retaining all previous content becomes increasingly expensive and redundant, making effective historical selection essential. Existing approaches often determine historical relevance based on the current content. However, information relevant to the present is not necessarily useful for future generation, while seemingly less relevant history may become important later. Our key insight is that historical information should be selected according to its relevance to future information needs. Capturing these needs does not require generating the full future; instead, a compact representation of what becomes important next is sufficient to guide historical selection. Building on this insight, we propose FrameMorrow, a prospective frame selector that predicts a small set of prospective tokens representing future information needs and uses them to identify relevant information from history. FrameMorrow selects explicit historical frames rather than model-specific internal states, enabling plug-and-play integration across diverse generators, including closed-source models, with little additional inference cost. We evaluate FrameMorrow across five benchmarks and 11 generative models spanning long-video generation, interactive generation, and action-conditioned world models. Extensive experiments demonstrate consistent improvements in long-range consistency, visual quality, and action alignment across diverse generation settings.

---

### Average-and Last-Iterate Lower Bounds for Optimistic Matrix Mirror-Prox in Quantum Zero-Sum Games

Authors: Yiheng Su, Emmanouil-Vasileios Vlatakis-Gkaragkounis, Pucheng Xiong
Published: 2026-09-30
Categories: quant-ph, cs.GT, cs.LG, math.OC
arXiv: https://arxiv.org/abs/2609.38835v1

Optimistic matrix mirror-prox (OMMP) computes $ε$-approximate Nash equilibria in quantum zero-sum games with an $O(1/\varepsilon)$ average-iterate guarantee (arXiv:2311.10859). We investigate whether this dependence on accuracy is tight and whether geometric last-iterate convergence can be guaranteed. We study these questions through explicit games with one qubit per player. First, we prove an $Ω(1/\varepsilon)$ lower bound for the uniform-average output that includes the maximally mixed initial state, independently of the regularizer and step size. Second, we construct a fixed game on which optimistic gradient descent-ascent (OGDA), initialized at the maximally mixed state, has last-iterate Frobenius distance to equilibrium $Θ(1/t)$ and duality gap $Θ(1/t^3)$ for every sufficiently small fixed step size. A separate fixed game exhibits arbitrarily long delays in reducing the initial error by a constant factor across a family of initial states. Finally, we give a fixed game with a unique, strictly complementary equilibrium on which optimistic matrix multiplicative weights updates (OMMWU) converge only polynomially from the maximally mixed state for every fixed positive step size. The last-iterate Frobenius distance and quantum relative entropy from the equilibrium to the iterates decay as $Θ(1/t)$, while the duality gap decays as $Θ(1/t^2)$.

---

### Optimal VC Dimension of Contrastive Learning with Margin

Authors: Dionysis Arvanitakis, Vaggos Chatziafratis, Yiyuan Luo, Konstantin Makarychev
Published: 2026-09-30
Categories: cs.DS, cs.LG, stat.ML
arXiv: https://arxiv.org/abs/2609.38834v1

Contrastive learning is a successful paradigm for learning $d$-dimensional geometric representations from a collection of ``anchor--positive--negative'' triplets $(i,j^{+},k^{-})$, indicating that ``item $i$ is closer to $j$ than to $k$.'' Despite its success, understanding why contrastive learning leads to representations of high \textit{generalization} quality---beyond the often pessimistic predictions from PAC-learning---remains a central question. Recently, \citet*{alon2024optimal} proved that, for PAC-learning $d$-dimensional Euclidean representations of $n$-point datasets, $Θ(\min(nd, n^2))$ triplets are necessary and sufficient, while they posed as an open question whether their VC dimension bounds for the more realistic setting of \textit{contrastive learning with a margin} can be improved. For a margin parameter $α>0$, a triplet $(i,j^{+},k^{-})_α$ is satisfied by the embedding $φ:(n)\rightarrow \mathbb{R}^{d}$, if $\|φ(i)-φ(k)\|_2>(1+α)\cdot\|φ(i)-φ(j)\|_2$. In this work, we resolve their question by proving that the VC dimension of contrastive learning under any margin $α\in(0,1)$ is in fact $O(n/α^2)$, improving on the previous bound of $O(n\log(n)/α^2)$. We also establish that the bounds are optimal up to constant factors, by providing a matching lower bound of $Ω(\frac{n}{α^2})$ (the previously known lower bound was $Ω(\frac{n}α)$), for $α\geq \max(n^{-1/2},d^{-1/2})$.

---

### ReSCENE: Server-Side Replay for Structural Mitigation of Catastrophic Forgetting in Federated Continual Learning

Authors: Sungmin Kang, Zhengzhong Tu, Sunwoo Lee
Published: 2026-09-30
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.38833v1

Federated continual learning must integrate new tasks over time without losing earlier-task knowledge. Most existing methods attach an anti-forgetting mechanism to the client-trained, server-aggregated loop of federated learning, which holds back new learning to preserve earlier knowledge and burdens resource-constrained clients. We propose ReSCENE, which structurally mitigates catastrophic forgetting by having each client upload a small condensed surrogate of its local data while the server keeps the surrogates of past tasks and trains the global model on them together with the current task surrogates. For efficient server memory, we introduce temporal herding, which selects the more recent surrogates from the pool accumulated over a task into a compressed buffer. Our study provides a theoretical analysis showing that this buffer can represent the original task data more closely than full accumulation of all surrogates. Across CIFAR-10, CIFAR-100, and TinyImageNet, ReSCENE achieves the strongest accuracy over seven baselines, by up to $31.1$ points of average accuracy, while requiring as little as $0.11\times$ of the client computation and up to $179\times$ less upload than the model-update baselines. ReSCENE further demonstrates its effectiveness when scaled to larger client populations and larger models while remaining efficient, which makes it a practical method for federated continual learning.

---

### Scaling Parameter and Context in Attention: Native Sparse Attention from Mixture-of-Head

Authors: Zizhuo Fu, Runsheng Wang, Meng Li
Published: 2026-09-30
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.38832v1

Scaling attention parameters can improve language model quality, but retaining full token histories makes additional heads costly at long contexts. Furthermore, since attention retrieves and combines contextual information, parameter scaling should also support longer contexts. We therefore ask whether attention parameter scaling can directly enable efficient and effective context scaling. We introduce NAMOH, an architecture-native sparse attention mechanism that activates $K$ of $H$ heads per token. Each head retains only its assigned tokens and performs causal attention within this subsequence. Head selection thus jointly determines active parameters and available context without scanning the full history. Under balanced assignments, increasing $H$ at fixed $K$ shortens head histories and reduces per-token key-value (KV) access without increasing total KV storage. We further support head-relative rotary position embeddings to shorten positional spans within routed subsequences, aiming to mitigate position-induced attention noise. Experiments show that NAMOH can outperform fully activated models with the same total parameters, while enabling more efficient long-context inference than smaller dense models with matched active parameter counts. It remains compatible with GQA and existing sparse attention mechanisms. We hope this work offers a new path for scaling attention, with parameter scaling directly enabling context scaling.

---

### Forging LLM Authorship Fingerprints with Targeted Rewriting

Authors: Haohan Yuan, Simin Chen, Xi Niu, Hanqing Guo, Depeng Xu, Haopeng Zhang
Published: 2026-09-30
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.38831v1

Model-attribution classifiers can often identify which language model produced a text, making model-specific writing patterns a signal of provenance. Accurate attribution on unmodified text, however, does not show whether the prediction still identifies the original source after deliberate rewriting. We formulate this problem as targeted fingerprint transfer: rewriting one model's output so that attribution classifiers assign it to a chosen target model. We study summarization, where different models receive the same document and express the same underlying content, providing a controlled setting for conditional generation. We introduce ForgePrint, a search-then-distil framework that first searches for rewrites that move attribution toward a target fingerprint, then distils the selected rewrites into a one-pass 4B Student model. On CNN/DM, the Student reaches 70.2% target success rate, outperforming both its Teacher (54.1%) and the strongest of six published rewriting baselines (39.3%), against held-out classifiers that are never queried by the attack. It also reaches 68.3% target success when transferring summaries from an open model toward chosen commercial models. These results show that fingerprint detectability should not be conflated with source authenticity, and that text-only attribution can provide misleading evidence of model identity under targeted rewriting, even when it is accurate on unmodified text.

