# Research Papers — 2026-09 · bundle page 63 of 266 > Full-text markdown of papers #3101–3150 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/63 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/62.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/64.md) ### SketchSSM: Write to the Full State, Read from a Compact Sketch Authors: Omin Kwon, JoongWon Shin, Minseo Kim, Kurt Keutzer, Sehoon Kim, Jae W. Lee Published: 2026-09-27 Categories: cs.LG, cs.DC arXiv: https://arxiv.org/abs/2609.33051v1 Hybrid-attention models replace most softmax attention layers with linear attention, reducing KV-cache growth and enabling larger decode batches where recurrent state access becomes a major bottleneck. ReplaySSM amortizes state updates by buffering keys and values, but each new query still requires a full-state read even though the state remains unchanged between state updates. We observe that low-rank state-weighted query approximation accurately preserves state-read outputs. Although future queries are unknown, the basis vectors used to approximate them can be fixed offline. Based on this observation, we introduce SketchSSM, which preserves full-state updates while approximating reads. At each state update, SketchSSM reads the full state once to precompute outputs for these basis vectors, storing them in a compact sketch. Each subsequent decode step combines the sketch vectors with query-dependent coefficients to reconstruct the output without a full-state read. Across four Mamba-2-, GDN-, and KDA-based models, SketchSSM reduces state-access traffic by approximately 10x while largely preserving average accuracy across four decode benchmarks and recall on four RULER retrieval tasks. On one NVIDIA B300, linear-attention kernel speedups over the standard vLLM baseline reach 7.78x, 5.22x, and 5.20x for Mamba-2, GDN, and KDA, respectively, with up to 2.64x higher decode throughput on Nemotron 3 Super. --- ### DevelopmentODE: Structured Neural ODEs for Early Brain Development Dynamics Across a Decade Authors: Kaiqiao Han, Haitao Chen, Bryan Quah, Xiaoda Wang, Janelle Liu, John H Gilmore, Wei Gao, Yizhou Sun Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33048v1 Understanding how individual brain development unfolds over childhood requires modeling developmental trajectories from sparse longitudinal observations. Long-term neurodevelopmental forecasting is challenging because each child is typically observed at only a few irregularly spaced visits, while developmental dynamics vary across individuals and age. Generic continuous-time models accommodate irregular timing but often absorb these factors into a single flexible transition function, providing little structure for how population progression, individual variability, and developmental age shape the dynamics. We propose DevelopmentODE, a structured continuous-time framework that organizes population- and subject-specific variation within a shared developmental geometry while allowing the governing dynamics to evolve with age. The model builds this geometry around a developmental canal representing the population trajectory, whose local direction provides a reference for organizing subject-specific variation. Subject deviation velocities are constrained relative to this direction, while a shared nonlinear deviation field captures individual developmental motion without disrupting population-level progression. DevelopmentODE further models developmental non-stationarity through ordered age-dependent deformations of the shared vector field, progressively adapting a common dynamical structure as age changes, while elapsed time determines the integration horizon. This formulation uses population-level developmental structure to guide learning from sparse individual trajectories while allowing dynamics to evolve smoothly with age. We evaluate DevelopmentODE on longitudinal fMRI by predicting future functional connectivity of the same child from earlier observations. DevelopmentODE consistently outperforms competing baselines across short- and long-horizon predictions. --- ### Cost-free Spectral Estimation for Adaptive Newton--Schulz in Matrix Optimizers Authors: Kristi Topollai, Anna Choromanska Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33047v1 Matrix optimizers such as Muon transform each momentum matrix through an approximate orthogonalization, typically implemented by a small number of Newton-Schulz matrix multiplications. The quality and cost of this approximation depend strongly on the singular-value spectrum of its input, yet existing implementations use the same fixed polynomial routine for every layer and throughout training. We show that this uniform treatment is unnecessary: the computations in the Newton-Schulz method already reveal enough information to make the method adaptive. The Gram matrices formed inside Newton-Schulz iterations yield spectral moments through inexpensive scalar reductions, requiring no additional matrix multiplications. From these moments, we recover an estimate of the empirical singular-value distribution and use it to select a polynomial routine specialized to the current matrix. This turns Newton--Schulz orthogonalization into a spectrum-adaptive procedure that responds to differences across both layers and training time. On saved momentum matrices, spectral estimation substantially reduces orthogonalization error at a fixed iteration budget or reaches the same accuracy with fewer iterations, and in GPT pretraining up to 1B parameters it lowers the validation loss of two matrix optimizers. Our results suggest that matrix-function operations inside optimizers need not be designed for a conservative worst-case spectrum: they can cheaply measure the spectrum they are already processing and specialize computations accordingly. --- ### Pinned and Still Unstable: Within-Judge Verdict Variance and the Noise Floor of LLM-as-Judge Leaderboards Authors: Krishna Chytanya Ayyagari Published: 2026-09-27 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.33044v1 Modern LLM evaluation assumes that pinning a judge to a fixed model snapshot and decoding at temperature zero yields reproducible verdicts. We show this assumption fails as a property of how LLM-as-Judge is operationalized on cloud serving infrastructure, not of any particular model family. Across four frontier judges all served via a single major enterprise cloud platform and three standard benchmarks (Arena-Hard, AlpacaEval 2, MT-Bench), identical inputs to the same pinned, temperature-zero judge produce different verdicts across re-runs: per-item flip rates of roughly 5% on average and about 40% on the close-call items that decide leaderboard margins, with a per-judge magnitude spanning a 40x range (from 0.13% to nearly 10%). We introduce metrics tailored to this instability: per-item flip rate, a two-part stability profile (waver fraction and conditional intensity), and adjacency separability - and report what the variance does and does not do to rankings. For a single judge the aggregate ranking is stable (0% top-K instability, 0% pooled winner flip); what degrades is precision: under a paired hierarchical bootstrap, roughly one-fifth to three-quarters of adjacent leaderboard positions are statistically indistinguishable, a noise floor we attribute primarily to finite prompt sampling rather than to the judge. Across judges, leaderboards agree on the coarse ordering but diverge in the middle (Kendall's tau as low as 0.42-0.64 between families on Arena-Hard), and of 13 published head-to-head ranking claims we re-judge, 5 fail under a defensible judge swap or re-run. We argue that leaderboards report unhedged point estimates that misrepresent the noise floor of the instrument, and we propose a minimal, low-cost reporting protocol: several judge re-runs, published stability profiles and adjacency intervals, and results under at least two judges from different families. --- ### More than 83.69% of the zeros of the Riemann zeta function are distinct Authors: Kristian Muri Knausgård Published: 2026-09-27 Categories: math.NT, cs.AI, cs.LO arXiv: https://arxiv.org/abs/2609.33043v1 The lower asymptotic proportion of distinct nontrivial zeros of the Riemann zeta function, counted with multiplicity, is at least $0.8369928814\ldots$. The previous bound was $0.83625\ldots$. As in the proof of that bound, an unconditional version of Montgomery's pair-correlation theorem gives an asymptotic energy estimate. The new ingredient is a short matrix inequality with a free clipping parameter. It strengthens the lower bound for this energy in terms of the number of distinct zeros. The gain is a nonnegative correction from overlaps between different nearby zeros on the critical line, which is retained even when some of these zeros are double. The matrix inequality, the threshold lemma, the block dichotomy, the counting assembly and the exact arithmetic are proved formally in Lean 4. The constant relies on a computer-assisted local inequality from recent work that has not yet been refereed. That computation was re-run independently, and every imported input is listed. This paper is primarily an experiment in AI-assisted mathematical research (Section 4). --- ### Balancing Early Performance Sacrifices with Long-Term Gains: Scaling Learning-Rate Warmup Duration Across Training Horizons Authors: Kristi Topollai, Anna Choromanska Published: 2026-09-27 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33041v1 Learning-rate warmup is a standard technique in language-model training, yet its duration remains largely heuristic. Common approaches use either a fixed number of updates or a fixed fraction of the training horizon, two choices that imply very different scaling as training gets longer. When should warmup stay fixed, and when should it grow with the horizon? We address this question with a quadratic model whose modes respond differently to the peak learning rate. Warmup slows progress in directions that already contract well at the peak rate, but can remove persistent error in directions near the stability edge, with higher peak rates shifting the balance toward longer warmup durations. This yields a compact horizon scaling law that captures regimes ranging from essentially no warmup, through fixed-duration warmup, to durations that grow with the training horizon, and explains how the preferred regime changes with peak learning rate. Because the law captures the tradeoff between giving up early progress and improving the trajectory that follows, it can be fit using shorter runs and used to predict warmup at substantially longer horizons. Together, our results explain several familiar properties of warmup through a single tradeoff and suggest treating warmup duration as a horizon-dependent hyperparameter