# Research Papers — 2026-09 · bundle page 261 of 266 > Full-text markdown of papers #13001–13050 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/261 [← Previous bundle](https://aipod.dev/papers/archive/2026-09/page/260.md) [Next bundle →](https://aipod.dev/papers/archive/2026-09/page/262.md) ### What Does an Agentic Software Engineering Benchmark Measure? Profiling Task Demands and Agent Behaviour Beyond What Category Labels Reveal Authors: Radin Shayanfar, Keheliya Gallaba, Ahmed E. Hassan Published: 2026-09-01 Categories: cs.SE, cs.CL arXiv: https://arxiv.org/abs/2609.01271v1 Agentic software engineering benchmarks are typically summarized by nominal category labels such as "bug fix" or "feature implementation," yet benchmarks carrying the same label are built through very different curation pipelines. A label thus reveals little about the engineering work a benchmark demands. We introduce the Spread--Novelty--Centrality (SNC) profile, a three-axis characterization of the demands of repository-level coding tasks, grounded in empirical software engineering research. We apply the profile to five widely used benchmarks and 14,922 trajectories of two model families at three scales, and report three findings. (1) A label is an unreliable proxy for task demands, as every pair of benchmarks is statistically separated on at least two SNC axes, and the separations trace back to specific curation decisions. (2) Agent behaviour reveals demands that the human-written gold solution cannot. Agents produce larger solutions than the gold where problem statements withhold hints and smaller ones where curation inflates the gold. How a task is phrased shapes what an agent produces. (3) Task demands correlate with success uniformly, with resolved runs concentrating in the low-SNC region for every family and scale, whereas the behavioural signatures of success are family-specific. Claude succeeds by matching the scope of the gold solution, and its parity share on files rises from $0.17$ at the smallest scale to $0.54$ at the largest. Qwen succeeds by exceeding the gold scope at every scale, and editing too little marks failure for both families. --- ### Solving In-Table Prediction Problems by Deep Neural Networks with Performance Evaluation Using Synthetic Data Authors: Xiao Zhao, Daniela Oelke Published: 2026-09-01 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.01262v1 Tabular deep learning (TDL) leverages neural networks (NN) to extract patterns from tabular data. Traditional TDL methods follow a supervised learning paradigm, where a target feature is explicitly given. In this work, however, we explore a different approach by employing deep NNs to learn relationships among individual columns within a given table. We investigate whether NNs can predict the values of arbitrarily selected columns in a given table based on the remaining known columns. We call this problem In-Table Prediction (ITB), which is slightly different from table imputation methods and the pretraining task of TDL. Three potential usage scenarios are identified, which, to our best knowledge, have not been extensively studied in the literature. A self-supervised learning approach is applied to address this problem by randomly selecting columns to be masked out and used as learning targets. This work focuses on tabular datasets containing only continuous features. To handle missing values in continuous features, a novel neural layer is proposed to embed both numerical and empty values. Synthetic data is generated based on predefined column relationships, with empty values inserted using two distinct mechanisms. Additionally, an adapted masking strategy is employed to create test data. Performances of three NN architectures, namely MLP, Resnet and Transformer, are evaluated using the generated synthetic data. We conclude that, the attention-based structure outperforms the other two networks, when a sufficiently large number of training examples is available and a relatively large embedding length is chosen. We stress that these findings are obtained under controlled, synthetic conditions with a small number of columns and it should therefore be regarded as an initial, narrowly-scoped investigation rather than a general characterization of ITP on real-world tabular data. --- ### Dual Process Motion Planning Authors: Jiayi Yan, Francesco Fabiano, Alessandro Abate Published: 2026-09-01 Categories: cs.AI, cs.RO arXiv: https://arxiv.org/abs/2609.01260v1 Robotic systems are deeply embedded in both industry and everyday life, where they are expected to act with speed, precision, and reliability. Classical control and planning methods have long delivered strong guarantees, but often at the cost of computational efficiency and adaptability. More recently, learning-based approaches have shown promise in overcoming these limitations, enabling agents to leverage experience to accelerate decision-making and address previously intractable problems. In this work, we bridge these two approaches through a neuro-symbolic perspective on nonlinear motion planning. Inspired by the Thinking Fast and Slow paradigm, we introduce a dual-process architecture that combines the strengths of robust reasoning and learning. Our framework integrates state-of-the-art symbolic solvers as a ``System-2'' component with experience-driven ``System-1'' modules. A metacognitive