---

### SparLeak: Privacy Leakage from Sparse Attention in LLM Inference on Shared GPUs

Authors: Fahao Chen, Linkang Du, Jinhao Zhou, Peng Li, Zhou Su
Published: 2026-09-30
Categories: cs.LG, cs.CR
arXiv: https://arxiv.org/abs/2609.38830v1

Sparse attention is widely used to accelerate long-context inference in modern large language models (LLMs), but its input-dependent execution behavior introduces previously unexplored privacy risks. We identify a new GPU micro-architectural side channel, termed Sparsity-Induced Memory Access (SIMA), which arises from secret-dependent key-value cache access patterns induced by sparse attention. Based on this observation, we present SparLeak, a phase-aware side-channel attack that extracts SIMA traces during LLM inference and enables two practical privacy extractions: query attribute inference from prefill-phase traces and autoregressive response reconstruction from decoding-phase traces. By reconstructing approximate token-level sparsity profiles from page-level observations and applying profiling-based learning, SparLeak accurately recovers sensitive information, including user-query attributes and private LLM response content. Extensive evaluation across three LLM architectures, three sparse attention mechanisms, and three privacy-sensitive datasets shows that SparLeak achieves average attack success rates of 90.9% for attribute inference and 87.3% for response reconstruction under real-world LLM serving settings, highlighting the significance to account for SIMA leakage when deploying sparse-attention-based LLM systems. We provide anonymized SIMA traces, trained attack models, evaluation scripts, and documentation as artifacts at https://anonymous.4open.science/r/Janus_artifacts/.

---

### Diversity Combining for Multi-Path LLM Reasoning

Authors: Guangsheng Yu, Litianyi Zhang, Qin Wang, Xu Wang, Mingyuan Li, Shaoxiong Ji, Ren Ping Liu, Massimo Piccardi
Published: 2026-09-30
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.38829v1

Multi-path reasoning methods such as self-consistency (SC) sample $K$ reasoning paths and choose the most frequent answer. However, their gains quickly plateau as $K$ increases, and existing methods do not predict when this saturation will occur. We formalize multi-path LLM reasoning as a diversity combining problem from wireless communications: each path is a noisy channel observation, and the pairwise correlation of path correctness caps the design-effect effective sample size of the vote at a finite ceiling. Generalized least squares (GLS) analysis shows that, under exchangeability, the optimal symmetric linear combiner of latent embeddings is uniform, supporting majority vote as the natural default in standard SC while leaving room for weighting or pruning under heterogeneous prompt-template branches. Across 5 models and 12 benchmarks, prompt-template diversity reduces path correlation in $55$ of $57$ valid cells, with the strongest effect on open-ended QA. We derive an Adaptive-K rule that uses a four-path pilot to select $K^*$, retaining $96$--$103\%$ of MV@$K{=}32$ accuracy across Math, QA, and NLU.

---

### More Choices, Fewer Decisions: Ordinal-Scale Bias in JEV-like Direct-Decision Models

Authors: Tianxiang Gao, Jinzhe Li, Zhiyuan Li, Yi Chang, Yuan Wu
Published: 2026-09-30
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.38827v1

Direct-decision models turn text into low-latency structured labels and scores, making them attractive for classification and automatic evaluation. Yet reliability requires more than accuracy: a model must also use the ordinal decision scale supplied by the user faithfully. We analyze JEV~1.13 and three open KEV models. Our investigation begins with ANLI, where JEV assigns 38.8\% of all predictions and 51.3\% of errors to Neutral despite 74.95\% accuracy, nearly balanced gold labels, and balanced candidate positions. Across 36 ordinal datasets, final decisions use only 67--76\% of the effective gold support, versus 87--102\% on four nominal tasks. Randomizing candidate order weakens but does not remove this compression. Holding items and source scores fixed while balancing gold support and positions, we refine scales from $K=2$ to $14$; utilization falls for every model and reaches 26--75\% at $K=14$, although candidate probabilities remain broad for most models. Targeted BA-LoRA post-training raises gold-relative utilization from roughly 47\% to 86\% on eight supervised scales at both KEV sizes, showing that the compression is learned and modifiable rather than an immutable architectural limit. We call this ordinal scale-utilization bias: decision-stage candidate-space compression distinct from accuracy, gold imbalance, fixed position, and candidate count alone. The code and data are available at https://github.com/Glax147/jev_ordinal_scale_bia

---

### DecoMoE: Decoupling Visual Propagation and Expert Computation for Efficient Multimodal MoE Inference

Authors: Xudong Tan, Peng Ye, Ming Xie, Chenyu Huang, Yaoxin Yang, Jiayuan Fan, Tao Chen
Published: 2026-09-30
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.38823v1

Multimodal mixture-of-experts (MoE) models combine sparse expert activation with visual-language capabilities, yet their inference remains costly because long visual-token sequences repeatedly incur attention, routing, dispatch, and expert-MLP computation. Existing methods typically compress either the token or expert dimension, leaving redundancy along the other. Our analysis reveals two complementary regularities: the depth required for visual propagation varies across inputs, while text-token routing exhibits concentrated and recurrent expert-importance patterns. Based on these observations, we propose DecoMoE, a two-dimensional structured compression framework that decouples visual propagation from expert computation. The Sample-Adaptive Visual Boundary (SAVB) predicts an input-dependent visual-exit layer at which the visual-token block is removed. The Routing-Calibrated Expert Prefix (RCEP) reorders experts offline using text-token routed mass and, from this predicted exit layer onward, retains at each MoE layer the shortest contiguous prefix covering a target routed-mass fraction. We evaluate DecoMoE on Qwen3-VL-MoE and InternVL3.5-30B-A3B across six benchmarks. On Qwen3-VL-MoE, DecoMoE retains 97.91% of dense-baseline performance while reducing computation from 27.06 to 16.73 TFLOPs and latency from 0.44 to 0.26 seconds, yielding a 1.69x speedup. Code will be available at https://github.com/ShawnTan86/DecoMoE.

---

### SkillSeek: Revisiting Agent Skill Retrieval at Marketplace Scale

Authors: Guanqun Yang, Wenlong Zhang, Tian Shi, Ping Wang
Published: 2026-09-30
Categories: cs.IR, cs.AI, cs.MA
arXiv: https://arxiv.org/abs/2609.38822v1

Anthropic's Agent Skills package reusable procedural know-how for an LLM agent into SKILL.md directories, and open-source aggregations have grown past 230,000 skills, making selection rather than authoring the bottleneck. The standing answer in the literature outsources selection to the agent itself: an LLM-mediated retrieval loop that rewrites queries and refines candidates inside the agent's decision loop, paying LLM tokens on every task. We present SkillSeek, an open-source two-stage skill retriever built from the standard IR recipe (a BGE-base bi-encoder feeding a small cross-encoder, exposed over MCP). Across a $4 \times 11$ grid of pool, backbone, and method on the 89-task SkillsBench benchmark, SkillSeek reaches observed parity with the LLM-mediated loop of Liu et al. at essentially no extra cost: plain bm25 alone records a pass rate at or above their refined loop on three of four settings, and a small cross-encoder covers the remaining difference on the fourth. A first-stage recall ceiling explains the pattern, and total per-trial spend drops from USD 51.30 to USD 27.54 (within fifty cents of the no-skill baseline). Under the SkillsBench tasks and OpenHands harness we tested, this positions the standard IR recipe as a strong default for agent-skill retrieval, with LLM-mediated alternatives a natural fit for cases where deterministic methods fall short.

