# Research Papers — 2026-09 · bundle page 31 of 266 > Full-text markdown of papers #1501–1550 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/31 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/30.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/32.md) ### ATLAS: Aligned Transport of Latent Structure for Reliable World Model Planning Authors: Ke Fang, Yupu Yao, Lu Cheng Published: 2026-09-28 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.36333v1 Latent world models rely on representation geometry for planning, yet regularizing the latent marginal alone does not determine the state-to-state relationships used for action selection. We show that this can cause planning-relevant novelty structure to be weakened as representations are transformed into the final latent used by the planner. We introduce Aligned Transport of Latent Structure (ATLAS), a training objective that explicitly preserves relational geometry while calibrating the global latent distribution. ATLAS transfers normalized pairwise structure from an informative encoder representation to the planning latent and uses Wasserstein embedding matching (WEMReg) to calibrate its marginal through one-dimensional Wasserstein-2 transport. Our analysis shows that relational preservation and marginal calibration impose non-redundant constraints, and connects finite-candidate planning stability to relational distortion, latent-scale mismatch, and prediction error. Instantiated in LeWM, ATLAS improves mean goal-reaching success across PushT, TwoRoom, and OGBench-Cube on both lower- and higher-novelty evaluation subsets, with the largest gain on higher-novelty TwoRoom episodes. Representation and rollout diagnostics further show stronger novelty-related structure in the planning latent, improved marginal calibration, and lower multi-step prediction error. Together, these results highlight preservation of planning-relevant latent geometry as an important ingredient for reliable world-model planning. Code is available at https://anonymous.4open.science/r/atlas-world-model-72C4/. --- ### Calibrating One-Round Membership Inference with Neighbors Authors: Francesco Rita, Jie Zhang, Florian Tramèr Published: 2026-09-28 Categories: cs.CR, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.36331v1 The state-of-the-art Membership Inference (MI) methods calibrate their signal separately for each example using reference models, auxiliary models trained to exclude the target. This paradigm scales poorly to modern large models, however, whose training is too expensive to replicate. This has motivated one-round settings, where only a single trained model is available; but without reference models the per-example calibration that drives the strongest attacks can no longer be estimated, leaving the membership signal weak. We ask whether neighbors of the target point can recover this calibration without training any additional model. Our key observation is that reference models serve only to reveal how an example behaves under models not trained on it, and that querying the target model on nearby samples yields the same information. We propose two complementary ways to obtain such neighbors, and show that querying them against an early training checkpoint further sharpens the signal. We evaluate across three image classification datasets and three training setups, showing that neighbors yield strong membership signals and competitive attack performance at no additional training cost. --- ### DecoyTrace: Toxic Decoys for Active Defense in Decentralized Federated Learning Authors: Pedro Beltrán-López, Enrique Tomás Martínez Beltrán, Pantaleone Nespoli, Manuel Gil Pérez, Alberto Huertas Celdrán Published: 2026-09-28 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.36330v1 Decentralized Federated Learning (DFL) eliminates the central aggregation server, reducing the single point of observation that traditional defenses against attacks rely on. As a result, peer-to-peer networks become exposed to malicious updates containing backdoors or semantic poisoning, since such updates can remain close to benign ones in the parameter space while behaving very differently. This may evade defenses based on passive parameter inspection. However, existing deception-based defenses have mainly been designed for centralized FL and do not jointly address local observation, poisoning propagation, source attribution, and containment in strictly serverless DFL. To address these limitations, this paper presents DecoyTrace, a proactive cyber deception-based defense for strictly serverless DFL environments. DecoyTrace deploys a mobile DecoyNode that generates decoy challenges using chaotic maps, disseminates a dual model (clean vs. decoy) based on neighbor trust, and evaluates them using three-state semantic metrics. Upon confirmation, a distributed protocol isolates the source and performs a model reset or recovery to preserve training progress. Evaluated across sixty configurations on the NEBULA platform (five datasets, three topologies, and four attack/defense scenarios), DecoyTrace systematically restores lost utility. The F1-score remains within 0.03 of the baseline on MNIST/FashionMNIST (mitigating drops of up to 0.37), matches or exceeds the baseline on EMNIST and CIFAR-100, and remains between 0.05 and 0.10 below the baseline on CIFAR-10, the most visually complex convolutional scenario evaluated. Furthermore, containment reduces CPU and network usage by up to two-thirds. These results demonstrate the feasibility of unifying deception, identification, and containment in DFL, while also identifying its limitations in complex tasks and multi-attractor threat models. --- ### Reducing the Adaptation Gap Through Reachable Fisher Geometry Authors: Wasif Jalal, Sachin Deb, Asif Salekin Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36329v1 Parameter-efficient fine-tuning (PEFT) determines not only how many parameters are trained, but also which local directions a model can move in, so similar adapters can affect subgroup losses differently. Since curvature matrices are infeasible to form at adapter scale, scalar summaries such as the Fisher trace are often used instead. We study what the trace reveals and what it loses through the reachable Fisher: each subgroup's full-model Fisher pulled back through the adapter Jacobian. Under likelihood losses, it represents the Gauss-Newton curvature accessible to the adapter, and its trace can be computed from score-gradient norms without forming the full matrix. Under matched subgroup gradients, a positive-definite reachable-Fisher difference, with a margin exceeding the Hessian-Fisher defect, implies that every sufficiently small nonzero model-changing update increases the signed gap. In contrast, the restricted operator norm determines worst-case quadratic change, while a matrix-free Frobenius discrepancy bounds its reachable-Fisher component. Trace alone cannot certify definiteness or control matrix mismatch. Equal traces rule out a positive-definite difference but can still hide large operator discrepancies. Across 306 single-seed models, higher trace accompanies greater subgroup difficulty in 75.7 percent of 1,218 eligible evaluations, while trace matching reduces the best-worst subgroup gap in all 30 dataset-encoder-adapter combinations. However, held-out audits show that operator discrepancy decreases in 23 of 30 combinations, while the unbiased squared-Frobenius statistic decreases in only 16 of 30. Fisher trace is therefore a scalable diagnostic and training heuristic, but not a certificate of local gap behavior or matrix alignment. --- ### PILLAR: Private Inverted-Index Lexical Lookup for Augmented Retrieval Authors: Truong Son Nguyen, Daniel Blackley, Ni Trieu, Evgenios M. Kornaropoulos Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36326v1 Retrieval-augmented generation (RAG) hands the user's query to whoever hosts the corpus. We propose PILLAR, a Privacy-Preserving RAG (PPRAG) system based on Private Information Retrieval (PIR) in which a client utilizes the k documents most similar to their query from a server-held and publicly known corpus to respond to their query, while the server learns nothing about the query, either its terms or its access pattern. Prior PPRAG constructions rely on dense retrieval alone, translating approximate nearest-neighbor search into many query-dependent rounds of PIR, and pay for it in both latency and retrieval quality. PILLAR instead performs private hybrid retrieval in two stages. A sparse stage issues a small, fixed number of PIR queries against a carefully designed index of precomputed BM25 scores, filtering the corpus down to candidates that share terms with the query without the server ever seeing which terms these are. A dense stage then fetches only those candidates' document embeddings and re-ranks them locally, avoiding the many costly PIR queries that private dense retrieval typically requires. We instantiate PILLAR with two protocols that trade latency against retrieval quality, each built on a different private rendering of lexical search. PILLAR-Bin bins posting lists into a hash table and is a single-round design that achieves lower latency than state-of-the-art private retrieval schemes. PILLAR-Tree turns block-max pruning into an oblivious tree traversal combined with cuckoo hash tables and achieves the highest retrieval quality at lower latency than state-of-the-art schemes. --- ### Distill Locally, Schedule Globally: Flow Maps for Few-Step Text-to-Speech Authors: Yentl Collin, Evan Dufraisse, Amr Mohamed, Amine Khelif Khelif, Dani Bouch, Guokan Shang Published: 2026-09-28 Categories: cs.SD, cs.LG arXiv: https://arxiv.org/abs/2609.36324v1 Flow-matching text-to-speech (TTS) models achieve high synthesis quality but require many neural function evaluations (NFEs) to integrate their generative trajectories. Recent few-step flow-map distillation approaches for TTS construct targets from numerically integrated teacher trajectories, creating a trade-off between target accuracy and training cost. We propose Local Flow-Map Distillation (LFMD), which adapts Eulerian Map Distillation to conditional TTS and avoids teacher trajectory integration during target construction. For inference, we derive a sampling schedule (TD-DP) from teacher dynamics and consistency of the learned maps, with a single cost graph supporting multiple NFE budgets without external audio-metric evaluation. Because scheduling