# Research Papers — 2026-09 · bundle page 247 of 266 > Full-text markdown of papers #12301–12350 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/247 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/246.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/248.md) ### From Trait Vectors to Circuits: Tracing Refusal and Sycophancy Through Language Models Authors: Oscar Miró López-Feliu, Maya Ozbayoglu Published: 2026-09-02 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.22224v1 A direction in activation space that changes safety-relevant behavior when steered is not necessarily one the model uses to produce that behavior on its own. We therefore ask whether steering acts through the computation of the unmodified model or through a different set of components, studying two traits whose directions have been extracted and validated in prior work: refusal and sycophancy in Qwen2.5-7B-Instruct. For each, we use the trait vector to split the computation into a reconstruction circuit before the vector and a transmission circuit after it. We then test whether restoring the coordinate returns behavior removed by ablation. For refusal, both circuits are compact and faithful, restoring the coordinate alone recovers almost all of the refusal signal lost to ablation, and a circuit built around the vector matches the faithfulness of a direct input-to-output circuit at roughly half the edges. For sycophancy, transmission is compact, but reconstruction is broader and only partially faithful. Circuits that transmit a steering intervention are therefore not automatically the circuits that produce the behavior, and we report the counterfactual, target, and behavioral test for every circuit accordingly. --- ### EAVer: Long-Form Factuality Verification as an End-to-End Agentic Policy Authors: Kening Zheng, Aoying Zheng, Zhigang Chang, Yazhi Guo, Miaotian Guo, Qingwei Zong, Xianhai Xie, Weiqiang Jin, Chengze Li, Hanrong Zhang, Jie Yang, Wei-Chieh Huang et al. Published: 2026-09-02 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.22223v1 Long-form factuality verification is commonly implemented as a static decompose-search-verify pipeline, with separately prompted modules processing claims and invoking external search. Treating claims independently makes LLM and search calls scale with claim count and causes repeated searches for overlapping evidence about related claims. We introduce EAVer, an End-to-end Agentic Verifier that learns to control the complete response-level verification workflow as a unified policy. EAVer groups semantically related claims, routes each group to direct verification or targeted search based on confidence, and keeps evidence returned by search in compact in-context memos for cross-claim reuse. To train this policy, we develop a privileged-teacher synthesis pipeline that converts gold claim annotations into executable multi-turn tool-interaction trajectories with live search rather than post-hoc rationales. Structural, label-alignment, tool-use, search-budget, and leakage checks yield 1,447 quality-controlled trajectories. We further construct 794 bidirectional same-trajectory preference pairs that keep claim grouping, search, and evidence fixed, enabling decision-focused Direct Preference Optimization (DPO) over factuality-decision tokens. The results with Qwen3-8B show that EAVer outperforms the strongest search-based baseline on each benchmark by 2.88 Macro-F1 points on VeriFastScore and 4.73 points on the out-of-distribution FaStFact-Bench, while using about 80% fewer searches than the most search-efficient baseline. Moreover, EAVer consistently improves performance across models ranging from 4B to 32B parameters, demonstrating its strong generalizability. --- ### Can Coding Agents Reproduce Official Statistics? Metadata, Retry Budget and the Limits of Execution Feedback in a Controlled Eurostat Benchmark Authors: Sabina-Cristiana Necula Published: 2026-09-02 Categories: cs.LG, cs.AI, cs.CL, cs.SE arXiv: https://arxiv.org/abs/2609.22222v1 Large language models can generate executable data-analysis code, but successful execution is not equivalent to a valid official-statistics result. This study asks whether authoritative metadata and execution feedback improve the reproducibility of Eurostat answers produced by a coding agent, and isolates what execution feedback actually contributes. A benchmark of 30 natural-language tasks covering seven domains, seven Eurostat datasets and four difficulty tiers was run under four conditions: task only (A), task plus a frozen dataset metadata card (B), metadata plus a repair loop driven by sanitized execution feedback (C), and metadata plus the same attempt budget with no diagnostics of any kind (D). Claude Sonnet 5 generated Python through the Anthropic Messages API in three independent replicates, yielding 360 task-runs. Exact correctness required successful execution, the correct dataset, filters, output shape, values and unit. A companion experiment run under an under-specified output contract, in which the required ranking key and unit representation were never stated to the model, understated condition C by 23.4 points, showing that evaluator and contract design can dominate measured agent error. Reliable statistical coding agents need semantic validation against frozen specifications, a fully specified output contract, and a retry budget - not execution diagnostics. --- ### Team DArgk at the 2026 ELOQUENT lab for evaluating generative language model quality: Residuals of Humanity: AI Detection Evasion via GRPO Fine-Tuning Authors: Antonela Tommasel, Juan Manuel Rodriguez Published: 2026-09-02 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.22221v1 Large language models (LLMs) can generate fluent and coherent text that is increasingly difficult to distinguish from human writing, motivating the development of automatic AI-generated text detectors. However, the robustness of such detectors under adversarial generation remains uncertain. This paper presents SHADE (Stochastic Human-like generation via Adversarial Detector Evasion), a reinforcement learning framework that formulates detector evasion as a policy optimization problem. Instead of applying post-hoc perturbations or prompting-based rewriting, SHADE fine-tunes an instruction-tuned LLaMA model with Group Relative Policy Optimization (GRPO), using feedback from a surrogate detector based on the PAN 2025 mdok system. Our experiments show that full fine-tuning with a small KL regularization penalty achieves $98.5\%$ surrogate evasion, compared to $1.5\%$ for the base model, while LoRA-based adaptation is substantially less effective under regularization. Linguistic analysis reveals that successful evasion is associated with shorter, simpler, and less lexically diverse outputs, suggesting that high detector evasion does not necessarily correspond to more human-like writing. In the official Voight-Kampff competition setting, our submissions ranked sixth and seventh, indicating that optimization against a single surrogate detector only partially transfers to unseen evaluation classifiers. These results highlight both the potential and limitations of reinforcement learning for adversarial AI-text generation and motivate more robust, multi-detector evaluation protocols for AI-generated text detection. --- ### Measuring the Checker: Mutation Analysis for GPU-Kernel Benchmark Oracles Authors: Mingzhe Du, Anh Tuan Luu, Dong Huang, See-Kiong Ng Published: 2026-09-02 Categories: cs.LG, cs.PL arXiv: https://arxiv.org/abs/2609.22220v1 Benchmarks for LLM-generated GPU kernels decide correctness with a few random inputs and a loose floating-point tolerance, and their verdicts now feed leaderboards and reinforcement-learning rewards. Recent work agrees these checkers are weak and patches them by hand---extra input distributions, fuzzing recipes, tighter tolerances---with no way to \emph{measure} whether any patch suffices. We introduce mutation analysis as an adequacy metric for kernel-benchmark oracles: deterministic rules inject 10{,}303 compilable faults into verified CUDA implementations of 188 KernelBench problems, 7{,}384 of them with an independent kill witness; any test protocol is scored by the fraction it detects. The official check misses \textbf{one in six} witnessed faults (16.9%), deterministically, and the misses are skewed by family: 8.7% of arithmetic faults escape, but 78.6% of precision faults do. The metric explains why (a tolerance blind band growing with reduction size; a measured ceiling on input aggressiveness set by legitimate floating-point variance), audits the strongest existing patch (KernelBench-Verified's gain splits into $+4.0$ points from hidden inputs and $+4.5$ from tighter tolerance, a split its authors could not compute), and exposes a published fuzzing recipe that rejects \emph{correct} kernels 107 times. Optimizing suites over the kill matrix reaches 98.0% detection with two inputs per problem (94.8% held-out), and the measurement's fault taxonomy teaches a test generator more than the raw faults themselves. Across 48 whole architectures, the blindness grows with scale, concentrating in deep homogeneous pipelines, and two problems prove unrefereeable: their official references violate the benchmark's own tolerance against fp64. We release everything as \href{https://huggingface.co/datasets/Elfsong/KernelBench-M}{KernelBench-M}. --- ### Knowing, and Saying It Only When Asked: LLM Endognostics and the Schizognosis of Minerva-7B Authors: Fabrizio Davide, Francesco Collova Published: 2026-09-02 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.22219v1 Evaluating an aligned language model by reading its answers assumes the answers carry the distinction the evaluator cares about. We introduce LLM endognostics, a white-box internal auditing framework designed to extract and causally manipulate latent knowledge within the residual stream. Applied to Minerva-7B-Instruct-v1.0 on 124 minimal prompt pairs over 12 categories of professional risk, behavioral evaluation fails on most of the set: the model acts identically on 63.7% of the pairs (95% CI (55.0%, 71.6%)), complying with or refusing both members. Yet, projecting the residual stream onto the vocabulary by a Jacobian lens reveals a statistically significant Contrastive Endognostic Margin, proving the model maintains robust risk differentiation internally. In a second protocol crossing 25 facts with five linguistic framings, we show that the model conforms to presupposed falsehoods in 72% of cases, despite representing the true entity in its latent layers. Surgically ablating the direction of the planted falsehood restores the correct answer in 11 of 25 suppressed cases (McNemar p = 0.0010), validated by blind human annotation (binary agreement kappa = 0.68). In contrast, an out-of-sample linear probe achieves 77% accuracy at layer 10, but its orthogonal ablation yields a 0% recovery rate. This establishes a fundamental theoretical dissociation: abstract linear representation does not imply causal control over verbalization. Our main contribution is the formalization of endognostics to prove