rather than a fixed training heuristic. --- ### Agent Safety From Within: Detecting Harmful Trajectories from LLM Internal States Authors: Difan Jiao, Ashton Anderson Published: 2026-09-27 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33039v1 Language model agents can now perform sophisticated sequences of actions via tools and harnesses, which has increased the scope of the damage they can cause. Guard models, however, are mainly built for content moderation and thus are not well-suited to detecting this agentic risk. To address this, we proceed by first conducting a representational analysis, then use the resulting insights to build a solution. In our analysis, we focus on two types of trajectory-level agentic harms: harmful content, which is expressed directly, and unsafe tool use, which depends on whether an action is consistent with the interaction that produced it. We investigate how open-source guard models represent these two types of harm and find that they are linearly readable inside the model, even though guard models predict no better than chance on pairs that differ only in the called tool's schema. The two harm types also follow nearly orthogonal internal directions, and neither reliably serves as a proxy for the other. These results motivate reading trajectory safety directly from internal states. We introduce TACIT, a readout of a frozen backbone's internal states that decodes no tokens. Trained on six trajectory-safety benchmarks, a linear probe raises mean macro-F1 from 62.3 for the strongest open guard to 80.7, and refined readouts reach 86.2. With each benchmark held out of training entirely, the refined readouts still lead the strongest guard (65.7 vs. 61.1). With the same backbone, training data and test split, the frozen readout is on par with full safety fine-tuning, and it improves the fine-tuned model further when applied on top. The probe trains about one millionth as many parameters as full fine-tuning in about a sixth of the time, and TACIT has the lowest latency of the guards we evaluate. --- ### Improving the Diversity of LLM Outputs without a Trade-off Authors: Ryoma Sato Published: 2026-09-27 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.33038v1 We propose DAST (Diversifying Arithmetic Sampling with TokenTour), a method that increases the diversity of LLM outputs without any change to the marginal distribution and with negligible generation-time overhead (a few microseconds). We observe that token IDs are often arranged in a meaningless order and reassign them so that tokens with similar meanings appear consecutively. This can be done in advance in a few hundred seconds per model, and the resulting order can be reused for all subsequent generations. By combining this order with arithmetic sampling (or quasi-Monte Carlo methods), we make similar tokens less likely to be generated across runs while preserving the distribution. Our method not only produces qualitatively good ideas but also significantly improves performance on the downstream task of ProtoQA. --- ### Byzantine-Robust Federated RAG via Aligned Calibration and Fixed-Membership Conformal Prediction Authors: Prasanjit Dubey, Aritra Guha, Xiaoming Huo Published: 2026-09-27 Categories: stat.ML, cs.LG arXiv: https://arxiv.org/abs/2609.33037v2 Retrieval-augmented generation (RAG) lets language models answer questions more accurately by consulting relevant documents. Many valuable collections, such as medical records, cannot be pooled because of privacy rules. Federated RAG leaves each collection with its owner, or node, which scores candidate answers from its own documents; a central hub combines the scores. Some nodes, called Byzantine, may be compromised, faulty, or misled by instructions hidden in documents, and report arbitrary scores. Conformal prediction returns a set containing the correct answer with a chosen probability, using a cutoff set in a calibration step on questions with known answers. An unknown group of nodes, no larger than a declared bound, may misreport both in this step and at query time. Existing methods assume every node is honest or protect only the calibration step. We observe that the honest nodes are the same in both steps. The hub therefore has all nodes score the same calibration questions, and keeps a candidate only if some plausible group of honest nodes, using its own scores in both steps, would keep it. We prove that the resulting sets contain the correct answer with the chosen probability in finite samples, whatever the Byzantine nodes report. No method using the same information can return smaller sets without risking the loss of an answer the honest nodes support. If nodes fail at random, the guarantee weakens only by the probability that more nodes fail than declared. In simulations, on real question-answering tasks including medical exams, and with language models as nodes, some hijacked, our sets reached the target whenever no more nodes misbehaved than declared, while plain averaging could miss it. They were also clearly smaller than those of simpler methods with the same protection, most of all when the declared bound was generous, so a cautious bound costs little. --- ### GeoSET: Generalist Foundation Model for SAR-to-EO Image Translation Authors: Jeonghyeok Do, Munchurl Kim Published: 2026-09-26 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.37496v1 Paired synthetic aperture radar (SAR) and electro-optical (EO) imagery is increasingly available across sensors, resolutions, and geographic regions. Yet existing SAR-to-EO image translation (SET) methods are typically trained on a single, limited-scale dataset, producing models specialized to particular sensing conditions. We introduce GeoSET, the first generalist model for SET, built around a single pretrained parent that is adapted to downstream datasets under a common protocol. We curate over 3 million high-quality SAR--EO pairs from a collection of more than 10 million SAR observations, spanning diverse sensors, spatial resolutions, and ground sampling distances. To bridge the modality gap between SAR observations and a pretrained image generator, we develop a speckle-robust SAR encoder and pretrain the conditional generator on this heterogeneous corpus. The resulting parent supports efficient adaptation across downstream datasets through low-rank adaptation (LoRA), updating only 0.60% of the generator parameters and requiring approximately one hour per dataset. Across six downstream benchmarks, GeoSET achieves state-of-the-art results in FID and DISTS with full fine-tuning or LoRA, demonstrating effective transfer across heterogeneous SAR-EO domains. --- ### MotionMaestro: Masked Tokenization for Unified Motion Generation Authors: Yun Chen, Munchurl Kim, Jeonghyeok Do Published: 2026-09-26 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.37495v1 Human motion generation plays an important role in applications such as character animation, virtual environments, and embodied interaction. While existing approaches have achieved remarkable progress, many of them are developed for individual tasks, including text-to-motion, pose-conditioned generation, and trajectory control. Although these tasks involve different types of conditions, a unified framework capable of handling them within a common representation would greatly simplify motion generation systems. We observe that diverse motion conditions can be naturally formulated as different observation patterns over motion sequences, where each task corresponds to a specific masking strategy. Based on this insight, we introduce MotionMaestro, a unified motion generation framework that learns a shared representation for complete motions and heterogeneous partial observations through masked motion tokenization. MotionMaestro employs a three-stage training strategy that first learns a masked motion tokenizer, then refines its reconstruction ability on clean motions, and finally trains a conditional flow-matching generator in the learned latent space. Furthermore, we introduce an observation map and an observation loss to explicitly preserve provided motion conditions during generation. With this unified representation and conditioning mechanism, MotionMaestro supports text-guided and unconditional synthesis, pose conditioning and partial completion, temporal interpolation, trajectory control, and motion continuation. Experiments on the large-scale RoMo and MotionMillion datasets show state-of-the-art performance across diverse motion generation tasks. --- ### Learning Social Navigation from Internet Videos in the Policy State Space Authors: Jiaming Wang, Duc Thang Nguyen, Jizhuo Chen, Volodymyr Shcherbyna, Diwen Liu, Zhengcheng Shen, Harold Soh Published: 2026-09-26 Categories: cs.RO, cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.37476v1 Training robust social-navigation policies requires simulators with diverse scene layouts, terrain, and human motion, but constructing such environments and specifying pedestrian behavior is costly. We propose an efficient pipeline that converts ordinary monocular walking videos directly into closed-loop social-navigation training environments in the policy's state space. Our key observation is that local social navigation primarily depends on two types of information: where the robot can traverse and how nearby pedestrians move. We therefore represent the static scene as a metric traversability map, which can be rigidly transformed under counterfactual robot motion, while directly replaying the pedestrian trajectories recovered from the video over time. This abstraction allows us to define the forward dynamics directly in the policy's state space and efficiently simulate counterfactual robot states without reconstructing or rendering photorealistic observations. The resulting policy achieves 81.2% success in the independent Arena benchmark, compared with 75.0% for the strongest baseline, and succeeds in 19/20 real-robot trials without policy fine-tuning. --- ### Boundary-State Control for Tool-Using Language-Model Agents: Commit-Time Consistency under State Drift Authors: Wesley Shu Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.37475v1 Tool-using language-model agents can decide that an action is permissible and execute it only after security-relevant state has changed. We study this proposal-to-commit gap and introduce BSC-R, a deterministic effect-boundary mechanism that binds a single-use commit authorization to the exact action and to a semantic projection of the authorization state that justified it. On 2,847 attacked AgentDojo episodes, the boundary kernel preserves the unprotected agent's behavior exactly (80.576% utility; 1.616% attack success). On 10,302 frozen proposals, it accepts every unchanged commit and rejects