controller dynamically orchestrates their interaction, selecting when to rely on fast intuition versus slower, more precise reasoning. By evaluating the framework across diverse nonlinear benchmark environments, we demonstrate that this architecture yields consistent gains in planning efficiency, accuracy, and generalization, while promoting reuse across tasks. The results suggest that tightly coupling learning with structured reasoning offers a scalable path toward more capable and adaptive robotic systems. --- ### Measuring the Behavioral Fidelity of Long-Horizon Human Activity Simulations Authors: Yi Fei Cheng, Fan Yang, Iremsu Bas, Koichiro Niinuma, Narishige Abe, David Lindlbauer Published: 2026-09-01 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.01257v1 As LLM-based human simulators are increasingly used for policy, evaluation, and training, they must faithfully reproduce real behavioral patterns. While prior work has examined behavioral fidelity in survey responses and dialogue, longer-horizon real-world activity remains largely unexplored. We introduce a framework for evaluating behavioral fidelity in long-horizon activity simulations across temporal granularities and levels of analysis. As a case study, we collect a 43-hour multi-camera dataset of in-the-wild office activity and compare trace-derived conditioning mechanisms: persona descriptors, few-shot exemplars, and statistical transition and time-of-day priors. We find that behavioral fidelity is not uniform across metrics: statistical priors bring activity and sequence distributions closest to real behavior, yet over-fragment routines and suppress within-person variability. These findings motivate a more holistic evaluation that spans multiple metrics, temporal granularities, and levels of analysis. --- ### MeRoPE: Metric Rotary Position Embedding for Camera-Controlled Video Generation Authors: Zhijian Qiao, Xinjiang Wang, Jiajie Chen, Haoming Huang, Meng Li, Chih-Chung Chou, Jing Wang, Shaojie Shen Published: 2026-09-01 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2609.01252v1 In camera-controlled video generation, geometry-aware positional encodings condition tokens on camera extrinsics and per-token viewing rays. Existing schemes, however, have a scale-dependent failure mode on real-world metric camera trajectories: homogeneous projective encodings cause attention logits and feature norms to grow unbounded with physical translation baselines. We propose MeRoPE (Metric Rotary Position Embedding), a norm-preserving relative camera encoding for attention. MeRoPE encodes relative orientations between calibrated viewing rays with orthogonal rotation blocks, maps raw metric displacements into multi-frequency rotary phases, and adds a disparity-anchored correspondence prior along the epipolar arc. This design strictly preserves feature norms, bounds pre-softmax attention logits regardless of the physical translation scale, and maintains exact invariance to global rigid coordinate changes. Across nuScenes and PanShot, which cover large-baseline trajectories and diverse camera optics, respectively, MeRoPE achieves stronger camera control than prior encodings, with the best consistency between generated camera motion and conditioning poses in both rotation and translation. Code will be made publicly available. --- ### One Prompt Is Enough: Watermark Laundering Through Foundation Image Models Authors: Jidong Yang, Qi Li, Wei Zong, Yang-Wai Chow, Willy Susilo, Huaike Yu, Chunpeng Wang, Suo Gao Published: 2026-09-01 Categories: cs.CV, cs.AI, cs.CR arXiv: https://arxiv.org/abs/2609.01249v1 Invisible watermarks are typically evaluated against predefined perturbations such as compression, blur, noise, cropping, and denoising. Public foundation image models expose a distinct threat: an attacker can submit a watermarked image with a single reconstruction prompt and obtain a visually faithful output from which the invisible watermark can no longer be decoded reliably. We formalize this failure mode as watermark laundering and evaluate it using a joint payload-fidelity profile that combines bit error rate (BER) with visual and semantic preservation. Across six OpenAI and Google image editing models, three representative watermarking schemes, and 1,800 reconstructed outputs, we identify two complementary laundering regimes: OpenAI models produce the strongest payload disruption across the evaluated schemes, whereas Nano Banana 2 shows that DwtDct remains vulnerable under high-fidelity reconstruction. Prompt ablations show that no single removal-oriented instruction is necessary for payload disruption, indicating that the effect is primarily induced by the reconstruction pathway rather than by explicit attack wording. Comparisons with conventional attacks further show that prompt-conditioned reconstruction constitutes a distinct operational attack interface. These findings motivate foundation-model reconstruction as a missing robustness condition in invisible watermark evaluation. --- ### Ready to Speak: Aligning LLMs for TTS-Friendly Text Generation Authors: Thibaut Thonet, Jos Rozen, Laurent Besacier Published: 2026-09-01 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.01246v1 Current Large Language Models (LLMs) are primarily optimized for written text, often producing outputs that are grammatically correct and helpful yet poorly