---

### BARRAC: Adaptation of an English Aspect-based Sentiment Analysis Approach for Classification Tasks in Arabic Dialects

Authors: Ali Almutairi, Gelareh Mohammadi, Imran Razzak, Aditya Joshi
Published: 2026-09-30
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2609.38820v1

With the rapid growth of Arabic NLP, several models, datasets and benchmarks have been reported. This paper asks whether approaches developed for majority languages like English can be adapted to Arabic tasks. We adapt an English aspect-based sentiment analysis framework to Arabic classification tasks and present the adaptation as BARRAC: Brainstorming Alignment and Replaced Representation learning for ArabiC tasks. BARRAC replaces consumer-review attribute pools with Arabic linguistic devices and markers for dialectal sentiment, sarcasm, and dialect identification, and replaces noisy self-training with two-stage training. Evaluated on five Arabic dialect datasets, BARRAC achieves a mean macro-F1 of 63.93\%, outperforming the best few-label SOTA by 3\%, and outperforming GPT-4o on four out of five tasks. Error analysis provides insights into remaining challenges. These results demonstrate that adapting task-specific approaches is a promising direction for Arabic NLP alongside adapting models, datasets and benchmarks.

---

### Future Video Generation Better Aligns with the Human Visual Cortex than Observed Video

Authors: Chang-Bae Bang, Hyungjin Chung, Byung-Hoon Kim
Published: 2026-09-30
Categories: cs.CV, cs.AI, q-bio.NC
arXiv: https://arxiv.org/abs/2609.38819v1

Studying the alignment between the internal representations of vision models and the responses of the visual cortex to the same observed visual stimuli has enabled us to better understand human visual processing. However, studies so far have largely overlooked the fact that the human brain not only processes observed visual stimuli, but also predicts upcoming stimuli based on what has been observed. Accordingly, we hypothesize that internal representations for generating future video frames are better aligned with the predictive nature of human visual processing than representations of the observed video itself. To this end, we compare the alignment between human video-watching fMRI responses in the visual cortex and the internal representations from two types of video diffusion models, an autoregressive (AR) model and its non-AR base model. We first conduct a within-model analysis of the AR video diffusion model and show that the representations for future video generation align better with the visual cortex than the representations of the observed video. We then compare the internal representations of the AR model with those of its non-AR base model and again show that the representations for future video generation align better with the visual cortex than the representations for observed video reconstruction by the base model. Specifically, the alignment of observed video reconstruction is concentrated in lower-order visual cortex, whereas that of future video generation is concentrated in higher-order visual cortex. Finally, we show in a human behavioral experiment that humans prefer videos generated by amplifying the contributions of individual layers that align better with the visual cortex.

---

### Whose Voice Survives the Summary? A Voice-Retention Audit of LLM Employee Listening

Authors: Thilo Tamme, Anton Hantel, Bijan Khosrawi-Rad
Published: 2026-09-30
Categories: cs.AI, cs.CL, cs.CY, cs.HC
arXiv: https://arxiv.org/abs/2609.38818v1

Organizations increasingly route employee feedback to leaders through large language model (LLM) summaries, an unaudited layer that silences already-spoken voice. We introduce a Voice Retention / Representation Ratio metric for representational bias in summarization and apply it to a bilingual (English/German) corpus of 2,586 free-text responses from a global professional service company. First, employees supply criticism more reliably than praise (withholding praise is 82 times more common). Second, across 45 leader-summaries the pipeline filters by popularity, not sentiment: criticism survives, yet a concern voiced once is dropped 86% of the time, with short and German-only content lost on the same axis (theme retention 0.14 vs 0.74; German directional). Controlling for frequency, sentiment has no independent effect; the harm is prevalence-driven, which sentiment-only audits miss. A targeted prompt recovers only named themes. We contribute the metric, field evidence, and a disaggregated voice-retention card.

---

### When Reasoning Goes Astray: Attention Dynamics of Uncontrolled Reasoning

Authors: Yuanhe Zhang, Ziwei Wang, Jie Ren, Haoran Gao, Zhenhong Zhou, Fanyu Meng, Cong Wu, Li Sun, Sen Su
Published: 2026-09-30
Categories: cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2609.38817v1

Large reasoning models (LRMs) improve performance on complex tasks through extended reasoning, yet the same process can degenerate into redundant verification and persistent generation loops. Such uncontrolled reasoning increases inference cost and creates risks of resource exhaustion and service degradation. However, existing mitigations largely truncate long outputs or react to surface repetition, and thus fail to distinguish normal thinking from uncontrolled reasoning or explain how benign reasoning degenerates into harmful behavior. In this paper, we operationalize LRM generation as four states and further introduce Reasoning-state Analysis via Dynamic Attention Responses (RADAR), which identifies the current reasoning state in real time and characterizes how effective reflection can develop into uncontrolled generation. Guided by RADAR's analysis, we further realign abnormal attention distributions toward patterns observed in normal requests and examine how this correction affects excessive reflection and persistent looping. Temporal analyses show that uncontrolled reasoning is characterized by attention distributions that deviate from normal generation, with abnormal trends becoming detectable before repetition begins. Correcting these deviations through Attention Realignment consistently reduces looping while largely preserving benign performance. Together, RADAR provide a mechanistic account of how reasoning becomes uncontrolled, offering actionable guidance for identifying critical failure stages and designing targeted runtime interventions.

---

### You're Hired: Strategic Model Selection for LLM Collaboration

Authors: Zongwan Cao, Ziyuan Yang, Shangbin Feng, Michael Duan, Skyler Hallinan, Bingbing Wen, Lucy Lu Wang, Yulia Tsvetkov
Published: 2026-09-30
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.38816v1

While multi-agent and model collaboration algorithms gain traction to combine the strengths of diverse Large Language Models (LLMs), existing systems remain bottlenecked on pre-defined and hand-crafted model pools. In this work, we investigate the problem of model selection in multi-LLM systems. We propose and systematically evaluate a taxonomy of 9 selection algorithms ranging from diversity of model descriptions, capability-aware behavioral diversity, and LLM-based recruiters. We conduct extensive experiments across two candidate pools of 10 and 32 models, deployed in four model collaboration algorithms, and evaluated across tasks spanning math, coding, QA, and reasoning. Results demonstrate that successful selection algorithms greatly outperform random or heuristics-based teams such as merely selecting the models with top individual performance, by up to 36.1% across settings. Specifically, capability- and training-based selection strategies alleviate selection variance and achieve the best performance, which we recommend to employ before deploying real-world multi-LLM systems. Further analysis reveals that larger candidate pools pose greater challenges to shallow selection heuristics, while algorithms grounded in interacting with candidate models and understanding model capability robustly filter out misaligned, unsafe models, as well as generalizing to novel, out-of-distribution tasks. Together, we establish that principled and informed team selection is critical and present strong model selection algorithms for assembling effective multi-LLM systems.