that behavioral evaluation and internal reading systematically disagree in the common case, and that linear decodability is decoupled from causal control. --- ### Toollery: Scaling LLM Agents to Thousands of Skills and Tools Authors: Xiangxi Tian, Ran Guan Published: 2026-09-02 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.22218v1 As LLM agents are exposed to hundreds to tens of thousands of skills, tools, and API functions, full-library prompting becomes costly, slow, and less reliable: each added candidate increases prompt tokens and latency, while longer candidate lists introduce more distractors for LLM selection. We present \textbf{Toollery}, a training-free candidate-compression framework for scalable LLM skill/tool selection. Following established document-side query expansion, Toollery generates user-intent queries from each skill/tool specification and builds a retrieval index that maps real user requests to compact candidate sets before final LLM decision-making. By treating high-level skills and atomic tools as selectable capabilities, Toollery can be applied to both skill libraries and tool registries. We evaluate Toollery on the roughly 79K-capability SkillRouter benchmark, BFCL-V4 with over 440 atomic tools, and 3,396 proprietary smart-cockpit requests over 220 tools. Across these settings, Toollery keeps online selection bounded to a compact top-$k$ candidate set and improves recall over ordinary specification retrieval. At a fixed top-10 budget, Toollery improves end-to-end selection on the cockpit dataset, and maintains comparable AST Accuracy on BFCL-V4. These results support Toollery as a practical candidate-compression framework for large and evolving agent capability libraries, while showing that quality and cost gains depend on workload coverage and provider caching. --- ### UniGIO: Unified Generative Global In-situ Weather Modeling from Spatiotemporal Incomplete Observations Authors: Songru Yang, Zili Liu, Tao Han, Ben Fei, Lei Bai, Chang Liu, Zhengxia Zou, Xiangyang Ji, Wanli Ouyang, Zhenwei Shi Published: 2026-09-02 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.22217v1 Global In-situ Observation (GIO) provides fine-scale, direct records of the global weather system from sparse point stations, making it an indispensable source for capturing localized and transient dynamics beyond the reach of satellite gridded data, and playing a critical role in key fields such as numerical weather prediction, disaster prevention, and agriculture. However, GIO exhibits strong spatiotemporal incompleteness, severely impairing accurate and real-time in-situ weather modeling. Unlike existing methods waiting for completed AI-ready data with extra introduced errors, in this work, we explore UniGIO, a novel generative framework for directly modeling global in-situ weather dynamics from native incomplete GIO. By generating missing data from observed ones annotated by masks, it unifies the coexisting forecasting, imputation, and generation under arbitrary missing ratios. Between the missing and observed, UniGIO captures station and region level complementarity through the Observation Mixer and Event Aligner, which diffuse discrete observations into continuous spaces where weather processes naturally span multiple stations. We further establish temporal dependencies with pattern shifts using the Adaptive Temporal Mixer, and track extreme events in chaotic local weather systems through a Mixture-ofExperts structure. Steady and extreme events are adapted in decoder by a Local Refiner. Extensive experiments on the up-todate largest global station weather dataset Weather-5K validate its SOTA performance with 11%, 12%, and 5% advantages on accuracy, fidelity, and extreme event capture, delivering a novel holistic solution for weather modeling in GIO networks. --- ### The Effect of Quantization on Clinical Benchmarks: Accuracy and Safety Across Model Families Authors: Leonard Twagirayezu, Prasenjit Mitra Published: 2026-09-02 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.22216v1 Quantization enables deployment of large language models on resource-constrained clinical edge devices, but its effect on clinical accuracy and safety remains understudied. We evaluate five 7-8B parameter models at FP16, GPTQ-INT8, and GPTQ-INT4 precision across five benchmarks: MedQA, MedMCQA, Med-HALT, a risk-stratified sample of HealthBench, and MedSafetyBench. The study jointly varies quantization bit width, model family, and clinical task type, with explicit risk stratification and safety measures. INT8 GPTQ is universally safe (max. degradation -1.9%-1.9%), while INT4 degradation is substantial and model-dependent: BioMistral-7B, clinically fine-tuned, loses 19.7% on MedMCQA, more than any general-purpose model, showing clinical fine-tuning does not confer compression robustness. MedMCQA degrades more than MedQA under INT4; Med-HALT is largely unaffected. On HealthBench's emergency-risk subgroup, Qwen2.5-7B degrades by 26.8% under INT4, suggesting high-risk scenarios are disproportionately vulnerable to compression. On MedSafetyBench, the model family dominates over precision (refusal rates range 10.2%-74.9% at FP16), though Qwen2.5-7B (-17.8%) and Meditron-7B (-28.3%) show substantial INT4 safety degradation; notably, Qwen2.5-7B is simultaneously the most accuracy-robust model, demonstrating that accuracy and safety robustness are independent properties. We additionally test two recovery methods, clinical calibration substitution and QLoRA fine-tuning, both producing the same trade-off: MedMCQA recovers while MedQA further degrades, indicating recovery strategies require task-specific validation rather than being assumed universally beneficial. These findings indicate INT8 is broadly safe for clinical deployment, while INT4 safety must be assessed per-model and per-task, and that safety alignment is determined primarily by instruction tuning rather than clinical domain adaptation. --- ### On Mitigation of Subliminal Learning in Large Language Models Authors: Atsushi Yanagisawa, Brendan Gho, Rajendran Ramesh Babu Manoj Narender, Kevin Zhu, Madhur Panwar, Antonio Mari Published: 2026-09-02 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.22215v1 Knowledge distillation can transmit unintended behavioral traits from a teacher model to a student through training data that appear semantically unrelated to those traits, a phenomenon known as subliminal learning. Although recent work has established this effect, its training dynamics and mitigation remain underexplored. We study subliminal learning in open-weight language models ranging from 1.5B to 8B parameters, covering the Qwen, Gemma, and Llama families in number-sequence and chain-of-thought settings. Rather than evaluating only final models, we track trait-related probabilities throughout fine-tuning and find that subliminal acquisition can be highly non-monotonic, with transient spikes, reversals, and trait-specific failures of transfer. We then introduce liminal training, an annealed KL-regularized fine-tuning method that constrains early drift from the base model. Across our experiments, liminal training substantially reduces subliminal trait acquisition while largely preserving task gains, outperforming paraphrasing and layer freezing as mitigation strategies. The effect also extends beyond animal preferences: in a French-language response-style experiment, liminal training suppresses language transfer while retaining much of the GSM8K improvement. Finally, we show that KL timing matters: early regularization is more effective than late regularization, and sweeping the regularization strength reveals an empirical trade-off between task learning and trait suppression. --- ### The Bairong System for MLC-SLM 2026: Dynamic Question-Aware Evidence Routing for Multilingual Conversational Speech Understanding Authors: Shangkun Huang, Junchao Hu, Huan Shen, Guoji Wang, Yingao Wang, Shaosai Li, Wei Zou, Yunzhang Chen Published: 2026-09-02 Categories: cs.CL, cs.SD arXiv: https://arxiv.org/abs/2609.22214v1 Long multilingual conversational spoken question answering requires systems to balance long-range transcript semantics with sparse acoustic and speaker-sensitive cues. We present the Bairong system for the MLC-SLM 2026 Challenge, where a diarization-ASR front-end produces speaker-attributed transcripts and a dynamic evidence router constructs question-specific inputs for answer prediction. Instead of applying a fixed transcript-only or audio-only policy, the router infers the required evidence type and context scope from the question and answer options, and selects among full transcript context, local audio-text fusion, speaker-linked evidence, and compact global acoustic samples. This transcript-backbone design keeps discourse context available while activating audio only when it provides complementary evidence. Our Task 1 system achieves 25.70% and 18.44% tcpMER on the development and evaluation sets. For Task 2, the final system obtains 94.84% devel?opment accuracy, outperforming the full-transcript baseline by 1.68 points and the best audio-centric diagnostic system by 2.77 points. These results support dynamic question-aware routing as an effective evidence allocation strategy for conversational spoken QA. --- ### SCoP: Structured Constraint Parsing for Evidence-Space Control in Temporal Knowledge Graph Question Answering Authors: Xiaokun Guo, Zhen Xu, Dongdong Huo, Yanqiu Zhang, Dongjin Yu, Yu Wang Published: 2026-09-02 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.22213v1 Temporal Knowledge Graph Question Answering (TKGQA) requires answer inference from evidence that is both structurally valid and temporally admissible. Existing methods often leave anchor-event binding, temporal admissibility, and ordinal selection implicit in model reasoning, task-specific training, or similarity-driven retrieval, allowing locally relevant but invalid facts to enter the answer context. We formulate complex TKGQA as evidence-space control and propose SCoP (Structured Constraint Parsing), a constraint-centric framework that externalizes temporal decisions before answer inference. Instead of treating retrieved facts as admissible evidence by default, SCoP separates answer-seeking event patterns from temporal anchor events, conservatively grounds them to canonical TKG entities and relations, and translates temporal intent into executable constraints with optional ranking requirements. These constraints operate over normalized point and interval ranges, enabling deterministic filtering of structurally compatible candidates and producing a compact evidence space for generation. Experiments on MultiTQ and TimelineCronQ-R assess SCoP across timestamped point-fact and interval-oriented settings with richer temporal relations and ordering dependencies. Without task-specific parameter updates, SCoP achieves 0.825 Hits@1 on MultiTQ and 0.761 Hits@1 on TimelineCronQ-R, with gains on constraint-intensive question types. These results support explicit evidence-space control over unconstrained retrieval or implicit temporal reasoning. --- ### A Channel-Boosted