# Research Papers — 2026-09 · bundle page 27 of 266 > Full-text markdown of papers #1301–1350 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/27 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/26.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/28.md) ### Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models Authors: Ayan Sengupta, Vaibhav Seth, Tanmoy Chakraborty Published: 2026-09-29 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.36734v1 Knowledge Distillation (KD) trains a smaller-capacity student model to imitate a larger-capacity teacher model by matching output distributions, implicitly assuming the teacher to be a reliable oracle. In large language models (LLMs), this assumption often fails: teacher predictions can exhibit high entropy and hallucinations, causing standard KD to degrade well-calibrated student priors. We propose CaRE-KD, a confidence-gated distillation framework that replaces static objectives with uncertainty-adaptive optimization. CaRE-KD has two components: a token-level loss (CaRE-Divergence) that adaptively switches between Forward and Reverse KL divergence based on teacher--student confidence, and a batch-level epistemic rejection mechanism (Revival) that suppresses updates when the teacher is more uncertain than the student. We provide a gradient-level analysis showing how this dual-granularity design induces a conditional calibration mechanism that prior static divergences cannot reproduce. Empirically, across eight teacher--student pairs and eleven benchmarks spanning instruction following, chat alignment, code generation, and mathematical reasoning, CaRE-KD delivers consistent gains over strong baselines (Skewed-KL, $α$--$β$ divergence). Highlights include up to $+3.2$ average ROUGE-L on instruction-following tasks, $+2.1$ pass@1 on MBPP, $+1.7$ accuracy on GSM8k, and $+1.8$ accuracy on CollegeMath over the strongest baseline, with consistent gains in LLM-as-a-judge factuality (up to $+2.5$ per task over Skewed-RKL). Revival further acts as a principled, loss-agnostic plug-in that systematically strengthens existing distillation objectives by filtering epistemically unreliable teacher supervision. --- ### Can Agents Design Libraries for Agents? Authors: Gabriel Orlanski, Alex L. Zhang, Avi Trost, Vincent Sunn Chen, Frederic Sala, Aws Albarghouthi, Ludwig Schmidt Published: 2026-09-29 Categories: cs.AI, cs.CL, cs.SE arXiv: https://arxiv.org/abs/2609.36730v1 Agents increasingly build on code written by other agents, and they reimplement rather than reuse, growing the codebases later agents must work in. To measure how well agents design libraries for other agents, we introduce LibraryDesignBench, a two-phase benchmark in which an agent implements a full-featured library from a specification that defines required capabilities and potential use cases without prescribing the design. We evaluate the library through the correctness and simplicity of programs written by three user agents from different model families. The benchmark spans 242 expert-validated programming problems across 15 library-design tasks in four languages. On eleven of the fifteen tasks, agent designers reproduce the abstractions of the human-written production library. Downstream agents adopt agent- and human-written libraries alike but underuse them, reimplementing capabilities the library already provides. Our failure analysis finds that downstream agents write extra code mainly because agent-written libraries are rigid or hard to use, not because capabilities are missing. We also experiment with giving designers more prescriptive, agent-first guidance and having them test their library with subagents; this improves downstream scores and yields simpler programs. LibraryDesignBench provides both a testbed for evaluating library-design practices for agent users and an initial design baseline that improves downstream reuse. --- ### Can AI Scientists Change Their Minds? Prior-Evidence Conflict in Synthetic Universes Authors: Kargi Chauhan Published: 2026-09-29 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36726v1 Can a scientific agent distinguish a law it inferred from evidence from one it merely recognizes? We introduce Synthetic Universes, a controlled benchmark that pairs canonical famous worlds with matched twisted twins governed by nearby noncanonical mechanisms. We evaluate each reported law twice: by executing it on held-out continuations and transfer settings, and by independently checking whether it recovers the generating mechanism. In the current checkpoint of a pre-specified 60-cell study, 22 trials were graded and one additional run ended in infrastructure failure. Among 20 twin trials, 8 pass predictive verification while 5 recover the generator. The dissociation is bidirectional: six parsable outputs predict successfully while missing the mechanism, whereas three recover the mechanism but fail predictive rollout. Drag exhibits the first pattern (5/5 predictive pass, 1/5 mechanism recovery); Gravity exhibits the second (1/5 predictive pass, 4/5 mechanism recovery). Because matched famous controls, the corrected identifiability sweep, and the Evidence Ladder remain incomplete, we do not claim a confirmatory causal prior-conflict effect. Instead, the completed runs establish a narrower verification result: predictive adequacy and mechanism recovery are distinct scientific claims and require distinct tests. --- ### Routing in Gradient Space: Balanced Usage Is Not Expert Specialization Authors: Yuchen Li, Mingyu Du, Zongqi Fan, Nguyen H. Tran, Ken-Tye Yong Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36724v1 Sparse expert models can distribute traffic evenly while still grouping incompatible training signals within the same experts. We study routing as a gradient-partitioning problem and introduce gradient-aligned routing (GAR), whose load-normalized router objective rewards grouping observations with aligned gradients. On five multi-task text-classification mixtures, we compare GAR with task-loss-only routing, gradient-combination and gradient-conflict methods, and load-balancing losses. With a fully trainable RoBERTa backbone and classification-head experts, GAR has the highest aggregate validation accuracy, 1.07 percentage points above task-loss-only routing. With frozen DeBERTa and Qwen3-1.7B backbones and low-rank adapter experts, it again ranks first, 1.10 points above task-loss-only routing, with better-balanced expert load and higher gradient-mass purity, the share of each expert's gradient-norm mass from its dominant task; the load-balancing losses flatten load further but leave this purity near its task-loss-only level. Top-1 routing, trainable full-parameter feed-forward network (FFN) experts, and a larger backbone also show positive aggregate gains. The results distinguish expert-load balance from gradient-based routing organization and indicate the predictive value of gradient-informed routing in multi-task text classification. --- ### ATTUNER: Recomputation-Free KV Cache Reuse via Query-Side Adaptation Authors: Xinghao Chen, Junnan Dong, Cai Ke, Chak Tou Leong, Haocheng Sun, Keyu Chen, Siyu An, Ruizhi Qiao, Xing Sun, Wenjie Li, Xiaoyu Shen Published: 2026-09-29 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.36722v1 Large language model (LLM) agents repeatedly load reusable content, such as skills, documents, and memory entries, into the current context. Re-encoding this content for every request wastes computation. Position-independent caching (PIC) alleviates this by encoding each artifact independently and reusing its key-value (KV) states at arbitrary positions, but it incurs a quality loss relative to full-context prefill. Existing methods repair this loss by restoring global position IDs or recomputing selected tokens. In this work, we isolate the source of the loss, finding that the positional mismatch has minor effect, and independently cached artifacts retain faithful representations: reading a provided artifact stays largely accurate, and performance degrades only when the model must select among multiple artifacts. Moreover, replacing PIC's attention scores with full-prefill scores recovers performance with the cached KV unchanged, localizing the failure to the attention rather than KV recomputation. Motivated by this, we propose \textsc{Attuner}, a query-side adaptation method that learns to read a frozen artifact cache. \textsc{Attuner} inserts low-rank adapters into the query projections and is trained by distilling full-prefill distribution into the student. It trains fewer than 0.05\% of the model parameters and, at inference, requires neither cache recomputation nor a full-context reference. On Qwen3-4B and Qwen3-8B across seven benchmarks covering skills, documents, memory, and code, \textsc{Attuner} substantially outperforms prior PIC baselines in both in-domain and out-of-domain settings, matches full-context prefill quality while providing up to $3.73\times$ speedup. --- ### CALIBUDGET: Calibration-Guided Source Allocation for Fixed-Budget Mixed-Reasoning Adaptation Authors: Yupeng Chang, Yuan Wu Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36721v1 Fixed-budget adaptation from heterogeneous data sources requires deciding not only how much data to use, but how much exposure each source receives. Size-proportional rules can crowd out small sources, whereas difficulty-only rules can chase noisy estimates or allocate residual budget to nearly saturated pools. We introduce CALIBUDGET, a floor-protected, reliability-aware integer allocator that treats source exposure as an explicit adaptation variable. From small train-internal calibration splits, it combines model need, post-floor availability, and bootstrap stability, then produces exact capacity-respecting quotas without changing the model, objective, or total budget. In a controlled setting combining mathematical and commonsense data, CALIBUDGET improves CommonAvg, FragileAvg, and MacroAvg over validation-error-with-floor, the strongest matched comparator, in all three paired LLaMA-2-7B LoRA+ runs. The respective mean gains are 0.56, 0.46, and 0.41 percentage points (pp). Overall increases by 0.18 pp, whereas MathAvg decreases by 0.20 pp, exposing a coverage-retention boundary rather than a uniform gain. CALIBUDGET changes only 1.14-1.42% of the source budget but improves performance in 15 of 24 comparisons across commonsense tasks and seeds. These results suggest that small changes in source quotas can matter; example-level selection can then determine which examples fill each quota. --- ### Information-theoretic receding-horizon active learning of nonlinear dynamical systems Authors: Juncal Arbelaiz, Anushri Arora, Jonathan W. Pillow Published: 2026-09-29 Categories: eess.SY, cs.LG arXiv: https://arxiv.org/abs/2609.36712v1 Accurately learning nonlinear dynamics from a finite-duration experiment requires the efficient collection of informative data. We address this challenge for stochastic controlled nonlinear dynamical systems whose state is observed along a single trajectory. Our goal is to reconstruct the unknown controlled state-increment map over a prescribed compact subset of state-input space. We construct a parametric estimator of the map using fixed nonlinear features, so that the model is nonlinear in the state and input, but linear in the unknown parameters. A Gaussian prior over the parameters yields recursive Bayesian posterior updates as data stream in, enabling online quantification of predictive uncertainty in the reconstructed dynamics over the target set. We formulate an optimal adaptive-design problem over an information state, using a prediction-oriented acquisition criterion based on the mean marginal mutual information between candidate future trajectories and the reconstructed dynamics over the target set. We then approximate the resulting adaptive-design problem by a non-myopic receding-horizon formulation, evaluate its remaining expectation using a scenario-based sample average, and solve the resulting deterministic program with the cross-entropy method, leveraging parallel candidate-scenario evaluations. Numerical experiments on a noisy multistable system demonstrate that the proposed adaptive information-seeking strategy reduces predictive uncertainty and reconstruction error more efficiently than common excitation baselines under comparable experimental constraints. --- ### Into the danger zone: stable extrapolation in high-dimensional function and operator learning Authors: Ben Adcock, Simone Brugiapaglia, Xuemeng Wang Published: 2026-09-29 Categories: math.NA, cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.36709v1 Out-of-distribution (OOD) generalization is a central challenge in scientific machine learning. We study regression problems in which the test distribution differs from the training distribution and ask: under what assumptions on the target function or operator is stable extrapolation possible, and how far beyond the training domain can one extrapolate? Existing theory controls the test error through additive penalties measuring the discrepancy between the training and test distributions. Such guarantees show robustness to small distribution shifts, but can very pessimistic in comparison to OOD performance observed empirically. We identify classes of holomorphic functions and operators for which the OOD generalization error converges at algebraic rates even in the presence of large distribution shifts. This phenomenon stems from the increasing smoothness of higher-index coordinates, leading to what we term a `blessing of high dimensionality'. For learning with either polynomials, deep neural networks or deep neural operators, we derive explicit rates for arbitrary test measures supported on suitable domains and quantify how the admissible domain depends on the underlying regularity of the function or operator. Our extrapolation guarantees are independent of the test distribution, depending only on its support. We also present a series of numerical experiments across a range of functions and operators that support the main theoretical findings. --- ### LAURA: Knowledge Distillation for Interpretable Ambiguous Clause Identification in Legal Contracts Authors: Amrita Singh, Aditya Joshi, Jiaojiao Jiang, Hye-young Paik Published: 2026-09-29 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.36707v1 Legal contracts contain ambiguities that expose enterprises to financial and legal risks. Some ambiguities allow flexible interpretation without triggering disputes, while others lead to significant legal conflicts. This makes identification alone insufficient, and interpretable rationale analysis essential. We propose LAURA, a post-training framework for interpretable ambiguous clause identification. LAURA leverages knowledge distillation with an IRAC-Unlearning prompting technique to transfer knowledge from a teacher LLM to an open-weight student model (<=1B parameters), which is then trained using a joint objective combining classification and rationale generation losses. The framework supports both legal and non-legal stakeholders in making informed decisions about which ambiguities require further attention. Extensive experiments across 7 baselines and 7 open-weight models demonstrate that LAURA with Flan-T5 (250M) delivers state-of-the-art interpretability over all interpretable baselines while matching the identification performance of the best-performing opaque baseline. --- ### JudgeProfile: Understanding and Steering Subjectivity in LLM Judges Authors: Qi Cao, Kangning Liu, Xuan Kan, Shunwen Tan, Yang Pei, Dake Chen, Yatai Ji, Zixuan Ye, Yuanpeng Tu, Daniel Li, Junbiao Tang, Pengtao Xie et al. Published: 2026-09-29 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36705v1 LLM judges are inherently subjective, often favoring different responses in pairwise comparison when neither option is objectively wrong. To study this subjectivity, we introduce JudgeProfile, a framework that dissects LLM evaluation into perception (how a judge compares two responses across specific attributes like clarity, correctness, and detail) and prioritization (how much each attribute influences the final choice). We curate SubjectiveSet, a dataset of 50,013 response pairs from 17 public data sources, evaluated by 21 LLM judges across 87 attributes. We find a hidden consensus in perception: judges frequently agree on attribute judgments even when their overall choices diverge. Building on this separation, we first characterize each judge's prioritization using attribute weights estimated from its own overall choices. These weights differ across judges even when estimated from the same attribute judgments. We then learn new weights from reference labels to adapt their decisions to a target evaluation standard. Reweighting perceived attributes improves average held-out agreement with reference labels from 66.48% to 71.97%, outperforming fine-tuning and rubric prompting. Our findings show that understanding and steering the subjectivity of LLM judges requires attention not only to what they perceive, but also to how they prioritize it. --- ### When Is Coarse Supervision Worth It? Cost-Aware Learning under Unknown Aggregation Authors: Jianyu Xu, Smriti Jha, Aarti Singh, Bryan Wilder Published: 2026-09-29 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.36704v1 Modern learning systems often acquire supervision at multiple resolutions, trading annotation cost against information content. We study cost-aware two-resolution learning, where expensive fine labels reveal a vector response and cheaper coarse labels reveal a scalar aggregate formed with unknown weights, while the target remains the full response. The challenge is that unknown aggregation changes which directions coarse data can identify, so the value of coarse supervision depends jointly on cost, noise, and identification. We characterize this information geometry and develop an estimate-and-track policy that learns the aggregation rule and tracks the optimal resolution mix. We derive a closed-form break-even condition for coarse supervision and prove that the online policy attains the optimal leading cumulative-risk coefficient, with a matching local asymptotic minimax lower bound. Synthetic experiments support the predicted all-fine/mixed transition, show the online learner approaching the oracle-share benchmark, and demonstrate a finite-budget gain over all-fine acquisition when coarse supervision is sufficiently favorable. Our results provide a principled way to balance information and annotation cost across supervision resolutions. --- ### Quantum Fidelity Landscape-Guided Prior Calibration for Single-Circuit QGAN Image Generation Authors: Xue Yang, Rigui Zhou, Dax Enshan Koh, Siong Thye Goh, Yitao Tang, ShiZheng Jia, Young-Wook Cho, Hongyu Chen Published: 2026-09-29 Categories: quant-ph, cs.CV arXiv: https://arxiv.org/abs/2609.36702v1 Quantum Generative Adversarial Networks (QGANs) have emerged as representative generative models in the Noisy Intermediate-Scale Quantum (NISQ) era and have attracted increasing attention in quantum machine learning. However, most existing QGAN methods rely on patch-based decomposition strategies, which weaken the global consistency of generated images and increase quantum resource overhead. In this work, we investigate a simpler approach: pixel-level, end-to-end image generation using a single-quantum-circuit QGAN. By analyzing the structural matching relationship between the quantum prior and the target data distribution in Hilbert space, we provide a new theoretical perspective for understanding the training behavior of naive end-to-end QGANs. Specifically, we introduce the Quantum Fidelity Landscape (QFL), defined as the pairwise-fidelity structure induced by an ensemble of quantum states and preserved under shared unitary transformations of the quantum generation process. We show that, under a fixed Lipschitz readout, this invariant imposes a one-sided bound on decoded sample separation, motivating calibration of the prior-induced QFL before adversarial training. To validate this theoretical insight, we propose BasicQGAN, a QGAN framework incorporating quantum prior calibration. Before adversarial optimization, BasicQGAN aligns the prior-induced QFL with the data-induced QFL. Experimental results on small-scale grayscale image datasets show that BasicQGAN achieves stable and effective end-to-end pixel-level image generation while requiring fewer qubits and trainable parameters than representative patch-based quantum generators. Furthermore, experiments with different initial quantum-state ensembles show that QFL-calibrated ensembles achieve better generative performance. --- ### Lost in Conversation or Lost in Translation? Diagnosing Multi-Turn Degradation in RAG Authors: Pranav Handa, Ariful Azad Published: 2026-09-29 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.36700v1 When conversing with large language models (LLMs), users often begin with a simple question and build towards a multi-hop question through follow-up turns. Retrieval-augmented generation (RAG) and its graph-based variant (GraphRAG) have become the dominant approaches for grounding LLM responses in external evidence, yet both are evaluated almost exclusively on single-turn, fully specified queries. We