---

### Learning Chaos Without Seeing Chaos: Extrapolation of Global Dynamics in Autoregressive Transformers

Authors: Yilun Liu, Yi Zhang, Ganyu Wu, Sikuan Yan, Mengyue Wang, Alois Knoll, Volker Tresp, Yunpu Ma
Published: 2026-09-30
Categories: cs.LG, nlin.CD
arXiv: https://arxiv.org/abs/2609.38814v1

Autoregressive models are trained to predict a system's behavior one step at a time, and recursive generation allows the learned dynamics to unfold over long horizons. To what extent can such dynamics learned from local observations recover broader organization of an underlying system that was only partially observed during training? Here we study small autoregressive transformers trained from scratch on trajectories sampled from restricted parameter regimes of several non-linear dynamical systems, including logistic and sine maps, the Lorenz system, and the generalized Hopf system, with control parameters and state trajectories represented as sequences of continuous tokens. Under closed-loop evaluation at parameters far outside the training distribution, the models can recover self-similar period-doubling cascades, chaotic dynamics, and attractor structures with remarkable visual and numerical fidelity. For the logistic map, a transformer reproduces successive period doublings up to period 128, yielding a finite-order scaling ratio of 4.6687, matching the Feigenbaum constant to within $5\times10^{-4}$. We further investigate how these structures emerge over the course of training, and reveal with causal interventions how control-parameter information is processed through attention into state prediction and shapes the resulting closed-loop dynamics. These results suggest that a surprisingly narrow window into a system's local behavior may suffice for autoregressive transformers to generalize to its unseen global dynamical organization.

---

### Can Terminal Agents Trust Their Own Verification? Diagnosing and Improving Self-Verification

Authors: Yingfeng Luo, Shaowei Wei, Daixin Wang, Dingyang Lin, Kaiyan Chang, Weiqiao Shan, Tong Zheng, Zhiqiang Zhang, Jingbo Zhu, Tong Xiao
Published: 2026-09-30
Categories: cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2609.38812v1

Terminal agents rely on self-verification to assess and correct their solutions as they solve tasks through interaction with command-line environments. Yet how trustworthy such self-verification is remains poorly understood. To investigate this question systematically, we introduce a diagnostic framework that identifies the first complete solution in each trajectory, determines whether it is objectively correct, and uses this ground truth to quantify the agent's subsequent verification and recovery behavior. Applying it to ten terminal agents on TerminalBench2.1, we find that verification is nearly universal after a complete candidate is formed, yet only 61.43\% of incorrect candidates are detected and only 49.36\% of detected errors are successfully repaired. These results show that the main weakness in self-verification lies not in initiating verification, but in detecting and repairing errors. Motivated by these findings, we propose Student-Conditioned Verification Distillation (SCVD), which lets the student first produce a candidate solution and distills a stronger teacher's subsequent verification and recovery from the same interaction context. Across three Qwen3.5 backbones, SCVD improves \textsc{Pass@1} on TerminalBench2.1 by 9.74--16.85 percentage points over the corresponding base models and by 4.49--8.61 points over the standard full-trajectory distillation, while avoiding the pronounced out-of-distribution degradation of full-trajectory distillation on SWE-bench Verified.

---

### DCM-SAM: Defect-Conditioned Mixture of LoRA Experts for NPU-Deployed AM Defect Segmentation

Authors: Md Mushfiqur Rahaman, Md Mahedi Hasan, Imtiaz Ahmed, Srinjoy Das
Published: 2026-09-30
Categories: cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2609.38811v1

Metal additive manufacturing parts are inspected by X-ray computed tomography, where labelled data is scarce, the pores and inclusions that matter span a few pixels, and inspection must happen at the machine. We present DCM-SAM, a defect-conditioned adaptive mixture of LoRA experts: one frozen Segment Anything backbone carries a separate Conv-LoRA expert bank and mask decoder per defect class, each trained in its own pass, without prompts, on synthetic slices alone, updating only 4.4% of the parameters. On benchmarks that XCT-SAM reports, DCM-SAM improves on every baseline for both classes from a ViT-B backbone against their ViT-H, and reaches 64.2% pore IoU on real NIST scans having seen no real images during training. Deployment then exposes what adaptation work rarely measures: on a Qualcomm Hexagon NPU, ViT-H and ViT-L compile yet cannot allocate at 1024x1024 image resolution, since activations rather than weights exceed the device ceiling, and quantizing weights does not help. ViT-B alone runs, but the adapted encoder then fails to allocate where the stock one succeeds, until a numerically identical rewrite of the attention lets the complete DCM-SAM run in FP16 at 1024x1024, with no operator falling back to the CPU, masks within 0.01% of pixels of the FP32 reference. Code: https://github.com/MushfiqShovon/DCM-SAM.

---

### CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models

Authors: Chunzheng Zhu, Jiaqi Zeng, Hongbo Zhao, Yihang Chen, Yijun Wang, Jianxin Lin
Published: 2026-09-30
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2609.38810v1

As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at GitHub repository.

---

### StateTree: Enhancing Long-Term Dialogue Reasoning via Reinforcement Learning

Authors: Naen Xu, Wanqing Cui, Yibo Hu, Shixin Hong, Hengyu An, Meiguang Jin, Junfeng Ma, Tianyu Du
Published: 2026-09-30
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2609.38809v1

Large language models deployed as personalized assistants must reason over long, evolving interaction histories. However, in long-term dialogue reasoning, relevant evidence is scattered across sessions, preferences may be revised over time, and standard long-context training fails to address these challenges under data scarcity and prohibitive computational costs. We propose StateTree, a data-driven RL method that constructs a challenging auxiliary task from scarce dialogues with verifiable ground truth. StateTree augments multi-session dialogues with a tree-structured path-tracing task: key-value records are embedded across sessions to form a binary tree. Solving the task requires the model to traverse from root to leaf by retrieving records across sessions and comparing timestamps to resolve branches, then recover the hidden target question among distractor leaves. We apply curriculum RL training progressively increasing tree depth and introduce a compositional variant whose edges carry step-level reasoning fragments, training the model to compose partial cues into coherent queries. Trained on 10K-token contexts, StateTree generalizes to 128K tokens without full-length RL costs and exhibits capabilities including cross-session retrieval, temporal reasoning, knowledge update, and compositional multi-hop reasoning. StateTree outperforms both SFT and RL-based baselines while preserving short-context general reasoning. StateTree-7B achieves gains up to +23.60% on LongMemEval (128k), and StateTree-14B reaches 59.00% accuracy on LongMemEval, surpassing QwenLong-L1-32B (45.20%).

---

### Blackboard Intelligence Can Surpass Autoregressive on Globally Constrained Problems

Authors: Woosang Jeon, Jaeyeon Kim, Sham Kakade, Yilun Du, Amrit Singh Bedi, Arun Kumar Chithanar, Chul Lee, Taehyeong Kim, Sitan Chen
Published: 2026-09-30
Categories: cs.LG, cs.CL
arXiv: https://arxiv.org/abs/2609.38806v1

Next-token prediction has driven remarkable progress in large language models, yet a growing body of evidence suggests that they can struggle on problems governed by complex global constraints. In this work, we focus on this regime and ask whether some of these limitations arise from the inference interface induced by next-token prediction itself. We study this question through blackboard intelligence: an inference-time perspective in which a model works on a fixed, revisable canvas and searches over candidate solution states rather than committing to a causal, left-to-right trajectory. We instantiate this idea with diffusion language models, whose any-order prediction interface naturally exposes predictions over partially filled solution states. Our key observation is that mean confidence, a simple model-internal quantity available from the standard masked diffusion objective, provides a useful proxy for global coherence and can guide inference-time search and revision. Empirically, across ZebraLogic, Nurse Rostering, and Job-Shop Scheduling, Blackboard consistently improves inference while holding the fine-tuned LLaDA-8B-Instruct checkpoint fixed and substantially outperforms same-scale autoregressive baselines, reaching 90.4% accuracy on ZebraLogic-Hard, 76.4% exact feasibility on Nurse Rostering, and 80.2% optimality on JSSP. Stronger autoregressive search and refinement also fail to close the gap on ZebraLogic-Hard, while Blackboard surpasses tested frontier LLMs there and on JSSP despite their substantially greater scale and strong test-time reasoning. We open-source our codebase at https://github.com/jwoosang1/blackboard-intelligence.