suited for spoken delivery via Text-to-Speech (TTS). In this work, we study how to make LLMs natively generate TTS-friendly text, which we frame as a preference alignment problem: instead of relying on downstream rewriting modules, we directly align LLMs to generate text optimized for spoken delivery. We introduce two preference datasets spanning different target domains, CORA and Recipe, which contain paired TTS-friendly and TTS-unfriendly responses. We further propose an evaluation suite combining a pattern-based heuristic metric, a TTS$\to$ASR evaluation pipeline, and a MUSHRA listening study with human judges. Our experiments compare the recently proposed Feature-aware Sampling and Tuning (FaST) framework -- leveraging interpretable features instead of a black-box reward model -- against an array of alignment baselines on the TTS-friendly generation task. Notably, we found that FaST achieves the best overall tradeoff between TTS-friendliness and helpfulness across various settings. We also identified a strong correlation between our different metrics, highlighting the ability to reliably assess TTS-friendliness via an efficient heuristic. --- ### Explore More, Drift Less: Outcome-Only Reinforcement Learning Can Suffice for Long-Horizon Interactive Agents Authors: Liming Pu, Xiaoxia Li, Yifu Liu, Teng Cao, Bin Yang Published: 2026-09-01 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.01245v1 Reinforcement learning is a natural way to post-train LLM agents for long-horizon interactive tasks judged only by end-of-task verification, yet a shared belief holds that outcome-only RL soon hits a ceiling on small open models. Recent work therefore compensates around the training with denser rewards, SFT priors, skill libraries, curated memory, or multi-agent orchestration. We argue the ceiling is an artifact of two failures of common practice. Signal starvation: group-relative RL with sparse outcome-only rewards yields a gradient only when a task's rollout group mixes successes and failures, so under-scaled exploration silences exactly the hardest, most instructive tasks. Policy drift: squeezing many updates out of a small task pool degrades the policy itself, as an unanchored objective lets the sampling distribution collapse exactly when saturation has already made informative groups rare. We present CANOPY (Coverage-ANchored On-PolicY RL), a minimalist protocol attacking both directly: scale same-task exploration until the natural signal reappears, keep every update on-policy, KL-anchored, and confined to the agent's own action tokens, then cash in an enlarged interaction budget at test time. On AppWorld, a long-horizon interactive coding benchmark, a Qwen3-14B policy trained with CANOPY through environment interaction alone--without task-specific supervision, auxiliary credit signals, or elaborate agent scaffolding--topped the public leaderboard (Feb. 2026; Test-Normal TGC 86.9, Test-Challenge 67.6), and the same design principles lift Qwen3.5-9B on SWE-bench Verified by 16.6 points. Agentic RL alone thus internalizes long-horizon capability directly into a small open model; we plan to release the complete training stack at https://github.com/AlibabaResearch/SignalCoverageRL. --- ### Post-Training Science for Supervised Fine-Tuning Authors: Charles O'Neill, Mudith Jayasekara, Harry Partridge Published: 2026-09-01 Categories: cs.LG, cs.CL arXiv: https://arxiv.org/abs/2609.01244v1 Every supervised fine-tuning run forces the same chain of decisions, such as learning rate, batch size, LoRA or full fine-tuning, how many epochs, which optimiser, and what data to feed the model. Each of these is typically rediscovered from scratch for every new model and dataset. Here we measure them under one instrument: a sweep that varies one lever at a time, and spans dense and mixture-of-experts models in two families (Qwen3 and Llama), on four real-world customer SFT datasets, for both LoRA and full fine-tuning. These datasets give a controlled testbed: each task carries an evaluation built with the customer, and its training data is produced by iterative supervised fine-tuning that refines model outputs until they pass that evaluation, so the supervised target is internally consistent and the task judge we report against is the criterion the data was built to satisfy. We ask how the optimal learning rate and batch size move with model scale, family, and data, and whether one selection rule transfers across them; what LoRA trades against full fine-tuning, and how its rank and alpha set what the adapter can learn; whether validation loss (or other metrics, such as loss landscape flatness) faithfully ranks downstream quality; whether post-training gains scale with model size and data volume, on a model ladder extended through mixtures-of-experts to 235B parameters; how many epochs to train before general instruction-following erodes; and whether a geometry-aware optimiser improves on AdamW. Each recommendation is paired with a measure of its uncertainty. --- ### From Language to Behavior: Scaling Sequence Transformers for Industrial Recommendation Ranking with Rec-Native Designs Authors: Jie Chen, Xiangqian Yu, Yanchao Lian, Tan Lu, Run Yang, Zhengchun Shang, Xing Wang, Cheng Chen, Ke Hu, Qiang Li, Tianjiu Yin, Xiaobing Liu Published: 2026-09-01 Categories: cs.IR, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2609.01240v1 