systematically investigate this evaluation mismatch through a large-scale simulation study. Building on prior work on multi-turn LLM evaluation, we transform questions from multi-hop question answering (QA) benchmarks into underspecified conversations and evaluate ten LLM assistants with eight retrieval systems across 1.5 million simulated conversations. Our findings reveal that multi-turn interaction causes widespread performance degradation, incurring relative performance drops of up to 21% and increasing unreliability by 47%, making RAG systems simultaneously less accurate and less reliable. We identify two distinct failure modes behind this degradation. Systems are either lost in translation, where conversational rephrasing distorts the retrieval query, or lost in conversation, where retrieval succeeds but the LLM fails to synthesize evidence distributed across turns. --- ### Learned Queries and Keys Are All You Need: Replacing the Value Projection with Structured Transforms Authors: Ene Meco, Emadeldeen Hamdan, A. Enis Cetin Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36698v1 To reduce the number of parameters and cache memory requirements of transformers we introduce dual-headed transformers instead of three heads. We studied Walsh-Hadamard Transform (WHT), Discrete Cosine Transform (DCT), Discrete Fourier Transform, filterbank based Shearlet Transform, and Multiplication-Avoiding (MA) operators to construct dual heads. We combine spatial patches and their orthogonal transforms (or Shearlet and MA operators) in a structure similar to the attention block. We obtained better results than triple headed transformers in ImageNet. Extensive simulation examples are presented. --- ### Know Thyself, Teach Thyself: Internal Information Flow for Selective Self-Distillation Authors: Rui Wang, Ruijie Wang, Bo Chen, Jiangxuan Long, Yingyu Liang Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36695v1 Self-distillation turns knowledge distillation into a closed learning loop and offers a path toward recursive self-improvement. Without an external teacher, however, the model must determine both what information can improve its supervision and which induced changes should be learned. Existing methods typically improve teacher-generated data or select training examples in isolation, leaving the information transferred between these stages unmeasured. We introduce InFlow, a retrieval-guided on-policy self-distillation framework that models this process as potential-to-realized information flow. InFlow first retrieves potentially informative sources using certainty-calibrated hidden-state trajectories, then measures their realized effect through the Jensen--Shannon divergence between the teacher's initial and retrieval-conditioned answer beliefs. Examples with larger belief shifts are selected for on-policy distillation. Our analysis formalizes the information optimized by retrieval and selection and relates the answer-level shift to the teacher--student distillation gap. Across four open-weight language models and three knowledge domains, InFlow achieves the strongest cross-model average among the compared selection methods, with ablations supporting both stages of the framework. Our code is available at https://github.com/1240148048/INFLOW. --- ### AESplat: Advancing Pose-Free Feed-Forward 3D Gaussian Splatting via Decoupled Appearance Modeling Authors: Shiwei Ren, Zhiang Liu, Yongchun Fang, Hongwei Chen Published: 2026-09-29 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.36693v1 Pose-free feed-forward 3D Gaussian Splatting (3DGS) has demonstrated remarkable potential for generalized novel view synthesis. However, existing methods typically predict Gaussian appearance attributes represented by spherical harmonics (SH) in the same manner, overlooking the fundamental distinction between view-independent and view-dependent appearance, which results in suboptimal rendering quality. In this paper, we present AESplat, a novel and general framework for pose-free feed-forward 3DGS that introduces an effective decoupled appearance modeling strategy based on an analysis of SH, enabling higher-quality rendering. Specifically, AESplat directly derives the zeroth-order SH coefficient, which represents the base view-independent appearance component, from the input images without training. The higher-order SH coefficients are subsequently predicted by a shallow multilayer perceptron equipped with two efficient 3D-aware inductive biases to model view-dependent appearance variations. Extensive experiments across multiple datasets demonstrate that our method significantly outperforms state-of-the-art approaches, achieving a $0.8$ dB improvement in PSNR over the pose-free method NAS3R and a $1.1$ dB improvement over the pose-required method DepthSplat on the RealEstate10K dataset. Project page: https://aesplat.github.io/. --- ### Normalize-Then-Precondition: A Hierarchical Approach to Marginal Scale and Interaction Geometry for LLM Training Authors: Zixuan Gong, Zeyu Gan, Jiaye Teng, Yong Liu Published: 2026-09-29 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.36692v1 Matrix optimizers have emerged as a promising direction, with Muon standing out as a prominent design. Revisiting Muon through its full-Gram representation, we observe that it jointly processes marginal-scale and interaction information. This opens an alternative way to organize geometric information hierarchically, motivating the Normalize-Then-Precondition framework. Specifically, it first uses diagonal-Gram information to construct a marginally normalized update, then applies spectral preconditioning to its directional interaction geometry. Building on this framework, we develop NormPre with NormPre-G and NormPre-L adopting global and localized spectral preconditioning, grounded in spectral-norm steepest descent and a regularized formulation followed by leading mode selection, respectively. To enable large-scale training, NormPre-G uses Newton-Schulz iterations and NormPre-L employs randomized sketching to approximate the leading interaction eigenspace. Theoretically, we establish $\mathcal{O}(T^{-1/2})$ convergence guarantees for simplified versions of NormPre. Across extensive pretraining experiments on GPT-2 Small, LLaMA and Qwen3, both variants consistently outperform AdamW, Muon and MANO under matched training budgets. Further efficiency and spectral analyses reveal the complementary strengths of two variants and characterize their performance-efficiency trade-off. We open-source our code through a GitHub repository at https://github.com/zx-gong/NormPre. --- ### Video2Skill: From Streaming Experience to Reusable Embodied Skills Authors: Jianshu Zhang, Ce Zhang, Xiyuan Yang, Chenwei Xu, Haoran Lu, Yijiang Li, Yaqi Xie, Katia P. Sycara, Han Liu Published: 2026-09-29 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.36691v1 Manipulation behaviors vary widely across objects and scenes, but they share a small set of reusable skills, and planning with these skills helps embodied agents generalize to new tasks. Yet an agent can only plan with skills it knows. Recovering skills from observed experience, the inverse of planning, builds this knowledge over time and yields skill data for training future agents. Vision-Language Models (VLMs) describe individual manipulation events well, but can they organize a stream of events into reusable skills? We formulate this problem as Streaming Embodied Skill Discovery (SESD): a model watches videos in sequence and maintains a persistent skill library that shapes its later decisions. To systematically measure this ability, we introduce Video2Skill, a benchmark that covers robot tabletop manipulation and human kitchen activity and tests three core capabilities: (i) locating manipulation events in time, (ii) grouping events of the same transformation, and (iii) deciding when to reuse an existing skill or create a new one. Across 19 open-source VLMs, many models group events at near-chance level, and scale does not consistently help. Their errors depend on how perception and library updates are coupled: joint models merge distinct transformations into one skill, while models that update the library from text descriptions duplicate recurring ones. Supervised fine-tuning, including our counterfactual library-state rebalancing (CLaRe), improves grouping but exposes a deeper bottleneck: trained models consolidate familiar skills yet rarely expand the library. Their libraries stall below half the reference size, and transformations unseen in training are located in time but almost never given a new skill. Recognizing when existing skills are insufficient thus emerges as the central challenge. --- ### CHAIN: Calibrated LLM Forecasting via Causal-Temporal Hypergraph Inference Authors: Wenjin Liu, Chenxi Wang, Yue Lu, Zhe Cui, Haoran Luo Published: 2026-09-29 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.36689v1 Large language models have achieved significant progress in event forecasting, yet their probability outputs exhibit systematic calibration bias that varies heterogeneously across different domains and question types, undermining the trustworthiness of probabilistic outputs for decision-making under uncertainty. However, existing calibration methods typically correct probability outputs after prediction is complete, without modeling the structural sources of bias within the prediction process itself. To address this challenge, we decompose probabilistic prediction over causal-temporal hypergraphs into three stages, evidence weighting, evidence aggregation, and source fusion, and propose CHAIN, which designs stage-specific mechanisms to mitigate bias at each stage: (i) modulating the temporal decay function by causal topological distance, (ii) aggregating approximately independent causal chains via Noisy-OR after direction-aware deduplication, and (iii) driving adaptive fusion by causal coverage and directional balance. Experimental results on cross-domain forecasting benchmarks show CHAIN outperforms existing methods in expected calibration error, Brier score, and accuracy. Our project is available at https://github.com/QwenQKing/Chain. --- ### MyoCodec: A Streaming Neural Codec for Electromyography Authors: Jihwan Lee, Kleanthis Avramidis, Junhyeok Lee, Tiantian Feng, Najim Dehak, Shrikanth Narayanan Published: 2026-09-29 Categories: eess.SP, cs.AI arXiv: https://arxiv.org/abs/2609.36687v1 Neural codecs encode continuous signals into compact sequences of discrete tokens, providing an interface for efficient transmission, storage, and token-based sequence modeling. This paradigm has been widely adopted in modern speech and audio