---

### Explicit Trajectory Diversity for RL-Based Post-Training of LLM Agents

Authors: Huaiyu Fu, Heng Cao, Hao Wang, Jian Ya, Tao Chen
Published: 2026-09-30
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.38805v1

LLM agents often admit multiple high-quality solutions to the same task, differing in reasoning structure, tool-use pattern, or interaction trajectory. Yet existing notions of diversity in LLM post-training are mostly implicit, arising from general stochasticity and regularization mechanisms rather than explicitly targeting task-relevant behavioral variation. While such implicit diversity can be useful, it does not directly specify which forms of behavioral variation should be encouraged for a given task. In this work, we study explicit trajectory diversity in RL-based post-training for LLMs. Our key idea is to define diversity through user-specified, task-specific trajectory descriptors, which map each sampled trajectory to an interpretable behavioral representation, and then measure diversity as a set-level functional over the resulting descriptor matrix. Building on this formulation, we introduce Trajectory-guided Joint Policy Optimization(TJPO), a single-policy framework that optimizes explicit diversity over sampled trajectory groups, avoiding the need for population-based policy training, and instantiate it within group-based policy optimization through trajectory-level learning signals. This design makes the diversity objective both interpretable and controllable. Experiments on Sokoban and ALFWorld show that TJPO improves task-specific trajectory diversity while maintaining competitive task performance. Descriptor and trajectory analyses show that the learned variation follows the specified behavioral dimensions and includes distinct successful strategies. Extra experiment results suggest that explicitly shaping trajectory diversity can help LLM agents satisfy user requirements and remain effective when task conditions change.

---

### Uncovering Uncontrolled Repetition through Residual Stream Dynamics

Authors: Yuanhe Zhang, Xinyao Zhou, Haoran Gao, Yuyao Zhang, Zhenhong Zhou, Fanyu Meng, Li Sun, Sen Su
Published: 2026-09-30
Categories: cs.CL, cs.AI
arXiv: https://arxiv.org/abs/2609.38802v1

Uncontrolled repetition can prolong autoregressive generation in large language models (LLMs) and enable resource consumption attacks. Prior analyses of repetitive generation have identified strongly activated features in intermediate and late layers. However, how uncontrolled repetition activity emerges and develops before becoming prominent in these layers remains insufficiently understood. In this paper, we investigate this question primarily in large vision-language models (LVLMs), which support a richer set of uncontrolled repetitions through both visual and textual inputs. We propose Tokenwise Residual Comparison (TRC), a method that identifies and localizes anomalies associated with repetition from residual dynamics during generation. TRC compares attention and multilayer perceptron writes to the residual stream across generated tokens to identify patterns associated with repetition. It then selectively suppresses coordinates in the residual stream at the identified layer. Experiments show that TRC effectively mitigates uncontrolled repetition, reducing loop rates by 57\% on average. Our analysis further shows that repetition semantics emerge in shallow layers and propagate through the residual stream, disrupting normal representations. TRC also generalizes to large language models (LLMs) and large reasoning models (LRMs), where it consistently captures analogous repetition dynamics and achieves effective mitigation. Our work broadens the study of repetitive generation from its prominent internal representations to earlier opportunities for intervention, providing insights for mitigating resource consumption attacks.

---

### Overlap, Unique and Conflict: Can LLMs Extract What They Can Recognize?

Authors: Eftekhar Hossain, Santu Karmaker
Published: 2026-09-30
Categories: cs.CL
arXiv: https://arxiv.org/abs/2609.38799v1

Understanding multi-perspective alternative narratives requires identifying how their information agrees, conflicts, or differs across sources. Existing work on cross-text relations largely focuses on categorizing relations between predefined text pairs, such as entailment or contradiction, rather than directly extracting such information from full narratives. To address this gap, we introduce Overlap-Unique-Conflict (OUC) extraction, a cross-narrative task that extracts all overlapping, conflicting, and unique clauses from two narratives. To support this study, we construct a benchmark of approximately 22K narrative pairs and 140K OUC instances spanning factual, argumentative, and political discourse. Evaluating 14 open-source LLMs (0.6B-35B), we find that unique information is far easier to extract than overlap and conflict: the strongest model, Gemma-4-31B, reaches only 61.13% F1-score on overlap and 48.58% on conflict, against more than 75% on unique. Further diagnostic analysis reveals that this difficulty does not stem from relation recognition alone, but rather from a failure to pair and extract the corresponding clauses from full narratives, especially in smaller models. Nevertheless, learning these extractions with task-specific supervision narrows the gap considerably: a fine-tuned Qwen-3-8B gains 15-28% absolute over its baseline and surpasses models roughly four times its size (e.g., Qwen-3.6-35B) on several tasks. Even so, overlap and conflict remain well below satisfactory, leaving cross-narrative clause extraction an open challenge.

---

### GraphCert: Bootstrap Agentic Graph Reasoning with Certified Evidence Rubrics

Authors: Weiqi Jiang, Yuchen Ying, Rui Wang, Kaixuan Chen, Bingde Hu, Shunyu Liu, Yu Wang, Tongya Zheng
Published: 2026-09-30
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.38798v1

Graph agents extend large language models (LLMs) with the ability to actively explore and reason over knowledge graphs through multi-step interactions with graph tools. However, training capable graph agents typically requires large collections of question-answer pairs and reasoning trajectories, whose manual construction is costly and difficult to scale. Moreover, employing proprietary LLMs to generate such supervision further risks exposing sensitive graph data to external services. Therefore, we propose GraphCert to bootstrap agentic graph reasoning with certified evidence rubrics during post-training. Specifically, the Bootstrapped Graph Quizzer guided by generation controls produces graph-grounded QA pairs and marks supporting evidence, which undergo execution certification and semantic curation. The accepted evidence is then canonicalized into certified evidence rubrics that later reward Graph Solver evidence alignment alongside answer correctness during GRPO training. Experiments on five graph reasoning domains in GRBENCH demonstrate that GraphCert consistently outperforms substantially larger LLM agents and post-training method. Furthermore, our analysis demonstrates that the learned policy transfers robustly across heterogeneous graph domains, suggesting that GraphCert acquires reusable graph-reasoning capabilities rather than domain-specific patterns. These results establish executable self-certification as an effective approach to self-training compact graph reasoning agents. Our code will be made publicly available.

---

### Evaluating Persistent Calibration under Evolving Model Knowledge

Authors: Victor Wang, Thomas Hofweber, Mohit Bansal, Elias Stengel-Eskin
Published: 2026-09-30
Categories: cs.LG, cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2609.38797v1

As AI systems move from static repositories to agents that are capable of continual adaptation and learning, maintaining their trustworthiness means equipping the models backing them with the ability to produce confidence estimates that dynamically reflect their changing skills and knowledge. We introduce the problem of persistent calibration, which requires a confidence estimator to faithfully reflect the knowledge contained in a model as that knowledge changes, without recurring supervision. We operationalize this by examining persistent calibration across checkpoints of open models, asking whether confidence estimators trained on earlier checkpoints can generalize to later ones. Specifically, we aim to shed light on whether confidence is dependent on knowledge, a question with implications for the reliability of confidence estimates. To measure this relationship, we define and evaluate calibration on knowledge contrast sets: subsets containing questions that one checkpoint answers correctly and another checkpoint answers incorrectly, reflecting a change in knowledge. We show that both inference-time and fine-tuning methods fall short on contrast-set calibration compared to oracle methods trained on future checkpoints, even for methods that are well-calibrated on the full dataset. We provide evidence for the hypothesis that persistent calibration is challenging because there is a vast space of possible confidence functions that are well-calibrated on a given checkpoint, out of which only some rely on meta-knowledge features that would generalize to other checkpoints. Towards improving contrast-set calibration, we show that multi-checkpoint training helps, suggesting an avenue for identifying confidence features that remain robust across changing knowledge.