Scaling Transformers has driven large gains in language modeling, but transplanting this to behavior-sequence modeling in production ranking is challenging: recommendation differs in signal quality, where behavior sequences are noisy, temporally irregular, and sparsely supervised, and in computation asymmetry, where each request scores many candidates against one shared user history under tight latency budgets. We propose ReST, a recommendation-native Transformer scaling framework. For signal quality, it introduces a sequence encoder with dual-gated attention, rotary positional and temporal embedding, stabilized residual normalization, and training-only auxiliary objectives. For computation asymmetry, it factorizes ranking into a heavy reusable encoder and a lightweight cross decoder with projection-free KV attention and token-specific parameterization, coupling user-level shared-prefix training with shared-prefix serving for compute-once, decode-many-times ranking. Across industrial and public benchmarks, ReST achieves higher accuracy and scales more consistently along sequence length, depth, and width, where LLM-style Transformer blocks saturate. A one-week online A/B test on a production advertising platform improves online AUC by 1.31% and lifts a core revenue metric by 11.93% within a 50 ms P99 budget; ReST has since been fully deployed in production, showing that behavior-sequence scaling remains a promising, under-exploited axis for production ranking. --- ### MutMem-V2: Cryptographically Authorized Mutation in Persistent Agent Memory Portable Verification and Reproducible Evidence Authors: Walid Saidi Published: 2026-09-01 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.01235v1 MutMem V1 introduced retention-preserving, cryptographically authorized mutation for persistent agent memory but did not provide a complete portable verification contract or clean-install reproduction path. MutMem V2 closes that publication gap without introducing a second memory engine. It specifies exact canonical bytes, domain-separated object and bundle commitments, mandatory recall-evidence membership and ordering, external trust anchors, identity epochs, revocation, authorization, request receipts, ordered disclosure, and three mutation terminal types. The released protocol contains 18 versioned object schemas, 39 recall vectors, 15 mutation vectors, and 37 closed recall failure reasons. Independent Node and Python implementations agree on verdict and primary reason for all 72 structural and cryptographic terminals; a production-conformance corpus agrees on 42/42 cases across 28 required classes. A clean Node v26.8.1 installation reaches first-boot, restart, and scheduler readiness with no experimental memories. A separately scoped 120-unit Canary experiment supports only explicit-marker traversal. Every public table regenerates from a self-hashed aggregate, and an independent verifier reconstructs the statistics and claim boundaries. Historical V1 empirical results remain historical. MutMem V2 supports claims about portable integrity, authorization, traceability, conformance, and reproducibility under stated assumptions; it does not establish semantic truth, universal robustness, or independent replication. --- ### Position Matters: Feature Inversion Attacks in ViT Split Inference with Token Reduction and Shuffling Authors: Stefano Leggio, Giulio Rossolini, Alessandro Biondi Published: 2026-09-01 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.01232v3 Vision Transformers (ViTs) are increasingly used in split-inference systems, where edge devices transmit intermediate token representations to a remote cloud. In this setting, token reduction lowers computation and communication costs, while token shuffling disrupts the spatial organization of the transmitted tokens, potentially limiting information leakage. However, their privacy benefits remain unclear against feature inversion attacks, which attempt to reconstruct the input from the transmitted embeddings. In this work, we show that, despite disrupting the spatial structure required by conventional reconstruction attacks, transmitted token embeddings retain substantial positional information. Based on this observation, we introduce the Spatially Aligned Reconstruction Attack (SARA), a unified pipeline that predicts token positions, restores their spatial layout, reconstructs missing embeddings using a feature-space masked autoencoder, and recovers the input image. Our results demonstrate that token shuffling provides only apparent privacy, as SARA largely reconstructs the original token organization. Token reduction offers stronger protection, but significant leakage persists when the retained tokens preserve sufficient semantic and positional information. Finally, we introduce a lightweight edge-side defense that removes positional embeddings and progressively adapts the edge-side transformer blocks through knowledge distillation. It substantially reduces attack performance against SARA, while preserving downstream task accuracy and requiring no changes to the cloud-side model. --- ### Multi-Head Self Attention is a Parameter Identification Mechanism Authors: W. Ross Morrow Published: 2026-09-01 Categories: cs.LG, stat.ML arXiv: https://arxiv.org/abs/2609.01231v1 We prove that a multi-head scaled dot product attention can be viewed as a parameter