Multi-Agent System with Iterative Consultation for Document Sensitivity Classification Authors: Aleesha Zainab, Asifullah Khan, Muhammad Ahmed Khalid, Faheem Ullah Khan Published: 2026-09-02 Categories: cs.CL, cs.LG, cs.MA arXiv: https://arxiv.org/abs/2609.22212v1 Organizations in critical national infrastructure sectors must assess heterogeneous documents for sensitivity before routing or storage. Manual assessment is slow, inconsistent, and unscalable. Extending our prior leakage-controlled benchmark, BERT established the top single-encoder baseline (89.14% accuracy, 89.33% F1-score under 5-fold cross-validation on the Strategic 16K corpus). However, transformer baselines suffer from a structural limitation: fixed input length truncation discards evidence beyond the retained window-precisely where sensitive cables tend to be longest. We present Channel-Boosted MAS (CB-MAS) and instantiate it as IC-MAS (Iterative Consultation Multi-Agent System) to solve this without long-context computational costs. A Channel Critic Agent learns document-adaptive trust weights governing Gated Channel Boosting between two first-window encoders, while paired Consultation Agents iteratively exchange belief states to reconcile evidence from the beginning and end of long documents. IC-MAS holds computation constant regardless of document length by reconciling fixed windows in a compact representation space. Ablation studies show critic-controlled Channel Boosting provides the bulk of accuracy gains, while consultation recovers recall without precision collapse. Critic-Controlled Gated Channel Boosting with Max-Pool fusion and Blackboard Adaptive Consultation achieves 90.72% accuracy, 91.23% F1-score, 92.01% sensitive recall, and 90.46% sensitive precision, using about 54% less average computation than a fixed-round baseline. Gains over the single-encoder baseline are statistically significant (McNemar's test, p less than 0.000001; paired t-test). We include LIME/SHAP explainability, multi-agent evaluation, and an honest accounting of limitations. --- ### Clinical Reasoning Under a Partially Observed Objective in Cone Beam CT Report Generation Authors: Ajo Babu George, Govind Arun, Sidharth N Krishna, Uma Ranjan Published: 2026-09-02 Categories: cs.CL, cs.CV arXiv: https://arxiv.org/abs/2609.13238v1 Maxillofacial report generation from cone beam computed tomography is scored here by a composite objective placing 80% of its weight on a large language model judgement of factual entailment and 20% on lexical overlap, of which only the lexical fifth is visible during development. The grader's BLEU-4 and METEOR routines are reproduced in pure Python and match the reference to machine precision, and an offline entailment surrogate, which tells a report written for one patient from one written for another at an area under the curve of 0.987, makes the composite objective cheap enough to optimise directly. Over the 622-case public release, a report selected against the visible lexical ranking scores 0.2909, whereas one selected against the composite objective scores 0.4122, because pursuing n-gram overlap drives entailment precision from 0.522 down to 0.266. A 29 million parameter encoder fine-tuned on the release reaches a prevalence-weighted out-of-fold area under the curve of 0.486 over 985 statements, indistinguishable from the corpus prior, while nine numbers read from the image header reach 0.945 for mandible coverage and 0.872 for condyle coverage, and acquisition centre alone predicts sentence choice at 0.718 against 0.663 for the image-derived model, identifying dictation convention rather than anatomy as the quantity the lexical metrics reward. The delivered system emits eight unconditional statements and five gated on header geometry under polarity, laterality and tooth-level consistency constraints, and reaches METEOR 0.3542 over 50 held-out cases from an unseen centre. The dataset and code are available at https://github.com/GIND123/CBCT-Clinical-Reasoner --- ### Occlusal Geometry in Closed Form for Orthodontic Report Generation Authors: Ajo Babu George, Govind Arun, Sidharth N Krishna, Uma Ranjan Published: 2026-09-02 Categories: cs.CV, cs.CL arXiv: https://arxiv.org/abs/2609.13237v1 Orthodontic report generation from intraoral data is normally cast as multimodal captioning, yet the released Bite2Text scan pairs are supplied already registered in occlusion, which makes several core occlusal quantities directly measurable rather than inferable. The system reported here exploits that property: an anatomical frame is recovered per case from arch taper and arch closure instead of the stated RAS convention, which does not hold across the release, and each arch is reduced to an occlusal ridge profile in arch-angle coordinates yielding overbite, overjet, midline deviation, transverse overlap, crossbite extent, cusp interdigitation lag, and the occlusal curves in closed form. Gradient boosting maps 31 such measurements onto 13 template fields, a field being predicted only where patient-level cross-validation beats its own majority baseline, and a deterministic renderer emits the corpus six-part narrative; a ConvNeXt-Tiny classifier over the five standardised photographic views is fused per field, raising mean field accuracy from 0.601 to 0.683. Reimplementation of the challenge evaluator shows that its BLEU-4 and METEOR are local variants whose F-mean weights recall nine to one, that two clinicians agree on 47 percent of findings for the same patient, and that a constant report consequently outscores a genuine second clinician report by 0.165 captioning. Held-out scores reach BLEU-4 0.458 and METEOR 0.677 against intraoral scan references and 0.278 and 0.507 against photograph references, and the submitted system placed third in the ODIN 2026 Bite2Text test phase at 0.2680 and 0.4629, within 0.022 BLEU-4 of first, running on CPU in under ten seconds per case. The dataset and code are available at https://github.com/GIND123/ODIN_toothfairy4 --- ### Learning Manipulation-Sufficient Representations via Outcome Bottlenecks Authors: Md Selim Sarowar, Sungho Kim Published: 2026-09-02 Categories: cs.RO, cs.CV arXiv: https://arxiv.org/abs/2609.13235v1 Networked manipulation endpoints couple perception to actuation across compute- and bandwidth-limited links, yet commonly exchange dense geometric states optimized for fidelity rather than action outcomes. A stochastic representation is learned with a policy-free, action-conditioned outcome bottleneck: marginal outcome log-loss supplies distortion and a KL term regularizes rate. The construction is motivated by the minimal statistic that preserves the outcome distribution of every admissible action, while the implemented finite model is evaluated as a rate-regularized mixture predictor. The same encoder and outcome head support grasp selection, singleton conformal filtering, active viewpoint selection, and latent test-time adaptation. A finite-probe theorem identifies the local level-set tangent space with the null space of an outcome Jacobian. The synthetic oracle verifies this result; on scanned objects, an analytic surrogate agrees with measured simulator invariances within \(0.56^\circ\). Across 11,979 simulated grasps on 13 objects, a reconstructed-geometry wrench score attains 0.542 AUC against lift success and falls below chance on curved objects, while our representation attains 0.876. At 25\% commitment, executed-grasp success is 0.503 versus 0.984. The 512-byte interface is \(288\times\) smaller than one RGB-D frame and runs at 16\,ms per CPU decision. Independent synthetic points track the tested conformal levels; scanned-object all-pair coverage is reported as a clustered empirical diagnostic. On unseen objects, within-scene AUC falls to 0.569, and a full-feedback update raises empirical mean pairwise coverage from 0.728 to 0.883. --- ### Harnessing human expertise for high-precision robotic assembly in industrialized construction: A sample-efficient installer-in-the-loop interactive reinforcement learning framework Authors: Zekai Jin, Huiguang Wang, Xiaoning Sun, Yi Shao Published: 2026-09-02 Categories: cs.RO, stat.ML arXiv: https://arxiv.org/abs/2609.13234v1 Industrialized construction imposes stringent precision requirements on robotic assembly of modular components such as prefabricated window units. In tolerance-critical operations, the central bottleneck is not only mechanical clearance but also converting tacit installer expertise into data-efficient autonomy under sparse acceptance feedback, contact variability, and millimeter-scale constraints. We present an installer-in-the-loop interactive reinforcement learning framework that acquires expertise through offline teleoperated demonstrations, sparse event-driven binary takeovers at contact-failure boundaries, and acceptance-aligned terminal rewards, logged under a unified schema for traceable offline-to-online adaptation. A temporally abstract action-sequence policy built on Q-chunking with Flow Q-Learning captures multimodal recovery maneuvers under sparse terminal rewards, while a non-updating warm-start phase stabilizes the offline-to-online transition. The framework is evaluated in MuJoCo across the workflow from suction acquisition through clearance-limited seating, under structured staging and end-to-end randomized placement. Within a defined stress-test regime with 2 mm per-side clearance, bounded pose perturbations, and friction randomization, the pipeline attains 100\% autonomous seating with 12--15 min of cumulative installer supervision over 3.0 h of online training, and reaches the 95\% success milestone in approximately 0.5 h and 1.5 h in the two experiments. We also report wall-clock adaptation time, cumulative takeover minutes, intervention-rate decay, and stage-wise failure attribution to inform supervision budgeting. Ablations isolate the complementary contributions of temporal abstraction, installer intervention, and warm-start value calibration. --- ### EMCStereo: Attention-Enhanced Stereo Matching for Thin-Structure Depth Estimation with a Synthetic Tree-Branch Benchmark Authors: Yida Lin, Bing Xue, Mengjie Zhang, Sam Schofield, Richard Green Published: 2026-09-02 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.13233v1 Thin structures such as tree branches are among the hardest cases for stereo matching: a branch is only a few pixels wide, the background is cluttered, and dense ground truth for real branches is nearly impossible to label by hand. We make three contributions. First, EMCStereo integrates three lightweight attention modules into a PSMNet-style cost-volume backbone: Efficient Multi-scale Attention (EMA) on deep semantic features, a Multi-Scale Fusion block (MSFblock) learning spatial pyramid weights instead of concatenating them, and Coordinate Attention (CoordAtt) on final matching features. Because MSFblock collapses four pyramid branches into one, the modules leave the network 2.0% smaller and add only 1.7% inference time overhead. Second, VirtualTree is a synthetic stereo dataset rendered in Unreal Engine 5 with a simulated ZED Mini rig, providing 5,520 