---

### Recovering Off-Policy Supervision for Speculative Decoding

Authors: Jungseob Lee, Chanjun Park, Sugyeong Eo, Hyeonseok Moon
Published: 2026-09-30
Categories: cs.CL, cs.CV, cs.LG
arXiv: https://arxiv.org/abs/2609.38795v1

Block drafters for speculative decoding are commonly trained on corpora written by external models, where a single off-policy token invalidates supervision for all subsequent slots in a block. Existing approaches discard these divergent slots, resulting in severe supervision loss. To resolve this problem while preserving the training corpus, we propose a rollout-based training framework that recovers full supervision through two complementary components. The first component, Anchor-Label Relabelling (ALR), replaces corpus labels with distributions from greedy target rollouts, restoring valid supervision across all predicted slots. The second component, In-Rollout Anchors (IRA), places draft blocks directly inside these rollouts to expose the drafter to target-generated context, reusing precomputed rollout features at no additional target cost. Across fixed vision-language and text corpora, our framework increases greedy accepted length by up to 36.5% over DFlash and consistently outperforms erasing baselines. Notably, a single epoch of our method surpasses the best erase schedules. After three epochs, it matches the acceptance length of training on target-regenerated responses. These results show that our framework provides an effective and compute-efficient approach for training speculative drafters on fixed corpora without modifying the original text. Code is available at https://github.com/js-lee-AI/ALR-IRA.

---

### Training LLM Judges from Language Feedback via Position-Selective Self-Distillation

Authors: Ilgee Hong, Changlong Yu, Zhenghao Xu, Xin Liu, Yuwei Zhang, Qin Lu, Bing Yin, Tuo Zhao
Published: 2026-09-30
Categories: cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2609.38792v1

We study training LLM judges from natural language feedback, especially for subjective tasks where the verdict depends strongly on which evaluation criteria the judge invokes and how it weighs them. The dominant approach, outcome-supervised RL (e.g., GRPO), credits every token in the rollout with a single scalar determined only by the accuracy of the final verdict, providing no separate credit at the criterion-choice tokens and ignoring the rich language feedback (e.g., preference rationales) that naturally accompanies preference labels. Self-Distillation (SD) is one natural way to use this language feedback: the same model, conditioned on this feedback, acts as a teacher providing dense, position-level supervision. However, not all positions carry equally useful signal. Using the per-position entropy shift between teacher and student, we identify two regimes: context sharpening, where the teacher concentrates probability on a particular feedback-aligned criterion expression, and context spreading, where the teacher distributes probability across multiple feedback-aligned alternatives. We interpret these patterns as follows: sharpening encourages memorization of a particular criterion expression, whereas spreading promotes semantic understanding by preserving these alternatives. Motivated by this asymmetry, we introduce position masking based on the entropy shift that retains the lower tail of the entropy-shift distribution. Experiments show that masking higher-entropy-shift positions improves out-of-distribution generalization over naive SD. The resulting self-distilled judges outperform judges trained with outcome-supervised RL by 2-9 percentage points on the evaluated subjective subcategories, while remaining competitive on objective ones.

---

### LEARN-TS: LLM-Enhanced Alignment and Reconstruction with Normality Guidance for Multivariate Time-Series Anomaly Detection

Authors: Jahyeob Koo, Kio Yun, Byoungmo Koo, Jun-Geol Baek
Published: 2026-09-30
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.38789v1

Reconstruction errors in multivariate time-series anomaly detection may not reliably distinguish abnormal behavior from benign deviations. Language-derived semantics offer complementary context, but existing multimodal approaches may rely on time-associated paired textual information that is difficult to obtain consistently and is not provided by standard multivariate time-series anomaly detection benchmarks. This setting poses two challenges: (1) conditioning masked reconstruction on window-specific semantics without exposing exact numerical targets or anomaly-specific cues, and (2) using a window-independent concept of normality as a complementary semantic reference rather than an independent anomaly detector. We propose LLM-Enhanced Alignment and Reconstruction with Normality Guidance for Time Series (LEARN-TS), which uses a frozen language model to construct two role-separated semantic representations without requiring temporally paired external text. Window-specific observation semantics encode temporal and cross-variable context without exact numerical values to guide channel-shared patch-masked reconstruction. A fixed, dataset-agnostic normality prompt provides a window-independent semantic reference for aligning normal representations and estimating normality discrepancy. At inference, masking each temporal patch once yields timestamp-level reconstruction evidence, conditionally modulated by discrepancy from a separate unmasked view. Across four benchmarks, LEARN-TS achieves the highest mean performance in 13 of 16 dataset-metric comparisons. Controlled ablations examine observation conditioning, joint normality alignment and scoring, and reference content, showing dataset-dependent ranking benefits and modest average gains from semantic over random references.

---

### Positive Ratings, Hidden Concerns: Employee Voice Disclosure in AI-Mediated Organizational Listening

Authors: Thilo Tamme, Michael Saatkamp, Alma Bonte, Daniel Weiss, Anton Hantel, Andrej Levin
Published: 2026-09-30
Categories: cs.AI, cs.HC
arXiv: https://arxiv.org/abs/2609.38788v1

Organizations started listening to employees through conversational AI agents alongside structured surveys. Little is known about what these channels change in what employees say when disclosure carries hierarchical risk. We report a field study inside a global management consulting firm whose process pairs a pre-survey with an adaptive AI voice interview on the same themes within one session. Across 44 first-session interviews (132 matched theme observations), 20-41% of sessions showed a favorable rating co-occurring with a substantive concern voiced later, depending on the favorability threshold. The Gioia analysis drew on 158 protective quotes from 65 eligible sessions. Disclosure rarely arrived unguarded: employees softened concerns, deflected accountability, and bounded how far they went, and this protective work tracked the perceived legitimacy of the listening structure. We develop a grounded model of bounded disclosure and derive four propositions for voice, channel and listening research. Silence, we argue, can persist inside expression.

---

### Same Loss, Different Gradients

Authors: Ningkang Peng, Xiaoqian Peng, Yifan He, Anjie Hu, Chao Tan, Peirong Ma, Yanhui Gu
Published: 2026-09-30
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.38786v1

Differentiable learning typically assumes that the scalar objective evaluated in the forward pass and the gradient supplied to the optimizer in the backward pass describe the same mathematical object. We show that this correspondence can fail when probabilistic objectives rely on finite special-function recurrences, custom backward rules, and numerical clipping. In high-dimensional von Mises-Fisher learning, real numerical implementations can produce identical forward scores and losses at the same learning state while supplying different gradients and following different optimization trajectories. We characterize the structure of this mismatch in finite-start Bessel recurrence and show that classwise radial mismatch can compose through probabilities into a locally nonconservative update field. Evaluating the accuracy of special-function values and derivatives separately is therefore insufficient to characterize the realized learning objective. Motivated by this observation, we introduce AR/FR, a fixed-depth analytic realization that constructs a potential and its derivative jointly, ensuring forward-backward coherence by construction. We establish a uniform cubic-order error bound relative to the exact Bessel ratio over the entire nonnegative concentration axis and propagate this guarantee to learning scores and objectives. As representation dimension increases, the original finite recurrence becomes sequentially deeper, whereas the worst-case AR/FR error guarantee tightens cubically, jointly providing coherence, certified fidelity, and fixed-depth computation. These results suggest that a differentiable numerical primitive is defined by both the values it realizes and the derivatives it actually supplies to the optimizer; together, they constitute the numerical realization of the learning algorithm.