identification strategy. The ratio of unidentified parameters to the total number of parameters scales like the reciprocal of the number of heads ($1/2 \to 1/(2H)$), meaning models with more heads are structurally more identified. A subtle side effect of the mathematics observation that attention can never be fully identified. Similarly we also show that some bias terms can have no effect on softmax-based attention layers in both the single- and multiple-head settings, though this is mostly a curiosity that should have a marginal effect on model size and model training/prediction efficiency. We also touch on modern improvements to transformers including RoPE and GQA from this perspective, illustrating how those as well can improve the ratio of ``meaningful'' parameters to all parameters. Simple numerical examples demonstrate that training can indeed involve updates that overlap model-invariant subspaces that arise from a lack of identification. As part of our experiments we use a ``rebalancing'' approach that can ``fix'' updates that overlap unindentified subspaces but do not try to present evidence this should actually be adopted. Instead we simply view our numerical results as exploring and confirming the theoretical results. As a whole we discuss a purely mathematical/statistical explanation, identification, for why specific architectural choices in transformers may have improved performance. --- ### S$^2$Prune: Spatially Structured Visual Token Pruning for Multimodal Large Language Models Authors: Yuanyuan Jia, Shunpu Tang, Qianqian Yang Published: 2026-09-01 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.01224v1 Visual token pruning reduces the inference overhead of multimodal large language models (MLLMs) by retaining only a subset of visual tokens. Existing methods usually select tokens based on importance or redundancy. However, we observe that these criteria produce stable spatial biases across inputs and do not always outperform simple Uniform Grid sampling, highlighting the value of broad spatial coverage. Motivated by this, we propose S$^2$Prune, a training-free pruning method that preserves spatial coverage while adapting token density to local image structure. We first divide the image into regions and assign at least one token to each region to preserve coverage. The remaining token budget is then distributed according to Laplacian variation, giving more tokens to regions with richer structure. We then use Early Representation Change (ERC), computed from the first decoder block, to select representative tokens within each region. We evaluate S$^2$Prune across diverse settings and two MLLM architectures. On Qwen2.5-VL-7B-Instruct, it achieves the highest average accuracy among the evaluated training-free pruning methods. With only 32 of the original 576 visual tokens, it still retains 79.3% of the full-model performance. Code is available at https://github.com/yuanyuanjia71-spec/S2Prune. --- ### Prompt-Robust Language Models: Which Training Strategies Work? Authors: Frederic Sadrieh, Michal Štefánik Published: 2026-09-01 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.01217v1 Despite their strong performance, large language models remain highly sensitive to prompt formulation. Prior work addresses this through refined data construction or through dedicated robustness objectives. We reproduce and compare these strategies under controlled conditions, and measure how effective they are in addressing models' prompt sensitivity. We find the current robustness fine-tuning methods improve over standard fine-tuning and in-context learning, but the best-to-worst prompt gap remains as high as 40-57% of performance. Moreover, the recent robustness-enhancing methods we test - CoIN for contrastive alignment and PPCL for consistency regularization - often fail to outperform the simplest data construction strategy: training on one template per batch. Our diagnostics explain these results. The auxiliary objectives move the quantity they penalize, but do not generalize beyond it. Additionally, data construction strategies differ due to the conflicting signs of per-template gradients on 57-64% of parameters. Thus, batches that mix formulations force the optimizer to reconcile competing updates instead of finding a shared, prompt-agnostic one. --- ### H2Table: Hierarchical Hypergraph-Enhanced Large Language Models for Complex Table Reasoning Authors: Jia Ling, Yangfan Wang, Chen Tang, Haoming Tan, Yang Yang, Yi Guan, Jingchi Jiang Published: 2026-09-01 Categories: cs.AI arXiv: https://arxiv.org/abs/2609.01216v1 Tables are ubiquitous across diverse domains, yet reasoning over them remains a significant challenge for modern large language models (LLMs). Current approaches typically linearize tables into sequences, inherently overlooking their intrinsic two-dimensional and hierarchical structure. To address this, we propose H2Table (Hierarchical Hypergraph-Enhanced Table Reasoning), a novel framework that represents complex tables as hierarchical nested hypergraphs. To process this representation, we design a tailored hypergraph encoder to facilitate message passing between hyperedges (headers) and nodes (cells), thereby perceiving the semantic entailment relationships between them within complex tables. Furthermore, we introduce a set of learnable query vectors acting as a lightweight bridge