# Research Papers — 2026-09 · bundle page 49 of 266 > Full-text markdown of papers #2401–2450 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/49 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/48.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/50.md) ### Text-Vision Synergistic Token Caching: A Training-Free Framework for Efficient Vision-Language-Action Inference Authors: Qianer Li, Chengjie Zhang, Jingwen Chen, Zanjia Tong, Jiyuan Zhang, Hong Zhang Published: 2026-09-28 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.34319v1 Vision-Language-Action (VLA) models enable generalizable robotic control but remain computationally expensive. Token caching provides a training-free, plug-and-play acceleration alternative. However, existing VLA caching does not fully exploit a key inductive bias of VLA models: text-vision synergy, wherein textual semantics guide the precise visual grounding of task-relevant regions. In particular, existing designs insufficiently account for head-wise reliability in attention aggregation and layer-wise stability in cache reuse. To address this, we propose Text-Vision Synergistic Token Caching (TVCache), a training-free framework for efficient VLA inference. TVCache filters attention heads based on text-vision information focus to improve task-relevant and physically consistent visual grounding. Concurrently, we introduce a reuse-layer selection mechanism guided by text-vision entropy differences to avoid caching unstable representations and improve cache resource allocation. Extensive experiments across four representative VLA models, two simulation benchmarks, and real-world robotic tasks demonstrate the effectiveness and generality of TVCache. At matched token-retention ratios, TVCache consistently improves task success over existing VLA caching with comparable computational cost. On OpenVLA-OFT, it improves average success by up to 14.5 percentage points over VLA-Cache at 12.5% retention while reducing FLOPs by 2.45x relative to full-token inference. --- ### Dynamical Parameters: An Interpretability Framework for Time-Series Foundation Models Authors: Kang Yang, Gaofeng Dong, Liying Han, Mani Srivastava Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.34316v1 This work studies a central gap in interpreting time-series foundation models (TSFMs): a dynamical property may be accessible in a hidden state even when the forecast fails to respond correctly as that property changes. We formalize these properties as Dynamical Parameters, including trend slope, oscillation frequency, and autoregressive dependence. We compare their representation accessibility, measured by recovery from hidden states, with their forecast response, measured by agreement with the expected forecast change. Across nine frozen TSFMs and thirteen laws, 42 of 63 model-parameter cells achieve accessibility above 0.95, whereas their median reference-aligned response relative to the conditional reference is only 0.46. To explain this gap, causal geometry compares the hidden-state change required to produce the reference response with the change induced by the parameter intervention. Directly modifying the hidden state recovers the reference response, but the parameter intervention often moves the state in a different direction. These results show that accessible parameter information need not be expressed in forecasts when input changes miss the required hidden-state direction. --- ### PlaylistEval: Can Video-Language Judges Be Trusted at Day Scale and Beyond? Authors: Shayekh Bin Islam, Hwanjun Song Published: 2026-09-28 Categories: cs.CV, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.34314v1 Video-language models are increasingly used as judges of video understanding, both for evaluating model outputs and for training reward models. Whether their judgments remain reliable when the evidence is buried in day-long videos has yet to be established. Existing benchmarks cannot answer this. Their videos are typically only a few minutes long, many answer pairs can be separated from the transcript alone, and collecting human judgments does not scale to ultra-long videos. We introduce PlaylistEval, an agentic framework that builds video-language judge benchmarks over 100-hour playlist collection without human annotation. It automatically generates questions with paired answers whose differences are controlled by causal degradation, so that every pair demands retrieval across the collection. The resulting benchmark contains 630 pairs across seven domains spanning both static and dynamic knowledge, and on a stratified subset of 152 pairs it agrees with human judgments 93.0% of the time (IAA 0.781). Evaluating 17 omnimodal and multimodal models from eight families reveals that frontier judges reach only 75.4% pairwise accuracy, while open-source judge models perform far behind. We further show that both retrieval and final judgment depend on using multiple modalities, and that judge accuracy degrades as the playlist set grows. We release our pipeline, benchmark, and evaluation code at https://playlisteval.github.io. --- ### ControlScope: Workflow Revision and Reliability in LLM Agents Authors: Jingjie Ning, Xueqi Li, Yibo Kong, Dongting Li Published: 2026-09-28 Categories: cs.AI, cs.CL, cs.SE arXiv: https://arxiv.org/abs/2609.34313v1 How much of a running workflow should a language model agent revise? ControlScope compares continuing generated code, editing the next tool call's data arguments, and replacing the unfinished workflow from the same public execution state. The nested permissions separate available repairs from the actions an agent selects. We evaluate one-time and repeated reviews across filesystem tasks, ALFWorld, and AppWorld. Across two source programs per task and three reasoning-reviewer draws on 20 filesystem tasks, FULL completes 15-16 tasks versus 13 for KEEP; across four fast draws it completes 10-13 versus 13. Fresh student-record confirmation reproduces a batch-read repair. ALFWorld fast panels yield KEEP/ARG/FULL scores of 85/86/87 on 87 tasks across 52 scenes and 134/134/127 on 134 tasks across four scenes; reasoning on the 87-task cohort also yields 85/86/87 with substantial review cost. An AppWorld V1 official-test panel of 585 task instances from 195 scenario templates shows small net differences. Frozen replays expose viable agent-written replacements interrupted by later revision in two failed file-organization runs. An offline source-trajectory midpoint comparison shows later reviews completing an insufficient repair. Five-call protection saves 19.4% of logged model output and loses one success across 20 fresh source runs. An argument-only shortcut shows that the broader sampled policy can overlook a cheaper successful edit available in both operation sets. These outcomes tie repair access to actual choices and subsequent execution. --- ### Riemannian Difference-of-Convex Optimization for K-Means Clustering Authors: Meng Xu, Bo Jiang, Hanfu Zhang, Ya-Feng Liu, Anthony Man-Cho So Published: 2026-09-28 Categories: cs.LG, math.OC arXiv: https://arxiv.org/abs/2609.34310v1 K-means is a widely adopted clustering approach in signal processing and machine learning. In this paper, we study K-means clustering through a cardinality-constrained formulation on a compact embedded submanifold. We replace the cardinality constraint with a difference-of-convex (DC) penalty and establish a global error bound to prove that the penalized and constrained formulations share the same global minimizers whenever the penalty parameter exceeds a finite threshold. To solve the resulting nonsmooth Riemannian DC problem, we reformulate it as a minimax problem and propose RADA-DC, a Riemannian alternating descent ascent method combining dual regularization with DC linearization. Under standard assumptions and suitable parameter choices, RADA-DC finds an $ε$-Riemannian critical point within $O(ε^{-3})$ iterations. We conduct experiments on synthetic and real-world datasets to demonstrate that the proposed method outperforms the tested baselines, including K-means++, in solution quality at competitive computational cost when the number of clusters is large. --- ### MaLiang-Harness: A Programmable Path to Image and Video Generation Authors: Haoyu Zhao, Zihao Zhang, Xudong Wang, Jiaxi Gu, Zuxuan Wu, Yu-Gang Jiang, Shuicheng Yan Published: 2026-09-28 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.34309v1 Executable programs offer explicit control over how images and videos are constructed, but generating runnable code is only the beginning of visual creation. A program can execute correctly while violating the requested composition, appearance, or motion. We define this discrepancy as the Program-to-Visual (P2V) gap and introduce MaLiang-Harness, a unified framework for organizing MLLM-driven visual generation into a persistent process of construction, inspection, and revision. Its central design is to make the evolving visual program, its construction history, and its verification share a common revision reference. We define the Persistent Executable Generation (PEG) state as preserving programs and task context. Traceable Generation Process (TGP) connects edits to rendered evidence, and Revision-aware Editing and Verification (REV) supports restoration and checks the current revision before completion. Together, these mechanisms coordinate planning, execution, and visual feedback across rendering backends. We evaluate 11 powerful closed-source MLLMs on MaLiang-IBench and four on MaLiang-VBench, measuring generation success, visual quality, and computational cost. GPT-6-Astra achieves 100% generation success on both benchmarks, with 96.0% of image tasks and 76.9% of video tasks meeting all quality thresholds. The comparison also reveals a mismatch between general capability scores and visual generation performance, with similarly scored models differing substantially in their ability to satisfy visual requirements. MaLiang-Harness provides a systematic basis for studying how MLLMs translate executable code into visual outcomes, exposing both the potential of programmable generation and the limitations of general benchmarks as predictors of this ability. The project is available at https://github.com/gulucaptain/MaLiang-Harness. --- ### CAT-Free: Multi-View Pedestrian Localization without Calibration, Annotations, or Target-Scene Training via Adaptive Geometric Filtering Authors: Taigo Sakai, Hiroki Kouno, Naoki Kato, Kazuhiro Hotta Published: 2026-09-28 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.34302v1 Multi-camera pedestrian localization is useful for wide-area monitoring in public and commercial spaces. However, deploying these systems often requires considerable setup for each new environment. Existing methods typically require camera calibration, position annotations, or target-scene training. CAT-Free removes all three requirements. It uses synchronized RGB video as its only scene-specific input. Camera configuration is estimated directly from the video. Pedestrian locations are then estimated by combining observations from multiple cameras. Automatic