---

### How Accurate Is Accurate Enough?

Authors: Ningkang Peng, Qianfeng Yu, Jingyang Mao, Xiaoqian Peng, Yanhui Gu
Published: 2026-09-30
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.38785v1

How accurate must a numerical approximation be within a learning system? Primitive error alone cannot answer this question: errors of the same magnitude can have very different consequences for losses, predictions, and gradients at different learning states. We study this question through the learning objective itself. The objective weights classwise numerical errors nonuniformly according to the current state, so the importance of an error depends not only on its magnitude but also on the class it affects and the weight that class receives. For softmax cross-entropy, we characterize this coupling between class weights and errors and derive the exact extrema of the signed loss change over pairings of fixed non-target probability and score-error multisets, with the target probability and target score error held fixed. Building on this structure, we establish finite-error guarantees that propagate primitive error to losses, probabilities, predictions, and feature gradients, then invert these guarantees to obtain a certified primitive tolerance for the current state under prescribed learning-level error requirements. We give a complete instantiation of the framework in high-dimensional von Mises-Fisher learning. Controlled interventions and a large collection of saved learning states show that identical primitive error can produce substantially different learning consequences, while certified numerical tolerances vary by orders of magnitude across states under the same learning-level requirements. These results show that the adequacy of a numerical approximation must be assessed in relation to the current learning state and the quantity to be preserved; numerical accuracy should itself be treated as part of the learning objective.

---

### Does a Shared Temperature Imply a Shared Angular Scale in Probabilistic Contrastive Learning?

Authors: Ningkang Peng, Qianfeng Yu, Jingyang Mao, Xiaoqian Peng, Tingyu Lu, Peirong Ma, Yanhui Gu
Published: 2026-09-30
Categories: cs.LG
arXiv: https://arxiv.org/abs/2609.38784v1

In probabilistic contrastive learning, a shared temperature is commonly interpreted as a shared similarity scale, but this interpretation does not hold for high-dimensional distributional class representations. We study the exact von Mises-Fisher (vMF) probabilistic score used by ProCo when representation dimension and class concentration grow jointly. We prove that the score retains a class-dependent leading angular gain $g_c=A_c/τ$, where $A_c$ is the mean resultant length. This gain enters Softmax competition, pairwise decision boundaries, and feature gradients. On real CIFAR-LT, ImageNet-LT, and iNaturalist representations, the theory accurately predicts boundary movements and local gradient changes under the full vMF score. Classwise temperature adjustment also changes the cosine-zero intercept and finite-dimensional response. We construct intercept-preserving and Pure Angular controls to separate the leading gain from these accompanying changes. Complete gain equalization yields a shared-scale cosine prototype rule at leading order; a finite-dimensional margin condition guarantees agreement of the two classifiers. Across 16 frozen representation settings, prediction agreement is 98.43-99.99%, with disagreements concentrated at small cosine margins. In controlled contrastive-only training with the training-frequency prior, Pure Angular editing improves both learned representations at all tested CIFAR-10/100 imbalance factors and retains positive changes on ImageNet-LT. Thus vMF concentration not only describes class distributions, but also forms a decision and learning scale in high-dimensional probabilistic contrastive learning.

---

### Persona and Persuasive Framing in AI Voice Agents: A $2\times2$ Field Experiment with Children

Authors: Thilo Tamme, David Steck, Anton Hantel
Published: 2026-09-30
Categories: cs.AI, cs.HC
arXiv: https://arxiv.org/abs/2609.38782v1

Conversational agents increasingly interact with children, yet evidence on how their design shapes children's susceptibility to persuasion comes almost entirely from the lab. We report a $2\times2$ randomized field experiment embedded in a public German Santa Claus telephone hotline. Children's calls were randomly routed to one of four LLM voice agents varying persona (Santa, high authority, vs. Helper, low authority) and framing (persuasive nudges toward prosocial wishes vs. neutral). Of 1,072 logged calls, 89 conversations (median age 6) met inclusion criteria. Persuasive framing raised the probability of a prosocial wish from 11.6% to 45.7%, robust to controls. Persona authority showed a near-zero effect: Santa did not outperform the Helper. Persona instead shaped engagement; children hung up on the Helper far more often within the first minute (65% vs. 39%). Where context already lends an agent legitimacy, how it speaks shapes children's compliance more than who it claims to be.

---

### ChartDensity-Bench: Benchmarking MLLMs for Numerical Data Reconstruction under Visual Density

Authors: Xinhe Wu, Yadong Jin
Published: 2026-09-30
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.38781v1

Multimodal large language models (MLLMs) offer a promising approach for recovering numerical data from scientific charts, but their ability to reconstruct chart data from visually dense figures remains poorly understood. Existing chart understanding benchmarks primarily evaluate question answering or chart-level reasoning and provide limited support for evaluating structured numerical reconstruction from scientific figures. We introduce \textbf{ChartDensity-Bench}, a benchmark for evaluating MLLMs on structured numerical data reconstruction from compound chart figures under controlled visual density. Built from charts paired with source-level ground-truth data, ChartDensity-Bench systematically varies the number of simultaneously presented charts ($k\in{1,3,6,9}$), enabling controlled evaluation of density-induced degradation. We further propose a multi-dimensional evaluation framework covering structural reliability, reconstruction completeness, parseability, and numerical fidelity. Experiments on five recent MLLMs show that numerical reconstruction generally degrades as visual density increases, while the magnitude of degradation varies substantially across models. Chart-level paired comparisons further show that the same source chart can incur higher reconstruction error when embedded in denser visual contexts. These findings highlight visual density as an important and previously underexplored factor in MLLM chart data reconstruction and provide a systematic benchmark for evaluating model robustness in this setting.

---

### RAST: Resolution-Aware Privileged Structure Transfer for Low-Resolution Audio Activity Recognition

Authors: Ji Hwan Park, Gautham Krishna Gudur, Yufei Shen, Dawei Liang, Edison Thomaz
Published: 2026-09-30
Categories: cs.SD, cs.AI
arXiv: https://arxiv.org/abs/2609.38780v1

Audio is increasingly used for human activity recognition (HAR) because it captures object interactions, environmental events, and contextual cues in everyday environments. High-resolution (HR) audio provides rich acoustic information for model development but incurs substantial energy and storage costs and may expose sensitive speech content. Low-resolution (LR) audio offers a more privacy-preserving and resource-efficient alternative for deployment, but reduced sampling rates can remove acoustic cues essential for activity recognition, leading to significant performance degradation. We formulate this training-deployment mismatch as sensor-resolution privileged learning, in which HR audio is available during training, while inference relies exclusively on LR audio. We propose RAST, a resolution-aware transfer framework that compresses HR teacher representations by preserving token-level information and neighborhood structure before performing localized HR-LR alignment. Experiments on the SAMoSA and AudioIMU datasets show that RAST consistently outperforms LR-only training and direct teacher-transfer baselines, improving LR-only recognition by up to approximately 7.8% while requiring only LR audio at inference.

---

### Action Conditioned Bisimulation For GUI Agent Memory

Authors: Hongbo Zhang, Liuyang Song, Quanquan Li, Daqian Yang, Yan Wen, Zhengtao Yao
Published: 2026-09-30
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.38778v1

An agent that remembers what it did on a web page must decide when two pages count as the same. Memories built on observation similarity merge pages that look alike but behave differently, and GUIs are full of such pages: two tabs of one widget or two rows of one menu answer the same click differently. We define the merge rule as an action-conditioned bisimulation over the empirical predictive state graph a frozen agent fills as it acts. Two states merge only when their shared actions lead to agreeing outcomes and successor blocks under an affordance label. Observation similarity never enters the rule, and nothing is trained. It replaces the merge rule of an existing outcome-value memory, so a closed-loop comparison isolates it. On MiniWoB++ it raises success rate over a memoryless agent, while a control taking identical exploratory detours, the prior successor-representation merge, and the same criterion without action conditioning change nothing.