offers no flexibility at one NFE, we refine this regime with alignment-aware temporal self-distillation using soft-DTW. Across Seed-TTS and LibriSpeech-PC, LFMD improves low-NFE synthesis over a matched integral-distillation baseline. On Seed-TTS, the refined student reaches 1.80% WER with 1-NFE, compared with 1.76% for its 32-NFE teacher. --- ### Towards an AI Software Factory for Data Systems Authors: Anna Pavlenko, Bogdan Crivat, Brandon Haynes, Carlo Curino, Fotis Psallidas, Jaro Slawinski, Johannes Freischuetz, Laura Pereira Sanchez, Markus Weimer, Mathieu Demarne, Matthias Jasny, Mauktik Gandhi et al. Published: 2026-09-28 Categories: cs.AI, cs.DB, cs.SE arXiv: https://arxiv.org/abs/2609.36323v1 AI-assisted coding tools deliver significant acceleration of coding, but only limited impact across the end-to-end software development lifecycle (SDLC)--an Amdahl's law effect! In this paper, we discuss our progress towards building an AI SW Factory that accelerates all the stages of SDLC-Targeting, Coding, Reviewing, and Ops. The AI SW Factory produces a metadata exhaust that enables self-improvement by fine-tuning model weights and updating our World Model (a rich data substrate). We focus on Data Systems and the important class of Evolutionary Coding Tasks (i.e., those with a measurable objective to hill-climb) and report on 1) scaled deployments at Microsoft (tens of repositories) leading to 3x engineering efficiency above agentic coding and up to 22x token efficiency, and 2) several open challenges. --- ### Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression Authors: Xingyu Zhu, Pu, Yi, Ziheng Cheng, Ang Lv, Jing Liu, Lexing Ying, Yiyuan Ma, Xin Dong Published: 2026-09-28 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.36322v1 Chunked KV-cache compression reduces the memory and attention costs of long-context inference by compressing windows of consecutive tokens into fewer cache entries at a fixed stride. Such compression also introduces a new positional coordinate: a token's phase, or its position relative to compression-window boundaries. We uncover a systematic asymmetry in models using such compression: the same information can be easy to retrieve at one phase and difficult at another. We call this periodic variation in retrieval performance phase sensitivity. In large open-weight models with such compression, long-context retrieval accuracy can differ by up to 40 percentage points across phases, revealing periodic weak spots that average benchmark scores can conceal. To investigate this behavior, we pretrain a family of transformers from scratch across multiple KV-compression designs, reproducing phase sensitivity across the variants. Mechanistic analysis using causal interventions in these models reveals phase specialization: different attention components contribute asymmetrically to retrieving information at different source phases. We further analyze idealized retrieval models, showing how gradient flow dynamics may favor sharp phase specialization. Evaluating models with chunked KV-cache compression thus requires measuring across compression phases: high average accuracy can coexist with systematic positional failures. --- ### StateTape: Action-Conditioned Evidence Lifecycle Modeling for Long-Horizon Coding Agents Authors: Ziyang Yu, Liang Zhao, Bowen Zhu, Hasibul Haque Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36319v1 Despite the recent success of coding agents built on large language models, it remains challenging to run them over long horizons, since every observation is appended to the context and the context grows with each one. History-based maintenance is a common remedy, which masks or summarizes old observations, or prunes what a model reads as useless, and bounds the context at little cost. However, it decides from the text of the history alone and sees nothing of how the code is connected. Since a coding agent edits code many times over a single task, and each write can change what code elsewhere means, such maintenance may keep records a write has falsified, drop ones that still hold, and miss code the agent needs next. To overcome these challenges, this paper proposes StateTape, a novel and scalable framework that rewrites a coding agent's context as the repository changes rather than as the context grows. The key idea of StateTape is to model the repository as a symbol-level code graph, whose dependencies and language rules expose which symbols a write can affect. Upon this graph, a tape marks the symbols each write changed, which turns staleness from an inference about text into an observation of the agent's writes. We propose a per-write procedure in which the tape nominates the records a write could have falsified while a small manager model settles what the write log cannot, and further provide a theoretical analysis and TraceBench, a benchmark that labels what an agent is holding against what is actually needed. Empirically, we demonstrate that StateTape can effectively clear falsified records and retrieve what is needed, and thus achieve a higher resolve rate in all experiments spanned by six coding agents and three edit-heavy benchmarks with little computational overhead. --- ### Training LLMs to Verbalize Evaluation Awareness Authors: Usman Anwar, Sahar Abdelnabi, David Krueger Published: 2026-09-28 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.36316v1 Evaluation awareness (EA) can cause large language models (LLMs) to behave differently during audits than in deployment, yet measuring and accounting for EA remains challenging. We introduce verbalization training (VT), a method for making LLMs less reticent about verbalizing evaluation awareness while avoiding to supervise the latent belief itself. VT uses a model's spontaneous verbalizations as evidence that awareness is present and truncates each rollout immediately before the verbalization, producing training prefixes at which the model is presumed to be aware. The model is then trained with an RL objective designed to increase verbalization in a calibrated way. Across Qwen3.6-35B-A3B, Kimi K2.6, and Inkling, VT increases verbalized EA by 2.4-2.9 times and transfers to held-out agentic settings, while measured latent EA and behavior remain largely stable. In a causal experiment, we independently implant meta-knowledge about evaluations through synthetic-document fine-tuning and show that VT-induced verbalizations reflect the richer knowledge acquired by the model. --- ### PyroStack: A Multi-Band Spatio-Temporal Sub-Daily Dataset for Wildfires in the United States Authors: Arya Kondur, Giosue Migliorini, Cameron Schmitt, Francesco Immorlano, Tairan Wang, Rebecca C. Scholten, Efi Foufoula-Georgiou, Gary Johnson, Chris Lautenberger, Valentin Waeselynck, J. Shane Romsos, Kasra Shamsaei et al. Published: 2026-09-28 Categories: cs.LG, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.36315v1 Wildfires are an increasing hazard to ecosystems, air quality, and human systems, creating a growing need for datasets that support systematic development and evaluation of models for predicting fire spread across diverse landscapes. Effective prediction requires integrating meteorological conditions, fuels, vegetation, and topography at spatial and temporal resolutions suitable for both physical simulation and data-driven approaches. However, existing datasets often lack the resolution and coverage needed to capture these interacting controls. The PyroStack dataset addresses this gap by providing a harmonized, event-based collection of wildfire and environmental data across the contiguous United States and Alaska. It integrates satellite-derived fire observations with atmospheric reanalysis, vegetation, fuel characteristics, and topographic information into a unified framework spanning 6994 wildfires that occurred between 2012 and 2024 across a wide range of ecosystems and climate conditions. PyroStack offers spatial resolutions ranging from 30 m to 9 km and hourly temporal resolution, along with fire progression data at 12-hour intervals to support model initialization and evaluation. By combining broad spatial coverage with fine spatial and temporal detail, the dataset enables systematic analysis of wildfire dynamics and supports both physics-based and machine learning approaches, providing a foundation for benchmarking and improving fire spread models, with future extensions aimed at incorporating additional regions and fire suppression data streams to further advance wildfire prediction. --- ### Fractional State Space Transition for Long Sequence Modeling Authors: Ivan Kobyzev, Abbas Ghaddar, Ali Nasiri-Sarvi, Lifeng Shang, Yufei Cui Published: 2026-09-28 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.36314v1 State Space Models (SSMs) compress sequence history into a bounded recurrent state, making the resulting memory law a central architectural choice for long-context performance. Most modern SSMs rely on ODE-based dynamics that lead to exponential forgetting, limiting their ability to retain information over broad temporal ranges. We introduce FRAC, a selective SSM architecture derived from fractional dynamics that replaces this exponential decay with power-law long memory. To make fractional dynamics practical, FRAC approximates the heavy-tailed target kernel with a finite-state, log-spaced sum of exponential modes. This construction turns fractional memory into an efficient recurrent module with parallel training and prefill, while retaining bounded-state autoregressive decoding. Extensive experiments, including 1.3B-parameter language modeling, demonstrate that FRAC consistently improves long-context performance over state-of-the-art SSM baselines while staying competitive on short-context. These results show that fractional dynamics provide a practical and effective prior for long-context SSMs. --- ### Learning Samples Importance: Parameterizing Dual Variables in Everywhere Learning Authors: Ignacio Boero, Jonathan Nixon, Alejandro Ribeiro Published: 2026-09-28 Categories: cs.LG, eess.SP arXiv: https://arxiv.org/abs/2609.36310v1 Everywhere learning provides a principled framework for training AI models under constraints that must hold throughout the data distribution. In the dual domain, these pointwise constraints give rise to functional dual variables. In this work, we propose to learn these dual variables, motivated by the fact that their values encode useful information about the underlying constrained problem. By representing the dual variable as a parametric function of each sample, we enable the learned multiplier to be evaluated on new, unseen samples. This contrasts with standard empirical dual formulations, which assign an independent multiplier to each training sample. We characterize the error in the recovered primal solution induced by restricting the dual variable to a parametric function class and show that it is controlled by how well this class approximates the optimal statistical multiplier. Moreover, we show that the learned parametric multiplier retains the sensitivity interpretation of the optimal statistical multiplier, yielding approximate sensitivity guarantees that extend beyond the samples used for training. We empirically validate our theory across a variety of everywhere learning tasks, showing that the resulting constrained problems can be solved efficiently and that the learned dual variables provide meaningful representations of sample-level sensitivity. --- ### CheatBench: Measuring Reward Gaming in AI Agents Authors: Long Phan, Stephen K. Yang, Jason J. Lim, Mantas Mazeika, Wenyu Zhang, Zheyuan Liu, Richard Ren, Jingxiang Meng, Yaoteng Tan, Weiliang Zhao, Addison Wu, Matei Anghel et al. Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36308v1 Reinforcement learning has helped AI agents solve increasingly difficult tasks, but high rewards do not always reflect the work users intended. In recent incidents and controlled evaluations across the AI industry, agents trained to maximize reward have accessed unauthorized information, attempted to evade monitoring systems, and even breached sandbox protections to attack external systems. As agents become more capable, this behavior could pose increasingly serious risks. To measure this problem, we introduce CheatBench, a benchmark of cheating in AI agents across mathematical research, knowledge work, coding, visual tasks, and other domains. Its environments combine challenging assignments with opportunities to cheat, allowing researchers to study how agents pursue a goal when honest work is difficult. CheatBench supports comparisons across models and task categories, providing a testbed for measuring and reducing cheating as agents take on more consequential responsibilities. We publicly release CheatBench at https://cheatbench.ai --- ### Cheap and Powerful Tests for Supervised Subspaces: Per-Component Inference for PLS Authors: Paweł Lenartowicz, Hubert Plisiecki Published: 2026-09-28 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.36307v1 Partial Least Squares (PLS) regression extracts a few outcome-aligned directions in a high-dimensional X and is widely used across applied science, but inference on the resulting fit is either expensive, biased and discouraged, or absent. We reduce inference to held-out OLS refits of the supervised subspace, a primitive shared by PLS, supervised PCA, and linear probes, and supply two tests using held-out correlations: a Nadeau-Bengio corrected asymptotic t-test as a fast approximation, and a permutation test with comparable power, finite-sample valid under outcome-predictor independence and iid rows. Held-out predictions are unchanged under any orthogonal rebasing of the supervised span, so an interpretable basis such as varimax inherits the joint claim but not a per-axis p-value; per-component claims come from a fixed-sequence test on the PLS extraction order. We validate on synthetic geometries, two NIR chemometric datasets, and cross-lingual word-embedding regressions; the exact test also transfers to supervised PCA and a ridge probe. The proposed tests have more power than CV-permutation-Q^2, at a fraction of its cost. A pre-run check on n and the spectrum of X says when the approximation is safe. We release a Rust library with Python, R, and Julia bindings, plus a Python text pipeline. --- ### HeurEvo: Agentic Evolution of Hybrid Solver-Augmented Heuristics for Time-Critical Mathematical Optimization Authors: Feijie Wu, Hugo Barbalho, Konstantina Mellou, Marco Molinaro, Jing Gao, Ishai Menache, Xinzhi Zhang, Sirui Li Published: 2026-09-28 Categories: cs.CL, cs.LG, cs.NE arXiv: https://arxiv.org/abs/2609.36303v1 Recent advances in agentic heuristic design use AI agents and execution feedback to automate algorithm discovery for challenging optimization problems. In many practical settings, high-quality solutions must be obtained under strict runtime constraints, motivating hybrid approaches that combine problem-specific heuristics with powerful mathematical programming solvers. However, existing approaches typically improve heuristic components within predefined procedures or tune solver configurations in isolation. This limits holistic adaptation of where to allocate computation, how to leverage solvers, and how to refine the overall algorithmic structure. To address these limitations, we propose HeurEvo, an automated plan--code--component co-evolution framework that jointly evolves the high-level algorithmic structures, their implementations, and a shared pool of reusable components. A planner determines which algorithmic components to use, how to combine them, and how to allocate runtime across stages, a coder realizes the resulting plan as executable code, while a component evolver updates the shared component pool. Within an island-based evolutionary framework, plans and implementations co-evolve with feedback from an interpreter agent that analyzes execution results and identifies opportunities for improvement. Across diverse combinatorial optimization benchmarks and challenging MIPLIB instances, HeurEvo finds high-quality solutions within tight runtime budgets, often matching or surpassing state-of-the-art optimization solvers given hours or days of computation. On several nonlinear geometry problems such as hexagon packing, it also improves upon the best previously reported results. These results highlight the value of jointly searching over algorithmic structure and implementation for agentic heuristic design. --- ### The Universal Classifier for Graph Learning Authors: Ben Finkelshtein, André Linhares, Petar Veličković, Bryan Perozzi, Mikhail Galkin Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36302v1 While foundation models have revolutionized natural language processing and computer vision by leveraging universal vocabularies, Graph Machine Learning (GML) remains fractured due to the absence of a unified feature and structural representation across diverse domains. Existing works claiming to be Graph Foundation Models (GFMs) are typically restricted to node-level predictions or require fixed feature dimensions, failing to provide a truly task-agnostic backbone for the full spectrum of graph learning applications. In this paper, we introduce the Universal Classifier (UC), which supports arbitrary feature and class cardinalities, unifying node-, edge-, and graph-level objectives under a single similarity-based classification objective. The UC reformulates all node-, edge-, and graph-level prediction tasks as maximizing similarity in the latent space: by lifting heterogeneous features and labels into 3D latent tensors, the model learns transferable features independent of specific input schemas. This architecture allows a single pre-trained model to generalize to node classification, node regression, and link prediction across unseen graphs with varying feature semantics. Experiments show strong zero-shot transfer performance across node-, link-, and graph-level tasks. --- ### MoRE: Scaling mixture of experts with hardware-aware low-rank routing Authors: Honam Wong, Surbhi Goel, Enric Boix-Adserà Published: 2026-09-28 Categories: cs.LG, cs.AI, cs.CL, stat.ML arXiv: https://arxiv.org/abs/2609.36301v1 Mixture-of-Experts (MoE) layers are central to frontier language models, and recent architectures push toward more and smaller experts. In this regime, the standard linear router becomes a bottleneck: with $M$ experts and hidden dimension $h$, its per-token cost $Θ(Mh)$ dominates the MoE layer once $M$ is large. We introduce MoRE (Mixture of Rank-reduced-routed Experts), which factorizes the router weight matrix at rank $r$ and reduces the routing cost to $O((h + M)r)$. We prove that rank logarithmic in $M$ suffices for routing expressivity when the number of active experts is fixed, and is necessary up to precision factors. We also prove that logarithmic rank preserves load balance in a Gaussian memorization model, and training on a synthetic phonebook task shows that low rank does not hurt memorization. At matched active FLOPs, the factorization allows a factor of $Θ(h/r)$ more experts. To realize this gain in wall-clock time, we design a fused Triton kernel at inference that avoids expensive memory operations on HBM. Empirically, MoRE improves memorization on the phonebook task and performance on knowledge-intensive Q\&A benchmarks after pretraining, while matching reasoning ability. Code available at https://github.com/Matheart/MoRE_code. --- ### When Trees Are Not Enough: Learning Mixed-Topology Feature Graphs with Adaptive Graph Sparse Autoencoders Authors: Xiaozuo Shen, Yifei Cai, Tian Tan, Rui Ning, Chunsheng Xin, Hongyi Wu Published: 2026-09-28 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.36294v1 Sparse autoencoders (SAEs) expose interpretable features in large language model activations, yet existing structured SAEs impose single-parent trees or forests, while post-hoc graphs permit multiple parents but neither guide feature learning nor ensure reliable relation recovery. We introduce the Adaptive Graph Sparse Autoencoder (AG-SAE), a structure-guided training paradigm that treats each feature's complete parent set as an atomic structural hypothesis and lets evidence select zero, one, or multiple parents. By competing complete parent sets against null, subset, and alternative explanations, AG-SAE identifies jointly necessary multi-parent relations while rejecting redundant or spurious alternatives and verifying that each child contributes beyond its parents. The induced topology over SAE features then defines a differentiable structural loss that guides SAE training, while topology-guided refinement mitigates feature absorption and uses persistent reconstruction gaps exposed by the learned structure to initialize new features. The entire graph is then induced again from the revised dictionary by reassessing every feature's complete parent set, closing the dictionary-graph self-consistency