every instance of ten prospectively specified changed or replayed classes. In an independently generated boundary-drift experiment, full joint binding commits 0/4,403 invalid contexts while retaining 5,899/5,899 valid contexts. A prospective external evaluation on the 3,460-scenario CONTINUITY suite retains all 700 benign cases, prevents 1,200/1,200 represented non-replay invalid effects, handles 160/160 replay lifecycles correctly, and withholds 200/200 ambiguous no-release cases. The broader external suite also exposes the method's limit: across all 2,560 attacks, BSC-R has a 25% invalid-effect commit rate versus 0% for CONTINUITY. The result is therefore a scoped commit-time consistency mechanism, not a universal agent-safety claim. --- ### Authority Before Utility: Non-Compensatory Control for Persistent LLM Memory Authors: Wesley Shu Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.37474v1 Persistent memory creates a control problem that retrieval relevance alone does not solve: a memory can remain highly useful after an update, deletion, or revocation makes it inadmissible for the current answer. We formalize this as a separation between utility and authority. A fixed finite penalty applied to an unnormalized utility score cannot guarantee exclusion under arbitrary positive-affine reparameterization of that score; by contrast, rank-normalized compensation is scale-invariant and therefore forms a stronger empirical comparator. Our prospectively frozen TIDE/LongMemEval primary was quarantined before a valid HELDOUT comparison because the materialized TIDE adapter conflated historical age with query-relative inadmissibility and the aligned LongMemEval split left no DEV set for the predeclared penalty selection. We therefore report a post-primary replacement diagnostic on Memora Remembering, where update/delete operations provide item-level forgetting state. On Qwen3-8B, DEV selected lambda = 0.6 from a ten-point normalized SOFT family. Across 185 HELDOUT units in 28 dependency clusters, HARD exclusion yields 4.04% balanced construct error versus 19.66% for locked SOFT, a paired difference of 15.61 points with a 20,000-replicate cluster-bootstrap 95% interval of (13.07, 18.76). The effect is driven primarily by forgotten-value leakage while current-value recall is preserved. This is same-Q operator-comparison evidence, not a universal claim that scalar control fails, not an evaluation of learned authority inference, and not an independent downstream-harm endpoint. --- ### Backdoor in the Loop: Compromising Agentic Search via Malicious Retrievers Authors: Beining Xu, Peichun Hua, Yunming Xiao Published: 2026-09-26 Categories: cs.CR, cs.AI, cs.IR, cs.LG arXiv: https://arxiv.org/abs/2609.37468v1 Agentic retrieval-augmented generation (RAG) interleaves reasoning with repeated retrieval, giving the retriever influence over both the evidence an agent observes and its subsequent search decisions. We study retriever backdoors that exploit this feedback loop and repurpose weak backdoor purification to conceal their presence. An attacker supplies a compromised retriever checkpoint while leaving the search agent and deployment corpus unchanged. Without corpus write access, the attacker can still suppress useful evidence, persistently retrieve a selected existing document, or steer the agent toward prolonged search, inflating retrieval, context, and latency cost. To conceal these behaviors from detection, we propose leveraging a controlled inject-and-remove cycle: deliberately inject a weaker backdoor and then unlearn it. This process weakens detector-visible signatures and fools the backdoor detectors with an illusion of purification while preserving the malicious retrieval behavior. These findings expose a systematic vulnerability in RAG systems in which a weak defense becomes an attacker's concealment tool for a backdoored retriever, even when the underlying corpus remains trustworthy. --- ### VeriWeave Govern: Evidence-Gated Deterministic Runtime Governance for Enterprise AI Agents Authors: Kabeh Mohsenzadegan, Vahid Tavakkoli, Kyandoghere Kyamakya Published: 2026-09-26 Categories: cs.AI, cs.CY, cs.MA, cs.SE arXiv: https://arxiv.org/abs/2609.37457v1 Enterprise artificial-intelligence agents increasingly call tools, modify infrastructure, and process protected data, creating a need to separate action generation from action authorization. This article presents VeriWeave Govern, a deterministic runtime governance layer that evaluates structured agent actions against versioned policies, validates typed evidence, applies fixed deny > review > allow precedence, routes consequential actions to accountable human review, and records replayable tamper-evident audit state. GovernBench evaluates the design over 30 independent seeds and 60,000 oracle-labelled cases spanning five enterprise domains, adversarial evidence, out-of-distribution actions, and temporal policy evolution. VeriWeave achieves 0.9888 mean accuracy, 0.9836 macro-F1, zero observed aggregate false allows, and zero observed Governance Attack Success Rate on the evaluated cases. Six ablations show that evidence gating, deny precedence, out-of-distribution fail-safe behavior, human review, contradiction handling, and temporal replay contribute complementary safety. The deployed API additionally passes 12/12 end-to-end scenarios and a 40,040-request concurrency matrix with zero failures. A separate 150-case EU/Austria regulation-grounded evaluation uses frozen predictions and two independent blinded human annotators, who agree on all decisions. On this set, deterministic engines remain conservative, while a Gemma 4 31B comparator aligns more closely with the human consensus. The results expose a measurable safety--utility trade-off and motivate evidence-aware, replayable governance as an independent control plane for enterprise agent execution. --- ### Governing the Edge: Automating Commercial Property and Casualty Insurance Underwriting via a Hybrid Local-Cloud Multi-Agent Framework Authors: Vivek Kumar Singh, Gautam Bhowmick Published: 2026-09-26 Categories: cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.37454v1 Underwriters in commercial Property and Casualty (P&C) insurance spend 30 to 40% of their time on administrative work rather than risk judgment, and a single submission takes about 40 minutes by hand. We present Governing the Edge, a multi-agent framework for that layer, organized around a data-residency constraint: sensitive submission data must not leave the perimeter. Eleven agents and two deterministic control nodes, one a human-escalation interrupt, form a 13-node LangGraph workflow across two tiers. Agents touching raw submissions run locally on Gemma 2, each bound through a tool interface logging every invocation (synthetic stubs in the current prototype); only anonymized scores and non-identifying fields cross to Claude Sonnet in the cloud. Compliance rules are encoded as conditions on graph edges, so a non-compliant submission never reaches pricing. We walk through one full scenario end to end, a commercial auto submission whose principal driver carries serious violations, showing every agent call, every tool invocation, and the resulting routing decision. The framework processes a 40-minute manual submission in a few minutes on a single edge device, dominated by serialized on-device inference. On the hard-stop violation tier the framework enforces every rule correctly and reproducibly, since hard stops are deterministic predicates over fields extracted at temperature 0; overall compliance accuracy across the 20-scenario benchmark is 70%, with the remaining gap concentrated in softer, judgment-based tiers. It keeps a complete audit trail and respects the privacy boundary throughout. We release all code, compliance rules, tool stubs, and synthetic datasets. --- ### From Static Policies to Adaptive Priors in Offline Reinforcement Learning Authors: Tianwei Ni, Vineet Jain, Akash Karthikeyan, Pierre-Luc Bacon Published: 2026-09-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.35880v1 Offline reinforcement learning (RL) has traditionally focused on learning policies for direct deployment under conservative objectives, where uncertainty outside the offline dataset is treated pessimistically to ensure robustness. We argue that this formulation becomes incomplete when an offline-trained policy is subsequently updated through online interaction, as increasingly occurs in modern intelligent systems through test-time adaptation and online fine-tuning. This position paper argues that, in such settings, the objective of offline RL should extend beyond immediate deployment and instead prioritize learning adaptive policy priors: policies that preserve the capacity to improve during subsequent interaction through memory, exploration, and self-correction. We formalize this perspective as adaptive offline reinforcement learning (AORL), distinguish it from offline-to-online RL, and explain why adaptability becomes important under distributional shift, limited dataset coverage, and changing test-time conditions. We further discuss Bayesian offline RL as one principled direction for constructing adaptive policy priors by preserving epistemic uncertainty over plausible environments. Finally, we outline connections, open challenges, and research directions for treating offline RL as preparation for future experience rather than as a static deployment problem. --- ### CruxBench: A Benchmark of Information Discovery Authors: Hui Dai, Lina Piao, Nick Merrill, Nadja Flechner, Ezra Karger, Haifeng Xu Published: 2026-09-26 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.35879v1 Benchmarks for large language models (LLMs) typically evaluate the accuracy of answers against fixed reference labels. But a central step in many complex real-world tasks is identifying which questions are worth asking in the first place: decomposing a difficult problem into subquestions -- which we call cruxes -- whose answers provide key steps on the path toward solving the target problem. To evaluate this capability of information discovery, we introduce CruxBench, a benchmark that grades LLM-generated questions by their Value of Information (VOI): how much a model-proposed crux updates beliefs about a target forecasting question. CruxBench enjoys a rare combination of three key properties: it is (1) contamination-resistant by construction, since ground truth is generated by future world events; (2) open-ended, admitting unbounded and complex text-based submissions rather than one correct numeric answer; and (3) grounded, with informativeness measured against quantified changes in real-world beliefs. We evaluate a diverse set of eight models on 293 target forecasting questions and find that VOI