---

### Distill the Visual Evidence, Not Just the Answer: Cross-World On-Policy Distillation for Vision-Language Models

Authors: Yuanhao Sun, Huawei Ji, Jiaxin Ding, Luoyi Fu, Xinbing Wang
Published: 2026-09-30
Categories: cs.CV
arXiv: https://arxiv.org/abs/2609.38777v1

A central goal of vision-language model (VLM) distillation is to transfer both the teacher's language capabilities and its visual understanding. However, existing methods primarily supervise the student's output, leaving visual understanding implicit. Our analysis reveals that a student can match the teacher's answer without relying on the same visual evidence, raising the question: how can we ensure the student responds to the visual information that actually determines the answer? To this end, we propose \textbf{Cross-World On-Policy Distillation (CW-OPD)}, which explicitly supervises the student's response to changes in visual evidence. For each example, CW-OPD constructs two visual worlds that share the question and scene context but differ in answer-critical evidence, yielding different answers. We perform on-policy distillation in both worlds and distill the teacher's cross-world belief transition, encouraging the student to match not only \emph{what} the teacher predicts but also \emph{why} its prediction changes with the evidence. A gradient analysis shows that this term is invariant to errors shared by both worlds and supplies a corrective signal invisible to endpoint matching alone. In this way, CW-OPD makes reliance on the relevant visual evidence an explicit distillation target rather than an implicit consequence of output matching. To diagnose whether a model truly grounds its answers in visual evidence, we introduce CWBench, which measures cross-world consistency via Cross-World Pair Accuracy (CWPA). Experiments on Qwen3.5-4B show that CW-OPD outperforms the strongest baseline by \textbf{1.2} points on average, and the 4B student exceeds DeepSeek-V4.1 (552B) by \textbf{22.4} CWPA points on CWBench. Code is released in https://github.com/baokou-fw2/CWAD.

---

### Distilling Diffusion Score Discrepancy for Efficient Training Data Attribution

Authors: Shixuan Liu, Joan Serrà, Kin Wai Cheuk, Jinju Kim, Woosung Choi, Yukara Ikemiya, Wei-Hsiang Liao, Jiaqi W. Ma, Yuki Mitsufuji
Published: 2026-09-30
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.38776v1

Training data attribution for diffusion models aims to identify the training samples that influence a generated instance, but existing methods either require costly per-sample gradient computation or query-specific model optimization. Moreover, most methods attribute changes in a proxy loss rather than changes in the actual model's generative behavior. We address these limitations by formulating attribution directly with a local score discrepancy measure, which applies to any diffusion variant (including DDPM, EDM, and flow matching), and by showing that such measure can be estimated without retraining, as a preconditioned gradient similarity. We instantiate this estimator as Training-data Influence via score Discrepancy (TID), which uses Kronecker-factored curvature to avoid random projections and per-sample gradient storage. We then distill TID into TIDE, a forward-only student trained online to reproduce the teacher's rankings from the diffusion model's internal activations. Under counterfactual evaluation on CIFAR-10, ArtBench-10, and MS-COCO, TID matches or outperforms state-of-the-art approaches, while TIDE retains most of TID's accuracy at four to five orders of magnitude lower per-query cost, attributing generated samples in milliseconds and faster than the generation itself.

---

### Learning Under Forgetting: Statistical Support-Selective Retention in Stochastic Training Dynamics

Authors: Fujie Gao, Zuyue Zhang, Gang Sun
Published: 2026-09-30
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.38768v1

Prior work has shown that neural networks exhibit implicit biases toward low-complexity structure (e.g., spectral bias), memorization dynamics, and compression-like effects during training, but a unified dynamical account of selective retention remains incomplete. We propose Repeated Reinforcement with Persistent Forgetting (RPF) dynamics, a minimal framework in which repeated exposure reinforces patterns and structures that recur in the data, while persistent forgetting attenuates learned information. This view treats forgetting not merely as a failure mode, but as a selection mechanism. We build the theory in three successive layers. First, in an independent-feature model, we derive an exposure-selective survival law and a support-dependent retention boundary characterizing which patterns persist under forgetting. Second, in a shared-parameter model, we show that forgetting induces spectral filtering over covariance modes, preserving strongly supported shared components while suppressing weak ones. Third, under small-step and norm/coding approximations, we show how RPF dynamics induce an implicit trade-off between data fitting and the cost of stored information, yielding Minimum Description Length (MDL)-like compression. Controlled experiments provide evidence for this reinforcement--forgetting selection mechanism in scalar memories and a nonlinear shared network. Joint reinforcement and attenuation interventions shift conditional retention, while matched exposure counts reveal forgetting-dependent effects of reinforcement timing and changes in the composition of the retained set. Together, these results show that repeated reinforcement and persistent forgetting jointly provide a controllable source of inductive bias beyond neural architecture and scale.

---

### dattri-LLM: A Unified and Efficient Library for Training Data Attribution at LLM Scale

Authors: Shixuan Liu, Tongli Zhou, Junwei Deng, Pingbang Hu, Jiaqi W. Ma
Published: 2026-09-30
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2609.38767v1

Training data attribution (TDA) estimates the contribution of individual training examples to model outputs. Most scalable TDA methods rely on per-example gradients, whose computation and use at LLM scale pose challenges in efficiency, compatibility, and extensibility. We introduce dattri-LLM, a TDA library that makes gradient-based attribution more practical at scale. For efficiency, dattri-LLM uses compact gradient representations and dynamically routes gradient operations based on a cost model. For compatibility, its capture mechanism collects per-example gradients from existing training loops that call backward(), without requiring changes to the loop or its configuration. This includes distributed training with DDP and FSDP and pipelines built with HuggingFace Transformers, TRL, and OLMo. For extensibility, dattri-LLM exposes reusable gradient operations and training-time callbacks for implementing attribution methods and applications. These interfaces support a variety of attribution methods, including gradient similarity, curvature-based influence, and trajectory-based methods, as well as applications that act on gradients during training, such as online data selection. On the same hardware and workload, dattri-LLM achieves 3.2x the throughput of the fastest competing library on average, scales multiple attribution methods to 110B-parameter models across four H200 GPUs, and offers superior attribution fidelity-cost trade-offs across a range of models with different model families and scales.

---

### PathAnchor: Path-Structured Evidence for Scientific Agents

Authors: Qiuhui Chen, Jiafan Lu, Shuaimin Tang, Tao Dai, Suyuan Wang, Chenrui Ji, Zhenglei Zhou, Weimin Zhong
Published: 2026-09-30
Categories: cs.AI
arXiv: https://arxiv.org/abs/2609.38766v1

Scientific agents can retrieve relevant passages yet still lose functional order, mix evidence across sources, or state conclusions that exceed the retrieved record. We introduce PathAnchor, a bounded scientific reasoning system built on path-structured evidence workspaces. Instead of treating passages or extracted concepts as independent units, the system retrieves source-linked Material-Sensor-Signal-System trajectories that preserve role, direction, and the evidence supporting each transition. A controller uses three read-only tools to search paper-specific trajectories, trace paths across candidate sources, and open exact evidence before producing a claim-cited answer and an explicit evidence boundary. On 120 single- and cross-paper flexible-sensor questions, PathAnchor scores 82.6% and leads six evaluated systems. Under a matched controller, corpus, and six-call budget, replacing unordered concept graphs with path-structured records raises source recall from 61.3% to 82.9%, increases answers whose claims all cite opened evidence from 69.2% to 90.0%, and reduces tool calls. These results show that evidence organization affects retrieval and citation completeness under fixed agent resources.

---