# Research Papers — 2026-09 · bundle page 3 of 266 > Full-text markdown of papers #101–150 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/3 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/2.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/4.md) ### Efficient Active Auditing of Multi-Group Fairness with Bias Probes Authors: Ayoub Ajarra, Debabrota Basu Published: 2026-09-30 Categories: cs.LG, cs.AI, cs.CY, stat.AP, stat.ML arXiv: https://arxiv.org/abs/2609.40034v1 Over the past decade, Machine Learning (ML) has been trained under dual objectives: minimizing prediction error via Empirical Risk Minimization (ERM) while controlling unfairness bias. In practice, however, fairness-aware training often yields limited improvements over standard ERM, making reliable post hoc auditing essential. Existing auditing approaches for black-box models either rely on model reconstruction --exposing systems to extraction attacks-- or directly estimate fairness metrics, offering limited insight into which regions of the data distribution drive bias. More fundamentally, property-specific auditing --aimed at extracting only targeted fairness information without reconstructing the model-- remains poorly understood. In this work, we introduce the bias probe framework, which enables targeted and adaptive querying to reveal bias structure while preserving model confidentiality. Building on this framework, we propose ALeBi, an active auditor that learns such probes to efficiently estimate multi-group fairness metrics. We establish novel sample complexity guarantees governed by a property-specific complexity measure, resolving a previously posed open question, and extend our analysis to adversarial settings where the model owner may strategically obscure bias. Our results uncover a fundamental trade-off between model confidentiality and reliable auditing, and show that property-specific probing enables both accurate estimation and interpretable identification of high and low-bias regions. Extensive experiments support our theoretical findings and demonstrate the practical effectiveness of our approach. --- ### WARP: A Unified Benchmark for Invisible Image Watermarking -- Robustness and Protection Against Attacks Authors: Khaled Abud, Aleksey Yakushev, Aleksandr Akimenkov, Irina Serzhenko, Kirill Aistov, Egor Kovalev, Dmitry Obydenkov, Sergey Lavrushkin, Anastasia Antsiferova, Dmitriy Vatolin, Yury Markin, Kirill Lukianov Published: 2026-09-30 Categories: cs.CV, cs.AI, cs.MM arXiv: https://arxiv.org/abs/2609.40031v1 Digital image watermarking is increasingly critical in media contexts, as emerging regulations and industry practices require marking AI-generated content and ensuring traceable sources to prevent manipulation or misuse. Recent advances in invisible watermarking methods highlight the need to update existing benchmarking practices to reflect current techniques and evaluation criteria. We address this by introducing WARP -- a unified framework and benchmark for evaluating the robustness of invisible watermarks. WARP incorporates 32 recent classical, deep, and generative watermarking methods, as well as 34 different erasing techniques, ranging from traditional distortions to more sophisticated adversarial, purification, and re-embedding attacks. It provides standardized, reproducible, and easily scalable protocols for evaluating perceptual quality, watermark readability, and attack resilience. Using WARP, we extensively evaluate current invisible watermarking techniques, collecting the largest robustness benchmark in the field. Results identify the most robust approaches under both distortion and adversarial conditions, and reveal consistent relationships between watermarking methods and the attack strategies most effective against them. Our experiments also highlight that some of the watermarking methods considered are highly vulnerable to reembedding, even if they are robust to standard distortions. The code is made available at https://github.com/ispras/wibe. --- ### Fenchel Tilting: Weighted Correction for Efficient Finetuning of Generative Models Authors: Maksim Bobrin, Maksim Zhdanov, Dmitry Dylov Published: 2026-09-30 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.40030v1 Adapting a pretrained generative model to an arbitrary preference expressed as a utility function underlies reward alignment, guided design, and constraint satisfaction, enabling diverse applications. Existing fine-tuning methods trade off generality against computational cost: they either restrict the family class of supported preferences to keep optimization simple or preserve generality at the expense of efficiency. We introduce Fenchel Tilt Flow Control (FTFC), which decouples utility optimization from generative-model fitting. FTFC first optimizes for a target distribution by jointly fitting an effective reward and density-ratio weights on pretrained samples. Method combines the utility's variational structure with Fenchel duality, supporting general $f$-divergence penalties that determine how rewards are transformed into an distribution-correction weights. These weights are then frozen and used to modify a diffusion or flow model in a single stage of importance-weighted denoising or flow matching, without differentiating through sampling trajectories. We establish exact duality for concave utilities under suitable conditions and show that weighted fitting reproduces the optimal target distribution for a given utility. Across image and molecule generation benchmarks, FTFC improves over baselines on diverse preference functions, while also being up to $20\times$ more efficient. roposed method enables adaptation beyond expected-reward maximization without complex optimization, while preserving robustness for more general class of the utility functions compared to baselines. --- ### Who Verifies the Graph? Misspecification Attacks on Causal Action Verification for Language Agents Authors: Fabio Rovai Published: 2026-09-30 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.40027v1 Causal action verifiers gate an agent's state-changing tool calls by checking whether each proposed intervention is identifiable against a committed action-state graph, and they issue a certificate that carries the identification argument and a one-sided lower confidence bound. One such verifier, CIVeX, reports zero false executions on a confounded tool-use benchmark. We red-team it by corrupting only the committed graph. Omitting a single bidirected edge takes it from zero false executions to 15.3% at the benchmark's published confounding strength, with 91% of its executions harmful and utility falling from +2.27 to +0.35. Reversing one arrowhead, so that a mediator is committed as a confounder, gives 48.9% false executions and no correct ones. Every one of these actions carries an internally valid certificate. An attestation step that tests each observationally certified execution against a bounded randomised sample detected both attacks, with 2 false alarms in 555 executions on a truthful graph; refusing what fails the test, or cannot be tested, gave zero false executions in every setting we measured. It does not restore beneficial execution: at the published strength 97.1% of beneficial actions are still never executed, because the same misspecification rejects them before attestation runs. Those rejections carry certificates too, and auditing them works, but its cost scales with the number of rejections rather than the number of executions. Recovering safety costs 127 experiments per 1,050 actions; recovering the lost value costs 614 more, at which point the audited verifier makes the honest graph's decisions on every instance and spends exactly its experiment budget. An audit that inspects only executions protects against wrongful action. Wrongful inaction has to be paid for separately. --- ### Amortized Bayesian Inference on Multilevel Models of Arbitrary Structure Authors: Daniel Habermann, Andreas Bulling, Stefan T. Radev, Paul-Christian Bürkner Published: 2026-09-30 Categories: stat.ML, cs.LG, stat.CO arXiv: https://arxiv.org/abs/2609.40024v1 We develop a general method for amortized Bayesian inference on multilevel models of arbitrary structure. Given a generative model specified as a directed acyclic graph, our method automatically derives valid factorizations of the joint posterior and matching neural network architectures. The key steps, graph expansion and graph inversion, yield an inverse graph that determines how inference networks are stacked and conditioned, producing factorizations that amortize over the number of groups and the number of observations within each group. Unlike approaches that simplify the dependency structure to speed up learning or inference, our method preserves all conditional independence and exchangeability assumptions of the generative model. Across three case studies, it closely matches gold-standard samplers on models with more than 6,500 parameters while reducing inference to a near-instant forward pass once trained. --- ### Component-Weighted Centroid Search for Exact Incremental BPE Authors: Harshit Verma, Rex Ying Published: 2026-09-30 Categories: cs.DS, cs.LG arXiv: https://arxiv.org/abs/2609.40016v1 Exact incremental BPE maintains the canonical tokenization state after every appended byte. The recent algorithm of Jiang and Gong (2026) does this in $O(\log^2 t)$ worst-case time, where $t$ is the maximum canonical token length. Its centroid search visits $O(\log t)$ components and can pay another $O(\log t)$ for ordered point location at each one. Within Jiang and Gong's normalized/proper merge-stage model, we change only that local search. Each interval is weighted by the size of the recursive component it selects, so a move from size $m$ to size $m'$ costs $O(1+\log(m/m'))$. These charges telescope, giving $O(\log t)$ time per append and $O(n\log t)$ over an $n$-byte stream, with the same BPE semantics and asymptotic space. We also construct a normalized proper BPE family over a fixed alphabet where count-balanced search uses $Θ(\log^2 t)$ probes on a reachable update, while the weighted search uses $Θ(\log t)$. A Rust implementation matches the predicted probe counts on every tested instance. On ordinary vocabularies the queried degrees are small, however, and the improvement is a worst-case guarantee rather than an average-speed result. --- ### Can We Anticipate Violence? Multimodal Learning from Pre-Incident Behavioral Cues Authors: Sindhuja Penchala, Mohammed Yusuf Mujawar, Noorbakhsh Amiri Golilarz, Sudip Mittal, Shahram Rahimi Published: 2026-09-30 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.40014v1 Detecting violence after it begins is important from recognizing behavioral cues that appear immediately beforehand. This work studies short-horizon pre-incident risk recognition from multimodal video signals. We construct a binary Normal-versus-Risky setting from temporally annotated XD-Violence clips, using 443 samples with source-level separation across training, validation, and test sets. Each sample consists of a variable-length pre-incident