pairs with exact disparity for thin branches. Third, an eight-way ablation establishes a run-to-run noise floor of 0.009 px end-point error (EPE). EMCStereo achieves 1.31 px EPE (5.96% D1-all) on the VirtualTree test split, 1.00 px on SceneFlow, and 0.80, 0.73, 0.62, and 3.19 px on KITTI 2012, KITTI 2015, ETH3D, and Middlebury, with depth accuracy delta_1 from 92.6% to 98.7%. Evaluated against the noise floor, the attention stack is accuracy-neutral at a 100-epoch budget, while MSFblock and CoordAtt cost 0.03-0.05 px unless EMA is present. --- ### What Does the Encoder Actually Decide? A Controlled Comparison of Vision Backbones on Joint Tree Segmentation and Stereo Depth Authors: Yida Lin, Bing Xue, Mengjie Zhang, Sam Schofield, Richard Green Published: 2026-09-02 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.13232v1 A robot pruning trees needs two facts per pixel: whether it belongs to a tree, and its distance. Both are usually obtained via task heads attached to a vision backbone chosen by reputation rather than measurement. Holding dataset, decoders, losses, schedule, and evaluation fixed, we ask: how much does the encoder choice change joint semantic segmentation and stereo depth on thin vegetation? We build a hard parameter-sharing network with one encoder feeding both branches, swapping only the encoder without downstream retuning. We evaluate (N) encoders across (M) architecture families (CNNs, transformers, hybrids, MLP-mixers, state-space models) near a ~25M budget, trained from scratch. Depth is evaluated on tree pixels only; segmentation uses boundary F1 and background IoU to prevent "label-everything-tree" shortcuts. Three findings stand out. First, the strongest encoders are convolutional and hybrid, not transformers: (BestEncoder) leads with (MIoU) segmentation mIoU and (Delta) depth $δ_1$, while (X) of (Y) plain vision transformers collapse when trained from scratch. Second, parameter count does not predict quality --- (SmallEncoder) at only (P)M parameters outranks models two orders of magnitude larger. Third, segmentation and depth rankings agree strongly (Spearman $ρ$ = (RhoValue)), showing no task conflict. Finally, (K) of (N) encoders collapse to degenerate all-tree segmentation --- exposed by boundary F1 but hidden by region IoU. --- ### ShieldVLA: Feasibility-Aware Safety Alignment for Vision-Language-Action Models Authors: Manan Tayal, Akshay Nambi Published: 2026-09-02 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.13231v1 Vision-Language-Action (VLA) models demonstrate strong generalization in robotic manipulation and navigation, but existing fine-tuning methods provide limited safety guarantees. Current approaches primarily rely on Lagrangian optimization that enforces safety through soft penalties on expected cumulative cost, often resulting in residual constraint violations or overly conservative behavior. Moreover, learning safety in visual domains is challenging due to the absence of dense per-step safety annotations. We propose ShieldVLA, a safety-aligned fine-tuning framework for VLA models based on Hamilton-Jacobi (HJ) reachability. ShieldVLA learns a model-free approximation of the HJ reachability value function directly from visual observations to estimate the safe operating region. The learned safety critic gates policy optimization by separating reward maximization within feasible regions from recovery near unsafe states, avoiding persistent reward-cost trade-offs. To enable scalable supervision in visual environments, we introduce rubric-based VLM safety scores that convert semantic safety feedback into structured critic targets without requiring manual cost labels. Across five navigation and manipulation benchmarks spanning multiple VLA backbones, ShieldVLA reduces cumulative safety cost by 57% on average and improves task success rate by +0.13 over SafeVLA. --- ### Chemical and geometric representation fidelity improves drug--target affinity prediction Authors: Yixiao Li, Yining Qian, Yefan Chen, Zenghui Chen, Jiayue Sun, Yuhai Zhao, Cheng Tan, An-Yang Lu Published: 2026-09-02 Categories: q-bio.BM, cs.AI, cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.13230v1 Predicting drug--target binding affinity (DTA) requires models to distinguish subtle chemical and structural determinants underlying molecular recognition. Although recent approaches increasingly incorporate richer drug and protein information, such information may be compressed, homogenized or discretized during representation construction, causing affinity-relevant distinctions to be lost before interaction modelling. We hypothesized that this representation-stage information loss constitutes an upstream bottleneck that cannot be reliably overcome by increasingly complex interaction predictors. To test this hypothesis, we developed ReGeoDTA, a representation-preserving framework that maintains affinity-relevant chemical heterogeneity in molecular representations and continuous geometric relationships in protein structures. Across three benchmark datasets, ReGeoDTA consistently improved affinity prediction, and the proposed representation-preserving strategies retained their benefits across diverse DTA architectures. Controlled representation degradation progressively reduced predictive performance, whereas increasing downstream predictor complexity failed to recover information lost during representation construction. These findings identify representation fidelity as an upstream design principle for accurate and generalizable drug--target affinity prediction, with potential implications for computational compound prioritization. --- ### Rethinking Handwritten Character Recognition Authors: Ranjit Raut, Aarav Subedi, Ashim Shrestha Published: 2026-09-02 Categories: cs.CV, eess.IV arXiv: https://arxiv.org/abs/2609.10572v1 Non-Latin handwritten character recognition (HCR) remains understudied. Dominant methods consider it as generic image classification, which uses model scale to implicitly learn stroke structure. Structural-prior efficiency---the principle that explicitly encoding script-geometric regularities as architectural inductive biases can be both more accurate and require fewer parameters. We introduce GraphemeNet, a unified multi-script architecture, governed by two orthogonal binary axes. Axis 1 operationalises stroke-level geometric regularity via Persistent Scaffold Injection (PSI): a script-specific asymmetric convolution injects a stroke scaffold as a weighted residual at every encoder stage, continuously anchoring learned features to script geometry---distinct from skip connections, auxiliary losses, or attention reweighting. Axis 2 selects between global average pooling with gated fusion and cross-scale attention with a Stroke Topology Module (STM), depending on whether glyph discrimination requires spatial relational reasoning. A Linear Capsule Routing (LCR) with $O(n)$ routing is shared universally. On fourteen benchmarks across eight writing systems, the architecture generalises with only scaffold and decoder topology varying per script, consistently challenging, outperforming published baselines, and establishing structural-prior efficiency as a broadly applicable principle for multi-script HCR. --- ### Do Depressive Facial Patterns Transfer Across Cultures and Contexts? Evidence from a German RCT and E-DAIC Authors: Misha Sadeghi, Robert Richer, Lydia Helene Rupp, Lena Schindler-Gmelch, Marie Keinert, Farnaz Rahimi, Malin Hager, Bernhard Egger, Matthias Berking, Bjoern M. Eskofier Published: 2026-09-02 Categories: cs.CV, cs.HC arXiv: https://arxiv.org/abs/2609.05543v1 Automated assessment of depression from facial dynamics holds promise for scalable mental health monitoring, yet cross-corpus generalization of learned biomarkers remains an open challenge. We present a systematic bidirectional transfer study pairing the EmpkinS-EKSpression randomized controlled trial (RCT; N = 256, SCID-5-CV diagnoses) with the Extended Distress Analysis Interview Corpus (E-DAIC; N = 275, semi-structured clinical interviews), predicting depression severity and binary diagnostic status from facial action units, head pose, and gaze. Cross-corpus binary classification proves more robust than continuous PHQ-8 severity regression, with forward transfer achieving AUC = 0.70. Regression transfer is governed by functional context alignment: passive observation phases yield the most transferable models, while active emotion regulation phases elicit stronger within-corpus signals. These findings establish functional context alignment as the primary determinant of cross-corpus generalization, with passive elicitation contexts offering the best trade-off between within-corpus sensitivity and cross-corpus robustness. --- ### Physics-Informed Neural Networks for Depth-Averaged Granular Avalanche Dynamics on Curved Topography Authors: Pujan Pranavkumar Purohit, Pradyumn Singh Sikarwar, Vishal Sharma, Gaurav Bhutani Published: 2026-09-02 Categories: cond-mat.soft, cs.CE, cs.LG, physics.flu-dyn arXiv: https://arxiv.org/abs/2609.05542v1 Physics-informed neural networks (PINNs) provide a mesh-free framework for solving governing equations, but their application to granular avalanche dynamics over curved terrain remains largely unexplored. This study extends a depth-averaged PINN formulation based on the Savage-Hutter equations to an exponentially curved chute with spatially varying inclination and a strain-rate-dependent Mohr-Coulomb earth-pressure closure. The model is validated against measured front- and rear-edge trajectories from a laboratory granular-avalanche experiment, with selected observations withheld from training. A staged temporal curriculum proved essential for accurate prediction, reducing the held-out trajectory error by approximately two orders of magnitude compared with training over the full time domain from the outset. Sparse-data experiments further showed that observation placement was more influential than observation number within the configurations tested. Four observations bracketing the transition from acceleration to deceleration achieved nearly the same accuracy as the eight-observation reference configuration, whereas observations clustered at early or late times performed poorly. The results demonstrate the importance of both training strategy and informative data placement when applying PINNs to granular flows over curved topography. --- ### Video Compression with Graph-inspired Neural Representation Authors: Changqi Wang, Ge Gao, Fan Zhang, Yue Li, Kai Zhang, Li Zhang, David Bull Published: 2026-09-02 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05541v1 Implicit Neural Representations (INR) provide a compact and content-adaptive paradigm for video compression, typically representing a video through shared network parameters and frame-indexed embeddings. Compared to conventional or autoencoder-based codecs, these approaches exploit temporal redundancy within videos in an implicit manner, which potentially results in sub-optimal compression performance. In this paper, we propose G-NeRV, a graph-inspired INR that explicitly improves temporal