camera estimation is not always accurate. This can produce unreliable pedestrian locations. CAT-Free therefore introduces two adaptive geometric filters. They remove unreliable position estimates. Their thresholds are estimated from each input sequence. CAT-Free achieves 82.5, 84.5, and 65.7 MODA on WildTrack, MultiviewX, and GMVD. It uses no supplied calibration, position annotations, or target-scene training. Published methods using such scene-specific information report 88.2--95.0 MODA on WildTrack and 83.9--96.5 on MultiviewX under their respective protocols. CAT-Free also transfers without retuning. It reaches 74.9 MODA on four additional sequences and 78.6 on an unseen 8-camera installation. Finally, localization uncertainty predicts MODA with $r=-0.98$. This provides a label-free estimate of localization reliability. --- ### One Sequence, Many Decodings: CAGenMol-2 Recasts Drug Design as Masked Molecular Inference Authors: Yanting Li, Enyan Dai, Lei Wang, Wen-Cai Ye, Li Liu Published: 2026-09-28 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.34301v1 Drug design couples property evaluation, conditional generation, structure-based design, and local optimization, yet machine learning systems typically address these capabilities with separate task-specific models. We introduce CAGenMol-2, a masked diffusion molecular language model that represents molecules, continuous scalar properties, and 3D protein pockets within a single wrapped sequence. Within this pretrained interface, downstream operations are selected by which sequence regions are observed or masked at inference, allowing one checkpoint to perform property prediction, property- and pocket-conditioned generation, and partial-constraint design without task-specific architectures or backbone fine-tuning. We further propose Adaptive Fragment Optimization (AdaFO), a gradient-free mask-and-refill search that turns the masked decoder into an iterative local molecular optimizer. On CrossDocked2020, AdaFO increases Success Rate from 30.2\% to 70.8\%, the best reported under this protocol, while largely preserving drug-likeness and diversity. Finally, scaffold-preserving directional editing and CRBN/VHL case studies demonstrate its use in compound design workflows spanning local molecular editing, structure-based prioritization, and downstream simulation-based screening. --- ### When World Models Lie: Adaptive Safety Analysis Under Wrong Imaginations Authors: John Cao, Somil Bansal Published: 2026-09-28 Categories: cs.RO, cs.AI, eess.SY arXiv: https://arxiv.org/abs/2609.34300v1 World models offer a powerful substrate for safety reasoning in high-dimensional robotic systems, but they are also fallible: their predictions can be biased, miscalibrated, or confidently wrong. This creates a central challenge for latent-space safety filters, which often learn Hamilton-Jacobi safety value functions on the dynamics of a world model. If the world model is incorrect, the resulting value function can inherit its errors and produce overconfident safety estimates. Existing latent safety filters often rely on auxiliary signals such as ensemble disagreement or value-target consistency residuals for adaptation, but these signals can remain small even when the world model's predictions deviate from observations. We propose an adaptive latent safety filter that calibrates safety reasoning using directly observed world-model error. Our method uses Adaptive Conformal Inference to construct online uncertainty sets from discrepancies between predicted and observation-inferred latent states, then evaluates safety pessimistically by minimizing the learned value function over these sets. This allows the filter to remain minimally conservative when the world model is accurate, while becoming more cautious when observations reveal model mismatch. We provide a finite-time coverage guarantee for the adaptive uncertainty radius. Through simulation and hardware experiments, we show that our method significantly reduces failures relative to state-of-the-art latent safety filters while preserving task completion. --- ### PSM: Dataset Distillation Based on Precise Statistical Matching by Difficulty Authors: Hongxu Ma, Guang Li, Shijie Wang, Dongzhan Zhou, Suorong Yang, Baoli Sun, Takahiro Ogawa, Miki Haseyama, Zhihui Wang Published: 2026-09-28 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.34299v1 Dataset distillation (DD) condenses a large original dataset into a small distilled dataset with high training utility. Decoupled statistical matching methods substantially reduce distillation time and memory overhead while achieving strong performance. However, they typically supervise all distilled samples using running statistics estimated from the entire original dataset. These statistics mainly capture the average feature distribution while overlooking differences in sample difficulty, limiting their ability to characterize the difficulty structure of the original data. To address this issue, we propose Precise Statistical Matching (PSM) by difficulty. After pretraining, PSM uses the Global Precision Score (GPS) to estimate image difficulty, ranks the samples within each class, and partitions each class into IPC (images per class) difficulty groups. During distillation, Statistics Updated Again (SUA) updates the teacher's batch normalization (BN) running statistics through forward passes on original samples from each group, providing difficulty-specific supervision for the corresponding distilled batch. Meanwhile, Initial Sample Screening (ISS) initializes distilled samples using original images from the corresponding difficulty group, providing an effective starting point for precise matching. Experiments across multiple datasets and model architectures demonstrate that PSM broadens the difficulty range of distilled samples and improves downstream performance in most evaluated settings. Code will be released. --- ### Dr.Credit: Rubric-Grounded Process Credit Assignment for Deep Research Agents Authors: Yingjian Zhu, Zhenyi Wang, Jiaxin Guo, Kun Ding, Ying Wang, Shen Huang, Xunjie Zhu, Pengjun Xie, Shiming Xiang Published: 2026-09-28 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.34296v1 Rubric-based tasks are increasingly addressed through reinforcement learning (RL), with rubric scores used as training rewards. However, these rewards typically supervise final answers without distinguishing the contributions of intermediate decisions. Many existing credit assignment methods rely on ground-truth answers to define process rewards, limiting their applicability to open-ended tasks without canonical solutions. To address this limitation, the proposed rubric-grounded credit uses task requirements as a shared reference for final answer evaluation and process supervision. The information returned by tools is assessed for the additional support it provides toward satisfying each rubric relative to that rubric's history of accepted support. By referencing these histories, credit distinguishes new support from evidence already present in the trajectory while recognizing partial support for each rubric. Dr.Credit uses rubric-grounded credit to supervise intermediate tool turns in an RL framework for deep research agents. The resulting process advantages are combined with GRPO outcome advantages to guide research decisions while retaining supervision of final-report quality. Evaluations on four in-domain and out-of-domain benchmarks show that Dr.Credit outperforms the evaluated open deep research baselines on every primary metric and submetric. Meanwhile, with an 8B-parameter backbone, the trained agent achieves average performance competitive with the evaluated frontier proprietary models. Further analyses suggest more efficient evidence acquisition and higher-quality reports under limited research-turn budgets, motivating the extension of rubric-grounded process supervision to a broader range of rubric-based tasks. --- ### Semantic Modality Compensation for Unsupervised Visible-Infrared Person Re-identification under Unpaired Settings Authors: Duanning Chen, Ke He, Bin Yang, Yongxiang Yao Published: 2026-09-28 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.34294v1 Unsupervised visible-infrared person re-identification (USL-VI-ReID) learns person representations that can be compared across modalities without identity annotations. In the unpaired setting, however, identity correspondences between modalities are often incomplete, leaving many identities without an observed counterpart in the other modality. Existing unpaired methods bridge this gap by generating or mapping features for the other modality, mainly by exploiting the statistics of visual features without explicitly separating content that is discriminative for identity from style that is specific to modality. Consequently, the generated features may distort identity cues or inherit bias from the source modality, undermining the reliability of supervision across modalities. We formulate unpaired learning across modalities as a semantic compensation problem and propose Semantic Modality Compensation (SMC), a framework based on prompt composition that decouples identity semantics from modality style within a shared visual semantic space. SMC first constructs a discriminative ReID space through augmented dual contrastive learning, yielding pseudo labels, cluster prototypes, and memory banks for each modality. It then learns visible and infrared modality prompts in the CLIP semantic space and maps clusters obtained from pseudo labels to identity semantic tokens. For each cluster lacking a reliable match in the other modality, SMC combines its identity token with the prompt for the target modality to synthesize a semantic counterpart in the missing modality. The synthesized counterpart is then projected back into the ReID space and injected into a compensation memory through confidence gating. Extensive experiments under both paired and unpaired settings demonstrate that SMC consistently outperforms state-of-the-art methods, with particularly large gains when identity mismatch is severe. --- ### Pre-registered tests of solid-state-physics-inspired LLM compression: a cluster-level negative result at small-language-model scale Authors: Jun-qiang Lu Published: 2026-09-28 Categories: cond-mat.dis-nn, cond-mat.mtrl-sci, cond-mat.other, cs.LG arXiv: https://arxiv.org/abs/2609.34292v1 We report a three-month autonomous research-agent program testing five solid-state-physics-inspired compression mappings on pretrained language models, with predictions committed to git before any pilot data and a 3-sigma gate deciding PASS or SHELVE. The common anchor -- area-law / Kohn-nearsighted decay of the one-particle density matrix -- has a distance face (P001 Wannier, P002 tight-binding) and a rank face (P003 DMRG-truncated MLPs, P005 Wilson-RG, P011 tensor-train embeddings). P005 was pre-empted at Phase 1; three of four Phase-3 pilots were falsified. On the attention face, GPT-2-medium attention-versus-distance is best fit by a stretched exponential in 12 of 16 median-layer heads once probe padding is