clip, with its duration determined by the observable behavioral context preceding the incident. The inci- dent itself is excluded from all input clips. We evaluate three complementary information sources: facial-region appearance, temporally aligned audio, and body-motion features derived from tracked keypoints. Controlled ablations are performed with Swin-Tiny, ViT-Tiny, and DeiT-Tiny to measure the contribution of each modality under the same split. Results show that combining all modalities is more effective than using any other combination alone. The best configuration, Deit-Tiny with audio, facial appearance, and motion, achieves 91.21% accuracy, 88.96% balanced accuracy, 93.65% F1-score, and 96.38% ROC-AUC on the held-out test set. These results suggest that complementary appearance, acoustic, and kinematic cues provide useful evidence for recognizing elevated pre-incident risk. --- ### PINNing the pion: conformal deep learning for $F_π(s)$ and the $(g-2)_μ$ hadronic contribution Authors: Mayank Goel, Subhadip Mitra, Monalisa Patra Published: 2026-09-30 Categories: hep-ph, cs.LG, hep-ex arXiv: https://arxiv.org/abs/2609.40008v1 Extracting the pion electromagnetic form factor $F_π(s)$ through phenomenological curve-fitting models introduces model dependence, unphysical artefacts, and kinematic inconsistencies. We introduce a Physics-Informed Neural Network (PINN) embedded in a conformal $z$-plane that constructs $F_π(s)$ directly from first principles across spacelike and timelike domains: charge normalisation and Schwarz reflection are enforced by construction, while Cauchy-Riemann analyticity, dispersion relations, Watson's theorem, and perturbative QCD asymptotics enter through the loss functional. Thus, the fundamental S-matrix principles dictate the form factor's behaviour while data act as constraints. Mapping the cut complex plane onto the unit disk bounds the Hessian norm and prevents Neural Tangent Kernel spectral starvation, two known failure modes of deep-learning optimisation. Besides $e^+e^-$ scattering data, we also incorporate $τ$-decay data through a switch that isolates the pure isovector form factor natively, bypassing model-dependent isospin-breaking pre-corrections. The network organically yields an interior zero-free form factor, while the framework tests experimental tensions around the $ρ(770)$ peak against analyticity and dispersion constraints. We obtain model-independent estimates of the pion charge radius, $\langle r_π^2 \rangle = 0.435 \pm 0.008_{\text{stat}} \pm 0.007_{\text{cali}}$ fm$^2$, the second-sheet pole parameters, $m_ρ^{\text{pole}} = 761.72\pm 1.04$ MeV and $Γ_ρ^{\text{pole}} = 135.99 \pm 1.20$ MeV, and the two-pion contribution to the muon anomalous magnetic moment, $a_μ^{ππ} = (506.48 \pm 2.02_{\text{stat}} \pm 1.70_{\text{cali}}) \times 10^{-10}$. --- ### Multi-Link Safety Filtering for VLA Policies Around Moving Hazards Authors: Yatharth Agarwal, Vijay Raghunathan Published: 2026-09-30 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2609.40007v1 A vision-language-action (VLA) policy can finish a manipulation task while knocking over objects unrelated to it, so task success alone does not show that the policy is safe to deploy in clutter. We study how to keep a pretrained VLA policy clear of such hazards at run time without retraining it, which requires guarding more of the arm than the end effector, following the hazard as it moves, and sharing onboard compute with the policy. Our training-free shield covers the gripper, wrist, and forearm with five ellipsoids and filters every commanded motion through one barrier program against a keep-out ellipsoid fitted from RGB-D perception at reset. Sparse optical flow then carries that ellipsoid's center along with the hazard, with no repeated detection or refitting. Over six simulated hazard-motion conditions, the shield lowers collision from $65.62\%$ to $27.27\%$ and raises safe-success, task completion without collision, from $29.35\%$ to $50.43\%$. Ablations show that guarding the arm links protects beyond end-effector shielding, and that tracking recovers most of the protection lost when the hazard estimate is frozen at reset. On heterogeneous edge hardware, the five-ellipsoid barrier runs on the CPU in $2.2$~ms at the 99th percentile, and trimming the vision--language prefix and taking fewer flow-matching steps shortens each $π_{0.5}$ policy call on the integrated GPU from $343$ to $177.3$~ms. On a physical SO-101 arm across four tasks, the arm touched the hazard in 3 of 16 shielded episodes versus 11 of 16 unshielded ones. Project page: https://yathag.github.io/multilink-safety-filter/ --- ### DashVMC: Real-Time Discrete World Model Control in Geometry Dash Authors: Florent Tariolle, Florian Yger Published: 2026-09-30 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.40003v1 World-model agents are usually evaluated in simulators that can wait for the policy; live games impose the opposite constraint, requiring capture, prediction, and action before the next frame. We present DashVMC, which learns a compact, action-conditioned world model from approximately two hours of recorded Geometry Dash gameplay. To test whether the learned dynamics are actionable, a controller is initialized by behavioural cloning (BC) and refined with Proximal Policy Optimization (PPO) entirely in frozen-model rollouts, without further interaction with the live game. Across three controller seeds, the refined policies survive longer than their BC initializations on all three official levels and a held-out community layout. At deployment, the baseline skips visual generation and sustains a 60-Hz capture-to-action loop on a consumer GPU. Action-conditioned continuations and rollout diagnostics show that the model remains useful for control despite imperfect long-horizon fidelity. --- ### Learning to Explain While Planning: Rule-Aligned Diffusion Planning for Autonomous Driving Authors: Jiaxi Ye Published: 2026-09-30 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.39995v1 Diffusion planners exhibit strong capabilities in generating multimodal trajectories. However, existing methods primarily rely on expert demonstrations to fit trajectory distributions, learning statistical correlations among scenes, behaviors, and trajectories without explicitly modeling driving rules. In long-tail scenarios where expert data are scarce, the lack of behaviors to imitate may lead to trajectories that violate safety or compliance requirements. Moreover, their generation process lacks rule-level explanations, making it difficult to determine which rules drive trajectory adjustments, when they take effect, and how strongly they act, thereby limiting failure diagnosis, safety validation, and targeted improvement. To address these limitations, we propose the Rule-Aligned Diffusion Planner (RADP), which incorporates differentiable driving rules into the diffusion objective during training, turning rule knowledge into intrinsic behavioral principles beyond finite demonstrations. We further introduce Rule-Pressure Attribution (RPA), which constructs supervision signals from gradients of rule losses with respect to predicted trajectories and employs a lightweight attribution head to estimate the optimization pressure exerted by each rule online. To assess the closed-loop behavioral relevance of these attributions, we propose a temporal risk-alignment protocol that evaluates whether current rule pressures reflect corresponding risks during subsequent closed-loop execution. Experiments on nuPlan show that RADP improves closed-loop planning in challenging safety-critical scenarios, while RPA exhibits consistent temporal alignment with subsequent rule-specific risks, validating both intrinsic rule learning and rule-level interpretability. --- ### What Can Component-Replacement Evidence Establish? A Critical Scoping Review of Local Decisions in LLM Agents Authors: Shuyang Zhang, Jianshuo Chang Published: 2026-09-30 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.39989v1 Background. A component replacement in a language-model agent changes an execution trajectory, potentially altering later observations, resource use, and recovery opportunities. Different evidence is needed to assess its task-level benefit and the contribution of local decision quality. Methods. This critical scoping review maps 348 studies and examines 90 comparison records: 88 from 40 included studies and two from supplementary studies. Eight purposively selected cases structure the synthesis around the replaced decision, executed conditions, measurement comparability, controls, and remaining explanations. Results. Of 222 studies reporting local decision metrics, 142 also report measured task endpoints and 49 report proxies. These counts identify studies that report both types of measurement, without establishing that the measurements come from matched comparisons. Outcome Monitors reports a package-level completion gain whose attribution to detector quality remains limited; First-chunk selection reports a local improvement assessed against an offline proxy endpoint; Evidence-Carrying Termination reports fewer premature unsupported terminations and completion non-inferiority, without establishing completion superiority. Cross-case analysis identifies three candidate mechanisms involving recovery and disruption, intervention timing, and downstream use. Attribution and deployment depend on the comparison controls, label definitions, and information available to the controller. Conclusions. The review distinguishes the task-level benefit of a component replacement from the contribution of local decision quality and derives eight claim-specific reporting items. Neither online execution nor simultaneous gains in local and task metrics alone establish that better local decisions explain the task-level gain. --- ### Mid-Harness: Scaling Actions Between Model and Harness for Terminal Agents Authors: Minki Kang, Ryo Hachiuma, Shaokun Zhang, Subhashree Radhakrishnan, Yonggan Fu, Jindong Jiang, Mingjie Liu, Ehsan Hosseini-Asl, Yi Dong, Yu-Chiang Frank Wang, Byung-Kwan Lee Published: 2026-09-30 Categories: cs.CL, cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.39982v1 Terminal agents act through stochastic model generations, yet the ability to generate a useful action does not ensure its reliable execution. A poor command (e.g., wrong package install) can change the environment in ways that hinder subsequent progress, even when the model could generate a better alternative. We investigate whether allocating test-time compute at the model-harness boundary can improve action reliability and trajectory success, and what makes this allocation effective. To study these questions, we introduce Mid-Harness, which samples and verifies candidate actions before forwarding one for execution, while keeping the generator and harness unchanged. With a TMAX-9B generator, more action sampling yields little benefit under weak verification, whereas a capable verifier can exploit useful alternatives from the same generator. On TerminalBench-Lite, a GPT-5.6 