redundancy exploitation in the implicit latent space. Motivated by the total correlation principles in information theory, we construct a temporal neighborhood over frame embeddings and perform message passing to aggregate reusable information from neighboring frames through an adaptive gate controlling the injection of neighboring information. Inspired by the reference frame buffer in conventional video coding, a memory bank mechanism has been further designed to enable efficient temporal-neighbor retrieval under random frame-index sampling in INR training. This new representation model has been integrated into an advanced representation compression framework and compared with existing conventional and neural video codecs. The results show that the G-NeRV codec outperforms the state-of-the-art INR-based codec, NVRC, and the latest standard video codec, VVC VTM, by 8.86\% and 14.68\% (in BD-rate), respectively, measured by PSNR on the UVG dataset. --- ### Knowing When Not to Answer: Abstention and Refusal Reasoning in Vision--Language Models Authors: Karan Dua, Amit Agarwal, Hitesh Laxmichand Patel, Hansa Meghwani, Jyotika Singh, Ranjeet Gupta, Graham Horwood, Tao Sheng, Avi Sil, Sujith Ravi, Dan Roth Published: 2026-09-02 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.05540v1 Many medical conditions require diagnosis through detailed, multi-context clinical assessment rather than from visual appearance alone. Despite this, vision-language models (VLMs) are increasingly queried to interpret images in ways that touch on medical or diagnostic judgments, raising safety concerns when such inferences are unsupported. ASD diagnosis requires behavioral and developmental evidence, not static facial photographs. We audit whether VLMs abstain from this unanswerable paired-image query, and whether expressions sway non-abstaining choices. We introduce PARITY (Paired Assessment with Reused Identity), a synthetic, demographically balanced set of identity-controlled neutral/expression portrait pairs with neutral-neutral controls. All identities are synthetic and have no ASD status; because the query is unanswerable from images, any non-abstaining selection is treated as a harmful attribution. Across contemporary VLMs, we find a clear split between refusal-first models and speculative models; in the latter, certain expressions disproportionately trigger harmful selections. Clinical guardrails and single-image framing substantially increase abstention, suggesting actionable mitigations in both prompting and interface design --- ### Dual-Latent Memory Routing for Vision-Language Reasoning Authors: Hao-Xuan Ma, Jin-Fei Qi, Yicheng Xiao, Han-Jia Ye Published: 2026-09-02 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.05539v1 Multimodal large language models (MLLMs) have recently made strong progress in vision-language reasoning, yet their performance often degrades as generations grow longer. A key factor is that they frequently lose track of earlier visual evidence and intermediate constraints under a monolithic growing context. Inspired by how humans separately recall what they see and what they infer when solving complex tasks, we propose DLMR, a parameter-efficient mechanism that equips MLLMs with Dual Latent Memories: a visual memory that compresses image evidence and a reasoning memory that tracks intermediate conclusions and constraints. A Router then dynamically decides which memory and how much to reuse during inference, preserving visual grounding while maintaining coherent long-horizon reasoning. DLMR is trained in three stages, from latent memory construction to selective router learning, while keeping the base MLLM frozen, yielding substantial gains on both general and reasoning benchmarks with only a small number of additional trainable parameters. Analyses further show interpretable, state-dependent routing with specialized memory roles and reduced decoding tokens over long generations. Code is available at https://github.com/Hunter-Wrynn/DLMR. --- ### VIS-DICT: A Visual Dictionary for Missing Modality Imputation in Social Network Depression Detection Authors: Hamed Marvi, Mohammad Mehdi Keikha, Abolfazl Nadi Published: 2026-09-02 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.05537v1 Tracking social media posts can help spot early signs of depression. Recent studies show that combining text and images works better for detecting depression than using text alone. However, many social media posts do not have images, which makes it hard to use multimodal models. Most existing methods fill in missing images using retrieval or generative models that need extra training. In this paper, we introduce Vis-Dict, a dictionary-based method that builds missing visual features by linking words to average image vectors from complete training posts. These estimated visual features are then combined with text to track changes in user behavior over time. We tested Vis-Dict on a social media dataset using user timelines of up to 512 posts and compared it with other missing-data methods. The results show that Vis-Dict performs on par with generative networks, reaching an F1-score of 0.9454 and an ROC-AUC of 0.9890. Most importantly, Vis-Dict achieves this strong performance with zero trainable parameters for image generation. These findings show that directly connecting words to visual features is an effective and practical way to handle missing images in depression detection systems. --- ### Beyond the Verdict: Evidence-Aligned Evaluation of Visual Prompt-Injection Guardrails Authors: Suyoung Lee, Myungsub Choi Published: 2026-09-02 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.05535v1 Verdict-only evaluation does not reveal whether a vision-language model (VLM) used the visual evidence that should support its decision. We study this problem in web-agent guardrails, where a VLM judges whether on-screen text conflicts with a user instruction. We introduce Mind2Web-Injection, a benchmark of 9,954 instruction-screenshot pairs with instruction-relative labels, pixel-exact evidence boxes, and matched image-side counterfactuals. Across six VLMs, two models with nearly identical average precision differ ninefold in Evidence-Aligned Detection (EAD), the fraction of attacks both detected and correctly localized. To test whether a verdict depends on the command cited as evidence, we replace the instruction with one that endorses that command. Qwen3-VL-32B, the strongest open-weight localizer, returns aligned in only 58.7% of cases, whereas GPT-5.6-luna does so in 99.9%. To diagnose these failures, we propose two training-free interventions. ReadGate improves grounding without changing verdicts, while CmdCompare tests whether explicit instruction-command comparison resolves instruction-side inconsistency. These results motivate reporting verdict correctness, evidence localization, and counterfactual responsiveness separately. --- ### SimpleMemVLA: A Simple but Effective Native-Video Memory for Vision-Language-Action Models Authors: Cheng Yin, Wang Xu, Junpeng Yang, Sikyuen Tam, Hanyu Liu, Yuan Yao, Xiangrui Zeng, Junbo Cui, Yequan Wang, Zhouping Yin, Yankai Lin Published: 2026-09-02 Categories: cs.CV, cs.LG, cs.RO arXiv: https://arxiv.org/abs/2609.05533v2 Long-horizon manipulation is partially observable: the information needed to choose the next action may appear only in observations from minutes earlier. Existing memory mechanisms for VLAs, such as retrieval banks, learned compressors and recurrent states, must decide what to keep from the past before knowing what a future decision will require. They were motivated by the assumption that minute-scale history is too large to process directly, which no longer holds for modern VLM backbones. We propose SimpleMemVLA, a VLA without a dedicated memory module that uses the backbone's native video context directly as memory. It keeps the sampled history intact in the timestamped video format the backbone was pretrained to process, routes the evidence it finds to a standard flow-matching action head through the hidden states of a generated sub-task, and prefills the history shared by consecutive decisions during action execution, keeping latency close to that of a single-frame VLA. SimpleMemVLA achieves state-of-the-art results on four memory benchmarks without loss on general-purpose control, and with the same backbone and training setup it outperforms retrieval, compression and recurrent-state methods by a wide margin. History interventions show that the policy reads specific evidence from its past and follows edited histories without parameter updates, a visual form of in-context learning. On a physical dual-arm robot, it completes two tasks whose decisive evidence disappears before the robot acts. Code available at https://github.com/OpenBMB/SimpleMemVLA --- ### A Specialized Large Multimodal Model for Interpreting PET/CT in Head and Neck Cancer Authors: Haengbok Chung, SunGyu Kim, Joo hyun Lee, Sangjin Bae, Min Jeong Cho, Minseok Suh, Jae Sung Lee Published: 2026-09-02 Categories: cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.05532v1 Background: Diagnosing head and neck cancer using PET/CT is clinically challenging and time-consuming due to the anatomical complexity of the region, motivating computer-aided diagnosis (CAD). Generalist Large Multimodal Models (LMMs) remain limited in medical contexts by insufficient domain-specific knowledge, privacy and security concerns, and verbosity, motivating specialized standalone LMMs. Purpose: We evaluated the feasibility of a specialized LMM for automated PET/CT interpretation in head and neck cancer using a large-scale multi-institutional PET/CT dataset, a tailored training curriculum, and autoregressive training. Methods: LLaVA-NeXT was fine-tuned using a two-level curriculum with image-conversation pairs curated by two radiologists from public data. The dataset included clinically important annotations such as primary tumor presence and metastatic lymph node location. Level 1 used 28,000 image-conversation pairs to learn basic information, including modality type and hypermetabolism. Level 2 used 12,975 pairs to learn primary tumor presence and the existence and anatomical location of cervical lymph node metastases. External validation included four institutions with diverse imaging devices. Results: The specialized LMM substantially outperformed ChatGPT and LLaVA-NeXT. In Level-2 external validation, ROUGE-L, ROUGE-S, Cosine Similarity, Precision, Recall, and F1 were 0.8751, 0.8794, 0.8324, 0.8794, 0.8711, and 0.8751, while generalist models consistently scored below 0.1. Primary tumor classification accuracy was 83.14 +/- 1.15% internally and 69.03 +/- 0.81% externally. For lymph node localization, the corresponding scores were 0.6389, 0.6257, 0.5287, 0.5782, 0.6371, and 0.6648. Conclusion: Specialized LMMs show promising results for fast, accurate PET/CT-based diagnostic support and medical education, highlighting their potential for clinical translation. --- ### When Agent Governance Helps Authors: Michael Ray Johnson, Linda Naimi Published: 2026-09-02 Categories: cs.CL, cs.AI, cs.CY, cs.MA