excluded, and a tight-binding cutoff costs +96% perplexity (P002); on Pythia-160M the Wannier sparsity 0.054 +/- 0.004 is indistinguishable from PCA, random-Haar and identity baselines (P001). On the rank face, per-token tensor-train bond dimension does not track surprisal (r = 0.016 vs a pre-registered 0.65) and the format inflates rather than compresses (P011). P003 is mixed: its scaling claim shelved (r = -0.434), its MPO premise died at stage-0, and its cross-paper check, r = 0.523 as first written, collapses to 0.047 under the same correction, leaving both cross-paper checks null. The results invert the pre-registered prediction that most attention heads behave like Kohn-nearsighted insulators, pointing instead to critical, glassy or heavy-tailed regimes; the inversion is specific to the <= 350M scale tested, while the rank-face no-gain result held to 7-8B. We contribute the pre-registration + 3-sigma + cluster-framing + append-only-catalogue discipline -- including why our own enforcement gate was designed but not deployed -- four pre-registered negative results with full data release, and the inversion. The catalogue holds eighteen concluded studies, seventeen negative. --- ### ReScraper: Unified Scraping and Cleaning of Web Data for Effective LLM Pretraining Authors: Zichun Yu, Jiarui Yan, Shlok Sanghvi, Nihar Atri, Chenyan Xiong Published: 2026-09-28 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.34287v1 LLM pretraining corpora are normally cleaned by a stack of hand-written heuristics. A heuristic scraper extracts the main content from HTML, and dozens of rule-based filters then clean it, so corpus quality is capped by the coarseness and accuracy of the rules. In this work, we propose ReScraper, a unified language model of only 0.6B parameters that replaces this entire stack. To train ReScraper, we carefully curate supervised data from the outputs of three teacher models, so it learns to first extract the main content from raw data and then choose among four operations: keeping the page as extracted, editing out noisy lines and spans, deleting it entirely, or rewriting it when it is poorly written but informative. Based on the same crawled data pool, pretraining 400M, 1.4B, and 2.8B models on our curated data improves the DCLM Core score by a relative 3.8--4.7% over the strongest baseline at each scale, including the costly multi-agent curation. Our analyses show that each operation plays a distinct and complementary role, and that extracting and cleaning in one model outperforms a cascade of separate models. ReScraper also concentrates its operations on the pages that need them, raising the quality of poor pages the most while keeping the corpus diverse. These results demonstrate the feasibility and effectiveness of AI4AI for pretraining data curation, where a small learned model takes over an entire stage of the pipeline from hand-written heuristics. We open-source our code at https://github.com/cxcscmu/ReScraper --- ### Dexterous Tactile World Model Authors: Ziyao Zeng, Xiatao Sun, Hao Wang, Yueyang Pan, Zhengxiang Yu, Fengyu Yang, Tianyu Liu, Zhiwen Fan, Daniel Rakita Published: 2026-09-28 Categories: cs.CV, cs.LG, cs.RO arXiv: https://arxiv.org/abs/2609.34286v1 World models for manipulation are typically trained from video, yet the events that determine how manipulation unfolds, such as making and releasing contact, are difficult to observe visually and are often easier to sense through touch. We present the Dexterous Tactile World Model (DTWM), a video world model for future-frame prediction of egocentric manipulation from both observed video and tactile signals from a glove worn on each hand. We condition a pretrained video diffusion transformer on each hand's tactile signal through a zero-initialized residual at the corresponding hand location in the video tokens, while a causal mask prevents predicted frames from accessing future information. Compared with a vision-only model matched in architecture, parameters, and training, DTWM reduces the underestimation of hand motion from 23% to 9%, while reducing the perceptual error in the hand region by 7.4% across three training runs per model. The benefit also increases over the prediction horizon, with the improvement in the later predicted chunks being about 4.1x larger than in the first. DTWM also outperforms other visual-tactile world models under the same setting, and training with touch improves future-frame prediction even when no touch is available at inference. Ablations show that the model benefits from both the magnitude and spatial location of force: replacing the tactile signal with binary contact states, either per hand or per location, increases prediction error. The observed course of the force indicates whether the interaction will persist or change. --- ### Epistemic Learning from Imprecise Annotation Authors: Kaizheng Wang, Siu Lun Chau Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.34285v1 Imprecise annotations may support several plausible labelling distributions, yet learning methods often resolve this ambiguity into a single predictive distribution. This can obscure what the annotation evidence leaves unresolved. We introduce epistemic learning from credal supervision, a framework that uses convex sets of plausible labelling distributions, called credal sets, as supervision and learns sets of predictive distributions. We instantiate the framework with the pessimistic--optimistic credal classifier (POCC), which combines a shared backbone with two classification heads trained to minimise worst-case and best-case losses over the supervision sets. Their outputs define a predictive credal set whose spread provides an uncertainty score. We also show how credal labels can be obtained through a simple relaxation of existing probabilistic labels, reducing commitment to their precise probability assignments. This construction admits closed-form inner optimisation under cross-entropy loss, enabling efficient training. Assuming the supervision sets contain the true conditional label distributions, and other regularity assumptions, we establish a finite-sample generalisation bound for the averaged predictor with an explicit penalty for supervision imprecision. We evaluate POCC using human annotator disagreement and teacher predictions, alongside label smoothing as a controlled proxy for annotation imprecision. Across these settings, POCC achieves a favourable balance of predictive accuracy, calibration, and uncertainty-based selective classification versus competitive baselines. --- ### Over-Personalization Is a Decision Failure: Generation-Induced Apply Bias in LLMs Authors: Haeun Jang, Yonghyun Jun, Hwanhee Lee Published: 2026-09-28 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.34284v1 Personalized LLMs must decide, for each stored preference, whether the current context calls for applying or suppressing it, which we call its applicability. They frequently over-personalize, applying preferences the context rules out, yet existing benchmarks score only the final response and cannot tell where this failure arises. We decompose preference handling into three stages and measure each separately: (1) knowing whether a preference applies, (2) deciding on an explicit Apply/Suppress label, and (3) generating a response consistent with that label. Using linear probes, we first show that this applicability signal remains decodable from hidden states during generation. By making the decision explicit, we then find that in most settings wrong decisions faithfully followed outnumber correct decisions lost in generation. We thus locate the failure in the decision, which breaks once the model is also asked to answer. To determine whether this reflects lost sensitivity or a response bias, we propose ABIDE (Apply-Bias Investigation via Decision-score), which adapts signal detection theory to Apply-vs-Suppress decision scores read directly from logits. ABIDE reveals a generation-induced Apply bias: merely stating an answer-generation objective shifts the decision score toward Apply while sensitivity is largely preserved, and the shift persists under controls for prompt structure, cascades across preference slots, and prompt wording. Finally, we show that subtracting a single bias scalar, estimated on a held-out split, from the decision score at decoding time reduces leakage while largely preserving fulfillment. --- ### Agentic High-Dimensional Bayesian Optimization with Hypothesis- and Evidence-Guided Search Authors: Zhixuan Gao, Ke Xue, Rongxi Tan, Ming Chen, Chao Qian Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.34281v1 High-dimensional Bayesian optimization (HDBO) seeks sample-efficient optimization when the number of variables is large relative to the evaluation budget. Recent LLM-based and agentic BO methods incorporate task knowledge and adapt search decisions during a run, but have primarily been evaluated on low- and moderate-dimensional problems. We ask whether this paradigm can transfer to the higher-dimensional regime. Our experiments show that these methods do not remain reliable in the high-dimensional regime, where the challenge is not only where to evaluate, but also which modeling assumption and search geometry to use when the objective's useful structure is unknown. We therefore introduce HERA, a Hypothesis- and Evidence-guided Research Agent that uses task context, optimization feedback, and structural diagnostics to revise search hypotheses, select and configure HDBO strategies, and determine their execution length. PRISM, its numerical optimization engine, generates and evaluates candidates sequentially within each search block, updating numerical models after each observation. HERA remains competitive with strong numerical HDBO baselines and outperforms the evaluated LLM-based and agentic methods on four metadata-free synthetic functions. Across eight real-world tasks, HERA achieves the best mean final objective among all evaluated systems on most benchmarks. Further analyses show that structural diagnostics change strategy use, metadata effects vary across tasks, and adaptive search blocks reduce inference cost. --- ### MoSPR: Histology-to-Gene Expression Prediction with Morpho-Spatial Macrostates and Low-Rank Molecular Programs Authors: Dongmyung Shin, Geongyu Lee, Yesung Cho, Park Jong Bae Published: 2026-09-28 Categories: cs.AI, cs.CV, cs.LG arXiv: https://arxiv.org/abs/2609.34280v1 Predicting molecular profiles from histopathology remains challenging because whole-slide images contain spatially organized, heterogeneous tissue patterns, while gene expression comprises thousands of correlated targets. We introduce MoSPR (Morpho-Spatial Program Regression), a linear framework that couples an adjacency-informed histology representation with a low-rank molecular basis. MoSPR clusters frozen patch embeddings into morphology microstates, aggregates their spatial adjacencies across the training cohort, and groups microstates with similar adjacency patterns into shared macrostates. Each slide is then represented by global morphology and macrostate-specific deviations, which are