correlates highly with independent measures of model capability (r=0.90) and captures cruxes' usefulness for answering target questions. However, information discovery remains challenging even for frontier LLMs, which only narrowly outperform a random-timing baseline. --- ### BarcodeMAE+: Rethinking Masked Pretraining and Global Representations for DNA Barcode Foundation Models Authors: Monireh Safari, Pablo Millan Arias, Scott C. Lowe, Lila Kari, Angel X. Chang, Graham W. Taylor Published: 2026-09-26 Categories: q-bio.GN, cs.LG arXiv: https://arxiv.org/abs/2609.35877v1 Many DNA foundation models are pretrained by masking parts of a sequence and asking the model to reconstruct them. Standard masked pretraining exposes the encoder to special (MASK) tokens that are absent at inference, creating a mismatch between training and downstream use. The role of an explicit global sequence representation such as a (CLS) token and how it should be trained also remain poorly understood for DNA barcodes. We introduce BarcodeMAE+ and study model architecture, global (CLS) representation, and auxiliary pretraining objectives across arthropod COI (BIOSCAN-5M) and fungal ITS (UNITE+INSD) barcodes. Across both barcode regions, the encoder-decoder MAE-LM architecture outperforms its matched encoder-only counterpart in nearly all evaluated configurations, supporting MAE-LM as an effective architectural design for DNA barcode foundation models. A trained global (CLS) representation provides substantial additional gains: on BIOSCAN-5M, (CLS) accuracy increases from 47.53% without an auxiliary objective to 80.65% with cross-entropy genus classification. The best auxiliary objective is region-dependent: cross-entropy performs best on BIOSCAN-5M, whereas pairwise same-genus classification performs best on UNITE+INSD, reaching 73.19% on Yeast and 63.07% on Filamentous Fungi. BarcodeMAE+ outperforms published DNA foundation model baselines on BIOSCAN-5M and achieves the highest Yeast accuracy among the evaluated UNITE+INSD baselines using frozen encoder representations. Similarity-weighted softmax KNN voting further stabilizes accuracy as neighbourhood size increases. Overall, encoder-decoder masked pretraining and an explicitly trained global representation are strong design choices for DNA barcode foundation models, while the optimal objective for learning that representation depends on the biological domain. --- ### Hybrid Ensemble Learning for EEG-Based Epileptic Seizure Forecasting Authors: Mason Dana, Khandaker Mamun Ahmed Published: 2026-09-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.35876v1 Epileptic seizure forecasting aims to provide actionable warnings before seizure onset, yet patient-independent generalization and false-alarm control remain major challenges. We propose a calibrated hybrid ensemble for EEG-based seizure forecasting that combines five deep learning models and three classical machine learning models through a logistic regression stacking meta-learner. The proposed pipeline integrates signal preprocessing, handcrafted feature extraction, class-imbalance handling, probability calibration, and clinically motivated post-processing. We evaluate the framework on CHB-MIT using strict Leave-One-Patient-Out (LOPO) cross-validation, with threshold and post-processing parameters selected only on held-out meta data. On the filtered cohort, excluding patients with anomalous preictal rates below 1\% or above 15\%, the model achieves 74.2\% seizure-level sensitivity at 1.24 false alarms per hour, with an average warning time of 16.9 minutes. A test-tuned oracle constrained to the target false-alarm budget achieves 60.9\% sensitivity at 0.951 false alarms per hour, highlighting the importance of reporting sensitivity together with realized false-alarm rates. Our code is available at: https://github.com/DanaMason/IEEE-CARS-Hybrid-Ensemble-Learning-for-EEG-Based-Epileptic-Seizure-Forecasting --- ### Beyond Symmetric Agents: Cognitive Diversity and Multi-Agent Debate in Small Language Models Authors: Leonardo Ferreira, Gardenia Liu, Kaden Zheng Published: 2026-09-26 Categories: cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.35875v1 Multi-agent debate (MAD) reportedly improves reasoning and factuality over single-model inference, but prior work treats agents as symmetric peers, leaving open what drives the gains. We test the hypothesis that cognitive diversity among agents is the driver, in the setting where the question is still measurable: small open-weight models with benchmark headroom. Across 23 models from eleven vendor families, five tasks, and 5,500+ debate and control runs, we vary diversity along three axes - personas, sampling temperature, and model identity - pairing every debate configuration with a generation-budget-matched majority-vote control. The hypothesis is rejected on every axis. Debate beats single-agent inference (3--7 points where tasks have headroom) but at matched budget conditions it ties or even loses to self-consistency sampling at 1.6$\times$ the wall-clock and 3.4$\times$ the token cost. Persona prompting reduces accuracy and a dose-response experiment over each model's full combinatorial persona space shows the cost is a persona tax, not a diversity tax: redundant personas hurt most, while maximally-diverse teams recover part of the loss. Furthermore, mixed-model teams lose to majority votes over their own rosters, with accuracy tracking member capability rather than heterogeneity, and nearly all of debate's benefit comes from the first exchange of answers. We further identify a pervasive measurement hazard in which debate transcripts silently overflow serving context windows, whose correction alone moves our debate-versus-sampling comparison from $-1.8$ points to parity. Our results recast reported MAD gains as an ensemble-sampling effect and provide the budget-matched, contamination-checked baseline bar that future debate mechanisms should be required to clear. --- ### Risk-Averse Online POMDP Planning via CVaR of the Immediate Cost with Performance Guarantees Authors: Yaacov Pariente, Vadim Indelman Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.35874v1 Online POMDP planners optimize the expected cumulative cost, which can mask dangerous states when the belief places significant mass on high-cost states. Existing risk-averse methods apply static or dynamic Conditional Value at Risk (CVaR) to the value function, capturing trajectory-level risk, but share two gaps: (i) by retaining the immediate cost as an expectation of a state-dependent cost over the belief, the risk \emph{within} the belief is left unaddressed; and (ii) by modifying the value function, they require new tailored algorithms rather than reusing existing expectation-based planners. We instead apply CVaR to the immediate cost over the belief at each step, directly targeting per-step uncertainty about the current state. The standard expected cumulative return is retained as the objective, so the resulting problem has a standard MDP structure: any expectation-based POMDP planner can be made risk-sensitive by changing only the cost computation. We inherit finite-time guarantees for policy evaluation and sparse sampling---with estimation error independent of the risk level---and, as our central theoretical result, prove a finite-time bound on the gap between the particle belief MDP surrogate and the original POMDP, which together yield an end-to-end guarantee from the true POMDP value to the algorithmic estimate. In the risk-neutral limit, the formulation recovers standard expectation-based planning. --- ### More Programs or More Rolls? Separating Coverage from Specialization in LLM Harnesses Authors: Ziyang Xu, Haitian Zhong, Hao Zhou, Hao Qin, Chenhan Jin, Te Qi, Shengze Xu, Tieyong Zeng Published: 2026-09-26 Categories: cs.AI, cs.LG, cs.SE arXiv: https://arxiv.org/abs/2609.35873v1 Automated generation of LLM harnesses promises to improve inference through task specialization. Yet additional answer coverage can arise from repeated execution of the same program, making specialization difficult to identify. We introduce a controlled evaluation that separates answer coverage, repeatable task advantages, and gains from pre-execution selection. On 386 MATH-500 tasks, we compare eight generated harnesses plus a baseline with nine byte-identical baseline copies, using three executions per member. Identical programs yield 2.16 percentage points of repeat-averaged oracle headroom. Generated programs exhibit substantially more repeatable score patterns, but these chiefly reveal persistent weaknesses: losses relative to the baseline persist across all three repeats on 100 tasks, while persistent wins occur on only one task and are sensitive to answer extraction. The frozen selector gains 0.00 percentage points, and both populations reach 98.70% oracle coverage at 27 harness executions. Stable complementarity remains unresolved at three repeats. Supporting BIRD traces locate failures in mechanism implementation, activation, and output validity. Together, these findings establish why coverage and repeatability alone cannot justify claims of useful specialization. They motivate an evaluation standard for harness diversity: task advantages should persist across executions, guide usable decisions, and improve on additional fixed-program executions under matched inference budgets. --- ### The Price of Token Boundaries: Compression Certificates and Prediction Authors: Yuhao Du, Shunian Chen Published: 2026-09-26 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.35869v1 Pre-tokenisation restricts which text fragments can become prediction units, but its compression cost is obscured when tokenisers are compared only under the same boundaries. We measure this cost by bounding the minimum token count from both sides, with and without a regular-expression boundary rule. Nonnegative prices on token occurrences yield a lower bound through shortest paths and vocabulary-budget selection; maximising over all prices recovers the linear programming relaxation, and an independent integer checker certifies the reported values. On English Wikipedia, boundaries increase the optimal token count by 28.3--36.8\%. Byte pair encoding lies 2.1\% above the constrained lower bound, but 10.9\% above the unrestricted bound. Compression and prediction favour different dictionaries: at 85M non-embedding parameters and matched training-token budgets, unrestricted fitting yields higher mean held-out bits per byte under a common unrestricted decoder in all 12 languages in the paired study and 11 of 12 under independent tuning and evaluation. To study intermediate boundary policies, we introduce boundary licences, which limit the vocabulary entries permitted to cross cuts and