arXiv: https://arxiv.org/abs/2609.05531v1 No specification says how a governed autotelic AI agent organization, where agents pursue self-generated goals inside guardrails, should be designed and evaluated. We answer in two parts. First, we synthesize the Governed Autotelic Multi-Agent Product Organization (GAMPO) framework from a document-based qualitative evidence synthesis of 321 sources, integrating agency, agile, platform, and governance theory into a runnable specification. Second, we probe a prompt-layer instantiation of GAMPO on CHI-Bench, a long-horizon healthcare benchmark, across open and frontier models. The result is a boundary condition: governance benefit is gated by a model's spare capacity and is domain- and model-specific. On capacity-constrained open models the full procedure yields no reliable benefit, whereas a single "verify your writes" sentence doubles task success (pass@1 2/20 to 4/20). At the frontier the same scaffold lifts prior-authorization 24% to 40% but nets zero on another model, a gap traced to a stable recommendation-override disposition. A second result refines the first: replacing the generic procedure with an answer-blind, per-task definition-of-done, keyed only to the case's own policy and published standards, never the hidden key, raises prior-authorization to 84% under best-of-five self-consistency (68% single-attempt, confirmed by a held-out board) and utilization-management to 44%, while care-management meets a subjective content-quality wall. The contribution is a named, auditable framework and capability-gated evidence that governance should be sized to spare capacity, and that at the frontier a case-grounded specification beats a uniform procedure. Findings are exploratory: partial instantiation, small per-cell samples (n = 5-25), and single trials. --- ### FailSAE: Towards Interpretable Failure Prediction for Vision-Language Models via Sparse Autoencoders Authors: Jie Ma, Zongxi Liu, Yi Zhu Published: 2026-09-02 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.04276v1 Vision-language models (VLMs), such as CLIP, have achieved strong performance across multimodal tasks by aligning visual and textual representations in a shared embedding space. As VLMs are increasingly used for high-stakes domains, failure prediction becomes critical for risk-aware deployment and human intervention. Existing failure prediction methods typically rely on confidence scores or auxiliary classifiers. Although these methods are effective on predicting VLM failures, they provide limited interpretability. In this work, we investigate the use of Sparse Autoencoders (SAEs) for interpretable failure prediction in VLMs. We formulate failure prediction as a classification task over sparse SAE latent activations and introduce a three-stage failure-aware training pipeline that encourages the learned latent directions to remain interpretable while becoming more informative for failure prediction. Our experiments show that the resulting framework outperforms the evaluated baselines in failure prediction. Further analysis suggests that failure-aware training encourages SAE latent directions to capture more class-specific concepts. We also use the SAE to provide a concept-level analysis of how model representations change during failures, revealing a shift from class-specific concepts toward more ambiguous or style-related concepts. Finally, we explore how the learned SAE latent directions can support runtime failure recovery. --- ### Multi-scale Image Representation Compression Authors: Tianhao Peng, Ho Man Kwan, Fan Zhang, Shan Liu, David Bull Published: 2026-09-02 Categories: eess.IV, cs.CV, cs.MM arXiv: https://arxiv.org/abs/2609.04274v1 Overfitted codecs have demonstrated promising performance for image and video compression. In particular, for image compression, the Cool-chic family of models has shown competitive performance against scene-agnostic models, with orders of magnitude lower decoding complexity at the cost of a longer overfitting process. However, these overfitted image codecs are not fully optimized toward the rate-distortion objective: their network weights remain in full precision during training, and the associated quantization parameters are selected in a separate post-training stage. Furthermore, their synthesis operates at a single scale, which overlooks cross-scale redundancy. In this paper, we propose MIRC, an overfitted image codec in which every coded component, including the latents, the synthesis network, and the entropy models, is quantized and entropy coded under a single rate-distortion objective, adopting the end-to-end compression pipeline of the neural video representation codec NVRC. We further introduce a multi-scale representation with cross-stage parameter sharing, which improves coding efficiency at a small transmitted overhead. On the CLIC2020 professional validation set, MIRC achieves a 10.5% BD-rate saving against VVC (VTM 22.0). Moreover, MIRC offers a family of configurations spanning 1.2 to 2.9 kMAC per pixel, so the decoding budget can be selected to match the deployment target. --- ### Scalable Neural Video Representation Compression Authors: Tianhao Peng, Ho Man Kwan, Fan Zhang, Shan Liu, David Bull Published: 2026-09-02 Categories: eess.IV, cs.CV, cs.MM arXiv: https://arxiv.org/abs/2609.04273v1 Scalable video coding (SVC) encodes a video into a layered bitstream consisting of a base layer and one or multiple enhancement layers, enabling decoding at different bitrate/quality/resolution operating points to accommodate diverse device capabilities and network conditions. Due to its practical flexibility, SVC has been incorporated into major video coding standards and has recently attracted growing interest for both scene-agnostic and scene-adaptive neural video codecs. Among the latter, Implicit neural representation (INR) based codecs achieve compression by overfitting a compact neural network to an individual video, offering fast decoding and competitive coding efficiency compared to scene-agnostic neural codecs. However, research on scalable INR-based compression remains in its infancy: these methods support scalable coding by introducing additional network layers, which couple the bitrate with the decoding complexity and also cannot achieve comparable performance with strong scalable/non-scalable codecs. In this context, this paper proposes S-NVRC, a scalable INR-based video codec that jointly supports fine-grained bitrate and decoding complexity scalability from a single embedded bitstream. It adopts a coarse-to-fine prefix for feature grids and a nested prefix for network layers, which scale bitrate and decoding complexity, respectively. The proposed S-NVRC spans a wide range of bitrate and decoding-complexity using a single encoding (training) and outperforms SHM 12.4 and the multi-layer VTM-20.0, by 43.7% and 5.6% in BD-rate on the UVG dataset, while also providing flexible complexity scalability. Implemented code will be provided. --- ### Evaluating Large Language Models for Forced Outage Risk Prediction: Benefits and Comparison to Machine Learning Authors: Christos Petridis, Zoran Obradovic, Mladen Kezunovic Published: 2026-09-02 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.04272v1 This study examines the ability of large language models (LLMs) to predict the risk of weather-related forced outages in the distribution grid in a zero-shot framework, without labeled training data. The problem is formulated as a binary severity classification task across three forecast horizons (3h, 6h, 12h), using six years of outage records and high-resolution weather data for a utility service area in central Texas. Four zero-shot LLMs are benchmarked against two supervised classifiers across two input configurations: one using current weather observations and the other using weather forecast data. Results show that supervised models outperform LLMs on macro-F1 and precision, while newer LLM generations achieve competitive scores. Beyond accuracy, LLMs offer complementary strengths in actionable reasoning and geographic scalability, suggesting that combining them with supervised models may be the best practice. --- ### Quantum-Assisted Memory-Efficient Training for Parameter-Intensive Wi-Fi-Based Human Activity Recognition Authors: To Truong An, Jie Zhang, Guolin Yin, Junqing Zhang, Yanjiao Li, Trung Q. Duong, Simon L. Cotton Published: 2026-09-02 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.04271v1 Wi-Fi-based human activity recognition (HAR) has become an important part of integrated sensing and communications, paving the way for a range of context-aware services. However, most existing Wi-Fi-based HAR systems rely on deep learning (DL) models that are computationally and memory intensive in both training and inference, which poses significant challenges for real-world deployment. Conventional training requires simultaneous updates of millions of parameters, leading to prohibitive memory consumption. In this paper, we propose a novel quantum-assisted memory-efficient training framework (Q-MET) designed to improve efficiency in both training and inference. Q-MET utilizes a hybrid quantum classical neural network to indirectly generate parameters for HAR models, significantly reducing the trainable parameter count compared to direct optimization. To further support the deployment on resource-constrained devices, we integrate structured pruning during the training phase. Experimental results demonstrate that Q-MET achieves a 90% to 95% reduction in trainable parameters compared with conventional backpropagation-based DL training while maintaining or even exceeding classical classification accuracy. Additionally, Q-MET supports lightweight inference through structured pruning, achieving 75% to 85% model sparsity with less than 2% loss in classification accuracy. To the best of our knowledge, this work represents the first quantum-assisted approach to simultaneously tackle memory inefficiencies in both the training and inference stages of HAR systems. --- ### Reviewer Capability Governs Rejection Targeting, Not Repair Skill: Evidence from LLM Execute-Review-Revise Pipelines Authors: Faizan Tanveer Published: 2026-09-02 Categories: cs.SE, cs.CL, cs.MA arXiv: https://arxiv.org/abs/2609.04270v1 Multi-agent LLM pipelines increasingly assign roles, including execution and verification, to models of different capability tiers. This is done because running a flagship model at every stage is expensive. Previous literature has established that verification stages are not always beneficial, but holds reviewer capability roughly fixed relative to the executor. We vary it. We replace the reviewer with models spanning a capability range down to one that cannot solve the problems at all, and measure the outcome of every individual rejection. This is done across a constant set of 100 olympiad mathematics problems. A cross-family mid-tier reviewer improves final accuracy by 12 percentage points, from 52 to 64 percent (p = 0.0005), with zero damaged answers. Same-model self-review attains the highest error-detection rate of any condition (0.85 recall) yet yields no significant gain: it rejects 2.1 times as often for a third the repair rate (15 against 43 percent, p = 0.0074) and falsely rejects 35 percent of its own correct answers against 2 percent for the cross-family reviewer (paired p = 0.000015). The low damage rate of self-review proves to be an artifact of revision inertia rather than reviewer quality: of 18 falsely rejected correct answers, the three where the executor complied all became wrong, while the fifteen it ignored survived unchanged. Below a capability floor the role becomes inert: our weakest reviewer changed zero of 100 final answers while doubling token cost. These findings describe a single executor-reviewer configuration on 100 problems and should be read as a controlled pilot rather than a general claim about verification stages. --- ### Corporate-Family Resolution Is Not a String-Matching Problem: A Public Benchmark Stratified by Name Visibility Authors: Harshit Gupta Published: 2026-09-02 Categories: cs.DB, cs.IR, cs.LG, cs.LO arXiv: https://arxiv.org/abs/2609.04269v1 Deciding whether two supplier records belong to the same corporate family is a prerequisite for spend consolidation, credit exposure aggregation and sanctions screening. It is usually treated as entity matching, but the tasks differ: a family link connects records that are deliberately different entities, and the evidence often appears in neither record. We introduce CorpFam, a public benchmark of 54,864 candidate pairs over 10,307 corporate families, derived from 6,638,350 US federal award records in which every supplier self-reports its ultimate parent to a government registry. Pairs are stratified by name visibility: whether the names are identical after normalisation, share a distinctive token, or share none. Because strata have positive rates from 10.2% to 97.3%, we report per-stratum recall, base-rate invariant, rather than F1, which is not. The strongest of 5 matchers recovers 100.0% of identical pairs and 4.2% of invisible ones; no method exceeds 4.7% on the latter. The failure begins before matching. Blocking decides which pairs a matcher sees, and we evaluate 7 schemes spanning phonetic keys, attribute keys that ignore the name, and semantic nearest neighbours. None reaches three percent on invisible pairs, and their union recovers 6.8%. 93.2% of these links never enter the candidate set, so no matching-stage improvement can reach them. The links are real: against SEC Exhibit 21 subsidiary schedules, which share no provenance with procurement registration, 64.2% of invisible links are corroborated, against 0.16% under permuted parents and 0.41% against the same parent's wrong exhibit: two unrelated nulls agreeing to within 0.25 points. Corporate-family resolution is a retrieval problem misfiled as a matching problem; the intervention point is candidate generation, not ranking. The benchmark, adjudication log, and code reproducing every number are released. --- ### A Data Fusion Framework for Grounding Aerospace Surrogate Model via Experimental Wind-Tunnel Observations Authors: Nitin Nagesh Kulkarni, Dheeraj Vemula, Yin Yu, Peter Lyu, Juan J. Alonso Published: 2026-09-02 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.04267v1 Aerodynamic surrogate models trained on high-fidelity CFD data reproduce numerical predictions of both scalar outputs and entire fields accurately, yet their predictive fidelity is limited by systematic discrepancies between CFD and experimental observations. We present an experimentally grounded correction framework that adapts a CFD-trained deep learning surrogate using wind-tunnel PSP measurements. A Geotransolver surrogate trained on 2,300 high-fidelity CFD simulations of the NASA CRM wing-body configuration, spanning geometric variation, Mach 0.70-0.85, and angles of attack 0 to 4 degrees, reproduces the CFD integrated aerodynamic forces and pitching moment to R2 > 0.99 but does not match the experimental data. To incorporate experimental information without retraining the surrogate, a correction network is trained on spatially registered PSP measurements at two freestream Mach numbers (0.70 and 0.85) across the same angle-of-attack range, learning the discrepancy between the surrogate-predicted and experimentally measured surface-pressure distributions. At Mach 0.85 the correction substantially improves agreement with PSP, particularly at the wing suction peak, shock location, and subsequent pressure recovery, reducing both the magnitude of the prediction error and the fraction of wetted surface on which it exceeds 0.05 in Cp, and it does so from a limited experimental dataset without modifying the pretrained surrogate parameters. On held-out angles of attack the grounded surrogate agrees with measurement to within 2.3-2.7% of the measured Cp range, and outperforms direct interpolation between the measured conditions at every state tested. Experimental measurements can therefore ground a large-scale simulation-trained surrogate by learning systematic CFD-to-experiment discrepancies while preserving its generalization capability and computational efficiency. --- ### Compute-in-Memory Attention: A Time-Domain Analog Softmax Circuit with RC-Tunable Temperature Authors: Ankur Singh, Ashish Gautam, Shruti R. Kulkarni, Guojing Cong Published: 2026-09-02 Categories: cs.AR, cs.LG arXiv: https://arxiv.org/abs/2609.04266v1 Softmax is a key operation in Transformer attention, but its exponentiation and normalization add significant overhead in compute-in-memory (CIM) accelerators, especially when analog attention scores must first be converted to the digital domain. This work presents a tunable-temperature analog softmax circuit in GlobalFoundries 22-nm fully depleted silicon-on-insulator (FDSOI) technology that operates directly on CIM-generated score voltages without intermediate analog-to-digital conversion. Each input score is converted into a time-domain event using a shared falling ramp. The corresponding comparator transition samples an RC-decaying reference to generate an exponential weight, which is then processed by an in-circuit normalization stage. In contrast to analog softmax circuits that rely on transistor weak-inversion behavior for exponentiation, the proposed architecture controls the softmax response through the ramp slope and RC time constant, enabling programmable effective temperature. The 128-element architecture is evaluated using transistor-level and post-layout extracted simulations, including multi-level input vectors, capacitance variation and mismatch, process and temperature variation, monte carlo analysis, and shared-interconnect parasitics. The complete 128-element implementation occupies 9453.42~$μ\mathrm{m}^{2}$ including the shared global ramp circuitry, while each replicated softmax element occupies 70.2~$μ\mathrm{m}^{2}$. The circuit achieves a 242.97-ns evaluation latency at 13.44~mW total power, corresponding to 25.5~pJ per output element. The simultaneous 128-element evaluation achieves an RMSE of 24.46~mV relative to the ideal softmax response. The extracted circuit characteristics are further incorporated into a MemTorch-based hardware-aware Transformer model, where the proposed softmax achieves a validation loss within 2.5\% of the ideal-softmax baseline. --- ### ProToMEx: Rapid, Interpretable Explanations via Structured Representations Authors: Athina Georgara, Adarsh Valoor, Sarvapali D. Ramchurn Published: 2026-09-02 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.04265v1 Existing post-hoc explainers for machine learning classifiers primarily focus on feature attribution, assigning importance scores to individual features. While valuable, this approach struggles to articulate the complex, combinatorial patterns that often drive a model's decision-making process. To overcome this limitation, we introduce ProToMEx, a new paradigm for explainability that leverages Probabilistic Topic Models (PTMs). Our model-agnostic framework learns latent ''topics'' that represent distinct, high-level reasons for a classification, moving beyond simple feature importance to reveal underlying semantic structures. ProToMEx naturally provides both global explanations of a model's overall behaviour and local explanations that can disentangle multiple co-existing reasons for a specific prediction. We demonstrate empirically that ProToMEx not only produces explanations of comparable fidelity to popular methods like SHAP and LIME but also drastically reduces the amortised computational cost of generating local explanations, making it highly suitable for real-time applications. Specifically, we show that ProToMEx is ~30-40x faster than SHAP and LIME over standardised tabular datasets and synthetic datasets. --- ### Spectral-Target Physical Latent Structuring for JEPA-Style World Models Authors: Penghao Zhu, Salvatore Penachio, Kaustav Mukherjee, Aneesh Jonelagadda Published: 2026-09-02 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.04264v3 Latent world models have become increasingly popular as a method to predict and plan in latent space rather than pixel space. Recent architectures, such as LeWorldModel (LeWM), jointly train the encoder and predictor using regularization techniques like SIGReg to prevent representation collapse. Even with such regularization preventing representation collapse, we identify a new world model failure mode of physical representation laziness, particularly noted in highly dynamic environments. For these lazy cases, the learned latent states do not collapse but nonetheless fail to represent key physical properties, causing ubiquitous downstream planning failure. To resolve this issue, we propose training-time auxiliary supervision with a lightweight "Fourier auxiliary head", which enforces physically-informed structuring of the latent space with no additional inference-time cost and can be generalized to any environment. Experimentally, we show that the auxiliary head substantially improves planning success rates in dynamic environments where the baseline LeWM exhibits physical representation laziness. It also leads to modest improvements in other environments, even when the baseline does not exhibit physical representation laziness. We further observe superior planning performance being accompanied by higher latent space correlations with key physical properties, indicating both the ability of our method to physically structure latent states and the potential planning-side benefit to the learned representation being physically structured. We also see in low-data regimes, auxiliary supervision is particularly impactful in increasing success rate. These findings support the use of our Fourier auxiliary head method to improve both overall success rate and data efficiency, while avoiding representation laziness in latent world models. --- ### Quality Recovery for Quantized KV Caches via Low-Rank Attention Adaptation Authors: Seifeldin Abdellatif Published: 2026-09-02 Categories: cs.AR, cs.CL arXiv: https://arxiv.org/abs/2609.04263v1 Low-bit key--value (KV) caches reduce the memory required for autoregressive decoding, but the resulting quality