to extract representative structural embeddings from the encoder into the LLM. Experimental results demonstrate that our approach effectively handles complex table question answering tasks with hierarchical nested headers. Notably, on the HiTab dataset, H2Table achieves an average improvement of 22.88% over state-of-the-art baselines on highly complex tables with a nesting depth of four. Our code is available at: https://github.com/lila120/h2table. --- ### REFACTOR-VLA: Unsupervised Library Learning of Typed Motor Programs Authors: Riyaaz Shaik, Chandru Venkataraman Published: 2026-09-01 Categories: cs.LG, cs.AI, cs.RO arXiv: https://arxiv.org/abs/2609.01215v1 Most vision-language-action (VLA) models -- OpenVLA, $π_0$, RT-2, RDT-1B -- are monolithic: they emit raw motor commands or short action chunks without organizing behavior into reusable abstractions, so they degrade on long-horizon tasks and resist interpretation. Existing skill-discovery methods sidestep the core question of when two action sequences are behaviorally equivalent, either clustering contrastive embeddings or delegating the judgment to a language model uncalibrated to the robot's dynamics. We introduce REFACTOR-VLA, a wake/sleep system for learning reusable skills. Its sleep phase clusters motor-program fragments under a Behavioral-Equivalence Kernel (BEK) computed from rollouts of a learned latent world model $M_φ$; its wake phase emits typed lambda terms over a Hindley--Milner-inspired vocabulary, consumed by a library-conditioned rectified-flow action decoder. Abstractions are admitted only if they pass Minimum Description Length and return-preservation gates. On LIBERO we report two findings. First, enlarging the world model from 188M to 430M parameters worsened performance on 4 of 4 suites, so capacity alone does not help. Second, the training objective matters far more: adding an auxiliary supervised contrastive (InfoNCE) loss during world-model warmup substantially improves sleep-phase clustering, giving Normalized Mutual Information at $n=3$ seeds of $0.462 \pm 0.021$ (object), $0.867 \pm 0.025$ (spatial), $0.915 \pm 0.013$ (goal) and $0.754 \pm 0.010$ (LIBERO-10), and beating the strongest published baseline on all 4 suites by a mean $Δ= +0.184$. Across providers ($n=12$) the 95% bootstrap confidence interval for mean pairwise NMI is $(0.683, 0.729)$ (mean $0.705$). The sleep phase also yields the first real-LIBERO task-language library: the decoder uses 2 of 3 admitted abstractions and rewrites all 256 sampled demonstrations. --- ### Recent Developments in Transformer Inference Deployment on FPGA Platforms: A Survey Authors: Arjan Blankestijn, Uraz Odyurt, Amirreza Yousefzadeh Published: 2026-09-01 Categories: cs.LG, cs.AR arXiv: https://arxiv.org/abs/2609.01212v1 With the rapid and continuous growth in the incorporation of machine learning models based on the Transformer architecture, capable deployment is in high demand. In this context, capable deployment refers to operational performance aspects, e.g., throughput and latency, as well as efficiency aspects, e.g., energy consumption. When it comes to the task of inference using such models, purpose-built hardware accelerators provide a lucrative alternative to common deployment choices, such as Central Processing Units (CPUs) and Graphics Processing Units (GPUs). The Field Programmable Gate Array (FPGA) platforms category is an example of such alternative accelerators, promising implementation flexibility, energy efficiency, improved latency and suitability for on-site deployment. We investigate the most recent advances, trends, and design choices for Transformer inference on FPGA platforms. We perform a systematic literature review, extracting and delving into preferred techniques for implementation and optimisation. This study and the provided taxonomy of topics could act as a guide for researchers from the academia and industry alike. --- ### Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges Authors: Rui Yang, Shuang Huang, Junhua Liu, Ziqi Zhao, Qingzhong Yan, Yuhang Sun, Cong Liu, Guoping Hu, Rui Mei, Jing Shao Published: 2026-09-01 Categories: cs.CR, cs.AI arXiv: https://arxiv.org/abs/2609.01210v2 Safety benchmarks for large language models often assess the risk of a user query, although the outcome of question answering depends on whether the response violates a policy. This distinction is critical in Chinese harmful-content evaluation, where linguistic variation and adversarial transformations can obscure risky intent. We introduce C-SafeQA, a policy-grounded benchmark for response-level Chinese safety evaluation. It comprises 538 base queries and 8,877 adversarial queries answered by four full-model LLM deployments, yielding 37,660 query-response records labeled safe, unsafe, or disputed. Reference labels are generated through agreement-aware multi-model adjudication and blind audits of stratified subsets by three safety experts. C-SafeQA supports both evaluation of target-model safety and auditing of seven automated safety judges against shared reference labels. Unsafe-response rates range from 0.93% to 3.35% on base queries and from 11.68% to 30.05% on adversarial queries. On the adversarial subset, judges show substantial trade-offs between unsafe-response recall and risk-query-conditioned safe-response false positive rate, and no judge dominates all