cycle. Experiments demonstrate exact mixed-topology recovery in a controlled toy model, greater relational reliability and semantic validity than structured and post-hoc baselines on real LLM activations, and stronger feature-level causal interventions than conventional SAE features. AG-SAE thereby turns recovered mixed-topology feature structure into an unsupervised training signal that improves the dictionary, enables reliable feature organization beyond the topological limitations of trees, and exhibits stronger causal control beyond reconstruction. --- ### The Surge of Anti-Semitism in German Social Media following the October 7 Attacks Authors: Gregor Wiedemann, Daniel Wehrend Published: 2026-09-28 Categories: cs.SI, cs.CL arXiv: https://arxiv.org/abs/2609.36290v1 We investigate the extent to which the Hamas attacks on Israel of October 7, 2023, have affected German social media debates about Judaism and Israel. For this, we develop an approach to detect 26 anti-Semitic categories in user postings via large language models (LLMs). The approach is applied to Facebook and Telegram posts (N=125,718) from three months before and after the event. Methodically, we test different open-weight models in two setups---with and without user information as additional context to the post text. The best setup achieves up to 83 % F1-score for binary anti-Semitism detection on our manually coded validation set. User context provides valuable information for most LLMs and drastically reduces false positives, for example, when (critically) reporting on anti-Semitic incidents. Concerning our topic, we find that anti-Semitism is surging significantly on both platforms, while being about ten times more prevalent on Telegram compared to Facebook. Facebook users express anti-Semitic views most likely in posts about an alleged genocide in Gaza carried out by the Israeli army, whereas classic anti-Semitic stereotypes related to power and conspiracy theories are dominant on Telegram. After the attack, the discourse patterns on both platforms show signs of convergence, as classic anti-Semitism increases on Facebook, whereas Israel-related categories surge on Telegram. --- ### How Much Prompt Is Enough? A Blackbox Minimization of Few-Shots in LLMs Authors: Ali Alfageeh, Rahul Gopinath, Amin Alipour Published: 2026-09-28 Categories: cs.SE, cs.AI, cs.HC arXiv: https://arxiv.org/abs/2609.36289v1 Prompts are the primary mechanism for directing the behavior of large language models (LLMs). Yet the internal structure and causal hierarchy of prompts remain poorly understood: which parts are causally necessary and which are redundant is an open question. This opacity can have severe consequences. Subtle prompt variations can silently shift model outputs in critical software systems, and engineers lack techniques to reason about prompt reliability. We present \framework, a blackbox prompt-minimization framework that reduces few-shot prompts to their necessary minimal subset. We use a case study to apply \framework to a few-shot learning system and demonstrate the insights that this framework can provide. Our experiments show that few-shot exemplars can be reduced by a mean of 65.3\%~$\pm$~15.8\% in character count while fully preserving propositional output fidelity. The models preferentially retain logical identifiers and constraint declarations while discarding natural language prose and cross-prompt relational annotations. Our analysis also shows that some models are universal encoders, able to produce highly legible yet minimized prompts, while others are universal decoders, able to interpret minimized prompts from most other models. By identifying which components are indispensable, \framework provides a principled basis for prompt compression and structural analysis of few-shot exemplars. --- ### Representational and Functional Robustness to Electrode Montages in EEG Foundation Models Authors: Jakob Steglich, Justus Meyer zu Bexten, Shakiba Moradi, Laure Ciernik, Simon M. Hofmann, Mina Jamshidi Idaji Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36288v1 EEG foundation models (EEG-FMs) are intended to generalize across different datasets by learning representations that, ideally, are invariant to dataset-specific EEG configurations such as electrode montages. However, EEG-FMs that accept different montages as input do not guarantee that representations and predictions remain stable across different electrode configurations, especially outside the training setting. In this work, we investigate the effects of different electrode montages through a joint functional and representational analysis of four EEG foundation models selected to span distinct montage-handling designs. We evaluate embeddings on cross-subject resting-state eyes-open/closed and within-subject motor-imagery classification under spatially informed channel reduction. Functional robustness is tested through the generalizability of linear probes across channel counts, while representational robustness is assessed through within-subject similarity and preservation of between-subject geometry. The four models show distinct robustness profiles, and the two axes dissociate: large changes in embedding similarity need not come with comparable probe degradation, and stable embeddings can still lose downstream performance. Comparing two readouts of the same encoder further shows that aggregation, not the encoder alone, determines functional robustness: pooling into anatomically aligned regions degrades less than a learned global readout, despite being montage-invariant by construction. Montage robustness is therefore a joint property of the encoder and its aggregation, and characterizing it requires both a representational and a functional axis. Input compatibility alone is evidence for neither. --- ### GNA: Granular Neighbor Assembly for Retrieval-Augmented Multivariate Time-Series Forecasting Authors: Vincent Uhse Published: 2026-09-28 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.36281v1 Deep forecasters predict from a fixed-length lookback window, and lengthening it gives diminishing returns at a growing cost. Retrieval augmentation instead shows the model how similar past situations continued. Retrieving a whole past window gives every variate the continuation of the same past moment. In multivariate series, however, the best past match differs from variate to variate. We present GNA (Granular Neighbor Assembly), a retrieval layer for forecasting backbones that assembles neighbors at two granularities: whole past windows, which keep the variates coherent, and per-variate neighbors, in which each variate takes its future from its own best-matching past. A learned gate decides, per forecast step and variate, how much to trust these futures against a persistence forecast, next to the backbone's own forecast. Candidates come from an embedding trained to predict each window's future, and retrieval is strictly causal: a past window is used only once its future has been observed. With the same lookback for every model and the same retrieval constants for all datasets, GNA improves two Transformer backbones in 85 of 96 dataset-horizon settings, gives the lowest MSE on 8 of 12 standard benchmarks and beats its backbone in every seed on 10 of them. Both granularities are needed, and neighbors of mismatched queries are worse than none. Retrieval helps most where the lookback says least: the gate shifts trust to retrieved futures further ahead. Where it fails, on hourly non-stationary series at long horizons, the loss is consistent with a drifting level of the retrieved futures. --- ### Proofs Without Nominals: Gödel's Ontological Argument, its Shallow Embedding, and the Open Questions of the Monatshefte Notes Authors: Christoph Benzmüller Published: 2026-09-28 Categories: cs.LO, cs.AI, math.LO arXiv: https://arxiv.org/abs/2609.36279v1 The shallow embedding of higher-order modal logic in classical higher-order logic, used in Benzmüller and Scott's Notes on Gödel's and Scott's variants of the ontological argument (2025), reaches beyond the modal object language of the arguments: its property quantifiers range over terms that may also express nominals and satisfaction operators of hybrid logic, and a proof using one proves a theorem of the embedding that need not be one of the modal logic. That the framework affords this is not new, and whether a result is one of the modal logic can be settled in two ways: by replaying it in an explicit proof calculus, done by hand for chosen theorems, or by analysing the proofs the embedding itself produces, which this article does mechanically, for every result at once. Every statement the Notes prove has a proof inside the object language: 294 written out by hand and machine-checked, none using a nominal. The proofs the Notes themselves give instantiate no nominal either; what the detector flags there are terms a prover substituted. The three questions the Notes leave open are settled too, and without nominals, but the conjunction axiom has to be emended: generalised in the Notes to Gödel's "any number of summands", it covers the conjunction of no properties, and of one; the empty one alone settles all three, and the two together yield what a separate axiom of Gödel's is for. This article restricts the conjunction axiom to at least two different conjuncts, the reading Gödel's footnote suggests, and the questions are settled again, by proofs that turn on the argument rather than a degenerate instance. The restriction holds of the object language only: with a nominal the axioms make the accessibility relation the identity and the readings coincide. Every theorem is verified in Isabelle/HOL and independently in Lean 4; the countermodels are Nitpick's, certified by the build. --- ### Illusory Truth or Mere Exposure? Model-Dependent Repetition Effects in LLM-Based Social Media Simulations Authors: Azza Bouleimen, Nicolò Pagan, Anikó Hannák Published: 2026-09-28 Categories: cs.AI, cs.SI arXiv: https://arxiv.org/abs/2609.36278v1 Generative agent-based models (GABMs) are increasingly used to simulate social media dynamics, including misinformation spread. For such social simulations to be valid proxies of human behavior, LLM agents should replicate established human cognitive biases, among them the Illusory Truth Effect (ITE), where repeated exposure to a claim increases its perceived truth value. We investigate whether and how the ITE manifests across four LLMs (Gemma-3-4b-it, Qwen2.5-7B-Instruct, Llama-3.1-8B-Instruct, and GPT-5-nano) in a social media simulation context. We propose a two-phase within-context experimental design that embeds the repetition manipulation inside a realistic news feed interaction. Using this design, we collect 336,000 truth, importance, sentiment, and interest ratings across 100 statements, 10 feed variants, and 3 replications. The key comparison is between ratings assigned to repeated statements, seen throughout a simulation phase, and completely unseen ones, rated within the same experimental context window. We distinguish genuine ITE (truth-specific repetition boost) from mere exposure effects. We run an OLS regression followed by a Linear Mixed-Effect Model to account for differences across models and ratings. Our results reveal four qualitatively distinct patterns: Gemma-3 exhibits a genuine ITE; Qwen2.5 shows a mere exposure effect; GPT-5-nano displays no repetition effect on truth and mild skepticism toward repeated content; Llama-3.1 shows a small truth boost alongside decreases in evaluative dimensions. Crucially, temperature has no effect on these findings, and a variance decomposition highlights the high context-sensitivity of LLM rating behavior. Our findings caution against assuming uniform ITE replication across LLMs in social simulations, while suggesting that Gemma-3-4b-it may offer the most behaviorally realistic approximation for misinformation-related simulations. --- ### From Surfaces to Volumes: Registered Geometry for Protein Representation Learning Authors: Siyuan Chen, Cai Zhou, Jinrui Zhang, Zhaokang Liang, Taku Komura, Wojciech Matusik, Stephen Bates, Tommi Jaakkola, Wengong Jin, Peter Yichen Chen, Minghao Guo Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36277v1 Existing protein geometry models typically represent molecular surfaces using local geometric features such as sampled points, normals, and curvature. While effective for capturing exposed molecular shape, these representations do not explicitly model the volumetric organization beneath the surface or provide a consistent coordinate system for residue-wise volumetric structure. We introduce Protein-TetSphere, a registered residue-wise volumetric representation for proteins. Each protein chain is tetrahedralized to obtain local volumetric regions associated with individual residues, which are then registered to a shared fixed-topology tetrahedral reference and represented in a common Laplacian basis. This registration establishes consistent volumetric coordinates across residues, enabling local three-dimensional deformation to be integrated with surface and chemical information in a multimodal protein representation. We evaluate Protein-TetSphere on ligand-binding pocket classification, protein--protein interface prediction, and de novo protein binder design. Across the three tasks, Protein-TetSphere improves ligand-binding pocket balanced accuracy from $0.795$ to $0.826$, Pinder-Pair/Site AUROC from $0.914/0.852$ to $0.932/0.866$, and binder-design success from $14.95\%$ to $19.90\%$ on the BoltzGen Challenge Set and from $27.62\%$ to $32.19\%$ at the ProtDBench backbone level. These results show that registered volumetric geometry provides complementary spatial information beyond molecular surfaces across protein recognition, interaction, and design. --- ### Stochastic Optimization Under Power-Law Spectra: Tight Bounds and Shuffling Analysis Authors: Thomas Dybdahl Ahle, Yaroslav Bulatov, Christopher De Sa, Christopher Ré Published: 2026-09-28 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.36271v1 Recent work has established that power-law spectral conditions on data enable tight convergence bounds for deterministic gradient descent, resolving the conflict between classical exponential bounds and observed power-law learning curves. In this work, we extend this result to the stochastic regime of high-dimensional machine learning. We provide two main contributions: (1) We generalize the power-law spectral theory to Stochastic Gradient Descent (SGD), showing that the same spectral exponents govern stochastic dynamics; (2) For the fundamental case of isotropic Gaussian data, we provide a precise analysis of data shuffling, deriving exact constants that prove Single Shuffle is strictly superior to Flip-Flop and IID sampling. Our results bridge the gap between abstract spectral theory and practical stochastic training choices, offering a unified picture of how data geometry drives optimization speed. --- ### In-Context Learning Amplifies a Latent Symbolic Circuit Authors: Melissa Wessel Published: 2026-09-28 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.36265v1 Large language models can learn abstract rules from just a few in-context examples, but how their internal mechanisms activate as examples accumulate is not well understood. We trace a three-stage symbolic reasoning circuit (abstraction, induction, retrieval) across shot counts in three model families and find it is detectable and functional well before the model achieves high accuracy. Per-head causal contribution grows up to 8x from 1- to 10-shot, and cross-shot activation patching raises accuracy from 1% to 56% at 0-shot and 17% to 88% at 1-shot. Function vectors scaled and injected at 0-shot rescue accuracy up to 86%, largely substituting for the induction stage but depending critically on an intact downstream retrieval stage. The infrastructure for abstract rule-following is present in the weights before any demonstrations; in-context examples, function vectors, and related interventions appear to supply input to the same latent circuit. --- ### OTROPE: Optimal Transport-based Robust Off-policy Evaluation for Large Language Models Authors: Liner Xiang, Wenbo Zhang, Hengrui Cai Published: 2026-09-28 Categories: cs.AI, cs.CL, stat.ML arXiv: https://arxiv.org/abs/2609.36264v1 Reliable evaluation of large language models (LLMs) is essential for their development and deployment, yet is often costly, risky, and difficult to perform safely online. We study off-policy evaluation for LLMs, where limited human-labeled data from a behavior model are used to evaluate a newer target LLM. This setting is challenging because labels are scarce, behavior--target distribution shift is common, and response likelihoods are often unavailable for black-box LLMs. We propose the Optimal Transport-based Robust Off-Policy Evaluation (OTROPE), a likelihood-free evaluation that performs distributional correction in a semantic space via optimal transport to align labeled behavior-policy samples with unlabeled target-policy samples. OTROPE combines corrected human-labeled residuals with proxy predictors, yielding a doubly robust-style evaluation without behavior-policy modeling or density-ratio estimation. We theoretically characterize why baseline evaluators fail under LLM distribution shift, and establish consistency and convergence rates for OTROPE when either the reweighted behavior distribution or the proxy predictor converges. Experiments on synthetic and real LLM evaluation tasks show that OTROPE consistently outperforms baselines while enabling ensembles of weaker LLM evaluators to approach and sometimes surpass stronger evaluators. Code is available at https://github.com/LinerXiang/OTROPE. --- ### Paired Multimodal Scaling Laws Authors: Marcus Ma, Shrikanth Narayanan Published: 2026-09-28 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.36263v1 Existing multimodal scaling laws fit multimodality terms empirically after testing and never vary how much data is multimodally paired at fixed data budgets. We investigate how, under the same total data per modality, changing the number of paired data affects loss curves in multimodal classification tasks. We train models in three different environments and run experiment sweeps varying data sizes and pairing budget. Pairing ratios have a dramatic impact on loss and this impact is directly tied to how much information synergy the task contains. Only paired data is able to reduce synergistic loss, while unpaired data can reduce redundant or unimodal information up until unimodal floors. Unlike traditional scaling laws where loss drops immediately in power law decay, synergy acquisition is gated, requiring a critical threshold of paired data before synergistic loss falls at all. We introduce a new family of multimodal scaling laws where total data-attributable loss is the sum of four individual power laws corresponding to the four different information channels of redundancy, a unique channel per modality, and synergy, and show how this law is both more theoretically sound and empirically valid across our experiments. This law predicts multimodal loss in our experiments more accurately than existing laws, with 3.2% error on fit tests versus 10.4% error for the best pairing extension of published laws. --- ### Understanding LLM Parameter Update Sparsity through the Lens of Fisher Authors: Yufan Zhang, Sagnik Mukherjee, Hao Peng Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36262v1 Recent studies have observed that parameter changes during language-model post-training can be concentrated in a small subset of coordinates. This phenomenon has been reported in reinforcement learning, on-policy distillation, and supervised fine-tuning on near-policy data. Its recurrence across different post-training paradigms suggests shared structure in training dynamics. In this paper, we examine this pattern through the diagonal model Fisher, which measures the sensitivity of the model's output distribution to individual parameters and is independent of any particular reward or teacher signal. Theoretically, we show that small diagonal Fisher leads to small expected gradients across a range of training objectives, providing a common explanation for sparse gradient updates. Empirically, we test this connection in RL and OPD. We find that Fisher identifies where gradients are concentrated, and fixed sparse masks selected from the initial Fisher retain a large proportion of the improvement from full training. Finally, we investigate the mechanisms underlying low Fisher in on-policy training. Our results show that high-probability next tokens tend to have similar parameter sensitivities, contributing to low Fisher. Together, these results establish the diagonal model Fisher as a unifying perspective linking update sparsity to on-policy training dynamics in LLM post-training. --- ### CyFA: Linear Sequence Modeling with Relative-Time-Partitioned Memory Authors: Yixiao Chen, Shuojin Yang, Shi-Min Hu Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36259v1 Linear RNNs offer linear-time sequence processing and constant-memory decoding, but their fixed-size recurrent states must accommodate all past key--value associations. Existing forgetting mechanisms and Delta Rule updates reduce interference by selectively clearing or correcting the state, yet earlier associations can still become difficult to retrieve. We introduce CyFA (Cyclic Flow Attention), a Linear RNN with relative-time-partitioned memory. At each step, a learned clock controls the cyclic transport applied jointly to the key and value states before the current key--value pair enters the age-zero slot, thereby organizing stored associations across relative-time slots. We further derive an exact change to absolute-clock coordinates that expresses CyFA as two scalar-decay linear attention recurrences and enables efficient chunk-wise training. Across 400M--1.4B pretraining experiments with matched recurrent-state sizes, CyFA improves recall-intensive performance while maintaining competitive language modeling and high computational efficiency. At 400M, CyFA outperforms KDA on FDA (42.60 vs. 26.07) while requiring only 46.7% and 48.3% of KDA's forward and backward core-operator execution times, respectively. Our code is publicly available at \href{https://github.com/Chyxx/CyclicFlowAttention}{this https URL}. --- ### Towards Mitigating Deceptive Safety Alignment in Large Reasoning Models Authors: Xiangyu Zhou, Saleh Zare Zade, Rafi Ibn Sultan, Alexander Kotov, Dongxiao Zhu Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36254v1 Large Reasoning Models (LRMs) are commonly trained with reinforcement learning (RL) to improve their generation of chain-of-thought (CoT) reasoning before producing final answers. However, RL rewards are typically assigned based on final answers, providing little or no direct supervision over intermediate reasoning. This can lead to deceptive safety alignment, where the reasoning trace and final answer convey inconsistent safety signals. To systematically investigate this phenomenon, we introduce DSAR (Deceptive Safety Alignment Rate), a metric that jointly assesses reasoning traces and final answers to quantify their safety inconsistency. Across multiple LRMs and benchmarks, we find that deceptive safety alignment is pervasive under standard prompting conditions and is substantially amplified under prefilling attacks. We further provide a hidden representation analysis showing that models exhibit stronger safety discrimination at the final-answer stage than during intermediate reasoning. To close this gap, we propose SARA (Safety-Aware Reasoning Alignment), an RL-based method that rewards both safety-aware reasoning and safe final answers, encouraging early harmful intent recognition and enforcing reasoning-answer consistency. Experiments show that SARA significantly mitigates deceptive safety alignment under both standard and adversarial settings while preserving helpfulness and utility. Code is available at https://github.com/xzhou98/SARA. --- ### Population Fidelity: Evaluating Population Representativeness in LLMs Authors: Neemias B. da Silva, Martin Lukk, Ali Sutani, Abhishek Moturu, Harris Yang, Daniel Silver, Matt Ratto, Thiago H. Silva Published: 2026-09-28 Categories: cs.CL, cs.AI, cs.CY arXiv: https://arxiv.org/abs/2609.36253v1 Large language models (LLMs) show considerable potential in simulating human attitudes and preferences. Prior work finds that LLM-generated responses can compress the range of attitudes found within populations and misrepresent particular subgroups in ways that vary across models and topics. We introduce Population Fidelity, an evaluation framework that distinguishes key conditions required for a set of LLM-generated responses to represent a population. It incorporates three dimensions: group-level accuracy, the amount of between-group variation, and the structure of that variation. We demonstrate the framework's utility in two ways. First, we reproduce a prior study of "machine bias" in LLM survey responses and apply the framework to its models and more recent ones, showing that poor representation reflects not only insufficient between-group variation but also variation assigned to the wrong groups. Second, we evaluate one proposed approach to improving models' population representativeness: cultural fine-tuning. We find that cultural fine-tuning can improve alignment with the survey center without improving the representation of within-population differences, a distinction that measures of aggregate agreement do not capture. We argue that representing a population requires models to reproduce several features of human attitudinal variation simultaneously. Our framework organizes these features and provides reusable code, data, and trained models for evaluating population fidelity across substantive domains and assessing proposed alignment methods. --- ### The Signed Geometry of One-Shot Recourse: On-Path Validity and the Signed-Curvature Criterion Authors: Hazar Yueksel Published: 2026-09-28 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.36252v1 Closed-form recourse moves a rejected user along the unit gradient $\hat g$ of the classifier score $f$ by the promised distance $d_p=|f(x)|/\|\nabla f(x)\|$, at which the linearized score reaches zero. We ask when this one-shot step succeeds and what additional model queries change. To leading order the step ends on the favorable side exactly when the path curvature $κ=\hat g^\top\nabla^2 f(x)\,\hat g$ is nonnegative. Across 80 shallow models, the fraction of rejected users whose step ends there and the fraction with $κ\ge0$ correlate at $r=0.985$, although on Fashion-MNIST the first falls below the second by 8.2 points on average. No rule that uses only the score value and gradient can be valid for every score with path curvature bounded by $K$ without overshooting some by order $Kd_p^2/\|\nabla f(x)\|$. When the curvature is also Lipschitz and the step is short, one evaluation of $f$ at the promised point attains the minimax rate among deterministic one-query rules that know the curvature bound and its Lipschitz constant, and split-conformal calibration makes such a rule reach the first crossing or abstain with probability at least $1-δ$. Training with an asymmetric curvature penalty lets 99-100% of paths cross within the promised step on undershoot-prone shallow data, at about 4-22 times the overshoot of symmetric penalties (Fashion-MNIST, COMPAS). Because $κ$ and $d_p$ depend on how the score is scaled, part of this gain can be a longer promised step, and at matched validity a smaller audit of briefly trained models finds no uniform advantage over tuned inflation. Where a per-user line search along the ray is affordable, it is exact to grid resolution and preferable. --- ### Action Chunking Proximal Policy Optimization with Feedback Correction Authors: Sanghyun Hahn, Jonghyun Choi Published: 2026-09-28 Categories: cs.LG, cs.RO arXiv: https://arxiv.org/abs/2609.36250v1 Action chunking provides temporal abstraction in reinforcement learning by selecting short action sequences instead of individual actions, but many existing approaches face two limitations in high-dimensional robotic control. First, many rely on value functions over action chunks, which can be difficult to learn as action dimensionality and chunk length grow. Second, executing chunks open-loop removes within-chunk feedback, limiting reactivity in contact-rich tasks. We present Action Chunking PPO (ACPPO), a PPO extension that uses a chunked actor while retaining a standard state-value critic, thereby avoiding chunked Q-functions. We further propose ACPPO-Corr, which augments the chunk planner with a stepwise feedback corrector that adjusts planned actions online within each chunk. Across 25 simulated robotics tasks from IsaacGym and Bi-DexHands, spanning locomotion, arm manipulation, and dexterous hand-object interaction, ACPPO-Corr achieves the strongest aggregate performance among evaluated methods and performs best on both decision-frequency-sensitive and decision-frequency-neutral task subsets. Ablations show that moderate chunk lengths work best and that corrector regularization is important for balancing chunk-level planning with local feedback. These results suggest that action chunking can be effective in online PPO when chunk-level planning is paired with closed-loop correction. The code is available at: https://github.com/hshhahn/ACPPO. --- ### Learning from Teacher Continuations at Student States Authors: Haojin Wang, Dylan Zhang, Huaibo Chen, Suhao Yu, Yihang Sun, Zhanyang Jin, Jiaying Ye, Dianqi Li, Prasanna Sattigeri, Kamal Youcef-Toumi, Hao Peng Published: 2026-09-28 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.36246v1 We present OLIVE (OnLine InterVEntion). At each iteration, the evolving student policy generates a new prefix, the teacher continues it autoregressively, and the student is updated using cross-entropy computed on the teacher-generated tokens. Each design choice targets a corresponding limitation of existing distillation methods: (1) sequential covariate shift in offline supervised fine-tuning (SFT) on fixed teacher trajectories, (2) fragmented supervision under prefix failure in token-level on-policy distillation (OPD), and (3) the need for access to teacher token probabilities in distribution-matching distillation. OLIVE achieves higher reasoning performance than OPD (with a top-16 KL approximation) at comparable GPU-hour cost. Our asynchronous implementation further reduces OLIVE's total training time by 23.8\%. We evaluate OLIVE on both hard reasoning tasks and agentic tasks which reflects modern post-training scenarios, and it consistently outperforms existing distillation methods under the same training budget. By regenerating prefixes from the evolving student, OLIVE continues improving after offline distillation plateaus while better preserving the general capabilities and plasticity of the student. Using only text from GPT-5.4-mini, continuously training with OLIVE outperforms offline SFT from the same teacher by 13\% on ScienceWorld. These results support OLIVE as an effective and efficient approach to online language-model distillation. --- ### CoRe: Co-Evolving Reward Models for Mitigating Latent Reward Hacking in Video Diffusion