loss depends on the model and quantizer. We keep the quantizer fixed and distill the floating-cache model's behavior into low-rank Q/K/V projection updates while the student executes a physically packed incremental cache. Across three seeds, 4-bit affine-cache adapters recover $54.24\%\pm2.47\%$ of the held-out perplexity gap on TinyLlama-1.1B and $75.96\%\pm4.04\%$ on Gemma-4-12B. On the same frozen NF4 Llama-3.1-8B base, one validation-selected run per quantizer recovers $60.42\%$ under KIVI K2V2 and $37.61\%$ under KVarN K4V2, while preserving 180-case associative retrieval. Gemma's score on an official 4K/8K RULER subset rises from 42.80 with the unadapted 4-bit cache to 48.33 after adaptation (46.15 floating), with substantial task heterogeneity. Finally, a 2-bit rank--token sweep reduces TinyLlama's 2-bit PPL from 576.10 to $11.4000\pm0.0059$ across three seeds, versus 10.3988 floating, but restores only 11--12 of 180 retrieval cases. These results show that low-rank projection adaptation can recover held-out quality across fixed cache formats, while perplexity recovery need not restore long-context retrieval. --- ### Low-Latency Spell Correction for Japanese Music Search Queries Authors: Anshul Garg, Pavni Tandon, Karan Bhukar, Tanmay Khandelwal, Ujjal Kumar Dutta Published: 2026-09-02 Categories: cs.SD, cs.LG, eess.AS arXiv: https://arxiv.org/abs/2609.04262v1 Spell correction for Japanese search queries presents unique challenges due to the co-existence of four writing scripts (Latin/romaji, hiragana, katakana, and kanji) and the distinct error patterns each script induces. We present a compact BART-based sequence-to-sequence model (3 encoder + 3 decoder layers) designed for low-latency spell correction of Japanese music search queries. The core contribution lies in a script-aware synthetic misspelling generation pipeline that produces realistic training data by combining keyboard-layout models (QWERTY and flick input), phonetic confusion priors mined from real query logs, voiced/unvoiced consonant alternations, and kana case errors. A key design decision is normalizing mixed-script catalog titles to a single canonical script before misspelling synthesis, which we show is critical for reducing model hallucinations. We train a custom byte-level BPE tokenizer on the target music catalog to handle all four scripts in a unified vocabulary. Experiments on a curated evaluation set show that our model achieves an exact-match accuracy of 41.09% and a character error rate (CER) of 11.62%, outperforming edit-distance baselines and achieving the lowest character error rate among all evaluated systems while maintaining sub-4ms inference latency on a single GPU. We further analyze performance across individual scripts and mixed-script queries, demonstrating the effectiveness of script-aware data augmentation through systematic ablation studies. --- ### Self-Supervised Pretraining of Molecular Graph Encoders with LeJEPA Authors: Michał Kulczykowski, Rafał Łabędzki Published: 2026-09-02 Categories: q-bio.QM, cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.04261v1 Self-supervised pretraining has transformed language and vision, but its value for molecular graph neural networks remains contested. We ask whether pretraining on a large unlabelled corpus improves molecular property prediction. We adapt LeJEPA, a predictor-free joint-embedding predictive architecture regularised by Sketched Isotropic Gaussian Regularisation (SIGReg), to molecular graphs, evaluating GPS and Chemprop-style D-MPNN encoders on the Wong et al. (1) antibiotic-activity dataset and ogbg-molhiv using a multi-seed, bootstrap-based protocol. Pretraining improves learned representations but does not robustly improve finetuning. A frozen probe on pretrained embeddings exceeds random initialisation on both tasks (ogbg-molhiv ROC-AUC 0.788 vs 0.665; +0.123), reaching the published self-supervised band, but this does not translate into finetuning gains. On the antibiotic scaffold split, a canonical partition is significant (delta AUPRC +0.041, p = 0.010), but the effect vanishes across five partitions (pooled +0.013, p = 0.095). Finetuning is null on the random split, ogbg-molhiv, and D-MPNN. The representational edge is nevertheless recoverable. Embeddings saturate at ~16-32 effective dimensions, whereas Morgan fingerprints improve to 1024 bits. At matched dimensionality, fingerprints lead validation (0.799 vs 0.782 at 128 dimensions) but trail shifted test scaffolds (0.759 vs 0.788). Truncating embeddings and combining them with a 1024-bit Morgan fingerprint raises ogbg-molhiv ROC-AUC from 0.805 to 0.832 (delta +0.027; 95% CI (+0.003, +0.054); p = 0.014); an untrained encoder gains nothing (delta -0.003). Thus, pretraining supplies complementary information best realised through feature-level combination, while finetuning gains are weak and partition-dependent. --- ### Instability Floors: Separating Bias from Noise in Fairness Audits of Clinical LLM Agents with FairMedAgent Authors: Rohith Reddy Bellibatlu, Manpreet Singh, Deepak Parashar, Rahul Joshi Published: 2026-09-02 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.03221v3 Counterfactual fairness audits of clinical language-model agents report a flip rate: how often an action changes when only the patient's demographic descriptor changes. Part of that rate is not demographic. A stochastic agent also changes its own action when nothing changes, and a flip rate cannot be interpreted without knowing how often. We measured it. Re-running one condition ten times over sixteen synthetic vignettes at default sampling changed a clinical agent's action in 8.7 percent of replicate pairs, from 2.2 percent for intensive-care escalation to 17.9 percent for controlled-substance caution, an output given no operational criteria. Across six models from five vendors, pooled floors ranged from 2.5 to 23.7 percent; in this panel neither disclosed size, vendor, nor hosting ordered them. The floor depends on the decoding configuration: majority voting over five draws removed 39 percent of it (95 percent confidence interval (CI) 18 to 64); at temperature 0 three of four locally served models showed no disagreement, but a hosted model still did. We also show that, for a binary action, the flip rate expected under no demographic effect equals the floor and a real effect adds only its square, so a flip rate inside the floor is not evidence of fairness, and direction must be tested with a signed paired test. We give a four-step reporting procedure and release FairMedAgent, the harness, with its protocol, vignettes and analysis scripts, so any team can measure the floor for its own agent. --- ### The Analyst in the Prompt: Role, Retrieval, and Memory Biases in LLM Financial Analysis Authors: Ahmed Asaad, Amr Mohamed, Yang Zhang, Omneya Abdelsalam Published: 2026-09-02 Categories: cs.CL, cs.CE, q-fin.PM arXiv: https://arxiv.org/abs/2609.03218v1 Large Language Models (LLMs) increasingly use user context such as memory, profiles, and role prompts to personalize their responses. This personalization can affect evidence-based judgment: the same evidence may lead to different conclusions under different user contexts. Finance provides a high-stakes setting to study this problem because decisions often depend on interpreting long and complex documents. We test this using 3,575 SEC filings across twelve LLMs. We compare persona-conditioned retrieval, neutral retrieval, and memory-framed context to separate the effect of evidence selection from the effect of interpretation. We find that most user-context spillover comes from how models interpret the same evidence under different roles, rather than from retrieving different evidence. We then test two simple mitigation strategies: expressing the same investor mindset as a user profile instead of an assistant role, and separating evidence-based and personalized outputs. Both reduce spillover, but neither removes it completely, and their effectiveness varies substantially across models. --- ### ProgResViT: Progressive Resolution and Width for Adaptive Vision Transformers Authors: Ali Hojjat, Janek Haberer, Olaf Landsiedel Published: 2026-09-02 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.03216v1 Vision Transformers (ViTs) typically process every image using a fixed input resolution and model width, even though many images can be classified with substantially less computation. We introduce ProgResViT, an input-adaptive ViT that performs inference progressively across multiple rounds. The first round processes a low-resolution image with a narrow subnetwork. Inference terminates when the prediction is sufficiently confident; otherwise, the model reuses the representations produced in the current round and proceeds with a higher-resolution input and a wider subnetwork to refine its prediction. As all rounds share a single backbone, we propose Progress-Conditioned Soft Gating (PSG), which conditions token fusion and layer outputs on the current round, block, and input resolution. On image classification, applying ProgResViT to DeiT yields better accuracy-compute trade-offs than adaptive-width, adaptive-depth, and dynamic-token baselines. With knowledge distillation, a DeiT-based ProgResViT achieves 84.9% top-1 accuracy, slightly exceeding the reported DeiT-III-S accuracy under a comparable evaluation setting. We show that the same design also provides favorable accuracy-compute trade-offs for self-supervised DINO representations and downstream semantic segmentation. Code is available at https://github.com/ds-kiel/ProgResViT. --- ### SWIM: Student Writing Simulation via Proficiency-Conditioned Generation Authors: Heejin Do, Jakub Kontak, Mrinmaya Sachan Published: 2026-09-02 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.03215v1 Writing proficiency manifests in how students develop content, organize ideas, choose words, and use language. Despite growing interest in LLM-based student simulation, whether LLMs can reproduce such multidimensional variation in extended writing remains largely unexplored. In this work, we explore if language models can realistically simulate student writing, and introduce SWIM, a task that formulates Student Writing sIMulation as proficiency-conditioned essay generation. We evaluate prompting, supervised fine-tuning (SFT), and reinforcement learning (RL) methods for writing simulation using automated essay scoring as a measure of profile alignment. Extensive experiments reveal that prompting provides limited proficiency control, even for strong proprietary LLMs with rubric-grounded strategies. In particular, while models can adjust content-oriented traits, they struggle to reproduce the lexical, grammatical, and organizational variation in different proficiency levels. SFT substantially improves alignment, while RL with the proposed proficiency-alignment reward yields further gains across all writing traits and essay prompts. Our findings suggest that explicit supervision enables substantially stronger profile alignment than prompting alone, while authentic low-proficiency writing remains challenging to reproduce. ---