Sol verifier raises Pass@1 from 50.00% for the base agent to 68.03% with 8 sampled actions. When the same TMAX-9B model serves as the verifier, pairwise verification performs best among the evaluated verification mechanisms. Distilling responses from the stronger verifier into TMAX-9B further improves Pass@1, while leaving the action generator unchanged. With TMAX-9B on TerminalBench-Lite, combining action and trajectory scaling reaches higher success at lower estimated token cost than generating more trajectories alone. Mid-Harness also improves performance across additional models, benchmarks, and harnesses. These findings identify action scaling as a promising target for test-time compute scaling in terminal agents. --- ### Richard: Voice-First Mobile Interaction for Persistent Tasks Authors: Xinyang Chen Published: 2026-09-30 Categories: cs.HC, cs.AI arXiv: https://arxiv.org/abs/2609.39976v1 Mobile terminals need to provide application and network services while supporting users' control over their attention. We explore voice-first interaction organized around requests and delegated tasks, allowing users to leave a conversation and later inspect, revise, and retrieve the work. We present Richard, a system prototype that manages voice sessions, task execution, and result delivery separately, linking them through persistent request records. Conversation and task views provide visual feedback, while the backend coordinates immediate responses, dedicated service operations, and agent tasks. Request revisions, execution states, and notifications remain associated with the relevant task. We examine this design through Android functional records, controlled lifecycle verification, and execution records of a real programming request. Controlled verification reproduces revision, execution after confirmation, and result retention; deployed-service records show backend progress and failure feedback after client disconnection. These observations inform the design of task continuity, user control, and service integration in mobile voice interaction, providing an implementation basis for personal computing devices that accommodate intermittent user participation. --- ### Overview of BioASQ 2026: The fourteenth BioASQ Challenge on Large-Scale Biomedical Semantic Indexing and Question Answering Authors: Anastasios Nentidis, Georgios Katsimpras, Anastasia Krithara, Martin Krallinger, Miguel Rodríguez-Ortega, Eduard Rodriguez-López, Natalia Loukachevitch, Igor Rozhkov, Elena Tutubalina, Dimitris Dimitriadis, Vasiliki Patsiou, Grigorios Tsoumakas et al. Published: 2026-09-30 Categories: cs.CL, cs.AI, cs.IR arXiv: https://arxiv.org/abs/2609.39975v1 This paper presents an overview of the fourteenth edition of the BioASQ challenge, organized in the context of the Conference and Labs of the Evaluation Forum (CLEF) 2026. BioASQ is an international challenge series that supports progress in biomedical language processing tasks ranging from semantic indexing and information extraction to question answering and summarization. In 2026, BioASQ included six shared tasks: a) Task 14b on biomedical semantic question answering. b) Task Synergy14 on question answering for developing biomedical top- ics. c) Task MultiClinSum-2 on multilingual clinical summarization. d) Task BioNNE-R on extracting relations between nested named entities in Russian and English. e) Task ELCardioCC on clinical coding in cardiology. f) Task GutBrainIE on gut-brain interplay information extrac- tion. Across these six tasks, 87 distinct teams participated, submitting more than 1000 runs overall. As in previous editions, several submissions reached competitive performance, reflecting the continued progress of state-of-the-art methods across biomedical language processing tasks. --- ### UBTree: Parallel Tree Drafting via Unigram and Bigram Models for Speculative Decoding Authors: Chumeng Liang, Linxuan Wang, Xinyu Peng, Huabin Liu, Yuxin Chen, Ge Liu, Guang Lin, Qifan Song, Jianguo Li Published: 2026-09-30 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.39972v1 Speculative decoding accelerates language model inference by verifying multiple draft tokens in a single target-model pass. Recent parallel drafters have achieved breakthrough performance in frontier production models, but their effectiveness deteriorates as the entropy of target distributions increases due to insufficient draft diversity. To overcome this bottleneck without sacrificing parallelism, we introduce UBTree, a parallel drafter that couples a Unigram proposer with a Bigram selector to construct drafting Trees. The unigram proposer is trained with the standard cross-entropy objective to generate candidate tokens independently for each position, while a lightweight bigram selector predicts transition scores between adjacent candidate pairs. Unlike the proposer, the selector is trained with a renormalized KL objective on high-temperature data. This tree-native training broadens the supervision beyond the greedy path, encouraging plausible alternative branches that improve the chance of accepting additional tokens during tree verification. Across seven standardized benchmarks with Qwen3-4B and Qwen3-8B, UBTree achieves an average speedup of $5.84$--$6.94\times$ over autoregressive decoding and outperforms DARTree in all 28 comparisons. Production-scale evaluation further demonstrates UBTree's advantage over frontier baselines such as DSpark. --- ### When Instructions Retrieve Trajectories: Diagnosing and Mitigating Generalization Failures in VLA Models Authors: Hung-Jen Chen, Yu-Hsun Hou, Yan-Hong Chen, Yan-Fu Chen, Binghua Cai, Min Sun, Chun-Yi Lee Published: 2026-09-30 Categories: cs.RO, cs.LG arXiv: https://arxiv.org/abs/2609.39971v1 Vision-language-action (VLA) models can exceed 90% success on in-distribution tasks and withstand nuisance changes that preserve the required action, yet fail under counterfactual changes that demand a different action. Aggregate robustness scores can therefore conceal a more specific failure, in which a policy responds to both language and vision yet does not combine them to select the action the task requires. We call this failure instruction-action binding. Instructions cue familiar trajectory families, and visual feedback adjusts their execution. Behavioral analyses of fine-tuned $π_{0.5}$ and GR00T-N1.7 policies reveal that failed rollouts often retain the source behavior or switch to another demonstrated task. These switches show that language is not simply ignored. Readouts and interventions connect these choices to task-conditioned internal states. Our analysis of the imitation objective shows how narrow conditional action support can leave grounded and instruction-keyed solutions indistinguishable on the demonstrations. This motivates Equivariant Counterfactual Training (ECT), which acts at two levels. ECT data supply valid demonstrations in which the same instruction requires different actions in distinguishable scenes, while the ECT loss trains each demonstration with its counterpart in the same update. In a controlled LIBERO-PRO comparison, full ECT raises $π_{0.5}$'s mean position-swap success from 36% to 59%. On CALVIN, where counterparts already occur in the original data, the ECT loss improves five-task completion without new demonstrations. On a real UR5e under a fixed demonstration budget, full ECT raises unseen-position success from 8% to 88%. --- ### TACTIC: Temporal and Context-Aware LLM Tactical Planning for Roadside LiDAR Attacks Authors: Yiming Gao, Shaocheng Luo Published: 2026-09-30 Categories: cs.RO, cs.AI, eess.SY arXiv: https://arxiv.org/abs/2609.39969v1 Physical LiDAR attacks are often evaluated using fixed primitives and manually selected parameters, despite their strong dependence on surrounding traffic. We present TACTIC, a scene-aware framework that uses a multimodal large language model (MLLM) to coordinate state-adaptive roadside LiDAR attacks. Under a gray-box threat model, TACTIC relies only on an attacker-operated roadside perception stack, without accessing the victim LiDAR's native point clouds or internal processing. Local perception provides metric vehicle states, while the MLLM combines these measurements with roadside imagery to infer relational traffic context and construct a semantic scene graph. Based on this representation, TACTIC selects and configures two complementary primitives: \emph{push-away}, which shifts the perceived range of a lead vehicle, and \emph{phantom-obstacle braking}, which triggers emergency braking through obstacle injection. Measured traffic states and empirically calibrated constraints ground the generated tactics in physically feasible operating regions. To accommodate MLLM latency, TACTIC overlaps reasoning and execution asynchronously while high-rate local perception detects scene changes and triggers replanning. Across 280 randomized CARLA trials, the full policy achieves a 100% collision rate, compared with 35% for a fixed rule, 60% for random selection, and 75% for a restricted LLM using mode selection with default parameters. Joint physical-and-image input achieves 100% success, versus 65% with physical measurements alone and 75% with imagery alone, while asynchronous $Δ$ refresh reduces scene-mutation response from 7.4 s to 2.0 s. These results show that scene-dependent tactical planning can expose context-sensitive LiDAR failure modes that fixed attack policies may miss. --- ### What Limits Recursive Reasoning Models: Optimization, Architecture and Test-Time Scaling Authors: Yuliana Shakhvalieva, Dmitrii Kharchev, Viacheslav Bezrukov, Inessa Fedorova, Dmitry Bocharov, Ivan Oseledets, Valerii Ternovskii Published: 2026-09-30 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.39967v1 Recursive reasoning models apply a small shared Transformer block many times to refine a latent state. This gives them large effective depth with few parameters and makes them strong on algorithmic tasks. Such compact solvers are natural candidates for tools that an LLM can call on narrow algorithmic subproblems. However, existing models such as HRM, TRM and URM differ in architecture, gradient propagation and training procedure simultaneously. This makes it hard to tell what drives their performance, and their optimization is still poorly understood and often unstable. In this work we address both of these gaps. First, we study these questions under a unified experimental pipeline spanning six algorithmic domains. Individual controlled ablations are performed on representative domains, while the resulting recipe is evaluated across the full suite. The study reveals a surprisingly simple recipe for stable and generalizable recursive reasoning: an intermediate gradient horizon, large physical batches and controlled updates of the recurrent state. An explicit hierarchical architecture is not needed. Second, we combine these findings into a stable 13.6M-parameter model that achieves the strongest overall performance among the evaluated recursive baselines, with particularly large gains on out-of-distribution generalization. It raises Arithmetic OOD accuracy to 