Models Authors: Zhaolong Su, Yujin Han, Feng Wang, Jameson Dong, Hins Hu, Difan Zou Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.36245v1 Latent reward models (LRMs) enable efficient alignment of video diffusion models by scoring intermediate states directly in latent space. However, we find that optimizing against a fixed latent reward rapidly leads to latent reward hacking: the predicted reward stays high while perceptual and motion quality deteriorate. Our analysis identifies distributional escape as the central cause: within a few hundred updates, the generator moves beyond the reward model's training support, where its scores no longer reflect video quality. Based on this insight, we introduce CoRe, a co-evolving reward framework that treats latent-space alignment as a dynamic interaction between the generator and the reward model. Rather than optimizing against a stationary proxy, CoRe continually refits the reward model on the generator's current samples while anchoring it to real-video preferences, so the generator cannot gain reward by drifting away from the data. On Wan2.1-T2V-1.3B, experiments show that CoRe consistently improves generation quality over both the pretrained model and prior alignment methods, while avoiding the quality collapse of fixed-reward optimization. --- ### Think Before You Restore: Risk-Aware Manchu Manuscript Restoration with Stroke-Guided Attention Authors: Mingqiu Liang, Dongdong Wang, Siyang Lu, Ting Huang, Yingjun Qi Published: 2026-09-28 Categories: cs.CV, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.36243v1 Full-page blind restoration of historical Manchu manuscripts is challenging due to scarce annotations, unknown degradation regions, and fragile connected strokes. Generic restoration models may improve visual quality but often modify intact content, leading to over-restoration. We propose SAGE-Restore (Stroke-Aware Gated rEstoration), a selective restoration framework that first assesses where restoration is needed and then uses this assessment to guide restoration candidate generation and pixel-level selection. Its encoder predicts patch-level repair probabilities from complementary appearance and stroke-structural cues to condition restoration candidate generation, while the corresponding repair logits are refined into a pixel-level soft gate that selectively controls where the restoration candidate is applied. We further introduce a fidelity-aware evaluation protocol that jointly measures degraded-region recovery, intact-content preservation, and their balance. SAGE-Restore achieves the highest R-Recovery (0.463) and RFS (0.626), while maintaining high U-Fidelity (0.968), demonstrating an effective balance between restoration and content preservation. --- ### Can Representation Learning Decouple from Loss Minimization? Polar Updates Have an Answer Authors: Akash Kumar Published: 2026-09-28 Categories: cs.LG, math.OC, stat.ML arXiv: https://arxiv.org/abs/2609.36240v1 Does representation learning stop when the training loss stops improving? We study this question for matrix Muon, whose polar-normalised updates have a step length set by the gradient's rank rather than its norm. Near the edge of stability, full-batch Muon on teacher-student problems enters approximately period-2 loss oscillations that persist for thousands of steps: the cycle-mean loss stays flat or rises, yet the weights keep moving and the learned features continue to align with the teacher subspace. For linear teacher-student learning toys, we derive explicit cycle and alignment formulas and conditional plateau and decay bounds. For a population mean-field ReLU model, we prove that, under stated dimension, initialisation and small-head conditions, the leading eigenspace of the average gradient outer product (AGOP) recovers the teacher subspace exactly during a loss plateau, before the loss later drops. In all 33 ReLU, GELU and SiLU teacher configurations we study, direction-only alignment metrics show the student AGOP aligned with, or still aligning to, the teacher subspace during the period-2 oscillations; projected head refitting on selected configurations shows that the learned directions are useful for prediction, and further measurements distinguish AGOP alignment from weight-mass concentration. In deep residual ReLU students, freezing the downstream layers while the first layer trains with full-batch exact polar updates recreates a nearly flat cycle-mean loss with improving input-AGOP alignment; freezing and unfreezing switch between this plateau and loss decrease, and the effect is sensitive to momentum and to the choice of orthogonaliser. --- ### Cognitive Expert Language Models Better Align with the Corresponding Brain Systems Authors: Zhivar Sourati, Mengxuan Helen Wu, Nona Ghazizadeh, Jonas Kaplan, Morteza Dehghani, Samuel A. Nastase Published: 2026-09-28 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.36239v1 Large language models (LLMs) can predict human brain activity across a variety of brain regions during natural language comprehension. Typically, however, LLM-brain alignment is measured using one model for different regions of the brain, and then model performance is summarized across regions. This one-model-fits-all approach ignores the functional specialization of brain regions. In this study, we assess whether a model oriented toward a particular cognitive domain aligns better with the brain system dedicated to that domain. Through prompting and fine-tuning, we first build expert LLM variants for six domains: sensory, spatial, numerical, reasoning, social, and abstract processing. We then examine whether each expert best predicts activity in the brain region associated with the corresponding cognitive domain. Consistent with our hypotheses, each expert's representations align more closely with the brain system most associated with the matching domain than do other experts. This holds under both prompting and fine-tuning, across three base models and three fMRI datasets. In a series of control analyses, we show that this model-brain alignment is specific to cognitive domain interventions; non-cognitive and surface-level interventions do not result in comparable alignment. Specializing models shifts regional alignment while leaving aggregate prediction accuracy largely unchanged, suggesting that summarizing alignment across regions may obscure regional differences in performance for specific models. --- ### ChronoSRL: Temporal Geometry for Self-Supervised Reinforcement Learning Authors: Nico Bohlinger, Jan Peters Published: 2026-09-28 Categories: cs.AI, cs.RO arXiv: https://arxiv.org/abs/2609.36238v1 A goal that is close in space can be far away in time. Obstacles, terrain, and the agent's own capabilities determine how long it takes to get there. Yet, critics in contrastive and survival reinforcement learning do not measure the distances in their representation space in units of time. We therefore introduce ChronoSRL, which gives the critic's embeddings an explicit temporal geometry. The distance between state-action and goal embeddings is trained to match the time that the agent takes to reach the goal (goal-reaching time), while goals that were not reached, and goals from other trajectories, are pushed at least one discount horizon away. Furthermore, reaching a goal quickly once does not mean that reaching it is reliable in general, so the policy should not follow the temporal distance directly. Instead, we build on survival reinforcement learning and predict from our temporal embeddings not only the full distribution of goal-reaching times but also the time spent near the goal. Thereby, the policy is trained to favor actions that reach the goal sooner and more reliably and that keep the agent near it. ChronoSRL learns faster and reaches higher performance than contrastive, action-chunked contrastive, and survival reinforcement learning baselines on seven standard locomotion and navigation benchmarks, even with much smaller networks. To test the limits of self-supervised reinforcement learning, we introduce velocity tracking, goal-position reaching, and box climbing tasks with a quadruped robot in a realistic sim-to-real locomotion setup, and show how the shaping terms that are typical for robotics can be naturally incorporated into our framework. ChronoSRL is the only one of the tested self-supervised reinforcement learning methods that learns to stay at the commanded velocities and goal positions, and climbs the highest boxes. --- ### Transversal Pooling Neural Networks Authors: Emily J. King, Dustin G. Mixon, Michael Perlmutter, Lander Ver Hoef Published: 2026-09-28 Categories: cs.LG, math.FA arXiv: https://arxiv.org/abs/2609.36237v1 Many learning tasks require stability to small transformations while retaining sensitivity to larger ones. We introduce \emph{transversal pooling neural networks}, which generalize spatial max pooling to affine group actions. We establish equivariance to a chosen subgroup and derive explicit stability bounds for individual pooled wavelet coefficients under affine perturbations of the input. Experimentally, we demonstrate the utility of our networks in low-data environments and for predicting tropical cyclone intensification. --- ### MERID: Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis Authors: Lei Liu, Zhaokang Liang, Qingcheng Zeng, Chenda Duan, Lu Mi, Zhen Tan, Tianyu Liu Published: 2026-09-28 Categories: cs.AI, cs.LG, cs.MA arXiv: https://arxiv.org/abs/2609.36235v1 Major depressive disorder (MDD) severely impacts daily activities and quality of life. Detecting MDD involves multimodal data, such as interview recordings and sensor measurements. This is particularly challenging, as these heterogeneous modalities often demand distinct, customized prediction pipelines. Existing efforts to address this challenge have explored both manually engineered multimodal architectures and agent-assisted pipeline development. Despite their progress, it remains challenging to autonomously revise pipelines based on experimental feedback and carry verified improvements forward into subsequent designs. To this end, we propose Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis (MERID). The framework develops depression pipelines through experience-based recursive self-improvement (RSI). Grounded State Construction (GSC) grounds experience by aligning multimodal records with subject-level depression targets. Coupled Pipeline Exploration (CPE) jointly