linearly mapped to coefficients of a training-derived low-rank gene-expression basis. Across three cancer cohorts from The Cancer Genome Atlas, MoSPR achieves the highest mean gene-expression prediction scores among all evaluated methods. Without pathway-level supervision, pathway scores derived from its predicted expression profiles rank first in eight of nine comparisons across three pathway collections. Ablation studies on the breast cancer cohort show complementary gains from adjacency-derived macrostate representation and low-rank molecular prediction. Moreover, with half of the training data on this cohort, MoSPR exceeds the full-data gene-prediction score of the strongest competing baseline. Finally, its linear formulation enables exact decomposition of each predicted expression profile into global and macrostate-specific molecular contributions, providing an interpretable link between spatially coherent macrostate regions and their associated molecular programs. Our code is available at https://github.com/Radisen-Panthera/MoSPR. --- ### Direct Self-Evolving Optimization: Evolving LLMs without Challenger Training Authors: Yuyang Deng, Yu Wang, Jiayun Wang Published: 2026-09-28 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.34279v1 Self-evolving language models improve by generating tasks and learning from their own feedback, but adapting the task generator often requires a separate challenger-training loop. Can we generate tasks adapted to the current solver without explicitly training a challenger? We introduce \textbf{D}irect Self-\textbf{E}volving \textbf{O}ptimization (DEO), which replaces challenger parameter updates with solver-guided task sampling. The KL-regularized challenger objective defines an exponential tilt of a fixed base task distribution. DEO uses this distribution as a sampling target: a frozen LLM generates and mutates tasks, the solver scores them, and an approximate Metropolis selection rule refines the training pool. Only the solver is trained. Theoretically, for an idealized variant that samples exactly from the tilted distribution, and under regularity, local gradient-dominance, and initialization conditions, we show that DEO learns distributionally robust reasoning ability. In experiments, DEO achieves reasoning performance competitive with R-Zero while using over $50\%$ less wall-clock training time, and improves reasoning accuracy over a no-walk ablation. Replacing the task generator with a frozen API-only LLM further improves the local solver, illustrating a capability enabled by removing challenger training. --- ### See, Measure, and Reason: Learning Visually Grounded Reasoning in Pathology Authors: Chengyang Zhang, Wenchuan Zhang, Bo Li, Mengran Li, Xinyu Liu, Jiaming Yang, Jie Chen, Zhang Zhang, Yuhao Yi, Hong Bu, Jiancheng Lv Published: 2026-09-28 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.34277v1 Pathological assessment relies on recognizing fine-grained visual details in histological images. Vision-language models (VLMs) increasingly support pathology interpretation, yet their ability to perceive these details remains inadequate. This weakness leads to inaccurate cellular observations that can persist even when final answers are correct. In this paper, we propose ASPECT to improve visually grounded reasoning through explicit supervision of cellular appearance and abundance. ASPECT trains intermediate visual tokens through pathology feature reconstruction, cell feature alignment, and count supervision. Three-stage supervised fine-tuning teaches the model to perceive, generate visual tokens, and reason, followed by reinforcement learning that rewards answer correctness and consistency with reported measurements. We also introduce PathoVernier, a benchmark of 759 expert-reviewed questions from five pathology datasets covering four cellular composition tasks. It evaluates both final answers and intermediate measurements to expose errors hidden by answer accuracy. On PathoVernier, ASPECT achieves relative accuracy gains of approximately 19.2% over the strongest baseline, Gemini-3.1-Pro, and 99.3% over its Qwen3-VL-8B backbone, while reducing RAWR, which measures counting errors within correct responses, by 28.1% and 42.7%, respectively. ASPECT also improves over its backbone on three external pathology benchmarks covering classification and question answering beyond cellular composition tasks. --- ### BIABench: Evaluating AI agents on real-world bioimage analysis tasks Authors: Zixuan Pan, Davide Panzeri, Lukas Johanns, Marilin Moor, Yu Zhou, Hedi Peterson, Yiyu Shi, Jianxu Chen Published: 2026-09-28 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.34274v1 Artificial-intelligence (AI) agents hold promise for automating bioimage analysis, yet no benchmark evaluates whether they can carry out real-world analyses end to end. Such analyses are hard for agents because 2D images, 3D volumes and time-lapse sequences are often too large to read as context, so an agent must choose and run an analysis through code, specialized software and rendered views. Published studies make this capability testable, because each pairs raw images with a peer-reviewed result. We introduce BIABench, a benchmark of 16 tasks reconstructed from published biological studies that retain their scientific questions, imaging data and ground truth. The tasks span eleven analysis subtasks and modalities from H&E histology to single-molecule localization microscopy. Each submission receives an outcome score, which compares the output files with the ground truth using field-standard metrics, and a process score, in which a vision-language model judges method choice and quality control against an expert-written rubric. We evaluated general-purpose and biology-specific agents across several language models, with repeated runs of every task. Routine two-dimensional tasks were solved well, but on some tasks that added a third dimension or a time axis no agent scored above 0.19. Neither biological specialization, stronger models nor detailed expert instructions closed this gap. The agents were also unreliable, with scores varying more between repeated runs of one agent than between different agents, and without ground truth a correct run could not be told from a wrong one by its process score or by the time spent. Released openly with its data and code, BIABench provides a verifiable framework for evaluating, and eventually training, agents for reliable long-horizon bioimage analysis. --- ### Query Expansion and Key Specialization in Transformer Attention Geometry Authors: Vidit Gupta, Siddhesh Nadkarni, Mihik Chaudhari, Vinaya Sawant, Prachi Tawde Published: 2026-09-28 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.34273v1 The projection of queries and keys are central to the attention mechanism in Transformer architectures. While they are mathematically symmetric, they play different roles in attention mechanisms. The question of whether there is an effect from their functional distinction on their geometric development in training remains unanswered. We investigate the problem through the training of small GPT-like Transformers on character-level WikiText-103 for three different depths (4, 6, and 8 layers), three types of initialization for queries and keys, and four random seeds, resulting in 36 runs and 54 trajectories of average layers across seeds. We track the effective dimensionality of those layers using participation ratios and discover that effective dimension of queries expand while keys shrink, and that $PR_Q - PR_K$ is positive in all trajectories studied. In connection to attention, the shrinking of keys leads to a narrower spectrum of $QK^\top$ and more peaked attention weights. In order to determine if this connection is causal or coincidental, we directly control the spectrum of keys during training across five seeds: restricting it to make it shrink sharpens the attention with high directional confidence, while keeping it constant to the level of initial dispersion makes attention softer. Additional token-level checkpoint analyses show that the monotonic paired-contrast trend is not universal across pretrained families, but survives as an early-training regime that later decays over a full pretraining run, and the link between interaction-rank geometry and attention entropy remains visible in several models. --- ### Broken Symmetry in BF16 Attention: Why FlashAttention Gradients Blow Up Late in Training Authors: Junlin Chen, Daize Dong, Huanwei Di, Haolong Jia, Jiawei Wu, Haotian Xie, Mingkai Zheng, Yang Li, Leshang Chen, Huishu Wang, Eric P. Xing, Hongyi Wang Published: 2026-09-28 Categories: cs.LG, cs.DC, math.NA arXiv: https://arxiv.org/abs/2609.34272v1 BF16 is now standard in large-scale pretraining, including in fused attention kernels such as FlashAttention, and these kernels are widely trusted. When we used FlashAttention-3 to pretrain a 450M-parameter transformer on 50B tokens, however, we ran into a problem: training was healthy for 25B tokens, then the gradient norm grew a thousandfold and the loss ended 0.2 nats above FP32 attention, without a single NaN. Recomputing the attention backward of just two layers in FP32 removes almost all of the excess gradient. Part of the cause is known: a fused multiply-add in the forward softmax, so far treated as an extreme-input NaN case and never fixed in FlashAttention-3. Repairing it stops the blow-up, but the query gradient is still wrong by more than its own size, and training still drives attention logits to thousands of times their size under accurate gradients. The remaining error comes from a broken conservation law. The softmax score gradient sums to zero along every row, which makes the query gradient blind to where the keys sit as a group; rounding it to BF16 leaves a small nonzero sum that leaks the mean key into the gradient, and the leak grows exactly as late training makes keys large and attention sharp. We introduce GProj (gauge projection), which restores the zero sum after the cast with two rank-one corrections per row. It cuts the remaining median query/key gradient errors from 219%/13% to 0.34%/0.37%, on par with FP32 attention, for 4.7% more time per training step. In matched from-scratch runs it trains to the same loss as FP32 attention, while FlashAttention-3 and key smoothing both destabilize. --- ### Scaling Versatile 3D Assets Editing with a Million-Scale Dataset Authors: Badi Li, Tianxin Huang, Yu Zhou, Wei-Shi Zheng, Yi Ma, Shenghua Gao Published: 2026-09-28 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.34271v1 Although recent 3D generative models produce increasingly realistic assets, controllable 3D asset editing remains challenging. Existing methods are limited by scarce training data, insufficient source-aware modeling, and a lack of practical evaluation protocols. To address these limitations, we present Alchemy3D, a unified framework for training and evaluating versatile 3D asset editors that covers data construction, model architecture, and benchmark evaluation. Specifically, we curate Alchemy3D-1M, a large-scale 3D editing dataset containing 1.25M