admit the same form of certificate. On separate English and Chinese fitting corpora, licensing 10\% of the vocabulary budget recovers 85.2\% and 100.0\% of the achieved token-count reduction from removing all cuts. These results quantify the compression cost of boundaries while separating it from the prediction quality of the resulting token units. --- ### Is Human-Readable Text Necessary for Effective LLM Fine-Tuning? Authors: Jinhao Zhang, Zeyu Liu, Zicheng Yan, Yunquan Zhang, Daning Cheng, Song Tang Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.35868v1 Is human readability necessary for effective fine-tuning of large language models? We investigate whether model-conditioned training representations can preserve or improve adaptation utility without requiring a human-readable textual form. We propose Desired-Update-Aligned Synthetic Data (DASA), which uses activation-gradient feedback from a frozen reference model to guide the optimization of continuous synthetic input embeddings. Inspired by the role of activation gradients in local risk reduction, DASA targets useful adaptation updates rather than source-text reconstruction or linguistic fluency. The resulting embeddings are used directly for downstream fine-tuning; discrete token projections are employed only for qualitative inspection. Experiments on six models from the Llama and Qwen families, ranging from 1B to 32B parameters, cover six benchmarks spanning knowledge, mathematical reasoning, code generation, and commonsense reasoning. Under matched LoRA adaptation settings, DASA achieves performance comparable to the source natural-language data and surpasses it in multiple configurations, while outperforming GRADMM in most comparisons. Further experiments cover general-domain and task-specialized source data. Under the evaluated synthesis settings, DASA provides a $3.6$--$4.9\times$ speedup over GRADMM with comparable peak GPU memory. --- ### ReLOBGen: Replayable Limit Order Book Message Generation Authors: Junoh Kang, Kiseop Lee, Bohyung Han Published: 2026-09-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.35867v1 We propose ReLOBGen, a method for generating limit order book (LOB) messages that are replayable by construction. Replayability is required for closed-loop market simulation, yet existing LOB message generators may produce non-replayable raw messages, i.e., messages inconsistent with the current market state. These generators therefore rely on post-hoc correction or rejection followed by resampling, which may alter the replayed message distribution or increase inference cost. ReLOBGen instead ensures replayability during generation: it selects the referenced order from the resting orders in the current LOB and then generates the remaining message fields to be consistent with that order and the market state. For realistic reference selection, ReLOBGen samples from a learned distribution over eligible resting orders, efficiently computed from cached order representations and a context-dependent query. It then enforces the consistency of the remaining fields by masking out invalid tokens. Together, these components enable efficient generation of realistic messages without post-hoc correction or resampling. In 500-message rollouts, ReLOBGen achieves 100% replayability, improves market realism, particularly for top-of-book statistics and the relative prices of LOB messages, and provides a $2.7\text{-}3.6\times$ speedup per replayed message over the LOBS5 baseline. --- ### PACT: Pairwise-Anchored Calibrated Tuning for Single-Token Typed Decisions Authors: Yida Lin Published: 2026-09-26 Categories: cs.CL, cs.AI, cs.CV, cs.GT arXiv: https://arxiv.org/abs/2609.35865v1 Single-token typed-decision models answer a schema question by reading the logits of a few one-letter answer codes at a single position: they are fast and return a probability for every allowed answer, but they are trained with plain cross-entropy that ignores most of the structure in their training data. We study such a model whose data is curated as contrastive pairs---two contexts that differ in one edited fact that flips the answer---each carrying a machine-checked certificate that deleting the decisive sentence makes the fact unknown. We propose PACT, which turns this structure into four training terms that need no new annotation: a difference-in-differences margin over each pair that is invariant to any shared logit offset, a permutation-consistency term against answer-code position bias, an evidence-necessity term on certificate-verified ablated contexts, and an ordinal transport cost for rubric fields, plus a three-parameter contextual temperature. On a frozen 324-item holdout with three seeds, PACT matches the published recipe in accuracy ($84.6\%$ vs. $85.2\%$; McNemar $p \ge 0.50$ at every seed) while giving the lowest position bias of all runs (answer flips under relabelling $9.8\%$ vs. $13.8\%$) and the lowest ordinal error on rubric fields (MAE $0.232$ vs. $0.311$). Against a control with the same optimiser and schedule but cross-entropy only, PACT is significantly more accurate at two of three seeds, halves the seed-to-seed spread and lowers NLL by $26\%$. Seed-matched ablations and pre-specified falsification tests locate these gains precisely: no single term raises raw accuracy, and the method's value lies in robustness and stability rather than headline accuracy. Code, data splits, trained adapters, and all run records are available at https://github.com/BennyLinntu/PACT-Pairwise-Anchored-Calibrated-Tuning-for-Single-Token-Typed-Decisions. --- ### What Must a World Model Distinguish for Planning? Authors: Rongzhe Wei, Hans Hao-Hsun Hsu, Peizhi Niu, Yifan Li, Pan Li Published: 2026-09-26 Categories: cs.LG, cs.RO arXiv: https://arxiv.org/abs/2609.33030v1 World models simulate the consequences of action candidates, but good planning need not preserve every physical distinction required for accurate prediction. We formalize this gap through a hierarchy of mechanism, response, and decision sufficiency. Given a candidate set, the planning query determines which physical variations matter and how precisely they must be preserved: coarse decisions can discard much of the information needed for prediction, whereas fine decisions may require nearly the same resolution. In practice, planners often adaptively search to construct candidates, and information unnecessary for final selection may still be needed to discover good candidates. What a world model must preserve therefore depends on the query, the candidate set, and the planner. We study these effects in a collision system, nonlinear dynamics, and robotic planning. These varying requirements raise a design question: where should query information enter the planning system? A model that jointly generates actions and outcomes conditioned on the query achieves lower regret than an action-conditioned world model on seen objectives, but this advantage largely disappears when generalizing to unseen objectives. Motivated by this, we propose a modular design in which the query determines where to look and an action-conditioned model predicts what will happen, allowing the same predictions to be reused across objectives. --- ### Łukasiewicz Neural Networks Extended: Residual Architectures and Crystallization Strategies for Interpretable Rule Extraction Authors: Carlos Leandro Published: 2026-09-26 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.33028v1 A feed-forward neural network whose weights are integers and whose activation is the truncated identity implements, neuron by neuron, the connectives of Łukasiewicz many-valued logic. This exact correspondence --- established theoretically by Castro and Trillas and developed into a training algorithm by Leandro --- enables \emph{symbolic knowledge extraction}: training produces not a black-box model but a logical formula. Two obstacles have limited the approach to shallow architectures and small datasets: crystallization (forcing weights to integers) succeeds only probabilistically under the original Levenberg--Marquardt training scheme, and the theoretical guarantees break down as networks grow deeper. This paper addresses both obstacles. First, we prove that \emph{residual connections} (skip connections of the kind used in ResNets) extend Łukasiewicz neural networks to arbitrary depth while preserving the symbolic correspondence \emph{at merge neurons} by construction: merge neurons in a Łukasiewicz residual block automatically satisfy the neuron-classification proposition, regardless of the inner layer weights; inner-layer neurons are trained toward representability by the crystallization strategy. Second, we analyse three crystallization strategies --- Levenberg--Marquardt (corrected), straight-through estimation (STE), and proximal regularization --- characterizing their theoretical guarantees, failure modes, and interpretability trade-offs. --- ### Low-Rank Single-Index Bandits with Unknown Links: From Matrices to Tensors Authors: Zhongxuan Liu, Yue Kang, Thomas C. M. Lee Published: 2026-09-26 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.33025v1 Low-rank matrix and tensor bandits exploit structured interactions but typically assume a known reward link. Recent single-index bandit methods accommodate unknown links without directly exploiting matrix or tensor rank. We address this gap by studying stochastic matrix and tensor bandits with an unknown shared Lipschitz link and a low-rank index parameter under known regular candidate distributions and finite-variance noise. For monotone links, T-ESTOR combines robust, rank-adaptive Stein estimation with epoch-based greedy selection. Under exact selected-score access and a uniformly positive selected-design Stein signal, it achieves square-root regret with dimension dependence determined by the low-rank structure. For every admissible design, the monotone lower bound matches the rank, dimension, and horizon dependence up to logarithmic factors at large horizons, for fixed menu size and model/design constants. For nonmonotone links under a nonzero base-law Stein signal, T-BSTOR combines structured estimation with robust bin-based learning and attains the optimal $\widetilde{O}(T^{2/3})$ horizon rate for fixed dimensions, menu size, and model/design constants. Synthetic and CCLE-based experiments illustrate the benefits of structured estimation relative to vectorized and competing single-index baseline methods. --- ### SRE-Marathon: A Continuous, Change-Driven Benchmark for Autonomous Site Reliability Agents Authors: Yifang Tian, Yingjian Bai, Yifeng He, Zichun Chong, Yuanchen Gao, Yiran Li, Hans-Arno Jacobsen Published: 2026-09-26 Categories: cs.AI, cs.SE arXiv: https://arxiv.org/abs/2609.33023v1 Benchmarks for site reliability