frameworks; however, the biosignal domain still lacks a neural codec designed specifically for low-bitrate streaming and generalization across diverse downstream tasks. We present MyoCodec, a streaming neural codec designed for electromyography (EMG). Inspired by recent neural audio codecs, MyoCodec combines causal Transformers with residual vector quantization to encode continuous EMG signals into different levels of EMG representations spanning from continuous latent features to discrete tokens operating at 50 Hz. Trained on twelve public EMG datasets, MyoCodec achieves favorable performance in both intrinsic codec quality and representative downstream tasks, including typing (emg2qwerty), hand-pose (emg2pose), speech decoding (emg2speech), and speech-to-EMG synthesis (speech2emg). Across these tasks, MyoCodec exhibits strong performance against prior models while providing a compact and causal EMG representation. During streaming inference, it requires compute time of only 0.482 ms for each 20 ms frame, enabling real-time streaming. Also, the discrete token representation provided by MyoCodec has the potential to support integration into language-model based approaches, creating a path toward LLM-based interactive systems, where tokenized EMG representations are directly processed into such language or speech models. Code and model weights are released. --- ### Where Root Cause Analysis Fails: A Retrieval-Reranking Decomposition Authors: Hada Melino Muhammad, Luan Pham, Laure Barrière, Sachin Shetty, Leonardo Pulga, Flora D. Salim Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36686v1 Identifying the root cause of an anomaly among hundreds of sensors is critical for preventing safety incidents and costly downtime in complex monitored systems. Existing studies evaluate root cause analysis (RCA) methods using top@k accuracy. We show that this metric has a fundamental blind spot: it conflates two failure modes, retrieval failure, where the true cause is never considered, and reranking failure, where it is considered but ranked too low. In this work, we introduce a retrieval-reranking decomposition and audit four well-known benchmarks to expose this blind spot. Our experiments show that, on benchmarks with complex faults, statistical baselines mis-rank the true cause 79-100% of the time, and graph-based methods never clearly beat the best statistical baseline, whether their causal graphs are learned on short fault windows, on retrieved candidate pools guaranteed to contain the cause, or on multi-day normal-operation data. Meanwhile, on simple benchmarks where faults manifest significantly at their origin, retrieval is nearly solved (98-100%). Guided by the decomposition, we build a two-stage pipeline combining a multi-signal retriever with an LLM reranker that, as one fixed configuration, matches or exceeds the best baseline's top@1 accuracy on all six benchmark suites (by up to +12 points), with no causal graph or labeled data required. When all methods rank the same retrieved candidates with the true cause guaranteed present, adding a short system-description document lets the reranker lead the best baseline by +7 to +18 points on every benchmark. Code is available at https://github.com/cruiseresearchgroup/DecompRCA. --- ### When Semantics Matter: Reliability-Aware Semantic-Rhythm Control for Co-Speech Gesture Generation Authors: Zhirui Xing, Long Ye, Kaige Li, Ziyi Xu, Ming Meng Published: 2026-09-29 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.36685v1 Co-speech gesture generation aims to synthesize natural gestures that are both temporally synchronized with speech and semantically consistent with the spoken content. Although recent methods can generate rhythmically plausible motions, they often rely heavily on acoustic prosody while underutilizing textual semantics, especially when semantic annotations are incomplete, noisy, or unavailable. Consequently, the generated gestures may follow speech rhythm while failing to express the intended semantics. To address this problem, we propose a reliability-aware semantic-rhythm control framework for co-speech gesture generation. We first learn a discrete motion prior that represents continuous gestures in a compact and structured motion-code space. We then introduce a dual-branch semantic contribution estimation mechanism consisting of a full multimodal branch and an audio-only branch. Their distributional discrepancy is formulated as conditional information gain to quantify how much textual semantics changes the predicted motion. Based on this estimate, a controllable semantic-rhythm objective selectively strengthens semantic guidance in content-relevant segments while limiting unnecessary semantic intervention in rhythm-dominant segments. Furthermore, we treat background noise as an acoustic reliability condition and introduce noise-conditioned feature modulation together with beneficial latent perturbation to improve generation robustness under realistic acoustic environments. Experiments on benchmark datasets demonstrate that the proposed framework achieves a favorable balance among semantic expressiveness, rhythmic synchronization, motion diversity, and robustness, enabling reliable and controllable co-speech gesture generation. --- ### ProgressCompass: Embodied Progress Reward Models Are Lost Without the Right Context Authors: Jianshu Zhang, Keliang Wu, Chengxuan Qian, Xiyuan Yang, Ce Zhang, Ariel Tian, Anbang Liu, Haoran Lu, Han Liu Published: 2026-09-29 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.36684v1 Embodied agents now take on ever longer tasks. For long tasks, knowing only whether a task finally succeeds or fails says little; the steps along the way matter. Progress Reward Models (PRMs) score how far a task has come at every step, and serve as dense rewards, verifiers and monitors. Yet in long tasks the current frame alone often cannot tell how far the task has come, because progress depends on what happened before. We call this problem context-dependent progress estimation. Existing benchmarks on progress estimation mostly focus on short tasks whose progress can be read from the current observation, and whether PRMs can estimate progress when context is needed remains underexplored. We therefore build ContextProgress-Bench, with 24 manipulation tasks for 120 episodes. The benchmark covers three settings: (i) State Recall, where information needed for progress appeared earlier but is not in the current frame; (ii) Sequence Tracking, where steps follow a fixed order, so progress requires knowing which steps are done and which comes next; and (iii) Recurrence Disambiguation, where look-alike frames sit at very different progress. We then run a paired diagnosis: each PRM keeps the same input format in both runs, and in one run its instruction integrates the right context. Even PRMs that read the entire history get lost in estimating progress, yet with the right context the same five models cut their progress error by 77-82%. Embodied PRMs are thus not incapable of progress estimation, but lost without the right context. We therefore propose ProgressCompass, an autonomous agentic loop that reorients an existing PRM and uses current general-purpose VLMs to supply the context the PRM needs. Wrapped in the loop, the same frozen PRM cuts its progress error by 63% and raises its rank agreement by 76%. With such a compass, PRMs estimate progress far better on longer, more complex tasks. --- ### MARCO: Multi-Round Agentic Reinforcement for Conditional Molecular Optimization Authors: Shicheng Fang, Yuxin Wang, Zhuo Yang, Xiaohu Xu, Jiahao Lu, Chuanyuan Tan, Tong Zhu, Yining Zheng, Xipeng Qiu Published: 2026-09-29 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.36683v1 Molecular optimization is inherently iterative: a candidate is proposed, evaluated against several objectives, and revised while preserving a relationship to the source molecule. Most instruction-following models instead emit one edited molecule, forcing validity, property improvement, and similarity control into a single response. We introduce MARCO, an evaluator-grounded reinforcement-learning framework that trains molecular editors on bounded proposal--feedback--revision trajectories. MARCO aggregates shaped turn rewards into an undiscounted trajectory return for group-relative policy optimization. We evaluate two consequences of this training: Same-1 tests the trained policy under a one-response budget, while Same-5 tests whether the same policy can use verifier feedback when up to five responses are available. Across the three-objective MuMOInstruct benchmark, three Qwen backbones, and seen/unseen instruction splits, SFT-initialized MARCO obtains the highest product of property success rate and similarity in every reported primary setting. Same-5 further improves the observed score under the tested budget, while four-objective and public-checkpoint experiments test transfer across constraint sets and initialization regimes. --- ### Reprogramming Vision-Language Models via Structured Prompt Reparameterization Authors: Zizhao Li, Chengyi Cai, Mohammed Yaqoob Ansari, Feng Liu, Joseph West, Kourosh Khoshelham Published: 2026-09-29 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.36680v1 Visual reprogramming adapts pretrained models to downstream tasks by modifying their input and output interfaces while keeping the backbone fixed. In vision-language models, existing methods mainly rely on intra-class prompt aggregation and do not explicitly model relationships among classes. However, fine-grained categories often exhibit highly overlapping attribute descriptions and strong inter-class correlation in the text embedding space, where discriminative cues lie in subtle low-variance components. We propose Reparameterized Inter-Class Visual Reprogramming (RVP), a structured framework that aggregates multiple text prompts within each class and applies residual correction across classes. We also show that CLIP-based visual reprogramming with input-independent linear output aggregation can be expressed as a linear mapping from frozen image embeddings to downstream logits, and use this view to design a structured reparameterization that models shared semantic components and class-specific differences. RVP uses only a single visual prompt and can be reparameterized at inference into a frozen backbone followed by a linear classifier, incurring nearly zero computational overhead. Across 11 few-shot classification benchmarks and four CLIP backbones, RVP consistently improves over prior visual reprogramming methods with comparable or better inference efficiency. --- ### MLToolBench: Learning