assets and 1.38M editing pairs across seven editing types. On this data, we train a family of generative flow models for general-purpose 3D asset editing. The model family supports image- and text-conditioned editing, few-step inference, and transfer to multi-view 3D part segmentation. We further introduce GEdit3D-Bench, a large-scale, open-world benchmark with a multi-dimensional evaluation protocol. Across existing and newly introduced benchmarks, our method outperforms prior methods on most metrics of editing fidelity, source preservation, and visual quality. --- ### Bayesian Active Learning for Intent Disambiguation in Interactive Robot Planning Authors: Huao Li, Carson Sobolewski, Augustinos Saravanos, William Tan, John Karigiannis, Chuchu Fan Published: 2026-09-28 Categories: cs.RO, cs.AI, cs.HC arXiv: https://arxiv.org/abs/2609.34270v1 Interactive robot planning requires robots to infer and execute human intentions from natural language instructions that are often ambiguous, incomplete, or underspecified. Although large language models (LLMs) provide a powerful interface for clarification, relying on the generative model to drive an multi-turn conversation can introduce systematic failures. We propose a Bayesian framework that treats clarification as an active learning problem over grounded Signal Temporal Logic (STL) task specifications. Our method uses LLMs to initialize candidate formal specifications and translate informative contrasts into natural-language clarification questions, while Bayesian optimization maintains uncertainty estimation over user intent and selects queries that maximize information gain. After convergence, the inferred STL specification is passed to a formal planner to synthesize a verifiable robot trajectory. Across four simulated and real-world task domains, our approach generally achieves higher task satisfaction and requires fewer clarification rounds than LLM baselines, while helping smaller models close the performance gap against larger reasoning models. --- ### SAGE: Symbolic Action-Gating and Editing for LLM Task Planners Authors: Trung Minh Bui, JongSul Moon, YoungOuk Kim, Quang-Ngoc Phung, Se-Woong Jun, Dongin Shin Published: 2026-09-28 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.34268v1 Large language models (LLMs) are now the default cognitive core of embodied household agents, yet the plans they emit are rarely checked against a grounded model of the environment before execution, and the task-success they report is often measured on benchmarks so saturated that no method can be separated from another. We present SAGE (Symbolic Action-Gating and Editing), a single-LLM planner built from two lightweight mechanisms: a domain-agnostic symbolic gate (~250 lines of Python, zero tokens, $O(|π|)$) that blocks precondition-violating actions with typed reasons as a runtime safety monitor, and a local edit that regenerates only the failed sub-goal's suffix, keeping completed and untouched work intact; a hybrid seed+live memory store supports cold-start coverage. We evaluate under a leak-free protocol (leave-one-out retrieval) over five open-weight models and a 75-task AI2-THOR benchmark. On the standard benchmark goal-completeness saturates (52% of instances trivially solved) and SAGE ties strong hierarchical baselines. On a harder, method-agnostic multi-goal composition, SAGE's completeness lead re-emerges large (+0.06 to +0.23 across four models). Under injected mid-execution failures, SAGE recovers as reliably as whole-plan replanners at 2.4-3.3x fewer LLM calls. As a verify-before-execute gate, the symbolic monitor blocks unsafe actions before actuation and raises simulator-reported step-success for every planner tested (up to +0.11), a signal the verifier never sees (non-circular). Because the gate calls no model (0.008 ms/plan), it is a safety layer that runs essentially free on the edge: SAGE planning reproduces its quality on a Jetson AGX Orin, where small-model verification helps most. We release the benchmark, the leak-free protocol, the recovery and safety-gate harnesses, and a verifier-portability study (auto-induced on ALFWorld, 0.89 held-out). --- ### ZeroCode: On-demand Error-Correcting Code Construction from the Zero Matrix via Reinforcement Learning Authors: Ju-Hyeong Lee, Yongjune Kim, Sang-Hyo Kim, Dae-Young Yun, Hee-Youl Kwak Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.34265v1 Error-correcting codes (ECCs) are essential across diverse applications, from wireless communications and storage to quantum computing, yet each application imposes distinct design requirements on the parity-check matrix (PCM). To address these on-demand requirements in a unified framework, we propose ZeroCode, a reinforcement learning (RL)-based approach that constructs PCMs sequentially from the all-zero matrix. ZeroCode formulates construction as a discrete sequential decision-making problem and uses proximal policy optimization with action masking to select valid edges. ZeroCode achieves a gain of approximately 1 dB over the prior RL-based construction method at a bit error rate (BER) of $10^{-4}$ for the (32,16) code and outperforms existing genetic, differentiable, and classical code-design methods in our experiments. Beyond optimizing decoding performance, the masking mechanism allows on-demand structural constraints, such as a maximum degree, 4-cycle-free structure, and quasi-cyclic structure, to be flexibly incorporated. Moreover, a single policy rollout yields a library of PCMs with varying edge counts, offering trade-offs between decoding performance and complexity without retraining. Overall, ZeroCode addresses diverse code-design requirements within a unified framework, providing solutions with optimized decoding performance under given constraints. --- ### Rotated Manifold Optimization for Low-Rank Adaptation Authors: Yuhui Ding, Javier Zazo, James Hensman Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.34264v1 We propose a novel optimizer for low-rank adaptation (LoRA) that explicitly incorporates the gauge symmetry of low-rank factorization. Our optimizer extends recent matrix optimizers for full-parameter training to the manifold of fixed-rank matrices by interpreting them as normalization under a rotated basis. We show how rotation and normalization can be integrated with the fixed-rank manifold efficiently. Our optimizer converges faster to lower held-out loss and achieves better or comparable downstream performance on both supervised finetuning and reinforcement learning tasks. --- ### Maintaining Benchmarks Against Increasingly Capable Agents: Detection and Remediation of Unearned Passes Authors: Weijun Luo, Kelvin Luu, Xinyi Liu, Guangze Luo, Miguel Romero Calvo, Soham Dan, Daniel Yue Zhang, Ying Liu, Mohamed Elfeki Published: 2026-09-28 Categories: cs.AI, cs.LG, cs.SE arXiv: https://arxiv.org/abs/2609.34262v1 Agentic benchmarks guide model selection and training. Yet an agent can pass a task without demonstrating the intended capability. Such outcomes constitute unearned passes; their proportion among all passes defines the integrity gap. As agents improve, benchmark surfaces that once seemed harmless can become exploitable, making benchmark validity an ongoing maintenance problem. We introduce a process-verification framework that audits passing trajectories, distinguishes evidenced reward hacking from verifier weakness, and localizes exploitable surfaces for repair. Across 3,810 passing trajectories from 29 model-benchmark cohorts, confirmed violations often increase with model generation but not monotonically. On SWEBench Pro V1.0, confirmed violation rates rise from 24% to 73% between Opus 4.7 and Fable 5 on matched tasks; later cohorts fall to 11% for Fable 5.1 and 0% for GPT-6 Astra. These comparisons are descriptive: configurations were not normalized, and the latest models also pass fewer exploitable tasks. Violations concentrate around a small set of recurring surfaces, especially unintended access to reference solutions through git history. Three repair case studies across two benchmarks show why blocking a recorded exploit is insufficient: the same protected information can remain accessible through another route. Therefore, we combine minimal patches with exploit replay and fresh agent evaluation, auditing new passes under the original standard. No evaluated attempt against the final patches reached the protected channel, and every post-patch pass was judged legitimate. Benchmark integrity requires ongoing maintenance: audit passing behavior, repair the enabling surface, and re-evaluate both exploit access and legitimate solvability. --- ### RoboICL: Embodied In-Context Learning with GPT-6 Astra Authors: Fangcheng Liu, Yeqing Shen, Anda Cheng, Weishi Mi, Chao Tang, Chenyuan Liu, Yushun Xiang, Tingguang Li, Yong-Lu Li, Yehui Tang Published: 2026-09-28 Categories: cs.RO, cs.LG arXiv: https://arxiv.org/abs/2609.34261v1 General-purpose vision-language models offer a promising way to zero-shot robot control: \gptastra{} excels at open-ended and language- or image-conditioned manipulation but remains substantially weaker on high-precision and long-horizon tasks. We introduce \emph{RoboICL}, an in-context robot-control framework that narrows these gaps without robot-specific parameter updates or a learned VLA. RoboICL separates \emph{demonstration context}, which provides recorded examples when available, from \emph{interaction memory}, which accumulates the model's own actions and observed outcomes. Both use a shared observation--action--receipt--observation grammar. To preserve experience across task stages, RoboICL combines sampled demonstration blocks with bounded anchored memory. Fixed anchors keep earlier rollout interactions available for in-context learning, while the latest interaction supports immediate error correction. Across 30 RoboDojo tasks, using zero shot for Open and one demonstration elsewhere, RoboICL improves on official zero-shot \gptastra{} by 20--27 progress-score points in every category. It leads the leaderboard baselines on Memory and Open, achieves comparable performance to the strongest Precision baseline, and remains competitive on Long-Horizon. Its 30-task Overall score is 50.64, versus 33.68 for the strongest baseline. On a separate ten-task subset, RoboICL scores 60.60, within 2.00 points of the $π_{0.5}$ + \gptastra{} hybrid approach. On three real-robot tasks, mean progress rises from 14.45 at zero shot to 63.33 at one shot and 78.89 at three shots. On two development tasks, optional Jev-gated action reuse reduces \gptastra{} calls by 33--48\%. Code is available at \href{https://github.com/Mosi-AI/RoboICL}{https://github.com/Mosi-AI/RoboICL}. --- ### QuantaSpike: Short-Window Spike-Driven Quantization for Large Language Models Authors: Bang Hu, Guowei Zhu, Changze Lv, Xiaoqing Zheng, Fengzhe Zhang, Fan Zhang, Wei Cao Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.34259v1 Large language models (LLMs) achieve strong performance across many tasks but rely on dense multiply-accumulate (MAC) operations during inference, resulting in high energy cost. Spiking neural