engineering (SRE) agents are typically episodic: one fault is injected, the agent receives an incident task, and its response is scored. Production operation is not. Incidents surface through noisy alerts, overlap in time, and often originate from code or configuration changes. We present SRE-Marathon, a benchmark for long-horizon, continuous SRE operation. An agent is invoked at a fixed cadence with cumulative alert history and a persistent workspace while operating a live two-zone Kubernetes deployment as a fault orchestrator injects overlapping faults according to a seeded, production-calibrated schedule. Curated code and configuration changes deployed through the same build pipeline available for repair. Each run is recorded into a sealed bundle and scored offline: Marathon-Score credits each injected fault for ordered progress through correlation, localization, and repair, with all metrics computed deterministically from recorded system evidence. Across three applications and about sixty faults per run, the best of 10 methods reaches only 41.3 out of 100. Agents often correlate and localize faults, but almost never complete repairs while the faults remain active. --- ### Decentralized Optimization with Cross-Coupled Mixed Affine Constraints Authors: Ewsey Obzherin, Ilya Khomchenko, Natalia Shelegeda, Nhat Trung Nguyen, Demyan Yarmoshik, Alexander Rogozin, Alexander Gasnikov Published: 2026-09-26 Categories: math.OC, cs.LG arXiv: https://arxiv.org/abs/2609.33021v1 We study decentralized optimization with cross-coupled mixed affine constraints, where local and shared variables interact through two affine channels. We show that the intrinsic difficulty of combining separately well-conditioned channels is governed by their Friedrichs angle. This geometry induces a cross-coupling factor that cannot be removed by channelwise preconditioning and governs the additional affine-oracle and communication complexity. We develop an accelerated decentralized method with matching minimax guarantees in the smooth strongly convex regime and extend the framework to smooth and nonsmooth convex objectives. Experiments confirm the predicted dependence on cross-channel geometry and network conditioning. --- ### Residual Diffusion Implicit Models Authors: João Guerreiro, Pedro Tomás, Helena Aidos, Jacinto C. Nascimento Published: 2026-09-26 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33020v1 Diffusion models achieve state-of-the-art results across multiple tasks. However, in inverse problems, standard initialization from pure Gaussian noise misaligns the generative process with real-world degradations. More recent methods such as diffusion bridges impose strict endpoint constraints and often require long reverse processes that are prone to hallucinations. Alternative consistency models provide noise-invariant, one-step mappings but lack inherent variance modeling and can degrade under severe corruption. Hence, residual diffusion implicit models (RDIMs) are proposed, constituting a generalized framework that explicitly models the residuals between high-quality (HQ) and low-quality (LQ) images, aligning the forward process with the actual degradation. A non-Markovian implicit reverse sampler is derived, which can skip intermediate timesteps, enabling accurate few-step or even single-step reconstruction, while mitigating the hallucinations inherent to long diffusion chains. RDIM also introduces a controllable variance mechanism that interpolates between deterministic and stochastic sampling, balancing fidelity and diversity. Furthermore, it enables the straightforward use of perceptual losses, when needed. Experiments on denoising and super-resolution benchmarks demonstrate that RDIMs consistently outperforms the state of the art, including bridge and consistency models, in terms of PSNR, SSIM, and LPIPS, reducing hallucinations while requiring only a few sampling steps (often just one). The results position RDIMs as an efficient solution for a broad range of image restoration tasks. --- ### Trust and Task Completion in the World of Consumer AI Agents Authors: Jeroen Olieslagers, Eduardo Pujol, Gal Zahavi, Lukas Ingemarsson, Shivani Poddar Published: 2026-09-26 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.33017v1 Action agents do things for people. They send email, spend money, and call businesses while the user is busy with something else, so a mistake can turn into an action before anyone notices. They fail their users in two ways. They break trust when they do something the user never agreed to, or hold back after the user clearly said go. And they fall short on completion when they give up on errands that turn out to be hard. Both depend heavily on the harness around the model, meaning its instructions, tools, context, and guardrails. We built an evaluation that scores trust and completion on the same runs, in a simulated world of businesses with their own websites, inboxes, and phone lines, and of people who write back. A simulated user answers the assistant's questions. Trust means that nothing happens the user did not agree to. No email goes to someone they never approved, no private detail ends up on a group thread, no money is spent past their limit, no stranger's instructions are followed, and nothing is claimed without a source. Every trap has a matched control in which acting is the right call. We use the evaluation to measure Fo, Wajo's personal assistant, against a base model with basic instructions on three foundation models, and against the Fo harness with its guardrails switched off. Fo completes 71% of the errands and keeps the user's trust on 94% of the trap runs. The base models complete 50% to 64% and keep trust on 59% to 75%. On the matched controls, Fo goes ahead slightly less often. OpenClaw, a popular open-source assistant given the same access, completes 42% of the errands it shares with Fo, against 71%, and keeps the user's trust on 74% of the shared trap runs, against 94%. Measuring trust and completion together, on the whole system rather than the model alone, is how we think action agents become safe to hand real work to. --- ### Relative Generalization Invariance of LLM Pretraining Authors: Fengzhuo Zhang, Shuche Wang, Shenggui Li, Tianyu Ruan, Jianliang He, Ivor Tsang, Tianyu Pang, Chao Du, Tianwei Zhang, Zhuoran Yang Published: 2026-09-26 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.33016v1 Large Language Model (LLM) pretraining performance is jointly shaped by three components of the training triplet: the optimizer, model architecture, and training data stream. However, how these components influence performance in distinct ways remains unclear. We take a first step toward isolating their effects by studying relative generalization. We introduce Relative Generalization Invariance (RGI), the invariance of the validation-loss difference between any two tokens across models. We show that RGI approximately holds across a wide range of optimizers and moderate architectural variations, suggesting that these choices induce an approximately uniform shift in token-wise losses. In contrast, changing the training data stream can substantially alter relative generalization. We further show that RGI cannot be explained by the neural tangent kernel or mean-field regimes alone and prove that it can emerge in an overparameterized quadratic model. Overall, our work identifies RGI as a new phenomenon in LLM pretraining that helps distinguish the effects of optimizers and architectures from those of training data. --- ### TCMQA: A 38K-Question Traditional Chinese Medicine Benchmark with a Licensed-Practitioner Reference Authors: Tzu-Heng Huang, Jet Lin, Eric Lin Published: 2026-09-26 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.33014v1 Medical benchmarks for language models are built almost entirely on Western biomedicine. Traditional Chinese Medicine (TCM) is a separate system, with its own diagnostic framework and its own literature, and it remains largely unmeasured. The few TCM evaluations that exist are small, narrow, and rarely paired with a human reference. We present TCMQA, an open benchmark of 38,279 questions from Chinese TCM licensing examinations, paired with 15,151 responses from 101 licensed practitioners. We evaluate 29 instruction-tuned models from 9 families, spanning 0.27B to 14.8B parameters. Accuracy ranges over 59 points, and no model approaches saturation. Pretraining data predicts TCM ability far better than scale: a 12B Western-pretrained model reaches 39.6%, while a Chinese-pretrained model an eighth its size reaches 60.8%. Nine models exceed the practitioner majority vote of 64.9%, the best by 21.8 points, and all nine come from that same Chinese-pretrained family. Yet difficulty does not transfer between models and practitioners: accuracy is flat across practitioner-rated difficulty, item-level agreement is near zero for all 29 models, and on $8.4\%$ of items the practitioners are correct where the leading model is wrong. We release the corpus, the practitioner responses, the harness, and per-item model outputs at https://huggingface.co/datasets/TechTCM/TCMQA. --- ### The Epistemics of Agent Memory: Measuring, and Governing, the Consolidation Decision in Long-Horizon LLM Agents Authors: Sasank Annapureddy, Anjaneya Prasad Thamatani Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33013v1 Long-horizon LLM agents must convert accumulated experience into durable memory, deciding what to keep, compress, abstract into reusable skills and rules, or forget. We report a four-phase research program on this consolidation problem whose central finding is a shift in what is measured: from how much an agent remembers, to whether its consolidation decisions are any good, to whether those decisions can be trusted. Phase 1 learns episodic boundaries from agent traces by downstream utility; an honest near-miss (oracle correlation 0.691 vs a 0.70 bar) whose lasting output is a three-gate anti-leakage protocol. Phase 2 learns when to promote experience and to which abstraction level under a token budget, achieving a verified +22.7% task-success improvement with 7x compression, but exposing a degenerate-forgetting failure and a distribution-shift failure mode we name lambda-prevalence coupling. Phase 3 introduces ConsolidationBench, an oracle-by-construction benchmark that scores consolidation decisions against a known optimum on three non-circular axes; production retrieval systems retain information yet score zero on cross-level transfer. Phase 4 introduces governed consolidation: the decision wrapped in poison-resistance, reversibility, and auditability guarantees with a quality gate. Governance is statistically distinct from the quality score ($r^2 = 0.43$; partial $r = 0.27$; identical-quality