Tool-Augmented Agents for Machine Learning Development Authors: Xin Yu, Lizhu Zhang, Jiamu Bai, Yanhong Wu, Zellux Wang, Serena Li, Weiwei Li, Lingzhou Xue, Xiangjun Fan, Bo Peng Published: 2026-09-29 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36679v1 Machine learning engineering (MLE) agents have made substantial progress, but learning through ML experimentation remains costly in time and computation. Synthetic environments reduce these costs while introducing variations in data and experimental settings that require task-specific diagnosis. Access to diagnostic tools alone does not ensure that agents learn when to use them or how to act on their findings. We introduce ToolMLBench, a suite of executable tools for data inspection, code verification, and experiment diagnosis, together with an SFT and RL pipeline for learning their use. Diagnostic calls acquire evidence whose value depends on subsequent decisions, so final outcomes provide limited guidance on which calls to reinforce. We address this challenge with SPICE, which measures how privileged context changes the likelihood of a sampled tool action and uses this difference as a turn-level reward alongside the final outcome. We train on 80 synthetic tasks and evaluate on 25 in-domain and 10 out-of-domain tasks. Providing tool interfaces and descriptions alone yields inconsistent gains across unadapted models. With the same diagnostic interface, our training pipeline raises in-domain success from 24.8% to 52.4% for Qwen3-8B and from 35.6% to 69.2% for Qwen3.5-35B-A3B. The latter also improves from 31% to 48% out-of-domain, supporting learned diagnostic tool use on held-out sources and targets. --- ### Understanding Private Evolution as Learning-Augmented Clustering Authors: Audra McMillan, Kunal Talwar, Felix Zhou Published: 2026-09-29 Categories: cs.LG, cs.CR, stat.ML arXiv: https://arxiv.org/abs/2609.36678v1 Private Evolution (PE) is a differentially private algorithm for synthetic data generation. While it can be viewed as a Wasserstein learning algorithm, it performs much better in practice than worst-case Wasserstein analyses would predict. We recast PE as generative model-augmented Wasserstein learning. We show theoretically that when we take into account the use of a generative model that is able to capture something about the true distribution, then we can obtain much better performance bounds. For example, if the generator gives samples in the same low-dimensional space as the distribution, then sample complexity depends on intrinsic, not ambient, dimension. We also show that standard variants of PE can fail to converge on simple well-clustered instances, and propose a new geometry-aware version of PE with provable convergence on such instances. Experimentally, we show that our new algorithm is competitive with standard baselines and can improve recall. --- ### ReWorld-Track: A Recursive Event World Model for Language-Guided Multi-Camera Tracking Authors: Haoyang Wu, Shoudong Han, Chaoyue Li, Sijia Chen, Zhenyang Xie, Wang sihan Published: 2026-09-29 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.36677v1 Language-guided multi-camera tracking must preserve a target identity across unobserved gaps, where similar candidates and uncertain returns can make early associations unreliable. A wrong match can corrupt the history used to predict later observations and propagate identity errors across subsequent camera handoffs. We propose ReWorld-Track, a recursive event world model that carries association uncertainty into future predictions. Candidate matches and continued waiting define alternative target states, whose posterior probabilities are used to update a persistent recurrent belief. This representation preserves uncertainty about alternative trajectories through successive observations. This belief predicts the next camera, arrival time, and entry region, while appearance and language evidence guide association. By training across successive handoffs, the model learns to retain uncertainty that remains useful for later predictions and identity decisions. ReWorld-Track achieves HOTA scores of 65.19 on CityFlowV2 and 45.36 on MTMMC, with improved identity continuity across repeated handoffs. On MTMMC, its structured posterior update gains 0.50 HOTA points over a similarly sized generic updater and 0.94 points over fixed-moment soft association, raising next-camera accuracy from 86.03% to 87.41% and reducing median arrival-time error from 0.78 s to 0.71 s for subsequent target returns. --- ### Gödel Forest: Balancing Search Depth and Breadth for Data-Centric Recursive Self-Improvement Authors: Ziqi Zhao, Fanqing Meng, Haocheng Lu, Lingxiao Du, Qiguang Chen, Mengkang Hu, Xiao-Ming Wu Published: 2026-09-29 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.36675v1 Recursive self-improvement (RSI) aims to achieve compounding gains by having models improve themselves. While most existing RSI systems optimize external agent harnesses or prompts around a frozen base model, data-centric RSI directly updates the model's own parameters by training on agent-generated data. However, because validating data strategies requires expensive model training, existing methods face a fundamental dilemma: a single agent gets trapped in narrow directions and lacks exploration breadth, while naive parallel search or heavy trace sharing sacrifices long-horizon search depth. To address this challenge, we introduce G"odel Forest, a multi-agent framework that organizes recursive self-improvement as an ensemble of co-evolving search trees. In G"odel Forest, each agent autonomously grows a persistent tree, deepening, branching, or pruning data strategies based on model feedback to secure depth, while parallel trees explore distinct regions of the data space to expand breadth. Crucially, rather than leaving trees isolated or flooding them with heavy execution logs, a dynamically co-evolving memory connects the forest: agents continuously distill their successes and failures into compact procedural lessons anchored to a global leaderboard. Through this forest ecosystem, a dead-end in one tree instantly warns the whole forest against unpromising paths, while an empirical breakthrough quickly seeds new exploration branches in neighboring trees. Evaluated on RSIBench-Data across six diverse domains, G"odel Forest outperforms the single-agent baseline by an average of 10.70% while reducing wall-clock time on five tasks. Ablations confirm that co-evolving shared memory yields a +7.00% gain over independent parallel search, demonstrating that collective distillation is key to scalable self-improvement. The code is available at https://github.com/evolvent-ai/Godel-Forest. --- ### Human-inspired, Task-Dimension-Guided Exploration for Efficient Learning in High Dimensions Authors: Fanyu Zhu, Jiahui An, Ni Ji Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36672v1 Efficient exploration in high-dimensional decision spaces remains a central challenge for decision-making systems. Humans, in contrast, can navigate large decision spaces with remarkable efficiency. Recent behavioral studies suggest that humans reduce dimensionality in large decision spaces by probing candidate feature dimensions, identifying reward-relevant ones, and restricting the effective decision space. Inspired by this mechanism, we propose TDGE (Task-Dimension-Guided Exploration), a human-inspired, model-agnostic algorithm with an automatically constructed task-dimension--feature--item hierarchy. TDGE follows a top-down exploration strategy: it first selects task-relevant feature dimensions, then identifies informative features within those dimensions, and finally recommends concrete items based on the selected features. Experiments on MovieLens-20M, Last.fm, and Amazon recommendation datasets show that TDGE substantially improves exploration efficiency and cold-start adaptation over baseline algorithms. Comparisons with other structured algorithms and ablation studies attribute these gains to TDGE's hierarchical structure and semantic feature-space exploration, with robust results across clustering methods and hierarchy depths. Recommendation-trajectory visualizations also show exploration patterns similar to human dimension-guided behavior. --- ### FairDiff: Mitigating the Self-Reinforcing Matthew Effect in Diffusion Recommender Models Authors: Song-Li Wu, Xianquan Wang, Zhaocheng Du, Weinan Gan, Jingyi Wang Published: 2026-09-29 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36671v1 While the "Matthew Effect" and filter bubbles are widely recognized outcome-level biases in recommender systems, we reveal that Diffusion Recommender Models (DRMs) uniquely compound this issue through their generative dynamics. Rather than merely inheriting data imbalances, DRMs trigger a self-reinforcing amplification of popularity bias. We identify that this phenomenon is driven by two compounding mechanisms. First, while optimization loss is universally dominated by high-frequency items across recommenders, DRMs suffer from a unique structural prior mismatch during generation. Because the forward terminal distribution of long-tailed data deviates significantly from the standard Gaussian prior, reverse sampling trajectories inherently collapse toward high-density popular items, fundamentally suppressing niche item generation. To dismantle this self-reinforcing loop, we propose FairDiff, a plug-and-play fairness-aware diffusion framework. To overcome the popularity-dominated loss, we introduce Popularity Condition Guidance (PCG). Rather than altering the training objective, PCG acts as an inference-time distributional reweighting mechanism, mathematically reshaping the score-based gradient field to penalize high-popularity regions and guide trajectories toward niche semantics. Furthermore, we design a Semantic Calibration (SC) Module to bridge the prior mismatch, aligning the forward and reverse distributions via one-step optimal transport. Comprehensive evaluations demonstrate that FairDiff achieves state-of-the-art performance while effectively mitigating the self-reinforcing Matthew Effect, highlighting its value as a general framework for DRMs. --- ### FineSID: Scalable and Efficient Semantic Identifier Learning for Generative Recommendation Authors: Song-Li Wu, Weinan Gan, Zhaocheng Du, Xianquan Wang, Jingyi Wang Published: 2026-09-29 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36670v1 A critical prerequisite of generative recommendation is designing semantic identifiers (SIDs) that are both scalable to large item sets and efficiently learnable. Existing SID learning methods fundamentally rely on Top-1 hard assignment