networks (SNNs) offer an event-driven alternative in which synaptic integration uses lightweight accumulation. However, spike-driven LLM inference remains difficult because outlier-heavy activations typically require long firing windows or auxiliary non-spiking paths. We propose QuantaSpike, a short-window spike-driven quantization framework for LLMs built around Logarithmic Ternary Integrate-and-Fire (LTIF) neurons. LTIF uses ternary events with power-of-two membrane-response quanta, improving the information represented by each firing step while retaining shift-ACC-compatible computation. QuantaSpike combines this neuron with group-adaptive gain and selective outlier admission: normal values use residual LTIF steps, whereas admitted outliers receive one additional onset spike before entering the same residual dynamics. Across OPT and Llama-2, QuantaSpike achieves state-of-the-art or competitive perplexity and zero-shot accuracy among spike-driven LLM quantization methods. It also transfers to newer dense LLMs, remaining close to the FP16 reference on Llama-3-8B and Qwen3-8B under the same four-step firing window. Analytical linear-energy projections show that QuantaSpike reduces the energy of one linear transformation by about $80.0\%$ on OPT models and $67.1\%$ on Llama-2 models relative to SpikeQuant, providing an accurate and energy-efficient spike-driven path for LLM inference. --- ### Investigating Human--AI Discrepancies via Multiple-Solution Problems Authors: Zihao Wang, Francesco Insulla, Andrea Montanari Published: 2026-09-28 Categories: cs.CY, cs.AI arXiv: https://arxiv.org/abs/2609.34258v1 Frontier artificial intelligence (AI) models are benchmarked on whether they reach a correct answer. Yet many problems admit several correct answers and repeated attempts, by different people or by the same model resampled, trace out a distribution over them. In this work, we ask whether human and model reasoning lead to different distributions over valid solutions. Our testbed comprises 270 reasoning puzzles across five puzzle families. These multiple-solution puzzles each have 3 to 8 valid solutions and are simple enough that humans and models can solve them reliably. The resulting distributions differ markedly: models differ from one another, yet resemble each other far more than they resemble humans. Model distributions are, moreover, within every puzzle family, less diverse than human ones. We compare these discrepancies across puzzle categories, and trace how they respond to reasoning-effort settings, to prompting, and to perturbations of the puzzle that leave its solutions unchanged. Together, these results point at significant differences between human and AI problem-solving processes, and their choice among equally defensible solutions. As progressive deployment of AI systems in society comes into focus, evaluating such differences (beyond one-dimensional accuracy metrics) is increasingly important. Data and code are available at https://hai-discrepancies.github.io/ --- ### Recursive LLM Degradation in Biomedical Question Answering: A Cross-Generation Study Authors: Bibek Bhandari, Kshitij Lingthep Published: 2026-09-28 Categories: cs.CL, cs.LG arXiv: https://arxiv.org/abs/2609.34257v1 Repeatedly training language models on their own generated data may create a synthetic-data feedback loop in which errors and distributional biases are reintroduced into subsequent training datasets. This paper studies that process in biomedical question answering (QA) using PubMedQA and two Qwen2.5 model sizes, 0.5B and 3B parameters. The study compares a recursive synthetic-data condition, in which generation G(k+1) is trained on answers produced by G(k), against a Human-Control condition that repeatedly uses the original human training data. The study evaluates across four generations from G0-G3 with two random seeds (42 and 123) and a fixed evaluation set of 1,000 expert-labeled samples. The evaluation includes disease and chemical entity F1, context-supported rate, lexical and semantic similarity, answer length, repetition rate, and other evaluation metrics. The Recursive condition for both model sizes and both seeds showed larger declines than the Human-Control condition in disease entity F1, chemical entity F1, context-supported rate, ROUGE-L, and cosine similarity. Under the fixed no-repeat 3-gram decoding constraint, the main observed behavioral change was increased answer length, while the measured 3-gram repetition rate did not increase. The magnitude of the difference-in-change was larger for the 3B model than for the 0.5B model. This difference was particularly apparent in disease F1, context-supported rate, cosine similarity, and answer length. These results show domain-specific changes associated with using recursive synthetic-data training in biomedical QA, but do not establish clinical hallucination rates or universal model collapse. --- ### DreamingGoose: Staged Distillation from Autoregressive Transformers to Bidirectional Recurrent Diffusion Language Models Authors: Julian Boesch, Andrew Wee, Alexander Stranzl Published: 2026-09-28 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.34253v1 Pretrained autoregressive Transformers represent a large sunk investment in compute. Existing conversion methods reuse that investment by changing either the architecture (attention to recurrence) or the objective (next-token prediction to denoising), never both. We convert Qwen3 teachers at 1.7B and 8B into attention-free, bidirectional, gated-delta-rule diffusion students in three stages, so that each capability can be traced to the stage that kept or lost it. Language modeling transfers only partially and in-distribution; in-context retrieval does not transfer. On a multi-query recall probe where the teachers score 0.34-0.58, both converted students score 0.000, and diffusion pretraining alone does not restore retrieval. A retrieval curriculum in the final stage, which gradually lengthens the gap between a key-value table and the queries that address it, restores it only stochastically: on a fixed schedule, one seed in three learns to retrieve. Advancing the gap only while a running accuracy estimate stays above a threshold works for all three of those seeds, holds on real text, and carries unchanged to 8B, where two of three seeds succeed. The third had not learned within its fixed 16k-step budget: retrieval switches on abruptly at a seed-dependent step (6.5k and 11k in the other two), so a fixed budget can cut a late run off. One boundary survives every intervention: every model that learns retrieval scores 0.000 on tokens that never appeared in a retrieval episode, and an arm that resamples the key and value tokens every batch shows this is a coverage limit, not memorization of particular bindings. Separately, we convert a 7B code model into a 3:1 recurrent-attention block-diffusion hybrid over 85k steps and report two negative training results. --- ### WAM-OPD: Sharpening World Action Models via On-Policy Distillation Authors: Panjun Liu, Xiaohan Lei, Shiqi Zhang, Yikun Wang, Yongxin Zhang, Mingyi Hu, Shida Sun, Jiateng Shou, Wengang Zhou, Jiajun Deng, Zhiwei Xiong Published: 2026-09-28 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.34250v1 Pretrained world action models (WAMs) provide generalist capabilities across diverse robotic manipulation tasks, yet improving target-task performance to an expert level without degrading pretrained skills remains challenging. We explore on-policy distillation (OPD) for WAMs and introduce WAM-OPD. WAM-OPD inherits the advantage of OPD methods that transfer task-specific teacher knowledge under the student's own induced distribution, rather than directly fitting the student to a narrow task-specific data distribution. However, in closed-loop manipulation, the observation histories change as the student policy evolves, requiring fresh environment rollouts to remain on-policy. Applying OPD to WAMs entails repeated data collection, which is costly even in simulation and often impractical on real robots. To avoid repeated environment rollouts during distillation, we introduce prefix-weighted trajectory replay (PWTR). PWTR uses a fixed trajectory pool composed primarily of initial-student rollouts, supplemented with task-specific teacher rollouts to broaden trajectory coverage. For each trajectory replayed from this pool, PWTR conditions the current policy on successive stored histories to generate fresh denoising paths, along which the task-specific teacher provides supervision. Although these denoising paths are refreshed as the policy evolves, the replayed environment trajectories remain fixed. PWTR therefore reweights per-decision distillation losses using proxy importance weights derived from path scores accumulated over the trajectory prefix preceding each decision to mitigate the resulting shift in the history distribution. Simulated and real-world experiments demonstrate task adaptation without additional environment interaction during distillation. In both settings, WAM-OPD improves target-task performance while retaining near-initial performance on tasks excluded from adaptation. --- ### Evolving Support Priorities in Empathetic Reinforcement Learning Authors: Pengyu Huang, Zhiyuan Han, Wenwen Tong, Hewei Guo, Jiangnan Chen, Sirui Chen, Lewei Lu, Beier Zhu, Xun Yang Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.34249v1 We identify a fundamental mismatch in empathetic reinforcement learning: support priorities evolve with the dialogue state, yet existing methods typically optimize predefined reward specifications that remain fixed across turns. To model these evolving support priorities, we organize empathetic support along cognitive, affective, and proactive empathy, and propose Context-Adaptive Rubric Evolution (CARE). At each turn, CARE generates a context-adaptive rubric by adjusting both the weights of these three empathy dimensions and their fine-grained evaluation criteria. The rubric generator is trained with turn-level rubric supervision and human preference data through supervised fine-tuning followed by preference-based reinforcement learning, and then serves as an adaptive reward interface for online empathetic RL. Integrated with both RLVER and MICA, CARE achieves state-of-the-art performance across SentientBench, EQBench3, and EMPA under three independent LLM judges. Notably, on EMPA, CARE improves EPM-Idx over the strongest baseline by at least 13 points under all three judges, including an increase from 28.11 to 83.54 under Gemini-2.5-Pro. Further analyses show that learned rubric priorities systematically vary across dialogue stages and user emotions, demonstrating that CARE adapts what is rewarded as support needs evolve. --- ### SALMONN-duo: Adaptive Dual-System Coordination for Full-Duplex Voice Agents Authors: Wenyi Yu, Siyin Wang, Terumi Chiba, Xianzhao Chen, Xiaohai Tian, Jun Zhang, Lu Lu, Chao Zhang Published: 2026-09-28 Categories: cs.CL, eess.AS arXiv: https://arxiv.org/abs/2609.34247v1 Full-duplex speech large language models (LLMs) enable low-latency, natural voice interaction. However, real-world