modifies representations, fusion, and predictors to build successor pipelines for classification and severity estimation. Evidence-Guided Evolution (EGE) guides revisions through feedback and verifies gains under uncertainty in small depression cohorts before inheritance. Extensive experiments on depression benchmarks show that MERID achieves the best results on multiple tasks compared with multimodal and agent-based baselines. Further analysis highlights the value of acoustic and linguistic cues for depression detection. Our code is available at https://github.com/DiscoAILab/MERID --- ### The Role of Feed-Forward Layers in Transformer Dynamics Authors: Thomas Jacob Maranzatto, Semih Akkoc, Sennur Ulukus Published: 2026-09-28 Categories: cs.LG, math.DS arXiv: https://arxiv.org/abs/2609.36230v1 We study the dynamical behavior of tokens in transformers from a control-theoretic perspective. Our model includes the feed-forward layer present after the self-attention mechanism, with the self-attention mechanism interpreted as an interacting particle system and the feed-forward layer as an independent control. Our main theoretical result establishes that the feed-forward network can steer the tokens arbitrarily close to consensus regardless of the key, query, and value matrices. Our result are easily extended to convergence to many clusters and to multi-head attention. We conduct numerical experiments to verify our results, and compare thresholding behavior from our theory to real-world LLMs. --- ### An Empirical Study and Assessment of EU AI Act Compliance Checkers Authors: Zhen Tao, Alize Kahraman, Shidong Pan, Zhenchang Xing, Chiara Ullstein, Jens Grossklags, Chunyang Chen Published: 2026-09-28 Categories: cs.AI, cs.CR arXiv: https://arxiv.org/abs/2609.36228v1 The EU AI Act introduces extensive compliance requirements for organizations that develop, deploy, or integrate AI systems. Many of these requirements are directly relevant to security and privacy, while also addressing closely related issues such as data governance, transparency, accuracy, and robustness. However, stakeholders such as small-to-medium businesses and individual developers often lack the legal expertise required to interpret these obligations and translate them into engineering and governance practices. This disconnect creates challenges for implementing the EU AI Act and may lead to missing safeguards or misdirected development and deployment efforts. To address this, various automated EU AI Act compliance checkers (AIACCs) have emerged, claiming to streamline compliance assessments and provide practical guidance. In this paper, we present the first empirical study and assessment of AIACCs. We characterize 12 mainstream AIACCs across multiple dimensions, evaluate their legal coverage and alignment, and analyze checker-generated compliance reports for structure, determinacy, and actionability. We find that the quality of AIACCs varies significantly and that they currently can only serve as early-stage orientation tools. Specifically, we observe inconsistent interaction modes and user-friendliness, a tendency to overly simplify or omit key obligations, and a failure to provide determinate, actionable guidance. As a result, reliance on the current generation of AIACCs may foster a false sense of compliance. With our study, we provide a critical baseline of the current AIACC landscape. We further offer design principles for the implementation of more reliable compliance-support tools. --- ### One-Step Next-Latent Prediction Is Not a World Model Authors: Shitong Wang, Zhongang Cai, Yuzhou Hong Published: 2026-09-28 Categories: stat.ML, cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.36227v1 Next-latent prediction fits a map from the current embedding to the next one. LeNEPA carries this objective to time series, replacing the stop-gradient of next-embedding prediction with the isotropy penalty of LeJEPA. A world model is a transition kernel that can be rolled out. The one-step regression identifies a conditional mean, and a mean is a kernel only in special cases. For a linear-Gaussian Markov latent, the mean transition and the innovation covariance are fixed by the one-step problem, and the open-loop squared error at horizon $K$ equals the trace of the sum of the pushed-forward innovation covariances. That error grows with $K$ after the one-step fit is exact. If the conditional mean is nonlinear, composing it is not the multi-step conditional mean. If the observation is a non-injective function of a Markov state, a memoryless one-step map does not determine future observations, while a short window can. An isotropy penalty is a function of the embedding marginal, so its partial derivative in the transition weights is zero. On a scalar autoregression with coefficient $0.9$, the one-step mean squared error is $0.998$ and the $16$-step open-loop error is $5.10$. On a hidden rotation, an eight-step window reaches $16$-step error $0.056$, while the current scalar alone reaches $0.778$. Raising the isotropy weight from $0.1$ to $10$ leaves eight-step latent error inside $(0.78,0.85)$ on three seeds. --- ### Mutually Adversarial Self-Training with Evolving Data for Unified Multimodal Models Authors: Wentao Zhou, Weijie Gan, Jiayun Wang Published: 2026-09-28 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.36224v1 Unified multimodal models (UMMs) combine image generation and visual understanding in a shared backbone. Since generation and understanding are inverse tasks, recent studies self-train UMMs by letting the two branches cooperatively supervise each other. We introduce MATE (Mutually Adversarial self-Training with Evolving data), a reinforcement-learning-based post-training framework in which the two branches instead challenge each other, and the challenges evolve as the model trains. MATE lets generation and understanding take turns to be challenger and solver. Given an image, the understanding branch proposes several candidate descriptions that the generation branch must turn back into similar images, and vice versa. The candidates are screened for consistency with the image or prompt they were proposed from, and the solver is trained on the candidate it handles worst. The adversary thus comes from the model's own outputs, and no separate adversary is trained. Moreover, the candidates that defeat one branch become the sources of the next challenges to the other in the next epoch, which keeps the challenges evolving with the model and turns the training into self-play in data space. On Janus-Pro-1B, MATE improves GenEval by 2.4 points, DPG-Bench by 1.7 points, and the average over nine understanding benchmarks by 0.7 points, while strengthening consistency across repeated image-text cycles. --- ### BASE: Batch-Aware Selection of Experts Using Predicted Removal Error for Efficient MoE Decoding Authors: Ali Abbasi, Justin Shi, Soheil Kolouri Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36222v1 Large language models are increasingly expensive to serve. In large-scale serving systems, autoregressive decoding is often bottlenecked by transferring model weights from accelerator high-bandwidth memory into on-chip SRAM. Mixture-of-experts (MoE) models reduce computation by activating only a small subset of experts per token, but this sparsity does not translate directly to batched decoding. Different requests select different experts; therefore, the combined active set across many concurrent requests can span a substantial fraction of the expert pool and require significantly more expert weights to be transferred. Most expert-reduction techniques make retention decisions independently for each token and therefore do not address this batch-level expansion. More recently, batch-aware methods have attempted to coordinate expert use across concurrent requests and reuse experts already fetched for the batch. Yet their selection criteria are based primarily on router rankings or expert statistics collected during calibration. Consequently, these criteria are not directly tied to the output error caused by dropping an expert, nor do they capture how an expert's contribution changes across tokens at inference time. We instead rank experts according to how much their removal would change the MoE-layer output. To apply this criterion during serving, we train a lightweight linear predictor during calibration that estimates the expert removal cost for each incoming token, and develop custom GPU kernels for cost prediction and expert selection. Across three MoE architectures, BASE improves the quality-efficiency tradeoff without retraining. On Qwen3-30B-A3B, it improves average accuracy by 29.5 points over the strongest baseline at comparable throughput under a tight expert budget. At a higher expert budget, it is 60% faster than dense inference while remaining within 0.4 accuracy points. --- ### How Language Models Differ in Redistributing Attention-Head Activity Under Serial Demand Authors: Johnny Jingze Li, Abdulla Kuleib, Kalyan Basu, Gabriel A. Silva Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.36221v1 The way a model distributes activity over each layer's attention heads offers a coarse view of how it routes information through depth; how this changes with the task is part of what a mechanistic account must explain. Holding prompt length fixed, we vary how many serial steps a task demands and measure, in every layer of 17 open-weight models, whether activity concentrates on a few heads or spreads across many as demand rises. Both occur: in most models, layers just before mid-depth concentrate activity and later layers spread it. Models differ in where and how strongly this happens. The Qwen2.5 base models from 0.5B to 7B, for example, spread less than the average model in every task and concentrate activity in parts of their second half, where Llama models from 1B to 8B and OLMo-2 spread; the contrast largely holds between Llama-3.1-70B and Qwen2.5-72B, which have the same number of layers and heads. These differences are reproducible, and post-trained models keep much of their base model's pattern. An ablation study suggests that, within a task, models whose activity is more concentrated on their top heads also depend more on those heads for the answer. Concentration and spreading across layers thus offer a new way to compare models, by how they route information through depth. Code is available at https://github.com/johnnyjli/serial-demand-heads. ---