during vector quantization. While heuristic strategies -- such as clustering-based initialization or forced post-hoc collision resolution -- can artificially inflate codebook coverage, they often disrupt end-to-end semantic alignment and fail to address the underlying optimization bottleneck: sparse gradient propagation. In standard Top-1 assignment, gradients concentrate on a narrow subset of frequently selected codewords, leaving the majority inherently under-trained and causing severe SID collisions. To overcome this limitation natively without relying on complex initialization priors, we propose FineSID, a unified quantization framework that moves beyond Top-1 assignment by enabling fine-grained gradient propagation across the entire codebook. Instead of updating only a single selected codeword, FineSID distributes learning signals to all codewords in a soft, differentiable manner. This design promotes globally balanced codebook optimization while strictly preserving semantic consistency, effectively alleviating SID collisions and stabilizing training in large, high-dimensional codebooks. Extensive experiments on multiple public benchmarks demonstrate that FineSID is robust to initialization configurations and consistently improves both codebook utilization and recommendation accuracy. Our work provides a principled, initialization-agnostic solution for semantic identifier learning, advancing the practicality of generative recommendation. --- ### Stochastic Heavy Ball with Polyak Step Size and Armijo Line Search: A General Convergence Analysis Authors: Jiawei Zhang, Qitan Shi, Yuantao Gu Published: 2026-09-29 Categories: cs.LG, math.OC arXiv: https://arxiv.org/abs/2609.36668v1 Polyak step size (PS) and Armijo line search (ALS) have received increasing attention in stochastic optimization, with encouraging empirical performance and theoretical guarantees. However, their convergence theory for stochastic heavy ball (SHB) methods remains limited. In this work, we develop a unified convergence analysis for SHB equipped with PS and ALS. To this end, we introduce a modified Armijo rule that closely parallels the Polyak step size, together with a decoupling analysis that isolates the historical dependence induced by momentum. For SHB with standard PS and ALS, we establish expected convergence for strongly convex, convex, and non-convex objectives without interpolation or restrictive conditions on the momentum parameter. Under interpolation or strong growth, we further strengthen the results to almost sure rates and last-iterate convergence. Moreover, for general settings beyond interpolation, we prove almost sure convergence to the exact optimum or to stationarity for SHB with diminishing variants of PS and ALS. These results provide a more comprehensive theoretical view of Polyak step size and Armijo line search for stochastic heavy ball methods. --- ### GenLimitLib: A Formal Library for Language Generation in the Limit and AI-Assisted Mathematical Research Authors: Shuangping Li, Peng Zhang Published: 2026-09-29 Categories: cs.LG, cs.LO arXiv: https://arxiv.org/abs/2609.36663v1 We present GenLimitLib, a source-aligned Lean 4 library for language generation in the limit. Introduced by Kleinberg and Mullainathan at NeurIPS 2024, language generation in the limit studies a theoretical question motivated by LLMs: how to generate valid new strings from observed examples. This young and rapidly evolving field offers a natural testbed for studying large-scale formalization. GenLimitLib contains formal developments for 30 papers. It extracts shared definitions and reusable proof components while preserving paper-specific assumptions and statements, and records relationships across papers. In this way, GenLimitLib provides a concrete and structured view of the literature. We show through mathematical case studies and LLM experiments how our library can support both human mathematical research and AI-assisted research. Our Library: https://github.com/pengzhang91/generation-in-the-limit-lib. --- ### AutoLoCo: Communication Efficient Distributed LLM Training via Adaptive Synchronization Authors: Pengyu He, Yan Zhang, Ruien Li, Guangwen Yang Published: 2026-09-29 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.36662v1 The pre-training of Large Language Models (LLMs) is increasingly conducted across multiple data centers. As training scales to a larger number of accelerators, the fraction of time spent on computation decreases, while the fraction spent on communication increases. Therefore, frequent synchronization becomes a growing bottleneck. Local update methods reduce this cost by allowing workers to perform several optimizer steps between synchronizations. Most local update methods set the number of local optimizer steps between synchronizations before training and keep this interval fixed throughout the run. However, the best interval can change during the entire train process. If the interval and optimizer are adapted to the current training state, the communication frequency is reduced while maintaining the training performance. In this work, we introduce AutoLoCo, an adaptive training framework to reduce communication in LLM training. It adapts the local interval using scalar training statistics and corrects each outer update. Our method is motivated by two observations: 1) the appropriate local interval varies across training stages, and 2) changing the number of inner steps per interval creates a mismatch with an unchanged outer optimizer, requiring a correction to the outer update. We optimize this mismatch by correction of the outer optimizer for the momentum and the learning rate using the accumulated inner learning rate. Our experiments under communication constraints demonstrate that AutoLoCo reduces communication frequency by 27% relative to DiLoCo while maintaining training performance. --- ### You Only Reprogram Once: Rethinking Prolonged Training for Visual Reprogramming Authors: Zizhao Li, Mohammed Yaqoob Ansari, Xinyu Su, Jiayang Ao, Joseph West, Kourosh Khoshelham Published: 2026-09-29 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.36661v1 Visual reprogramming is a parameter-efficient method for adapting pretrained models, yet its training can remain computationally expensive: even with a frozen backbone, visual prompts are often optimized through the full model for hundreds of epochs. Before changing what the pretrained model sees, we ask whether we are fully using what it already tells us. We find that modeling the full source response can already yield strong downstream predictions without prompt optimization. Motivated by this observation, we introduce You Only Reprogram Once (YORO), which constructs a downstream predictor from the frozen response space in a single forward-only traversal. Its Bayesian Discriminant Mapping (BDM) derives a covariance-aware affine mapping from streaming class statistics, requiring no backpropagation, optimizer updates, or repeated visits to the training set. When further input adaptation helps, YORO-FP optionally refines the visual prompt for 20 epochs. BDM also extends naturally to CLIP by treating attribute-prompt similarities as source responses. Across three full-data settings, YORO improves average accuracy over the strongest prior gradient-free mapping by 18.4--24.4\%. On 16-shot CLIP, it raises the four-backbone average from 71.4\% to 77.2\%. YORO-FP provides further gains on selected tasks, while validation often retains the one-pass predictor. These results suggest a different default for visual reprogramming: read out the frozen response first, and optimize the input only when needed. --- ### Byzantine-Robust Federated Representation Learning Authors: Leonardo F. Toso, James Anderson, Rafael Pinot, Nirupam Gupta Published: 2026-09-29 Categories: cs.LG, math.OC arXiv: https://arxiv.org/abs/2609.36660v1 We study federated learning (FL) with adversarial clients, where the goal is to minimize the average loss of the honest (non-adversarial) clients without knowing their identity. Under heterogeneity, a single shared model parameter is statistically inappropriate: it cannot capture the distinct data-generating processes across clients, incurring an irreducible model-heterogeneity bias and severely limiting robustness to adversarial clients (a.k.a. Byzantine-robustness). We address this problem through representation learning, where each client learns a personalized linear head, while collaboratively estimating a shared nonlinear representation through Byzantine-robust aggregation. We demonstrate that the heterogeneity among honest representation gradients is controlled by the representation error and statistical errors that decay either with the number of data samples per client ($τ$) or the number of iterations ($T$). In particular, our non-asymptotic parameter recovery error bound reveals three terms: (i) an initialization-dependent error that goes away with $T$, (ii) finite-sample noise terms that decreases with $τ$ and the number of honest clients, and (iii) a stochastic gradient variance term that also reduces with $T$. Importantly, with no irreducible model-heterogeneity bias in our bounds. We extend the regression analysis to multiclass classification, and empirically validate it on CIFAR-10, FEMNIST, and School Exam Score datasets. --- ### On-Policy Parameter Update Direction Underlies Generalization in LLM Post-Training Authors: Shufan Shen, Zhongni Hou, Junshu Sun, Yufei Zhang, Wei Lin, Guojun Yin, Qingming Huang, Shuhui Wang Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36659v1 The strong generalization performance of on-policy post-training paradigms has motivated studies of their parameter update behaviors. However, these studies treat the observed behaviors only as byproducts in on-policy training, overlooking their potential to serve as optimization principles for improving the generalization of other paradigms such as supervised fine-tuning (SFT). To address this limitation, we investigate whether there exists a specific on-policy update behavior that can achieve such improvements. First, our analyses reveal that SFT updates parameters along consistent directions, while the on-policy paradigm continuously adjusts the direction during training. This difference inspires us to focus on the cumulative update direction of each parameter as a promising behavior. Then, we evaluate its effectiveness for improving generalization by proposing On-Policy direction-constrained Supervised Fine-Tuning (OPSFT), which constrains SFT updates to the direction identified by on-policy paradigms. The strong performance of OPSFT indicates that the generalization advantage