agents must also use tools and perform deliberative reasoning-operations whose variable latency and computational cost conflict with the stringent timing requirements of real-time conversation. To reconcile these demands, we propose SALMONN-duo, an adaptive dual-system voice agent inspired by dual-process theories of cognition. SALMONN-duo separates real-time interaction from deliberative computation by pairing an always-on, fast-thinking full-duplex speech LLM (system 1) with a powerful asynchronous slow-thinking LLM agent (system 2). Beyond handling real-time interaction, system 1 learns when to answer directly and when to delegate, remaining responsive during backend execution and seamlessly integrating returned information into the ongoing dialogue without exposing tool traces or losing conversational context. Evaluations on single-turn spoken question answering (QA) and multi-turn conversations demonstrate that adaptive delegation substantially improves accuracy on knowledge-intensive and multi-hop reasoning questions, while knowledge-boundary-aware training avoids unnecessary system 2 invocations. On a customized version of $τ$-Voice, SALMONN-duo further demonstrates its ability to complete environment-grounded, policy-constrained tasks through multi-turn interactions in realistic business scenarios. Finally, cost-aware reinforcement learning further enhances the trade-off between task performance and backend usage across the QA and conversation tasks, while improving task success and response safety on $τ$-Voice with an acceptable increase in the delegation rate. --- ### CasEm: A Cascade Architecture for Long-Horizon Neural Emulation Authors: Zhaoyi Li, Jingtao Ding, Shihua Li Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.34246v1 Autoregressive neural emulators can drift or diverge over long rollouts despite accurate short-term predictions. We introduce Cascaded Emulation (CasEm), a one-way rollout architecture that augments an existing full-state backbone with an independently evolving model of physically specified aggregates. Its forecasts guide corrections to full-state predictions, without feedback from the backbone to the aggregate model. Effective guidance requires aggregates that cover substantial backbone error, remain accurately predictable, and support useful full-state corrections. We derive a finite-horizon error bound that clarifies these three factors and use empirical diagnostics to guide subsystem selection. Across four ODE/PDE benchmarks, CasEm reduces long-horizon rollout errors across diverse backbones and suppresses the trend toward error divergence in both diffusion tasks using Fourier neural operator backbones. In global climate emulation, CasEm with a regional total-water subsystem reduces 10-year full-state time-mean error by 66.6% and 46.3% for frozen ACE and Spherical DYffusion backbones, respectively, while adding less than 3% to inference time. --- ### Certified Multi-Source Integrity for Structured Agent Actions Authors: Anmol Pandey, Aditya Jain, Liang Chen, Carsten Maple, Christo Panchev Published: 2026-09-28 Categories: cs.CR, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.34245v1 LLM agents increasingly take privileged, often irreversible structured actions, such as paying an invoice. They assemble each action from action-critical fields in documents and tool outputs that an adversary can corrupt, and indirect prompt injection can drive the model itself to extract attacker-chosen values. Current defenses gate on a source's trust label or certify free-text answer quality. None certifies the integrity of a coupled, policy-bound structured action under a corruption budget that accounts for shared upstream sources. We characterize when such an action is safely certifiable and give the maximally live safe certifier. It admits an action only when each field clears the rule its evidence structure supports: a bounded corruption radius over corruption-distinct evidence classes, counted by a minimum hitting set so that re-publishing or laundered copies cannot manufacture a quorum, deterministic reconciliation for complementary fields, and a trusted anchor where the evidence leaves a field single-sourced. We formalize two robustness notions, validate each mechanism by ablation, and measure how often the multi-source precondition holds on sanctions designations (70,966 entities) and software supply-chain provenance (450 packages). Under upper-bound proxies, genuine corroboration is a minority phenomenon in both, and naive attestation counting overstates it, since witnesses that look independent collapse to two corruption-distinct domains once shared origin is counted. Across five current models in a real agent loop, a realistic injection fools every model but one and a naive agent then executes the fraudulent action on most attacks. The certifier admits no unsafe action and recovers the correct value where corroboration permits, while action-gating and provenance baselines are broken in every world of our harness by some attack in its space. --- ### Stashbird: Efficient Speaker-Indexed Memory for Conversational Agents Authors: Chidera Biringa, Lucas Yannul, Xiaowen Wang, Marco Ayala, Nicholas Yi, Alex Moyse, Nishant Manchanda, Vivek Gupta Published: 2026-09-28 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.34242v1 AI agents require memory that preserves information across user-agent exchanges, user-to-user conversations, and group conversations with or without agent participation, while supporting updates as evidence changes or is removed. We present Stashbird, an agent memory system that links source episodes to derived memory state through explicit provenance. Stashbird organizes memory into episodic records, semantic relations, community summaries, and persisted graph state, with lifecycle operations for incremental updates and episode-level deletion. We evaluate question-answering accuracy and model-facing workload across four long-term memory benchmarks. On LoCoMo, Stashbird uses 76.4x fewer ingestion prompt tokens than Graphiti. Compared with reproduced Hindsight on the same benchmark, it uses 8.1x fewer retrieval prompt tokens, with accuracy 1.6 percentage points lower. It achieves higher accuracy than Hindsight on LongMemEval-S and GroupMemBench and comparable accuracy on EverMemBench. --- ### AdaGuard: An Adaptive Guard Model with User-defined Policies Authors: Yunhao Feng, Yifan Ding, Yuxiang Xie, Zheng Li, Mingrui Lao, Zeyuan Wang, Yanming Guo Published: 2026-09-28 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.34241v1 Guard models support the safe deployment of language model agents, but fixed risk taxonomies limit their ability to accommodate requirements that vary across applications and tasks. Under user-defined policies, detecting violations requires interpreting both the applicable rules and the agent's behavior, since identical actions can receive different judgments under different policies. To support learning this capability, we introduce AdaptiveSafety, a dataset of 10,939 training examples and 1,000 test examples covering policies with 1--100 rules. The dataset combines trajectories from multiple sources with policy and behavioral counterfactuals, pairing each example with an explanation and the complete set of violated rules. These counterfactuals expose changes that alter compliance, while structural augmentations provide supervision for consistency under rule reordering and identifier remapping. Building on this supervision, we propose SafePO, a reinforcement learning algorithm for refining violation identification while balancing explanatory reasoning and final verdicts. SafePO uses structured rewards to assess prediction correctness, retains group-relative advantages at the response level, and employs a separately trained value model to modulate token weights within explanation and verdict regions. Separate normalization controls their relative contribution to training despite differences in length. Through supervised initialization followed by SafePO, we develop AdaGuard, a family of 0.6B, 4B, and 8B guard models that assess agent trajectories under policies supplied at inference time. Our 4B model achieves binary accuracies of 89.30\% on AdaptiveSafety and 71.82\% on DynaBench. The project repository is available at https://github.com/Yunhao-Feng/AdaGuard --- ### Coherence-Aware Distributional Evaluation of Open-Ended Text Generation Authors: Jinnuo Liu, Junhao Zhu, Weifeng Jiang, Haoming Liu, Hongyi Wen Published: 2026-09-28 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.34240v2 Existing open-ended generation metrics measure likelihood, lexical diversity, or distributional similarity in generic representation space, yet can miss fundamental dimensions of quality. A prominent blind spot is global coherence: a generated passage may be locally fluent while remaining globally contradictory, causally inconsistent, or topically disconnected. We identify representation as a central bottleneck in detecting these failures and introduce CHORD (Coherence-aware Hidden-state Open-generation Reference Distance), a coherence-sensitive distributional metric. CHORD encodes generated and human-written corpora in the hidden-state space of a frozen LLM using a coherence-eliciting prompt, and compares the resulting distributions using RBF-MMD. To test coherence sensitivity and selectivity, we construct a counterfactual evaluation suite pairing graded coherence-degrading perturbations with meaning-preserving controls. CHORD selectively detects relation, discourse, structural, and mixture failures that perplexity, entropy, MAUVE, FBD, and MMD-based baselines either miss or cannot separate from benign rewriting. Factorial ablations show that representation is the primary source of coherence sensitivity, while RBF-MMD improves sample efficiency. Larger backbones capture finer-grained distinctions, but coherence prompting improves selectivity only when the backbone can follow the prompt. On unconditional generation and prefix continuation, CHORD yields model rankings that strongly align with human judgments of whether outputs make sense and appear human-written. Together, these results establish representation design as central to reliable distributional evaluation. Code: https://github.com/MAPS-research/CHORD. Experiments: https://github.com/MAPS-research/CHORD-Experiment. --- ### DecFlowEdit: Self-Localized Flow-based Image Editing via Guidance Decoupling Authors: Zheyuan Zhan, Can Wang, Jiawei Chen, Chun Chen, Siwei Lyu, Zeyu Zheng, Defang Chen Published: 2026-09-28 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.34237v1 Flow-based image editing (FlowEdit) enables inversion-free semantic changes through the difference between source and target velocities. In this paper, we observe that FlowEdit's default classifier-free guidance (CFG) configuration, with asymmetric source and target scales, causes substantial background leakage. Matching these guidance scales, for example by removing CFG, improves edit-relevant localization but severely degrades editability. To get the best of both worlds, we propose