71.2%, from 36.2% for the strongest baseline, while reaching 98.41% on Sudoku and 59.5% pass@2 on ARC-AGI-1. Our results show that, within the recursive architectures studied here, performance depends strongly on how recurrence is optimized and stabilized. More broadly, it shows how AI systems can be improved by optimizing their components one at a time. --- ### AIMS: An Agentic AI Framework for Sim-to-Real Multi-Modal ISAC Authors: Yijie Bian, Kai Zhang, Wei Guo, Zixin Wang, Shenghui Song, Jun Zhang, Khaled B. Letaief Published: 2026-09-30 Categories: cs.AI, cs.MA arXiv: https://arxiv.org/abs/2609.39964v1 Multi-modal integrated sensing and communication (ISAC) enables environmental perception and reliable connectivity for intelligent wireless networks. Data-driven multi-modal ISAC models depend heavily on annotated real-world data to learn relationships across sensing and wireless observations, thereby constraining scalable deployment. Although synthetic data generation reduces the burden, adapting existing simulation pipelines to a target deployment requires consistent scene, sensing, wireless, and learning configurations, while mismatches among these coupled components impair sim-to-real transferability. To address the challenge, we propose an agentic artificial intelligence (AI) framework for sim-to-real multi-modal ISAC, named AIMS. Given a natural-language deployment request specifying the target task, deployment conditions, and real-data budget, AIMS derives a deployment-specific sim-to-real configuration and coordinates its execution to produce a deployment-specific task model. A two-agent architecture coordinates scene construction with task learning. A scene construction agent generates geographically grounded, synchronized sensing and wireless records from shared physical states, while a scene understanding agent configures task-relevant modalities and mixture-of-experts (MoE) learning for zero-shot inference or few-shot adaptation. Structured domain knowledge guides dependency-aware planning, while validation evidence supports feedback-driven revision of affected decisions. Experiments on the real-world DeepSense~6G dataset demonstrate improved vehicle detection and beam prediction over the considered simulation and fusion baselines. A separate orchestration benchmark evaluates task interpretation, dependency reasoning, and feedback-driven replanning across diverse deployment requests, showing improved plan correctness with structured domain knowledge and validation feedback. --- ### Reconstructing the Dynamic World: A Representation-Centric View of 4D Scene Reconstruction Authors: Ziren Gong, Guo Chen, Yongjia Li, Yihua Shao, Fabio Tosi, Stefano Mattoccia, Matteo Poggi, Hao Tang, Fei Ma, Shuyan Li, Ziyang Yan, Nicu Sebe et al. Published: 2026-09-30 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.39960v1 4D scene reconstruction aims to recover the evolving geometry, appearance, and motion of dynamic environments from visual observations. Despite substantial progress in neural scene representations, reconstructing dynamic scenes remains challenging due to non-rigid motion, occlusions, temporal inconsistencies, and the trade-offs between reconstruction fidelity and computational efficiency. Recent advances in Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS) have introduced diverse approaches to representing and reconstructing dynamic scenes, yet their relationships, underlying design choices, and evaluation protocols remain fragmented. In this paper, we present a unified perspective on 4D scene reconstruction, organizing existing methods around their scene representations, temporal modeling strategies, reconstruction pipelines, and optimization objectives. Through this framework, we examine how different design choices affect geometric fidelity, appearance consistency, motion representation, and computational efficiency. We further consolidate commonly used datasets and evaluation metrics, identify limitations in current experimental practices, and discuss open challenges in reconstructing complex, dynamic real-world environments. By connecting methodological developments with their underlying assumptions and evaluation evidence, this work provides a structured foundation for understanding existing approaches and identifying future research directions. An evolving collection of relevant papers and resources is available at https://github.com/ZiyangYan/Awesome-4D-Scene-Reconstruction. --- ### Better Deck or Different Judge? Evaluating Agentic Harness Gains in Corporate and Investment Banking Authors: Ludovic Gibert, Matis Despujols, Andre-Louis Rochet Published: 2026-09-30 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.39958v1 Corporate and investment banking teams use presentations to support credit decisions and advise clients on financing and transactions. Producing these decks requires reconciling financial data, tracing sources and turning analysis into a recommendation. We retrospectively study the development of an agentic harness combining a 27B language model, financial calculations, narrative templates and validation checks. LLM judges guide engineering changes and assess the resulting decks, raising the question of whether higher scores reflect better documents or changes in grading. In shared-session text-only grading with template markers removed, five judges score the complete system 20.4 to 33.6 points out of 95 above the same model generating directly from a short prompt. Every judge scores the system higher on all seventeen development deliverables. Margins against direct Opus generation from a short prompt range from -4.7 to +0.8 points. Judges agree on broad progress across development rounds but agree less on final-deck rankings than on pooled scores. Repeated grading also shifts scores on unchanged decks, making small improvements difficult to distinguish from judge variability. --- ### Learning When and How to Intervene: A Hindsight-Distilled Sentinel for Coding Agents Authors: Jiangrui Zhao, Chenglong Li, Meng Zhang, Xiaoting Du Published: 2026-09-30 Categories: cs.SE, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.39957v1 Coding agents solve repository-level tasks through sequences of actions, where a single erroneous action can misdirect subsequent decisions and increase recovery costs. Existing approaches use execution feedback for recovery or specialized checks to block errors, but deciding before execution whether intervention will benefit eventual task completion remains challenging. To address this challenge, we propose HiSentinel, a hindsight-distillation framework that trains lightweight 0.6B and 1.7B sentinels to select pre-execution interventions aimed at improving task completion rather than correcting every imperfect action. A privileged teacher uses recorded execution outcomes as evidence for intervention judgments, which are distilled into a causal student that receives only the pre-action context and proposed action. Beyond identifying whether and when to intervene, the sentinel must also provide actionable feedback that helps the coding agent recover or obtain necessary human input. To support these capabilities, we introduce SWE-Intervene, an action-level dataset constructed from software-engineering trajectories that annotates whether an action should be allowed, autonomously redirected, or paused for human assistance, together with corresponding intervention feedback. Across SWE-bench Verified Mini and Ask or Assume, HiSentinel consistently improves task completion across Sentinel scales and coding-agent families, with gains of up to 14% and 10%, respectively, while maintaining competitive token consumption. These results demonstrate that lightweight pre-execution intervention can effectively prevent error propagation and improve the reliability of autonomous coding agents. --- ### Coverage Before Control: Route-Instruction Grounding and Steering for Controllable Retrosynthesis Authors: Xuemin Chen, Xiaozhuang Song, Xinjian Zhao, Yaoyao Xu, Tianshu Yu Published: 2026-09-30 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.39955v1 Single-step retrosynthesis models are commonly evaluated by their ability to recover recorded reactions. In practice, chemists may need to choose among several precursor sets for the same product, for example to preserve a particular motif. Recovering a recorded answer alone does not establish this ability to follow a preference. Satisfying such requests requires both coverage of relevant alternatives and control over which alternatives are favored. We introduce Route-Instruction Grounding and Steering (RIGS), a two-stage framework for instruction-conditioned retrosynthesis. Stage A trains a language projector, teaching it which alternatives an instruction favors or discourages. Stage B uses the projector learned in Stage A to steer a frozen generative model through lightweight residual adapters. We construct nested one-to-many training supports by pairing each product with increasing numbers of candidate precursor sets. Extensive experiments demonstrate that broader support helps the model generate a wider range of alternatives, and RIGS can learn to guide generation according to instructions. The relationship between coverage and control is consistent across model scales but non-monotone. --- ### Learning to Reason with Compressed Context: Ground-Truth-Free Adaptation of OmniLLMs via Self-Distillation Authors: Jianghao Wang, Ke Meng, Jian Li, Chi Cheng, Longyu Qi, Liyin Liang, Yifeng Qian, Chunbo Lai, Yutian Lin, Zeyu Wang Published: 2026-09-30 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.39953v1 Omni-modal large language models (OmniLLMs) enable unified audio-video understanding, but their long multimodal token sequences make deployment computationally expensive. Token compression reduces this cost, yet aggressive compression often lowers accuracy. Existing works predominantly focus on designing better compression mechanisms; however, adapting the underlying language model to reason effectively over the remaining compressed context remains under-explored. To address this, we propose CAFD (Compressed-Context Adaptation via Full-Context Distillation), a ground-truth-free self-distillation framework that adapts OmniLLMs to fixed compression pipelines without requiring reference answers, rationales, or correctness rewards. CAFD leverages the full-token view of the same multimodal sample as a source of privileged information: a full-context self-teacher provides soft target supervision to a compressed-context student along the student's on-policy trajectory. Evaluated on Qwen2.5-Omni-7B across five audio-video benchmarks, five compression pipelines, and five deployment budgets, CAFD demonstrates consistent gains, improving 120 out of 125 conditions with an average accuracy boost of 1.44 points and recovering 26.9% of the accuracy gap on average. These results demonstrate that the proposed ground-truth-free adaptation offers an effective and practical route to improving the accuracy-efficiency trade-off in deployed OmniLLMs. --- ### LEAP: Learned Block-wise