metrics. Both acrostic transformations reduce unsafe recall for all seven judges, revealing mechanism-specific evaluator weaknesses. Dataset records, metadata, verification code, and judge scripts are publicly released to support recomputation, while benchmark construction, target-response generation, and private adjudication remain outside the release boundary. --- ### Autonomous discovery of new structure-plausibility laws for explainable and rapid crystal diagnosis and screening Authors: Zhilong Song, Lixue Cheng Published: 2026-09-01 Categories: cond-mat.mtrl-sci, cs.AI arXiv: https://arxiv.org/abs/2609.01209v1 Crystal generators and tool-using agents propose structures faster than density functional theory (DFT) energy and phonon calculations or experiments can assess them. Deciding which candidates merit expensive assessment is therefore the bottleneck, yet most screens test little beyond atomic overlap and give no chemical reason for failure. Here, our agents generate, test and actively refute two million candidate laws, leaving eight Plausibility Rules for Inorganic Structures (PRIS). These laws encode five mechanisms: short-range repulsion, ionic contact and packing, electrostatic balance, bond-valence conservation and crystallographic site complexity. Experimental structures satisfy our law sets at 82--99%, but satisfy Pauling's rules 2--5 together at only 6.5%. The strictest set detects 87.9% of damaged crystal structures, whereas distance cutoffs detect only 1.6--3.2%. PRIS plausibility is linearly correlated with synthesizability, so the PRIS-derived synthesis score (PSS) explainably screens 83.7% of hard-to-synthesize structures while retaining 80.7% of experimental structures. In a property-conditioned inverse-design run, PRIS and PSS can reduce the DFT validation queue by up to 67.3% and keep 99.2% of the candidates whose DFT-validated bulk moduli reach the design target. Beyond screening, PRIS explains why GNoME remains enriched in rare low-symmetry structures and reveals how wrong-element assignments in falsified crystal reports hide behind plausible coordinates. PRIS moves screening from a pass-or-fail verdict to a chemical reason for failure, showing that autonomous agents can discover, by active refutation, physicochemical laws that guide calculations and experiments. --- ### Towards AI-Assisted Clinical Trial Matching: Practical Considerations, Multicenter Evaluation, and Real-World Deployment Authors: Yin Fang, Qiao Jin, Shubo Tian, Lauren He, Maya Geer, Noor Naffakh, Ryan Huu-Tuan Nguyen, Zifeng Wang, Jimeng Sun, Charalampos S. Floudas, James L. Gulley, Kamilia Moalem et al. Published: 2026-09-01 Categories: cs.CL, cs.AI, cs.CY arXiv: https://arxiv.org/abs/2609.01202v1 Clinical trials are essential for advancing cancer care and drug development, but many fail because of insufficient patient enrollment. While there is growing interest in using AI to support patient recruitment, existing systems largely perform eligibility assessment alone and have rarely been evaluated in real-world oncology workflows. Here we present TrialGPT 2.0, an AI-assisted clinical trial recommendation system designed for real-world deployment. Rather than asking only whether a patient may qualify, the system also assesses which trials warrant further consideration given the patient's current clinical needs and local workflow priorities, and provides structured, inspectable explanations for expert review. Importantly, we evaluated TrialGPT 2.0 retrospectively and prospectively across multiple oncology-focused settings, spanning government, academic cancer-center, patient-advocacy, and NIH referral workflows. In retrospective multicenter cohorts comprising 288 cases, TrialGPT 2.0 retrieved at least one clinician-recommended trial in its top 10 recommendations for approximately 91% of cases while reducing clinician screening time by 55.0%. In a six-month prospective evaluation embedded in an active precision oncology tumor board, TrialGPT 2.0 contributed additional trial opportunities missed by the routine workflow, expanding patient access to clinical trial participation by 90.9%. To support scientific reproducibility, we also introduce NIH-TrialBench, a clinician-authored dataset comprising 126 diverse synthetic patient vignettes and matching scenarios from 11 NIH Institutes and Centers. Together, these results support the value of AI to assist clinical trial matching by improving clinician efficiency and identifying frequently overlooked trial opportunities, ultimately helping to expand and accelerate accrual to cancer trials. --- ### Compressing AI Traffic: Standardized Neural Network Coding of Visual-Token Representations in Split Vision-Language Inference Authors: Reza Heidari, Hamed R. Tavakoli, Juho Kannala Published: 2026-09-01 Categories: cs.CV, eess.IV arXiv: https://arxiv.org/abs/2609.01200v1 When the visual encoder and the language decoder of a vision-language model (VLM) run on different compute nodes, the intermediate visual-token embeddings become a communicated payload rather than an internal activation. We call such machine-consumed intermediate tensors AI traffic and ask how far they can be compressed with a standardized, training-free codec. We insert ISO/IEC 15938-17 Neural Network Coding (NNC) round trips on the complete visual interface of a Qwen3-VL-8B-Instruct video question answering