DecFlowEdit, which decouples the optimal guidance scales for localization and for editing in flow-based generative models. In particular, DecFlowEdit first extracts an edit-relevant prior by temporally aggregating velocity differences evaluated without CFG, and then uses this prior to reweight the original updates under default CFG. Our method remains training-free and inversion-free, requiring neither external spatial masks nor attention manipulation. Experiments on PIE-Bench across FLUX, SD3, and SD3.5 show that DecFlowEdit improves background preservation, reducing structure distance by approximately 61 to 73 percent and background LPIPS by 68 to 80 percent relative to FlowEdit at comparable editing fidelity. --- ### SegBanana: Steering Unified Multimodal Models into Medical Segmenters Authors: Xiaoye Liang, Ye Yan, Mingze Yin, Shikun Feng, Mai Xu, Haiguang Liu, Lai Jiang, Yiheng Zhu Published: 2026-09-28 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.34235v1 Medical image segmentation remains challenging in practical deployment, as models often struggle to generalize beyond the distributions covered by their training data and high-quality pixel-level annotations are typically unavailable for adaptation. Inspired by the cross-task transferability of large language models, we investigate whether unified multimodal models (UMMs) can transfer their pretrained visual understanding, reasoning, and generation capabilities to medical image segmentation without task-specific post-training. By recasting segmentation as structured visual generation, we find that frontier UMMs (e.g., Nano Banana) already exhibit basic segmentation capabilities across diverse clinical scenarios, but still struggle with challenging tasks requiring specialized anatomical or domain-specific knowledge. We further show that these limitations can be effectively mitigated by incorporating visual anatomical knowledge from in-context exemplars, expanding candidate solutions through repeated sampling, and refining suboptimal predictions via targeted editing.Motivated by these observations, we propose SegBanana, to our knowledge, the first agentic visual generation framework for training-free medical image segmentation. SegBanana builds on a frozen UMM as the core generative model, augmented with Anatomy-Aware Knowledge Retrieval and Comparative Quality Critique to unlock its potential segmentation capability. A State-Aware Multimodal Controller maintains structured state and iteratively orchestrates these tools, repeatedly refining intermediate predictions toward higher-quality masks. Across eight medical segmentation datasets, SegBanana achieves an average mDice of 77.45%, outperforming representative generalist (SAM3 and SegGPT) and medical-specific (BiomedParse and MedSAM3) baselines by at least 14.93 points, while remaining robust to out-of-domain visual supports. --- ### MAS-OPD: On-Policy Distillation for Multi-agent Systems Authors: Qiyong Zhong, Mao Zheng, Mingyang Song, Houcheng Jiang, Jiajie Su, Huwei Ji, Li Zhang, Junfeng Fang Published: 2026-09-28 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.34234v1 Multi-agent systems (MAS) split a task across specialized roles and are promising on complex tasks, yet a prevailing approach relies on inference-time orchestration alone. General-purpose APIs are costly and hard to customize, while small models with role prompts rarely develop stable role competence or reliable collaboration, so post-training a MAS jointly is central. Most attempts use reinforcement learning, whose team-level reward leaves undetermined which step of which agent brought about the outcome, while local rewards need redesigning per task. On-policy distillation (OPD) gives token-level teacher supervision on trajectories the student samples, a denser signal needing no local reward, yet is underexplored for the interdependent agents of a MAS. Two difficulties arise: building complementary specialization from a judgement of which role a behavior belongs to while preserving the knowledge all roles need, and turning cross-agent collaborative information into supervision OPD can exploit. We present MAS-OPD, where Role-Advantage Specialization defines the role advantage as the difference between the teacher signals under target and non-target role conditions, and Privileged Attribution for Coordination attributes an interaction conflict to its source and supplies it to the teacher alone as privileged information. Extensive experiments on code and mathematics benchmarks show that MAS-OPD attains the highest mean score at both student scales and leads the agents to develop clearer role specialization and more effective collaborative behavior. --- ### Trustworthy synthetic visual media: Evidence across the media lifecycle Authors: Zexi Jia, Zhiqiang Yuan, Jie Zhou, Jinchao Zhang Published: 2026-09-28 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.34232v1 Images and videos have long helped people understand what happened and how a work came into being. Generative systems complicate that role. Realistic media can now be produced and revised without leaving a stable history, so appearance no longer reveals whether a scene was captured, synthesized, or altered along the way. Trust must instead come from evidence that explains the path an asset has taken and the circumstances in which it was used. Some of this evidence can be recovered from the media, while some must be recorded during production and preserved as the asset circulates. This review brings those approaches together and asks when their claims remain meaningful after ordinary processing or deliberate manipulation. We argue that trustworthy media do not depend on one universal marker of authenticity. The evidence must suit the question at hand, reach the person making the judgment, and remain open to correction when better information emerges. The larger goal is to keep the history of media intelligible even as the media itself continues to change. --- ### ReGDiff: Guided Diffusion in Regulated Latent Space for Exploring Metamaterial Voxel Geometry Authors: Wangzhi Zhan, Jianpeng Chen, Dongqi Fu, Dawei Zhou Published: 2026-09-28 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.34231v1 Metamaterials are artificially engineered structures whose mechanical and physical behaviors are strongly shaped by geometry rather than composition. Voxel representation provides a unified format for metamaterial geometry generation, as it can express diverse classes such as truss, shell, and porous structures within a single cubic discretization. However, voxel-based generation faces a plausibility-novelty trade-off: staying close to known geometries helps preserve geometric regularities, while moving away from them is necessary for novelty but may produce degenerate geometries. To address this challenge, we propose REGDIFF, a generative framework that couples voxel representation with latent space regulation and guided diffusion. REGDIFF introduces a repel-and-sink (RAS) mechanism to smooth the latent distribution of plausible geometries, and short-range repulsion (SRR) guidance to discourage generation overly close to known samples while maintaining geometric plausibility. We further contribute a voxel-based benchmark covering truss- and shell-type metamaterial geometries, together with an evaluation module for geometric plausibility, novelty, and diversity. Experiments show that REGDIFF outperforms voxel-based generative baselines, achieving +8.9% in geometric plausibility, +46.4% in novelty, and +128.6% in diversity on average across two datasets. These results suggest that REGDIFF is a strong geometry candidate generator for downstream evaluation. Our code is provided at https://github.com/wzhan24/ReGDiff. --- ### SleuthBench: Benchmarking Statistical LLM Evaluation Using Tabular Hidden Signals Authors: Jingyun Jia, Antoine Remond-Tiedrez, Aaron Alvarez, Joshua Shunk, Rich Caruana, Ben Lengerich Published: 2026-09-28 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.34228v1 Evaluating statistical discovery by large language model (LLM) agents requires verifiable analytical ground truth. Establishing such ground truth for real-world datasets is costly, and prior knowledge of public datasets can influence agent responses. We introduce SLEUTHBENCH, a benchmark that addresses both problems by injecting controlled data-quality problems and feature effects into public tabular datasets: the injected pattern determines the answer, so reference answers are computed automatically and memorized knowledge of the original table is insufficient, while the table keeps its background structure. The injected patterns are modeled on phenomena reported in real data analyses. The benchmark defines 17 question templates in two families: data-quality questions and feature-contribution questions. We evaluate six state-of-the-art LLMs that analyze the data using a Python coding tool, on data-science and business phrasings of 70 validated dataset-template combinations, yielding 1680 graded responses in total. The models detect data-quality problems reliably (83.8% accuracy) but recover feature contributions poorly (41.9%). Finding how features shape the target requires searching over both candidate variables and analytical procedures. To address this issue, we propose the Empirical Layer, a set of precomputed statistical artifacts comprising summaries, fitted feature and interaction effects, and dataset descriptions, which exposes candidate patterns for direct inspection. Access to these artifacts raises feature-contribution accuracy from 41.9% to 68.0%. --- ### When Does Selection Replace Extraction? A Pre-Registered Test of Agent Memory with a Typed Decision Model Authors: Rishabh Sharma, Rishika Lall Published: 2026-09-28 Categories: cs.AI, cs.CL, cs.IR arXiv: https://arxiv.org/abs/2609.34227v1 Does conversational memory need LLM-extracted facts, or is selecting the right raw turns enough? Published results disagree. Extraction-based systems report gains from distilled facts. Recent studies find raw history with good ranking does as well, but disagree about whether ranking matters. We ran a pre-registered study on held-out LoCoMo conversations and LongMemEval. At a tight budget on LoCoMo, raw turns selected by a single call to Jev, a typed decision model, are non-inferior to an LLM-extraction memory (one-sided 95% bound -3.0 points against a -5-point margin). Blind human grading narrows the margin but does not change the result. Raw turns cost 3,061 times less to write, and the result holds with a second answer model. Within this study, reranking's gain shrinks as the budget grows. It adds 17.4 points on LoCoMo and 9.1 on LongMemEval when three of 30 candidates are kept. At generous budgets it adds 1.5 and 1.1, and extraction systems are more accurate. This suggests why published results disagree. At matched context, Jev selects as accurately as an LLM reranker (non-inferiority bound -2.0) at a third of the latency, and more accurately than a multi-call graph traversal. Reranking lowers correct abstention. Plans, code and graded answers are released. ---