Evidence Retrieval for Long Audio-Video Perception Authors: Juyi Lin, Zhiqiang Lao, Jiali Cui, Lin Zhao, Pu Zhao, Dichang Zhang, Arman Akbari, Yu Qi, Xinru Jiang, Yanzhi Wang, Heather Yu, Liang Peng Published: 2026-09-30 Categories: cs.CL, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2609.39938v1 Hour-scale audio-visual question answering is constrained by a context dilemma: dense whole-recording encoding rapidly exhausts context limits, whereas uniform temporal compression severely dilutes fine-grained acoustic and visual evidence. We introduce LEAP, a framework where the model retrieves its own evidence without placing the whole recording in one context. LEAP divides a recording into fixed-duration blocks, applying a lightweight localization pass to each block to score short candidate windows. The highest-ranked windows are pooled and re-encoded in a single bounded answer pass. Consequently, the answer input and peak context remain independent of the recording duration. By decoupling evidence localization from reasoning, our framework can localize candidate temporal windows over pre-computed transcripts without decoding media frames, while preserving fine-grained visual and non-speech evidence by routing the final answering pass over raw audio-visual streams. LEAP trains both stages: a localization LoRA improves the selected windows, and an answer LoRA improves the answers read from the same windows. The block grid natively supports causal queries, enabling LEAP to support streaming inference without streaming-specific training. Across several AVQA benchmarks, LEAP improves over the Qwen3-Omni-30B-A3B baseline by 4.5-16.8%, and transfers to a second omni-modal backbone, MiniCPM-o 4.5, surpassing its published results by 3.1-13.0%. --- ### Reliability-Aware Checkpoint Selection for Domain Generalization Authors: Jinshi Liu, Jiahao Li, Pan Liu, Yanfeng Li, Rui Qian, Zhao Tong, Yue Sun, Tao Tan Published: 2026-09-30 Categories: cs.LG, cs.CV arXiv: https://arxiv.org/abs/2609.39934v1 Checkpoint selection in domain generalization often relies on source-validation accuracy, yet the selected checkpoint need not provide reliable probabilities on unseen target domains. Source-target distribution shifts can alter accuracy rankings, while accuracy alone does not measure predictive probability quality. We identify an empirical selection opportunity within fixed training trajectories: reselecting among checkpoints with near-optimal source accuracy can improve mean target probability quality with small observed changes in mean target accuracy. We study accuracy-constrained reliability selection (AC), which retains checkpoints within a tolerance of the best source-validation accuracy and ranks them by source reliability. Our reference rule aggregates within-set normalized negative log-likelihood (NLL) and class-wise calibration error (CwECE) using $D_\infty$. AC uses no target data and requires neither additional training nor weight averaging. We evaluate five domain generalization training algorithms on three benchmarks, using PACS to develop the objectives and a 0.5-percentage-point tolerance. In exploratory aggregation comparisons on 360 OfficeHome and TerraIncognita runs, the reference rule reduces mean target soft-bin squared-gap ECE and CwECE by 0.240% and 0.182%, respectively, and NLL by 0.030 relative to Source-Acc. Mean target accuracy changes by +0.213 percentage points. These results identify opportunities for reliability-aware reselection, while the additional benefit of joint over single-objective ranking remains unresolved. --- ### ConflictGuide: AutoResearch Improves When Competing Behaviors Are Made Visible Authors: Binqian Xu, Qiran Zou, Xiangbo Shu, Dianbo Liu Published: 2026-09-30 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.39933v1 When designing machine learning models, desirable properties are often in tension: improving one behavior can impair another, so task progress can depend on alleviating the conflict. LLM-based AutoResearch systems, which iteratively edit model code and retain edits based on scalar task-performance feedback, have largely ignored this trade-off. We find that scalar feedback supports broad exploration early in search, but it does not reveal how edits affect competing behaviors. In matched-budget experiments, introducing competing-behavior feedback as task gains diminish increases the share of proposals that improve both behaviors and sustains progress beyond scalar-only plateaus. Obtaining this feedback for a given model requires identifying its competing behaviors and designing probes to measure them. To make competing-behavior feedback actionable, we introduce ConflictGuide. Its reusable ConflictGuide-Skill combines a literature-grounded taxonomy with model-specific evidence to identify competing behaviors and specify probes for a code agent to implement as metrics. Evolution proceeds in two stages: Stage I explores with task feedback; Stage II uses probe feedback to steer proposals toward conflict alleviation and retains marginal-gain edits only when probes indicate sufficient alleviation. Across five diverse model families, ConflictGuide reduces task and conflict-related errors by up to 28% and 14%, respectively, relative to scalar-only AutoResearch, with gains extending to other code agents. --- ### RoPE at the End of Its Rope? Theory, Diagnosis, and Mitigation of Long-Context Failures Authors: Yuyang Wu, Yufeng Du, Hao Peng Published: 2026-09-30 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.39929v1 Long-context failures of RoPE-based language models can arise from RoPE's intrinsic tradeoff between maintaining stable token preferences and distinguishing nearby positions. Determining which weakness to address, and how, requires a more precise characterization of RoPE's behavior in trained models across context lengths. We address a key limitation of prior theory by allowing unequal query-key scales across RoPE frequencies, which aligns well with practical empirical observations. Our theory makes both vulnerabilities measurable for individual heads and inputs, and quantifies how high-frequency components support positional sensitivity while potentially disrupting semantic stability. We also derive a theoretical context-length bound beyond which, under specified conditions, a fixed attention-score comparison cannot jointly avoid semantic reversal and positional insensitivity. Guided by our fresh theoretical insights, we introduce RoPE Profiler, a lightweight, plug-and-play diagnostic toolkit that augments existing evaluations with zero additional forward passes by reusing cached query and key activations. Reusing activations collected during evaluation, the toolkit incurs little overhead. It supplements standard benchmark scores with two diagnostic scores that reveal semantic and positional weaknesses and help users prioritize which aspect to address. Crucially, our evaluations across 49 long-context task settings reveal a distinct pattern where reasoning tasks predominantly suffer from semantic reversal, whereas retrieval tasks are primarily vulnerable to positional insensitivity. Guided by our theory and diagnostic profiles, targeted high-frequency rescaling achieves immediate gains without additional training, improving task accuracy by up to 20 percentage points on Qwen3-8B and 25 percentage points on Llama-3.1-8B-Instruct. --- ### AdaGEPA: Adaptive Feedback Allocation for Reflective Prompt Optimization Authors: Junyang Chen, Zecheng Wang, Jingbang Chen Published: 2026-09-30 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.39927v1 Prompt optimization improves the performance of language-model systems on downstream tasks by refining their prompts. Classical methods evaluate prompts on task examples and use the resulting feedback to guide prompt revisions through reflection. However, when feedback selection does not account for the prompt's weaknesses, these revisions may improve performance on selected examples without yielding broader task improvements. To address this issue, we propose AdaGEPA, an adaptive feedback-allocation method that uses the prompt's performance and task structure to select examples for the next prompt revision. Our method replaces at most one example in each feedback minibatch to target an identified weakness while preserving the remaining feedback context. Across our main experiments on six downstream benchmarks, AdaGEPA achieves higher mean validation scores than non-adaptive feedback selection under matched rollout budgets. AdaGEPA also finds high-performing prompts earlier across several tasks. In the initial Schema-Guided Dialogue (SGD) study, its half-budget prompts outperform the non-adaptive baseline's full-budget prompts in joint goal accuracy on new dialogues from services seen and unseen during search. Overall, our findings highlight the potential of adaptive feedback allocation to improve both the effectiveness and rollout-budget efficiency of reflective prompt optimization. --- ### Super-Resolving Unseen Hyperspectral Sensors at Any Scale via Spatial Operators Authors: Ji-Xuan He, Guohang Zhuang, Bo Junge, Tingyi Li, Lingchen, Miaomiao Cai, Yanan Qiao, Xiujin Liu, Junfeng Fang Published: 2026-09-30 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.39926v1 Achieving cross-sensor generalization and arbitrary-scale reconstruction with a single model remains challenging in hyperspectral super-resolution (HSR). Although recent methods support arbitrary-scale reconstruction, applying them to new sensors or scales beyond the training range often requires additional data and computation to maintain reconstruction quality. To address these challenges, we propose OmniHSR, which predicts band-shared spatial operators rather than spectral values. Cross-Spectral Mapping (CSM) resamples inputs with any number of bands to fixed reference positions and predicts local operators with Gaussian supports. Continuous Operator-Field Reconstruction (COFR) composes these operators into a continuous field and applies them to all original bands for arbitrary-scale reconstruction. Experiments demonstrate that operator prediction outperforms direct spectral-value prediction on all seven datasets. Trained solely on ARAD with only 0.538M parameters, OmniHSR outperforms all directly transferred baselines on six unseen datasets without target-domain training data or adaptation. Across twelve upsampling factors from $\times2$ to $\times48$, it improves average PSNR on Pavia U and Chikusei by 0.55 dB over the strongest baseline. It also surpasses baselines trained from scratch or adapted on the target sensor and achieves up to $36\times$ faster inference. Our code will be publicly released soon. --- ### CoVisco: Codec-Native Vision Encoder with Native Token Compression for Unified Image-Video Understanding Authors: Yulong Liu, Xiaotian Han, Junyuan Shang, Yuchen Ding, Zhenyu Zhang, Shuohuan Wang, Guibo Zhu, Sirui Han, Dianhai Yu Published: 2026-09-30 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.39924v1 Vision-language models face a fundamental scaling bottleneck: the number of