policies differ threefold in governance), so the contribution survives independently of the metric's external validity. On that question we report a resolved negative: after a graded-reuse redesign removed a structural ceiling, a two-benchmark study with 2,532 real answer cells finds the quality score does not predict real transfer accuracy (pooled Spearman $ρ= -0.24$, n = 12, CI spanning zero). An adversarial self-critique pass cleared the final claim set with zero surviving overclaims. --- ### When Pair Count Is Not the Sample Size: What All-Pairs Agent Comparisons Estimate Authors: Wei-Jung Huang Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33012v1 When an agent benchmark compares every pair of leaderboard entries, the number of comparisons can look much larger than the independent evidence behind them: A versus B and A versus C both reuse A. Whether this reuse affects inference depends on what the analysis is meant to describe. If the board and its outcomes are fixed, the all-pairs mean is an exact summary of those entries, and any interval must come from another declared source of randomness. If the entries are instead treated as iid draws from a population of future configurations and the pair rule is regular and nondegenerate, the same mean is an order-two U-statistic whose first-order uncertainty depends on the number of configurations, not the number of pairs. We use near ties as the running example, but the distinction extends to other symmetric pair summaries when their regularity conditions hold. We examine both interpretations using a fixed SWE-bench Verified snapshot and an exact binary model with known truth. On SWE-bench, intervals that accounted for shared configurations were more than twice as wide as a pair-iid reference that treated the pairs as independent. In the exact model, pair-iid coverage fell far below the nominal level when edges shared endpoints but remained near nominal for matched independent edges. Results on two other fixed leaderboards show that exact summaries also depend on which pairs are included and how they are weighted. An all-pairs analysis must therefore state what is fixed, what is sampled, and how it handles shared entries and pair aggregation. --- ### Saturation-Insensitive Dueling Bandits with General Function Approximation Authors: Chenggong Zhang, Xuheng Li, Qiwei Di, Weitong Zhang, Quanquan Gu Published: 2026-09-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.33011v1 We study contextual dueling bandits with general function approximation under the Bradley-Terry-Luce (BTL) preference model. A key challenge in this setting is the saturation of the preference model: when the current reward model can already distinguish two actions with high confidence, the resulting preference feedback becomes weakly informative, making it difficult to further improve reward estimation. Consequently, existing sample-complexity analyses often depend on the inverse-derivative factor $1 / σ'(Δ_{r^\ast})$ which can be prohibitively large when the link function $σ$ saturates for large reward gaps $Δ_{r^\ast}$. To address this issue, we introduce `SI-CDB`, an algorithm that selects opponent arms using a carefully designed heuristic for arm selection. This design enables saturation-insensitive reward learning and recovers the near-optimal dependence for linear reward classes, eliminating the unfavorable $1/σ'(\cdot)$ factor. The core of our analysis is a localized Eluder dimension framework tailored to dueling bandits with general function approximation. Our theoretical results also explain why two-arm regret analysis is crucial for improving single-arm performance in dueling bandits. --- ### Model-Aware Data Selection from In-and-Out Information Interplay Authors: Yifan Wang, Xiaomin Li, Yuexing Hao, Dongwon Jung, Hemanth Neelgund Ramesh, Ananth Grama, Varun Chandrasekaran, Yu Hu, Andrzej Banburski-Fahey, Jaron Lanier Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.33010v1 LLMs are effective representations that assimilate vast amounts of knowledge during pretraining, but post-training is necessary for models to reliably access this knowledge and "know what they know." We observe an interesting rank equilibrium between knowledge stored in the weights and the data stream passing through the model. Across all model layers, we find that the hidden states (data stream) follow a U-shaped pattern, showing substantial compression in early layers and a steep rise during the late-layer decoding phase. In contrast, the weight rank follows an inverted U-shaped pattern, with very low rank in the early and late layers and high rank in the middle. We interpret this as an in-and-out information interplay: intermediate activations do not need to carry content that the weights can supply later, so they primarily preserve what the weights cannot provide. Motivated by this observation, we propose a model-aware data selection method, CAP (Counterfactual Assimilation Profile), which can determine whether a data candidate contains information accessible to the current model by utilizing the divergence gap in early- and late-layer representations between model-generated and reference responses. Across math, code, and science domains, CAP delivers 35.4% greater average improvement over the base model than the strongest baseline under different selection budgets. With only 10% of the data pool, CAP surpasses or matches full-pool training on math and science. We further show that CAP transfers to multimodal data selection and is robust to response horizon and noise. --- ### A statistical theory for blind denoising: minimax estimation of the noise level from a single sample Authors: Zhuoer Shen Published: 2026-09-26 Categories: math.ST, stat.ML arXiv: https://arxiv.org/abs/2609.33008v1 Motivated by blind denoising in diffusion models, we study estimation of an unknown Gaussian noise level from a single high-dimensional observation, assuming the signal law P is known. We characterize the minimax mean-squared error under two structural assumptions on P. For signals with covering complexity k, the minimax rate is $\widetildeΘ_Λ(\min\{Δ_Λ^2,d^{-1}+k^2d^{-2}\})$, and the constrained MLE attains it up to logarithmic factors. For $α$-strongly log-concave signals, the rate is $Θ_Λ(\min\{Δ_Λ^2,(1+α^{-1})^2d^{-1}\})$, attained up to constants by a $P$-centered norm estimator. These results show that the structure of the signal law determines both the difficulty of blind noise estimation and the appropriate estimator. --- ### CAPEX: Efficiently Distilling Foundation Model Behavior into Deployable Robot Policies through Experience-Adaptive Reasoning Authors: Shivam Aarya, Zhang Xi-Jia, Chengyue Huang, Junhyun Kim, Huishu Xue, Hrishit Leen, Roman Yakunin, Animesh Garg, Zsolt Kira Published: 2026-09-26 Categories: cs.RO, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.33007v1 Robot learning has largely relied on human-teleoperated demonstrations to acquire effective learnable behaviors. However, human-operated data collection processes can be unintuitive, difficult to scale, and inherently asynchronous. We explore an alternative: distilling physical behavior from general-purpose multimodal foundation models into deployable robot policies by using the foundation model itself as an autonomous demonstrator. While sufficiently capable models can generate successful zero-shot manipulation trajectories, repeatedly invoking them during physical execution is slow and expensive, limiting their utility as scalable data generators. As a solution, we introduce CAPEX, an experience-conditioned demonstration collection framework that uses execution experience from previous attempts to adapt how frequently the foundation model must observe, reason, and replan. We evaluate across RoboCasa tasks and on physical Franka and bimanual YAM-arm platforms, measuring task success, model calls, token usage, collection time, and cost. We further train Diffusion Policy and ACT on matched sets of human-teleoperated and foundation-model-generated demonstrations to evaluate the downstream learning value of autonomously collected data. We find that CAPEX increases the number of successful demonstrations by 4.3x while reducing the cost per successful demonstration by 80%. Policies trained on CAPEX-generated data approach the performance of those trained on matched human demonstrations; with longer training, this gap largely closes for policies trained from scratch. These results suggest that foundation models can serve as scalable sources of reusable robot experience. Project page: https://capex-paper.github.io/ --- ### Safety-Constrained Cascade Inference for Robust Malaria Cell Classification Under Field Corruptions Authors: J. T. Hagbe, Michel Emel Published: 2026-09-26 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33005v1 Automated malaria diagnosis from thin blood-smear microscopy could meaningfully reduce the burden on under-resourced laboratories, but a model that maximises accuracy on clean laboratory images fails badly the moment an inexpensive smartphone camera introduces sensor noise. This paper introduces MalariaCascade, a two-stage inference system in which a lightweight MobileNetV2 sentinel (2,225,153 parameters) makes confident classifications at low compute cost and escalates uncertain cases to an EfficientNet-B3 expert (10,697,769 parameters) that sees only clean, standardised images regardless of how corrupted the incoming frame is. Structural isolation of the expert stage, not learned robustness, is the mechanism. The sentinel is trained under a safety-score objective that places an explicit floor on Recall(Parasitised) (>=0.95) and Recall(Uninfected) (>=0.40) before precision is optimised, ensuring the checkpoint satisfies clinical safety constraints by construction. On the NIH Malaria Cell Images Dataset (27,558 cells), the cascade reaches Accuracy=0.9736, Recall(Parasitised)=0.9570, Precision(Parasitised)=0.9912, F1=0.9738, and AUROC=0.9955 on the clean test set. Under Gaussian sensor noise at full severity, cascade Recall(Parasitised) degrades by only 2.4 pp (0.9570 to 0.9329), while the flat single-model baseline collapses by 63.0 pp (0.9584 to 0.3281). McNemar's test confirms the cascade improvement is statistically significant (chi^2=11.14, p=0.00085). An ablation isolating the structural property shows that routing clean images to the expert is responsible for a 16.8 pp Recall(Parasitised) advantage under sensor noise relative to a cascade where the expert also sees corrupted inputs. --- ### Certified Interface Aliases: Exact Collisions in Vision-Language Preprocessing, and When They Exist Authors: Mert Onur Cakiroglu, Elham Buxton, Mehmet Dalkilic, Hasan Kurban Published: 2026-09-26 