of on-policy paradigms can be transferred to SFT through the parameter update direction. Once such a direction is identified, even SFT can generalize with its updates constrained to this direction. This finding offers two practical benefits by combining the strong generalization of on-policy paradigms with the advantages of SFT, including the high training efficiency and ability to leverage high-quality trajectories. For efficiency, we identify update directions that support strong generalization using a few on-policy training steps, and subsequently apply OPSFT to achieve high training efficiency. For leveraging high-quality trajectories, OPSFT can utilize these trajectories to continue improving a post-trained model along its update direction without disrupting the ability learned from on-policy training. --- ### Constitutional adapters: Inference-time interventions for misalignment and misuse Authors: Adam S. Lowet, Mark Kurzeja Published: 2026-09-29 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.36657v1 Training models to act in accordance with an explicitly defined set of principles, or "constitution," has shown promise as a robust and transparent mechanism for AI alignment. However, the generality and flexibility of such methods remain unclear. Here, we show that constitution-consistent behavior can be distilled from synthetic corpora into lightweight objects (low-rank adapters and steering vectors). Despite never seeing a harmful request or jailbreak during training, such objects increase jailbreak defense success and measured alignment -- particularly at long context lengths and against multi-turn attacks, where they outperform both prompted and steered baselines. Subtracting control-trained from constitution-trained objects further accentuates these effects, yielding defenses we call "constitutional adapters" (CAs). CAs can be trained on a base model, transferred zero-shot to its post-trained checkpoint, and scaled at inference time to predictably trade off defense for benign compliance. Taken together, these results recommend CAs as a lightweight, portable, and tunable lever for mitigating misalignment and misuse in API deployments. --- ### Not Every Correction Helps: Gain-Guided Continual Test-Time Adaptation Authors: Youjia Zhang, Huiling Liu, Soyun Choi, Jaehong Yoon, Sungeun Hong Published: 2026-09-29 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.36655v1 Continual test-time adaptation (CTTA) adapts a source model to an unlabeled test stream whose distribution may change over time. Existing TTA methods often assess prediction reliability using confidence or entropy, which primarily reflect the model's self-certainty for the current sample. In CTTA, accumulated target observations can provide complementary evidence for correcting the source prediction, but this history may become misaligned as the target distribution changes. The key question is therefore not how much the correction differs from the source prediction, but whether and how strongly it should be applied. This paper proposes Gain-Aware INtervention (GAIN), a backpropagation-free CTTA framework guided by a simple principle: history proposes, gain decides. GAIN maintains compact target statistics to form a correction proposal and a posterior-predictive evaluator that accounts for estimation uncertainty. The resulting source-relative gain estimates the proposal's benefit and determines a sample-specific intervention strength along a continuous path through efficient one-dimensional optimization. Gain-controlled predictions then update the target statistics online, limiting the propagation of unreliable corrections, all without backpropagation, sample storage, or replay. Across five benchmarks, our method achieves strong predictive performance, with favorable accuracy--calibration--efficiency trade-offs in continual adaptation. On ImageNet-C, for example, GAIN achieves 61.9% accuracy with near-source calibration. It remains stable under diverse and challenging continual shifts while running 15.9x faster than a representative optimization-based CTTA baseline. --- ### Replay the Curvature: Accurate and Scalable NVFP4 Quantization for Large Language Model Inference Authors: Ruiyi Ding, Jie Li, Kang He, Ziyan Liu, Chengru Song, Yuedong Xu, Yuan Cheng Published: 2026-09-29 Categories: cs.LG, cs.CL, cs.DC arXiv: https://arxiv.org/abs/2609.36654v1 Large language models make weight storage and memory traffic major inference costs, motivating low-precision formats that represent each weight with only a few bits. Such formats use a scale to map floating-point values into a small codebook; NVFP4 improves local range utilization by letting every 16 E2M1 weights share an E4M3 block scale. Choosing that scale is difficult in GPTQ because quantizing one column updates those that follow, so evaluating a block independently can misestimate its final reconstruction error. Large models pose a second challenge: full-precision weights, calibration activations, and second-order state cannot all remain on one accelerator, while assigning complete layers to devices leaves each time-consuming layer solve serial. We introduce \emph{Schur Replay}, a scale-selection algorithm that reproduces the GPTQ updates caused by each block scale and scores the resulting block error after accounting for compensation from unquantized columns. Separately, our execution infrastructure keeps only the active layer resident, tiers activations across device, host, and disk, retires full-precision layers after export, and distributes independent output rows across tensor-parallel ranks. Together, the algorithm and infrastructure attain $99.35\%$ and $100.84\%$ question-weighted recovery from BF16 across seven benchmarks on Qwen3.5-397B-A17B and Llama-3.3-70B-Instruct. On the 397B model, the infrastructure reduces measured per-layer time by $15.17\times$ over ModelOpt and $23.14\times$ over LLM Compressor, with lower memory used per GPU. --- ### Scheduling Recursive Reasoning in Looped Transformers Authors: Boyuan Wang, Chengyao Yu, Jiaxi Ren, Hongxin Wei, Bingyi Jing, Yuxin Tao Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36653v1 Recurrent reasoning models have attracted growing attention for scaling test-time computation, typically by iteratively refining latent states with shared parameters. However, these models apply each learned update with a fixed unit scale, which can be conservative when updates make persistent progress and overly aggressive when they fluctuate, limiting the benefit of additional loops. To understand how the scale should vary along the trajectory, we first analyze the sensitivity of terminal loss to recurrent update scale. We show that its temporal average admits an exact decomposition into persistent-progress and centered-fluctuation contributions. Based on this, we introduce the Trajectory Adaptive Progress-Fluctuation Scheduler (TAPS), which tracks their balance across recurrent updates and adapts the step size online. Theoretically, we establish sufficient conditions under which TAPS reduces expected terminal loss and reaches a target quality in fewer recurrent loops. Empirically, we show that TAPS improves terminal accuracy across structured reasoning tasks without retraining. By further incorporating the progress-fluctuation principle into training, TAPS yields additional accuracy gains with up to 1.56 times wall-clock speedup at matched baseline accuracy. The broad applicability of TAPS is supported by its effectiveness across diverse recurrent architectures and inference strategies. Together, these results establish update scale as complementary control axis of recurrent inference alongside architecture and depth. --- ### RankBuffer: Efficient Ranking-Based Rewards for Open-Ended Generation Authors: Zixuan Yang, Yiqun Chen, Qi Liu, Wei Yang, Erhan Zhang, Liyi Chen, Qimeng Wang, Yan Gao, Jiaxin Mao Published: 2026-09-29 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36652v1 Open-ended generation lacks canonical answers, making pointwise rewards difficult to calibrate for group-based reinforcement learning. Directly ranking same-query rollouts provides a more suitable relative reward signal, but existing ranking-based reward methods can incur substantial judging cost. We introduce RankBuffer, which maintains an ordered, query-specific buffer of previously judged responses as a reusable quality scale. Each rollout is first inserted into an anchor interval through an independent coarse judgment, after which only rollouts assigned to the same interval undergo local fine ranking. The resulting complete order is converted into bounded rank rewards, while boundary expansion, local refinement, and inactive-anchor pruning adapt the buffer as the policy evolves. Across four open-ended benchmarks, RankBuffer consistently outperforms all pointwise baselines. It also achieves nearly on-par performance with the strongest ranking-based reward baseline while substantially reducing judging cost. Ablations demonstrate the importance of both local fine ranking and anchor response content, while buffer analyses show that rollout-derived anchors progressively extend and refine the covered quality scale. These results establish response reuse as an effective approach to efficient relative reward construction. --- ### FocusVTC: Efficient and High-Performance Visual Text Compression with Adaptive Resolution Authors: FangZhi Zhong, Xuerui Qiu, Yuqi Pan, Ya Liu, Shaowei Gu, Bo Xu, Guoqi Li Published: 2026-09-29 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.36651v1 Long-context reasoning in large language models incurs substantial computation and memory costs. Visual text compression (VTC) reduces input length by rendering text as images, but fixed-resolution rendering creates a compression-performance trade-off: low DPI saves tokens at the expense of legibility, whereas high DPI spends tokens on irrelevant content. We introduce FocusVTC, which breaks this trade-off through adaptive resolution while preserving general multimodal capabilities. It combines compressed low-DPI global views with selective region enhancement, integrating enhanced views into ongoing reasoning. We construct 29.4K high-quality Reasoning-Evidence Localization (REL) chain-of-thought examples (REL-CoT) that link reasoning traces to page indices and bounding boxes. Multi-resolution REL supervised fine-tuning (REL-SFT) teaches the model to localize relevant regions, and Group Relative Policy Optimization learns when to enhance resolution and how to use the resulting observations, without a separate continual-pretraining stage. At 72 DPI on RULER v1, FocusVTC scores 87.4 at $2.9\times$ input