visual tokens grows with both temporal duration and spatial resolution, making long-video understanding expensive for the vision encoder and the language model. Existing methods often compress visual tokens after dense encoding, creating a mismatch between the representation used during training and the compact interface required at deployment. We present CoVisco, a codec-native vision encoder with native token compression for unified image-video understanding. By combining codec-native input support with segmented attention, CoVisco can encode long visual inputs in a single forward pass without forming dense patch-to-patch interactions across all frames. Each temporal segment is equipped with learnable abstract tokens that learn a compact segment-level representation, while fine-grained patch tokens remain available throughout the encoder. Alternating intra-segment and abstract-communication layers preserve video-level context through the abstract-token channel. A lightweight selector further exposes either abstract tokens alone or abstract tokens augmented with a runtime-selected subset of patch tokens, yielding a compact visual interface that reduces the visual context and prefill burden of downstream MLLMs while retaining fine-grained evidence when needed. Pretrained with contrastive objectives on 565M image--text pairs and 6.4M videos, CoVisco shows competitive performance on video-oriented embedding and multimodal understanding benchmarks. In the evaluated four-segment, 64-frame setting, abstract-only inference uses only 400 visual tokens while achieving video-understanding performance close to, and on some benchmarks exceeding, OneVision-Encoder. Selected patch tokens further improve fine-grained video reasoning. Project URL: https://github.com/ernie-research/CoVisco.git --- ### MCD: Causal Distillation of Multimodal In-Context Learning in Large Vision-Language Models Authors: Yanshu Li, Jiaqian Li, Canran Xiao, Xi Xiao, Tianyang Wang, Yongtai Liu Published: 2026-09-30 Categories: cs.CV, cs.CL arXiv: https://arxiv.org/abs/2609.39920v1 Large vision-language models (LVLMs) exhibit strong multimodal in-context learning (ICL) capabilities, yet this ability degrades substantially as model size decreases. Knowledge distillation offers a natural way to bridge this gap, but existing methods primarily align output distributions or hidden representations directly. Such alignment teaches the student what the teacher predicts without revealing which evidence in the complex context causally supports that prediction. Consequently, a student can imitate the teacher's answer while continuing to rely on language priors, prompt structure, or other spurious cues. To address this limitation, we introduce Multimodal Causal Distillation (MCD), a distillation framework that transfers how a strong teacher uses multimodal evidence during ICL. MCD uses structure-preserving token interventions to identify and verify causal evidence, then transfers how the teacher responds when that evidence is retained or removed. This design connects distillation to the causal patterns by which the model uses contextual evidence during multimodal ICL. Experiments across three LVLM families and seven benchmarks show that MCD improves student performance by 7.23 points on average and outperforms vanilla distillation by 4.68 points, while further analyses confirm the generalizability of these gains. --- ### NavHarness: Adaptive Goals for Agentic Vision-Language Navigation Authors: Haoxiang Shi, Zaijing Li, Muhe Ding, Xiang Deng, Yaowei Wang, Liqiang Nie Published: 2026-09-30 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.39915v1 Vision-Language Navigation (VLN) requires embodied agents to generate actions based on instructions and observations. General-purpose multimodal agents offer a promising basis for this task, but selecting plausible local actions does not ensure that execution remains consistent with the intended route, particularly in long-horizon tasks. Moreover, the accumulated interaction history increases the input required for subsequent decisions, resulting in a significant inference overhead. To this end, we introduce \method, an Agentic VLN framework that includes a Goal Agent that sets adaptive goals for local actions, a Verify Agent that dynamically verifies whether a goal has been completed, a Memory Agent for multimodal context compression, and a Visuomotor Agent to execute adaptive goals. Specifically, the Goal Agent formulates adaptive goals based on the instruction, current observation, and execution history. Then the Visuomotor Agent executes navigation actions to achieve each goal, while the Verify Agent uses a goal-specific verification question to dynamically assess whether the observed outcomes satisfy the intended completion condition. Verified goal completion then marks a boundary for the Memory Agent to compress the corresponding multimodal interaction history while preserving information needed for subsequent navigation. We evaluate navigation on R2R-CE and RxR-CE, examine framework variants across three model backbones, and study context evolution during execution. For Real-World evaluation, \method achieves 83.3\% success and 1.51\,m navigation error across eight challenging routes evaluated three times each. --- ### Cluster Attention Neural Operators for Solving Parametric Partial Differential Equations Authors: Ming Zhong, Antonio Colanera, Gianluigi Rozza, Zhenya Yan Published: 2026-09-30 Categories: math-ph, cs.AI, cs.LG, math.DS arXiv: https://arxiv.org/abs/2609.39914v1 Traditional simulations of parametric partial differential equations (PDEs) rely on repetitive computations for each parameter, which makes high-fidelity design impractical. Neural operators address this issue by learning solution operators, accelerating parameter-space mapping by orders of magnitude. Recent Transformer-based neural operators attempt to capture global dependencies, but often at the cost of quadratic attention complexity. Transolver resolves this problem by projecting physical states into a reduced slice space for attention computation. Although fast, this projection sacrifices fine spatial information. Moreover, by operating in this reduced space with shared weights across attention heads, it may constrain the model's flexibility, thereby limiting its capacity to capture complex phenomena. To address these issues, we propose the Cluster Attention Neural Operator (CANO), which reformulates attention via a novel cross-attention mechanism that dynamically clusters queries while preserving full-resolution keys and values. This avoids slice compression loss and removes weight-sharing limits. At the same time, the model remains fast without losing global interactions. Empirically, CANO achieves state-of-the-art performance across canonical PDE benchmarks, covering fluid and solid dynamics (e.g., Navier-Stokes, Airfoil, Plasticity), irregular unstructured geometries (e.g., Pipe Turbulence, Composites), and long-term temporal rollouts. Across solid deformation and turbulent flow benchmarks, CANO achieves lower errors than baselines and exhibits strong geometric adaptability and temporal consistency. --- ### The Concrete-Arbitrary Gap: Kinship Reasoning in LLMs Is Not Indifferent to Presentation Authors: Thomas Pashby Published: 2026-09-30 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.39913v1 We test whether large language models solve formally matched kinship problems equally well when relations are expressed in familiar vocabulary or by explicitly defined nonce predicates. Across 500 paired graphs, concrete accuracy exceeds arbitrary accuracy by 35.6 percentage points in local Qwen3.8-27B, 26.6 in Gemma 4 26B-A4B, 12.0 in Gemma 4 31B, and 5.4 in Qwen3.8-Max. All four paired gaps are statistically resolved. Reasoning budgets and prompt-language interventions can substantially reduce the difference, showing that it is modifiable rather than a fixed incapacity. The minimal conclusion is behavioral: on these tasks, the models' manifested relational competence is not indifferent to presentation. Explicit definitions provide the formal relations but do not make nonce predicates as usable as familiar vocabulary embedded in learned linguistic associations. --- ### TRACE: Trajectory Selection for Parallel Scaling of Search Agents Authors: Qisheng Zhou, Zhen Xiong, Qiaoyu Tan Published: 2026-09-30 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.39912v1 Parallel search may generate a correct answer that final-answer voting fails to select. We formulate this consolidation stage as trajectory selection and introduce TRACE (Trajectory Ranking with Aggregated Cross-Rollout Evidence), a lightweight learned selector that ranks completed trajectories using the search evidence behind their answers. TRACE preserves individual query and evidence occurrences, connects rollouts through shared content or document identity, and propagates information across these relations. Each candidate answer then reads the updated states of its own trajectory, preserving retrieval provenance while incorporating evidence from related rollouts. Trained with answer-level supervision over frozen text embeddings, TRACE returns an existing answer without additional search or autoregressive aggregation. One selector per search setting transfers across rollout policies and agent backbones without agent-specific fine-tuning, improving over voting across six WebQA policies and six long-horizon dataset-backbone combinations at $K=16$. On Qwen2.5-14B Base/SFT WebQA pools, TRACE achieves 45.2/49.2% EM, compared with 43.9/48.0% for the strongest Qwen3-32B generative aggregators. On long-horizon FRAMES, GAIA, and BrowseComp, it reaches 78.6% average accuracy, exceeding majority voting by 3.1 percentage points. On Base WebQA pools, TRACE with only 8 rollouts comes within 0.4 points of majority voting over 64. TRACE also achieves at least $10\times$ higher processing throughput than SolAgg, SummAgg, and AggAgent across all seven WebQA benchmarks. These results show that reusing cross-rollout search evidence provides an effective and efficient alternative to heavyweight generative aggregation for parallel search. Code is available at https://github.com/Jaasssoooonnnnn/TRACE. --- ### Patient-Centered Treatment Planning for Chronic Multimorbidity: A Hierarchical Reinforcement Learning Framework for Preference Modeling Authors: Nafiseh Payani, Soham Das, G. Anthony Wilson, Anahita Khojandi Published: 2026-09-30 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.39911v1 Patient preference, defined as a patient's demonstrated willingness and capacity to adhere to clinical recommendations, is a primary determinant of therapeutic effect yet remains structurally absent from existing computational treatment planning models. We address this gap by presenting patient-centered factored-action hierarchical option-critic (FAHOC), a hierarchical reinforcement learning (HRL) framework that jointly learns high-level options corresponding to therapeutic strategies and factored intra-option policies that decompose the joint action space into