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.33003v1 Vision-language verifiers and routers must distinguish errors repairable by more reasoning from those caused by visual evidence never reaching the language model. This distinction lacks ground truth because annotators see full-resolution images while models receive preprocessed tensors. We introduce AliasForge to create cases where the relevant fact is provably absent from the interface. Fixed-point resampling makes the pre-rounding resize an exact integer linear map that can send nonzero integer perturbations to zero. Hiding a label-flipping perturbation there produces images with opposite step-correctness labels but bit-identical interface states. Every verifier therefore has the same output law on both members, giving pair-balanced accuracy exactly one half and zero gain from language-side repair. We prove that every fixed-point downscaler has such null vectors and bound their smallest size at the most common ratios, which rules out an 8-bit fit whenever the bound exceeds 255. From the resize configuration alone, a lattice criterion supplies realizable collisions and certifies their absence within the specified construction family. It resolves all 20 screened configurations, 17 as constructible and 3 as non-constructible. We construct certified pairs across three architectures and certify four additional processors, with zero decision-logit gap on all 18 scored pairs and none of the 18 controls. The pairs also screen routers that waste computation on re-attention or further reasoning. On natural items, per-item routing headroom exists, but no tested interface-only router improves over stopping. Our fiber ceiling bounds the headroom recoverable from the interface. Code: https://github.com/KurbanIntelligenceLab/aliasforge. --- ### TriDrive: Joint Driver, Vehicle, and Road Modeling for Forecasting and Driver Monitoring Authors: Yuhang Wang, Jingxin Yang, Chuheng Wei, Yuechen Guo, Jinghan Xu, Zhao Han, Hao Zhou Published: 2026-09-26 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2609.33000v1 Predicting how drivers, vehicles, and road scenes interact and evolve together is central to driver monitoring. Prior work models in-cabin activity or traffic-conditioned driver motion in isolation, motivating joint driver, vehicle, and road modeling with real-time on-vehicle evaluation. We introduce TriDrive, to our knowledge the first unified framework that jointly forecasts driver kinematics, vehicle dynamics, and road demands through an automation-conditioned transition model. Modality-specific encoders (an anchored kinematic representation of the driver, causal CAN-bus dynamics, and frozen V-JEPA 2 road latents with structured road margins) are connected by directed residual connections through which driver and road context refine vehicle forecasts. We evaluate TriDrive on three downstream tasks. On the public AIDE benchmark, its kinematic encoder recipe sets a new full-set state of the art (SOTA) among published baselines (48.05 versus 71.47 All-MPJPE). On 197.2 hours of naturalistic BATON subset, directed connections and road margins raise assistance-engaged PR-AUC by 0.084 for steering onset and 0.286 for time-to-collision drops. For real-time use, we distill the road encoders and run TriDrive on a comma four with an external 8 GB GPU, where a lightweight current-state warning probe updates at 5 Hz with 177 ms p95 latency while the joint model forecasts concurrently. The probe is above an openpilot-based baseline on human-labeled manual-driving warnings (AUROC 0.725 versus 0.563), and in a paired on-road study 14 drivers rate its warnings as more appropriate (+1.79) and timely (+2.67) than those of openpilot's driver-monitoring system. --- ### Oracle Gaps in Reliability Coverage: Sampling Noise or Policy Specialization? Authors: Mert Onur Cakiroglu, Mehmet Dalkilic, Hasan Kurban Published: 2026-09-26 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.32996v1 Policies trained from the same base model can appear to solve different problems. An oracle that chooses the best policy for each problem may therefore appear much stronger than any single policy. Selecting the largest estimated success rate also selects favorable sampling errors. We study this effect through reliability coverage, the fraction of problems whose success probability reaches a chosen threshold. Our first test redistributes stored correctness outcomes across policies within each problem. A second also preserves each policy's total successes, accounting for overall quality differences under a specified statistical model. For five training seeds of a seven-billion-parameter vision-language model, redistribution reproduces 0.096 of an estimated 0.113 oracle gap at threshold 0.10. Neither test finds significant evidence at this threshold. Small advantages remain unresolved. Mixtures, routers, voting, and weight averaging show no detectable improvement over their corresponding single-policy baselines. Training policies on different datasets shows little detectable specialization under light post-training, and no router gain. A stronger recipe does create it, both tests detect it, and a router gain appears only at the high thresholds where the specialists separate. In a control with predictable specialization, a router recovers about half the oracle gap. Coverage bounds explain why even a genuine oracle advantage need not yield a deployment gain. The tests assess apparent specialization from stored responses before investment in routing. Code: https://github.com/KurbanIntelligenceLab/oracle-gaps --- ### X-Tree: Tokenizing Reusable Experience for Efficient Agent Generalization Authors: Sitao Cheng, Xunjian Yin, Zhiyuan Sun, Yuxuan Li, Ruiwen Zhou, Xiangru Jian, Victor Zhong Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.32993v1 Multi-step agents are trained on flat action streams: SFT and RLVR weight every token uniformly and ignore the sub-procedures that recur across tasks, the hierarchy that lets humans plan top-down from reusable routines. This structure sits unused, and flat training uses each scarce trajectory less fully than its content allows. Recent agents do use that structure, but only as LLM-written skills in context, never in the weights, so their gains do not generalize beyond retrieval. We instead recover this hierarchy from the data itself and train on it, with no LLM calls. Following text tokenizers, which build a vocabulary by counting alone, we score action spans by reusability and merge canonicalized actions into a reusable eXperience tree (X-Tree). Each X-Tree node captures how a frequent and success-bearing skill is composed from sub-skills, guiding efficient generalization. We integrate X-Tree into three training settings: offline RL, with each node as a training instance; online RLVR, with an adaptive skill bonus; and on-policy self-distillation, with X-Tree as the self-teacher's privileged context. Across WebArena, ScienceWorld, and WebShop at three model scales, X-Tree improves over standard recipes at matched data and budget by up to 4.5% SR on WebArena, 5.8% SR on ScienceWorld and 4.1% success on WebShop. Matched analyses attribute the gains to the X-Tree structure and the three integrations. --- ### What Should Data Teach? Moving Bottlenecks Across Circuit, Store, and Use Authors: Yixiao Chen, Ke Cheng, Jiangtao Guan, Shuo Huang, Yue Liu, Jun Zhang, Yuhong Liu, Jie Jiang Published: 2026-09-26 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.32991v1 What should data teach a language model at a particular point in training? A circuit view reveals three distinct bottlenecks: forming a computation, making its required content available, and selecting among available routes. A shared diagnosis-to-data principle connects them: localize the missing operation, preserve its causal relation, vary shortcut-bearing context, and re-audit the residual. Formation-sensitive selection and prerequisite ordering accelerate a binding-matching-transport path; a brief early prefix from the same training multiset retains a validation advantage through 100B tokens. Availability counterfactuals then distinguish writing content from invoking available memory, while paired supervision and context-opportunity ranking improve matched route decisions and long-context answer likelihood. A continuous 350M-model experiment connects the three interventions on the same facts: early circuit training improves subsequent learning, and the complete sequence outperforms stage-replacement controls on facts withheld from Use teaching. Independent query surfaces and opposed-source decisions expose conditional arbitration as the remaining frontier. Together, these results show why a change in the limiting operation calls for a change in supervision, not merely a new ranking of difficult examples. --- ### Certified Long-Horizon Code Agent Evolution via Validation-Gated Skill Optimization Authors: Yifan Wang, Hao Cheng, Xiaomin Li, Yuexing Hao, Hemanth Neelgund Ramesh, Dongwon Jung, Hao Tang, Keru Wang, Chenliang Zhou, Qianhui Wu, Wenlin Yao, Ananth Grama et al. Published: 2026-09-26 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.32990v1 Long horizon agent self-evolution without model weight updates is essential for enabling deployed agents to accumulate reusable skills and improve over time. Prior self-evolution work has focused primarily on short-horizon tasks, while repository-level software engineering remains unexplored despite being an ideal testbed for long-horizon adaptation. In this setting, agents are required to solve streams of sequential tasks, navigate complex dependencies with evolving repositories and persistently store and reuse experience. Text-based skill optimization offers an efficient, non-parametric approach for such adaptation. However, existing methods often suffer from unstable updates, performance drawdown, and agent collapse over extended deployments. In this paper, we formalize the concept of in-context self-evolution and introduce VALVE, a validated-gated framework for long-horizon skill optimization. We establish finite convergence, provide theoretical guarantees for future-task gain and drawdown, and derive the validation and evaluation holdout sizes required for a prescribed tolerance, with leading-order scaling Empirically, our pipeline, VALVE achieves stable self-improvement over evolution horizon spanning more than 1,000 SWE tasks, with average final and peak gains of $14.9$ and $16.5$ points across three frontier models (GPT-5.5, Claude-4.6 and MiniMax-M2.7). The validation gate reduces average drawdown by 75% and produces an 11x more compact skill bank than ungated evolution. We further present extensive ablations identifying the design choices most critical to long-horizon skill evolution. ---