compression, including tool observations, versus 57.5 for Glyph at $3.0\times$ input compression. It surpasses its text-input backbone on LongBench (56.40 versus 55.86), improves the MRCR macro-average by 13.91 points, and achieves a 51.19 macro-average on VTCBench. The MRCR latency evaluation also shows a $2.79\times$ online end-to-end speedup over Text. General multimodal capabilities are preserved, with MMMU increasing from 65.12 to 66.73 and MME from 2424.02 to 2457.62. --- ### VLM4Cluster: Benchmarking Deep Clustering In the Era of Vision-Language Pre-training Authors: Yuanwei Hu, Bo Peng, Yuheng Jia, Xinting Hu, Yadan Luo, Wenjie Zhu Published: 2026-09-29 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.36648v1 Vision-language pre-training has reshaped image clustering, giving rise to language-assisted image clustering (LaIC), which leverages textual semantics to complement visual representations. Despite the rapid proliferation of LaIC methods, it remains unclear how much LaIC has actually advanced image clustering, as existing studies generally suffer from major limitations, including inconsistent experimental settings, inadequate dataset selection, and limited evaluation dimensions. To address this gap, we introduce VLM4Cluster, a comprehensive benchmark for image clustering in the era of pre-trained vision-language models (VLMs). VLM4Cluster implements 17 representative methods spanning classical, deep, and language-assisted image clustering, and evaluates them on 20 datasets covering classical, challenging, fine-grained, large-scale, and out-of-distribution settings. Beyond effectiveness, VLM4Cluster systematically investigates image clustering along three complementary dimensions: robustness to adversarial perturbations, generalization under distribution shifts, and computational efficiency. Our study shows that LaIC substantially advances the clustering performance frontier on many semantically demanding benchmarks, generally exhibits stronger generalization under distribution shifts, and achieves a more favorable effectiveness-efficiency trade-off. However, its gains become less consistent on large-scale and fine-grained datasets, while language assistance does not systematically reduce sensitivity to adversarial perturbations. VLM4Cluster is released at https://github.com/YuanweiHuu/VLM4Cluster. --- ### Where Predictive Supervision Goes Shapes What VLA Policies Learn Authors: Hanseul Kim, Jewon Yeom, Youngjoon Jeong, Minsoo Jo, Taesup Kim Published: 2026-09-29 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.36645v1 Future prediction is increasingly used to improve vision-language-action (VLA) policies, based on the premise that anticipating scene evolution encourages representations useful for control. However, forecast quality alone does not establish that a policy has learned a better representation for action. This distinction matters under distribution shift, where successful control depends on preserving spatial state and likely scene change beyond familiar configurations. We study what determines whether predictive supervision improves the visual representation used by a VLA policy. Through controlled comparisons with matched target constructions, prediction horizons, and training conditions, we find that different prediction interfaces produce markedly different forecasts and visual representations, including in the spatial, dynamics, and action information that transfers beyond familiar scenes. We trace these differences to how predictive errors shape the policy's visual stream. Consistent with this controlled finding, VLA policies trained with more direct, scene-matched future supervision show stronger robustness under simulated and physical distribution shifts. Together, our results frame future prediction as a representation-learning design problem whose value for control depends on whether its supervision reaches the representations through which the policy acts. --- ### OCA: ODE-Driven Cross-Attention for Image-to-Point-Cloud Registration Authors: Pei An, Jiaqi Yang, Yulong Wang, Siwen Quan, Liangliang Nan Published: 2026-09-29 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.36644v1 Cross-attention is a crucial component in learning-based image-to-point-cloud (I2P) registration. Although existing cross-attention mechanisms have achieved promising progress, attention ambiguity remains a fundamental challenge that hinders the learning of discriminative 2D-3D correspondences. To address this problem, we revisit cross-attention and establish ordinary differential equations (ODEs) to model the ideal I2P feature interaction. Based on this formulation, we develop an ODE-driven cross-attention (OCA) module that refines feature representations and attention matrices through ODEs. In practice, OCA can be seamlessly integrated into existing I2P registration frameworks. To validate its effectiveness, we incorporate OCA into five state-of-the-art baselines and evaluate on four public benchmark datasets. Experimental results demonstrate that OCA improves registration recall by up to 5\%, 9\%, and 15\% under the standard, fine-tuning, and zero-shot settings, respectively. --- ### PR-OPD: Privileged Representation On-policy Self-Distillation for Agentic Reinforcement Learning Authors: Muyang Li, Jie Yang, Zhengyu Fang, Junchao Zhu, Zhengkun Xiao, Ruining Deng, Zhe Jiang, Shigang Chen Published: 2026-09-29 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36642v1 Language-model agents are usually trained by reinforcement learning from one reward per episode, and privileged self-distillation enriches it by letting the same policy, given a skill, teach its skill-free self through token probabilities. However, we identify two phenomena that question this channel. Invisible Advantage: a skill in context lifts WebShop success from 42.2% to 56.2%, yet changes the probabilities of fewer than a quarter of the sampled tokens. Much to Align: a skill changes the hidden states of over 80% of response tokens, in a way that linear probes can trace back to the specific skill. To exploit this, we propose Privileged Representation On-policy Self-Distillation (PR-OPD). After a GRPO warm start, the policy writes a hindsight skill for each trajectory, re-reads its own responses with that skill as a stop-gradient teacher, and aligns its projected hidden states to the teacher's at every layer alongside the reward objective, with no external skill library, separate teacher, or inference overhead. On ALFWorld and WebShop with two backbones, PR-OPD achieves the best overall results in every setting, improving over GRPO by up to 4.7 points in ALFWorld success and 14.0 points in WebShop accuracy. Code is available at https://github.com/balibata/PR-OPD. --- ### Inducing Process Supervision from Outcome-Only Reinforcement Learning Authors: Shengda Fan, Xin Cong, Zhong Zhang, Haotian Chen, Yankai Lin Published: 2026-09-29 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.36641v1 Process reward models (PRMs) have become a key component for LLMs, as their step-level feedback supports both post-training and test-time reasoning. However, training strong PRMs remains costly: human step annotation is difficult to scale, while Monte Carlo estimation is computationally expensive and can drift from the intrinsic correctness of steps. To get effective PRMs at low cost, we introduce TIPS (Thinking-Induced Process Supervision), an outcome-only reinforcement learning (RL) framework for training generative PRMs. In TIPS, the model generates a chain-of-thought (CoT) followed by step-level labels and an outcome label. The reward depends solely on whether the predicted outcome matches the ground truth, and the resulting group-relative advantage is used to optimize the entire generated response. Intuitively, when checking intermediate steps helps determine the outcome, more accurate checks can lead to better outcome judgments and higher rewards. Outcome-only RL can therefore reinforce step-level verification without explicit process supervision. We validate the effectiveness of TIPS across math and agent benchmarks and four backbone families. Notably, TIPS-Qwen3-4B-Thinking-2507 reaches 85.2 F1 on ProcessBench with only 3.2K outcome-labeled trajectories, surpassing all evaluated trained PRMs and strong prompt-only judges such as GPT-5.4-Instruct and Claude-4.7-Opus, while still trailing o1-mini. Code and data are available at https://github.com/RUCBM/TIPS. --- ### A Digital Simulation Toolkit for Physics-Based Generation of Realistic Experimental Scanning Tunneling Microscopy Images Authors: Huanhuan Zhao, Laxmi Bhurtel, Connor Vernachio, Fahmy Paiziah, Wonhee Ko, Arpan Biswas Published: 2026-09-29 Categories: cs.LG, cond-mat.mtrl-sci arXiv: https://arxiv.org/abs/2609.36639v1 Scanning Tunneling Microscopy (STM) is a widely used tool for characterizing surfaces of materials at the atomic scale, playing a crucial role in discoveries across condensed matter physics and materials science. Despite its extreme spatial resolution, STM is one of the most sensitive microscopy techniques and is highly prone to noise. While existing unsupervised denoising methods are very cheap to train, these are primarily focused on removing the noise with minimal recovery of key physical information. While supervised methods can offer superior performance, the major bottleneck is that a large amount of paired clean-noisy experimental images is required which are impractical to obtain. Thus, we developed a low-cost physics-driven digital toolkit to rapidly generate large volume of realistic STM images. Firstly, we simulate clean images from a chosen material system. Then, with prior knowledge of the physical characteristics of the artifacts and noise present in STM experiments, we formulate several artifact-noise functions such as Gaussian electronic noise, 1/f flicker noise, scan-line noise, background tilt and mechanical drift. These physically informed noise components are then added to the simulated clean images to generate realistic STM images. We demonstrated the capability of the proposed digital toolkit to generate AI-ready data for denoising images of the (111) surfaces of copper and lead, while preserving atoms, defects, and electron waves. We also validated the quality of the downstream image analysis of learning electron wave patterns induced by quantum interference from Cu(111) images. Results show that the supervised models trained on digitally generated AI-ready data can more effectively denoise and learn electron wave patterns on Cu(111) images than benchmarked unsupervised approaches, indicating that the proposed toolkit facilitate scientific discovery. ---