disease- and intervention-specific subcomponents, while imposing a cooperation-aware action masking mechanism. This enables structured exploration, improved credit assignment across hierarchy levels, and more interpretable decision pathways, while enforcing patients' preferences. Formal guarantees establish that cooperative patients achieve higher optimal expected health outcomes than non-cooperative patients, and that the factored Q-function approximation error is provably bounded. The framework is evaluated using longitudinal data collected from approximately 50,000 comorbid hypertension and type 2 diabetes mellitus patients from five hospitals in the Southeast U.S. FAHOC achieves a quality-adjusted life year expectancy equivalent improvement of 0.669 (vs -0.133 observed clinician practice), correctly identifies cooperative patients in 95.9% of cases and never violates a patient's preference in held-out test, demonstrating that HRL with explicit preference constraints can support preference-consistent, clinically safe decision-making in multimorbidity management. --- ### DoGBench: Can Agents Meet Expert Standards for User-Facing Documentation? Authors: Frances Liu, Manny Silva, Paige Calvert, Ayu Adiati, Sarah Sanders Published: 2026-09-30 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2609.39909v1 We introduce DoGBENCH (Documentation Generation Benchmark), to our knowledge, the first benchmark for generating and maintaining real user-facing software documentation. It asks whether an agent can produce documentation that experienced technical writers would accept in review. The benchmark contains 292 items from open source projects, including Helm, PostHog, and Mautic. Each item gives the agent a pre-change repository and a trigger, such as a code pull request or a reported documentation gap. The agent must first decide whether the documentation needs an update. For items that need one, the agent must produce an acceptable patch in one attempt. For items that do not need updates, the agent must abstain. Task-specific rubrics, validated with project maintainers, score each patch on accuracy, completeness, reader guidance, placement, and repository conventions. The composite score combines patch quality with correct abstention, and a score of 100 means an agent meets every requirement for the task. Scores should not be interpreted as a percentage of an expert's capability. We evaluated seven agents. The highest-scoring agent reached 47.3 out of 100 on the 117-item held-out split. In a separate audit of 1,267 patches, the most common failure modes were task-completion gaps (45.5%), technical inaccuracies (36.6%), and incomplete conceptual or reference coverage (32.5%). Analysis of the corresponding trajectories identified three key patterns associated with these failures: (1) describing interfaces without examining how readers use them (36.0%), (2) missing decisive evidence and filling the gaps with plausible assumptions (33.1%), and (3) stopping after finding the first plausible documentation surface and leaving other affected pages stale (30.1%). --- ### Understanding Parents' Complex Views of AI for Children's Pretend Play Authors: Sungho Oh, Mohammad Namvarpour, Maxi Heitmayer, Minahil Khalid, Afsaneh Razi Published: 2026-09-30 Categories: cs.HC, cs.AI arXiv: https://arxiv.org/abs/2609.39906v1 AI could support children's pretend play, but it could also direct the play on behalf of children. Whether AI should have roles in children's lives is controversial because its influence on children remains uncertain. We conducted semi-structured interviews with 10 U.S. parents, each with at least one child aged 4-15. During the interview, we described the concept of AI-supported pretend play and provided participants with two boundary-case storyboards. We analyzed the interview data through codebook thematic analysis, using inductive coding and affinity diagramming organized around the research questions, and then used qualitative systems mapping to examine relationships within and across themes. We found that the same characteristics of AI, e.g., ability to assume characters, responsiveness, and adaptability, were seen by parents as potentially useful but also concerning. Parents imagined that AI could make role-based play accessible to all children or help parents participate in family play. However, they opposed the idea of AI for children's play without a clear understanding of how it works and its long-term influence on their children. Parents worried about children's loss of imagination and creativity, emotional attachment to AI, reduced human interaction, inappropriate behavior by AI and/or children, and their inability to manage children's AI use. Parents viewed AI not only as a play tool but also as a social actor and a possible perturbation in the existing family dynamics. The appropriateness of AI and child--AI interactions therefore emerged as a requirement for AI in children's pretend play, in addition to technical safeguards and parental control. We contribute an integrated account of parents' interdependent judgments and emphasize the need for longitudinal research with children and their diverse families. --- ### OSWorld-Science: A Benchmark of Computer Use Agents for Learning and Using Scientific Software Authors: Dingyuan Dai, Heli Qi, Lei Liu, Yinxi Li, Baiding Chen, Zijun Dou, Qingcheng Zeng, Qi Kang, Oliver Sun, Eric Wang, Bo Zhou, Haixin Wang et al. Published: 2026-09-30 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.39903v1 Scientific software presents a demanding test for computer-using agents based on visual language models (VLMs): completing a research workflow requires interpreting specialized interfaces, manipulating scientific objects, and producing verifiable results. We thus introduce OSWorld-Science, a benchmark and evaluation environment that combines scientifically meaningful tasks, artifact-based evaluation, and an efficient agent harness for studying computer use in the scientific domain. The benchmark contains 12 VLMs and 146 high-quality tasks across several scientific domains and software configurations, covering workflows such as molecular drawing and retrosynthesis, pathology image analysis, statistical computing, and physical simulation. Tasks are developed through expert proposals and iterative human--AI co-design, with selection guided by scientific value and difficulty. Task-specific execution-based evaluators inspect application states and generated artifacts, including molecular structures, segmentation masks, plots, and numerical results, and award partial credit for incomplete outcomes. Our special harness integrates model adapters, interaction-loop control, and trajectory logging to support comparisons of models and interaction strategies. Our results show that current state-of-the-art VLMs with a strong harness still face challenges in addressing key questions in the scientific domains. We also analyze the benchmarking results across multi-linguistics, reasoning efforts, context length and other factors and derive several important conclusions and directions to assist future development. Overall, we provide an integrated framework connecting expert-defined scientific goals to verifiable software outcomes, enabling systematic evaluation of both agent capabilities and harness design in scientific workflows. --- ### CodeMimicry: Exploiting Safety Generalization Lag in Large Language Models via Structured Code Completion Authors: Zhen Liang, Hai Huang, Wentao Chen Published: 2026-09-30 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.39902v1 Large language models have achieved remarkable capabilities across diverse domains, yet their safety alignment remains vulnerable to jailbreak attacks. In this work, we identify a previously underexplored failure mode - safety generalization lag - where alignment trained predominantly on natural language fails to transfer to the code domain. We show that this lag induces a code-completion blind spot, allowing malicious intent embedded within syntactically valid code to evade safety mechanisms. To exploit this vulnerability, we propose CodeMimicry, a fully automated black-box jailbreak framework that generates structured, object-oriented code prompts to induce harmful outputs via code completion. Experiments on 8 state-of-the-art commercial LLMs demonstrate that CodeMimicry achieves a 96.25% attack success rate with 1.51 queries on average, significantly outperforming both template-based and optimization-based baselines. Beyond empirical performance, we provide a mechanistic analysis of code-based jailbreaks through latent space representations, including projection onto refusal-related directions and activation steering. This analysis offers an explanation of how CodeMimicry bypasses safety mechanisms in code-related domains. Our findings reveal a weakness in current safety alignment and highlight the need for robust alignments in structured domains such as code. --- ### Do Better Goal Representations Improve Goal-Conditioned Reinforcement Learning? Authors: Syed Nazmus Sakib, Abdul Monaf Chowdhury, Nafiul Haque, Shifat E Arman, Md Mehedi Hasan Published: 2026-09-30 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.39901v1 Goal-conditioned reinforcement learning (GCRL) relies heavily on how target goals are represented to the policy. While recent methods encode goals via temporal distance, occupancy, or controllability, it remains unclear how much downstream performance actually depends on representation quality. We study this in offline GCRL by constructing an exact temporal-distance goal representation in deterministic mazes. We then systematically corrupt its geometric quality while keeping the downstream learner fixed. Across OGBench navigation tasks and two algorithms, large changes in goal-representation quality produce almost no change in performance. However, applying the same interventions to the agent's current state more than doubles success, revealing the state pathway as the true bottleneck. Building on this insight, we show that simple random Fourier positional encodings substantially improve performance on the hardest navigation tasks without map information or objective modifications. Overall, our findings suggest that in state-based offline navigation, improving how the agent's current state is represented matters far more than refining the goal representation. Code will be released soon. --- ### Learning Where to Look: Anatomical Grounding and Guided Attention for Cardiac MRI Vision-Language Models Authors: Bangwei Guo, Xiao Chen, Boris Mailhe, Jia Yao, Yiqing Wang, Ankush Mukherjee, Yikang Liu, Zheyuan Zhang, Hang Yu, Terrence Chen, Shanhui Sun Published: 2026-09-30 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.39899v1 Cardiac magnetic resonance imaging (CMR) enables assessment of cardiac anatomy, ventricular function, and myocardial tissue characteristics. Clinicians interpret these images by identifying cardiac structures and focusing on the regions relevant to each clinical question, motivating anatomically guided vision-language models (VLMs). Yet CMR-specific supervision for anatomical localisation and clinical question answering remains limited. To address this gap, we investigate fine-grained CMR visual question answering through anatomical grounding and guided attention. We construct 128,915 anatomical-grounding and 42,799 clinical QA pairs across short-axis cine, late gadolinium enhancement, and long-axis cine. These datasets support anatomical recognition, localisation, and clinical assessment without requiring paired reports for individual training images. To help the model learn where to look, we introduce Cardiac Anatomy-Routed Attention (CARA), which selects predicted anatomical priors according to the question and guides decoder attention with learned task-specific strengths. Combining anatomical grounding pretraining with CARA yields our model, CARA-VL. Experiments demonstrate CARA-VL's strengths in clinical assessment and regional localisation across CMR imaging settings, with promising generalization to an external clinical cohort. Together, our data and method provide a practical framework for studying and advancing cardiac visual understanding in VLMs. We will release the QA data derived from public datasets upon publication. --- ### Spatial-Temporal Multi-scale Network for Screen Content Video Quality Enhancement Authors: Ziyin Huang, Sik-Ho Tsang, Xinyuan Qin, Yui-Lam Chan, Xueling Zhou, Feiyu Chen Published: 2026-09-30 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.39894v1 Different from natural videos, Screen Content Videos (SCVs) are characterized by abrupt motion, scene switches, and high-frequency details such as text and graphics. Conventional video enhancement methods, which rely heavily on temporal continuity, often suffer from performance degradation when processing SCVs due to the disruption of temporal correlations. To address these challenges, we propose the Spatial-Temporal Multi-scale Network (STM-Net), a novel framework specifically tailored for compressed SCV enhancement. Our approach integrates three complementary components: a Prior-Guided Spatio-Temporal Dispatcher (PG-STD) that routes input into three parallel streams to avoid feature contamination, a Bidirectional Temporal Feature Extraction (BTFE) module that adaptively handles abrupt transitions without explicit detection, and a Cascaded Multi-scale Feature Distillation (CMFD) module that preserves critical high-frequency details. Experimental results demonstrate that STM-Net outperforms state-of-the-art methods in both objective metrics and subjective visual quality, providing a robust solution for screen content artifacts. Code is available at https://github.com/HUANGZiyin1/STM-Net. --- ### Shared Weights, Selected Computations: How Looped Transformers Route What Each Loop Does Authors: Jiaju Wu, Yi Hu, Muhan Zhang Published: 2026-09-30 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.39892v1 Looped Transformers repeatedly apply the same set of Transformer layers, giving them a recurrent architecture for latent computation. Their strong performance on iterative reasoning and length-generalization tasks suggests an appealing explanation: recurrence may provide an inductive bias that lets the model reuse a learned algorithm across loops. However, weight sharing alone does not imply that every loop performs the same operation. This raises a basic question: is each loop actually repeating the same computation, and if not, what routes the shared parameters to different operations? We study this question using graph walks as a test case. In the model's native trajectories, decoded predictions can advance by different numbers of graph steps or remain at a reached target, showing that recurrent progress need not follow a fixed one-loop-one-step pattern. We then show that a frozen loop can be steered toward different transitions by modifying its entering hidden state: a learned linear layer $J$ selects the desired transition without changing the shared Transformer layers. To test how this steering works, we use activation patching and find that attention patterns can recover its effects and switch the selected transition. Across five matched pairs of graph models, changing intermediate supervision during backbone training changes which transitions $J$ can induce. This suggests that $J$ selects computations learned by the backbone rather than creating new algorithms. Together, these results show that the hidden state can control shared computation, with attention routing as a causal pathway. --- ### Markovian Dynamics Enforcer: Feasibility Preserving Correction on Learned Dynamics Manifolds Authors: Kevin Yu, Tao Guo, Constantinos Antoniou, Panagiotis Angeloudis Published: 2026-09-30 Categories: cs.LG, cs.RO, eess.SY arXiv: https://arxiv.org/abs/2609.39888v1 Neural trajectory predictors can reach low prediction error while violating dynamics, actuator limits, or state constraints, especially when controls are unobserved and dynamics are partially specified. We introduce the Markovian Dynamics Enforcer (MaDE), a time-invariant post-hoc operator mapping state-transition proposals onto a learned feasible dynamics manifold, trained on feasible states without ground-truth controls. For each transition it infers a control and recomputes the state through a completion model of known physics plus a learned residual. It then corrects that control by gradient-based inequality reduction, so inequality satisfaction is best-effort within an iteration budget. Since every correction iterate re-enters the completion model, the returned state is dynamically consistent by construction relative to that model and the supplied previous-state anchor. MaDE drives dynamics residuals to essentially zero on fully specified simulated systems, and on an underspecified system leaves a smaller true-dynamics residual than the baselines. Designed to attach to arbitrary predictors, the frozen operator is evaluated downstream of recurrent, structured state-space, and transformer predictors. On recorded vehicle trajectories the one-step residual against a kinematic bicycle model is 0.0071 to 0.0072 for MaDE and 0.1703 to 0.1714 for raw predictors. MaDE raises average displacement error by a factor of 1.57 to 1.83. --- ### OPSRD: On-Policy Self-Role Distillation Authors: Weijie Ren, Yanwen Zhang, Hao Li, Zhuolin Qi, Hengyi Zhang, Naibo Wang Published: 2026-09-30 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.39884v1 Role prompting elicits specialized behavior from large language models through an expert identity, offering a lightweight way to guide reasoning on demanding tasks. However, evaluating or distilling complete role-prompted answers can miss useful next-token preferences when the sampled solution remains incorrect. Transferring these preferences also requires an objective that reaches alternatives the student rarely predicts. We introduce OPSRD, which uses a fixed expert role as privileged teaching context for on-policy self-distillation without reference solutions. A role-free student generates a trajectory, and a frozen instance of the same base model supplies role-conditioned distributions on its exact prefixes, exposing alternatives beyond the sampled continuation. Teacher-weighted forward KL targets alternatives the student underestimates, with clipping to limit individual vocabulary contributions. Supervision is restricted to the highest-entropy half of student positions, concentrating learning where predictions are uncertain. Experiments on three competition-math benchmarks with Qwen3-1.7B, 4B, and 8B show improvements over the base models without role prompts at inference. Forward KL achieves the highest macro-averaged accuracy among the three evaluated divergences at every scale. Code is available at https://github.com/zhansan114514/OPSRD. --- ### Grounding with Confidence: Controllable Generative Video Temporal Grounding Authors: Jinhao Chen, Benlei Cui, Ruijian Jia, Ziheng Wang, Tianyu Wo, Pengfei Sun, Longtao Huang, Hui Xue, Yitong Yang, Haiwen Hong Published: 2026-09-30 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.39883v1 Video temporal grounding supports applications such as video search, content review, and automated editing by localizing events described in natural language. Yet existing generative models typically output timestamps without explicit interval-level confidence scores to guide candidate selection. We separate candidate generation from acceptance by scoring individual intervals within the original decoding pass. A lightweight confidence head reads pooled decoder states, providing an explicit score trained for interval selection. Offline verifier scores supervise the head on fixed candidate sequences, and temporal-overlap labels adapt it to current rollouts during reinforcement learning. GT-anchored candidate-pool supervision and set-level optimization train the generator. The resulting scores support ranking, threshold-based selection, and rejection without invoking an external verifier at inference. On a fixed OMTG-Bench candidate pool, confidence raises query-macro Recall@0.5 from 9.95% to 14.42% over generation order at a 10% global return budget, and from 26.48% to 31.12% at a 25% budget. The continuous scores let downstream applications adjust return budgets or acceptance thresholds to match their precision-recall preferences, without regenerating candidate intervals. --- ### LLM Persona Unlearning Authors: Kemou Li, Zhuan Shi, Qizhou Wang, Fengpeng Li, Negar Rostamzadeh, Golnoosh Farnadi, Jiantao Zhou Published: 2026-09-30 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.39882v1 Pre-training equips large language models (LLMs) with a broad repertoire of behavioral patterns associated with roles, styles, values, and goals. Post-training teaches conditional enactment and makes a helpful Assistant the default, but it does not erase alternative modes from the weights; explicit prompts can therefore elicit personas that repeatedly shape judgment, language, and action. In open-weight settings, runtime controls can be removed, motivating persona unlearning: a weight-level edit that makes a designated persona difficult to elicit and enact on unseen contexts. We introduce PersonaUnlearnBench, a model-specific paired benchmark spanning six LLMs from three families and five personas, with aligned forget/retain sets, held-out instruction paraphrases, and four-axis evaluation. The benchmark shows that standard unlearning methods cannot reliably erase the target persona without sacrificing meaningful generation or general utility. We therefore propose PaCE, which compares target and desirable responses to the same questions to locate an internal behavior direction, then trains target-prompt states away from the target mode and toward the matched desirable response. Experiments show that PaCE consistently suppresses target personas with high response quality and useful counterpart behavior, at moderate utility cost. These results establish persona unlearning as a distinct behavior-level editing problem and a practical route toward persistent control of latent LLM response policies. ---