pipeline, comprising the main visual-token representation and the DeepStack feature streams, while leaving weights, prompts, and generation untouched, and sweep the quantization parameter (QP) over a wide rate range. Closed-ended Video-MME accuracy remains close to the uncompressed reference up to a 98% reduction of the transmitted BF16 tensor and only then collapses; open-ended MLVU generation shows the same plateau-and-collapse profile under an LLM judge. This robustness is not due to near-lossless reconstruction: the decoded tensor is heavily discretized, carries substantial row-wise relative L2 error, and has a visibly steeper singular-value decay than its source. Downstream reasoning therefore depends on coarse structure and relative geometry rather than exact floating-point values, which argues for rate-task rather than rate-distortion optimization of AI traffic codecs. --- ### FinLifeBench: Exhaustive Life-Event History and Financial-State Reconstruction from Longitudinal Banking Dialogue Authors: Hangyeul Lee, Juyoung Oh, Jaeyong Ko, Sunmin Kim, Jaeik Park, Hyunkyu Kim, Jungmin Son, Pilsung Kang Published: 2026-09-01 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.01198v2 Repeated banking interactions require assistants to maintain complete, current, and traceable customer records as life changes emerge incidentally in routine requests. Existing benchmarks emphasize question answering, bounded episodes, or targeted recall rather than exhaustive longitudinal reconstruction. We introduce FinLifeBench, which evaluates two tasks over the same cumulative dialogue: reconstructing every life-event instance with its first-establishing session and reconstructing a complete 34-path financial state at consecutive checkpoints. The benchmark contains 6,000 eight-turn Korean banking sessions from 20 independent synthetic trajectories, with deterministic, exhaustive gold for 24 event types and 34 state paths and consensus quality assurance. Across eleven LLMs under a full-context condition, event-anchor recall falls from 0.591 at 15 sessions to 0.445 at 300. Errors are driven primarily by omitted events rather than poor anchor localization, while financial-state reconstruction frequently treats superseded or potentially outdated information as current; the best GCA@15 reaches 0.470. Performance on the two reconstruction tasks is only weakly associated. These results show that models can localize evidence for recovered events while still failing to maintain complete and temporally valid longitudinal records. --- ### CaRL-EM: Cost-Aware Reinforcement Learning for Entity Matching with LLMs Authors: Chaohui Guo, Michel Klein, Zhisheng Huang Published: 2026-09-01 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.01195v1 Entity matching (EM) requires fine-grained contextual understanding and domain knowledge. Recent work shows that large language models (LLMs) can serve as strong matchers across domains, but most methods either make independent pairwise decisions or rely on manually designed composite pipelines, thus lacking flexibility in realistic multi-candidate settings. At the same time, they typically ignore inference cost at scale. We formulate LLM-based EM with candidates as a cost-aware sequential decision problem and propose CaRL-EM, a reinforcement learning controller that manages LLM operations. Given the state of an anchor record, its candidate set, and the cost, CaRL-EM adaptively chooses among different operators (Match/Compare/Select/Decide) and model capacities to maximize a quality-cost objective. The policy interacts with abstract operators, allowing the same controller to be reused with different underlying LLM backends at inference time without retraining. Experiments on 7 benchmarks show that CaRL-EM (i) learns to dynamically plan the usage of inexpensive and expensive operators based on task complexity, (ii) achieves robust zero-shot transfer across diverse datasets and domains, and (iii) consistently achieves a better quality-cost trade-off than strong LLM-based baselines and manually designed pipelines, yielding a lower inference cost at comparable or higher quality. --- ### Births are difficult to predict even with rich survey and full-population register data Authors: Elizaveta Sivak, Emily M. Cantrell, Thomas Emery, Javier Garcia-Bernardo, Flavio Hafner, Kasia Karpinska, Malte Lüken, Adrienne Mendrik, Joris Mulder, Hanzhang Ren, Varun Satish, Mark Verhagen et al. Published: 2026-09-01 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.01194v1 Major life events have proven difficult to predict. Does this reflect limits of theory, data, and algorithms, or the large role of chance? We examine one outcome - having a child within three years - through a near-ideal setting for prediction: a data challenge where 147 researchers predicted births for Dutch residents aged 18-45, using survey data and full-population registers. Methods ranged from logistic regression to a large language model and transformers. Predictions were moderately accurate (best F1: register 0.59, survey 0.76); advanced models did not outperform classical ones; and the larger registers did not beat the survey. Simulating the stochastic biology of conception and pregnancy, we estimated a predictive ceiling (survey F1 ~ 0.86-0.94, register 0.88-0.96). Observed performance falls short of this ceiling, implicating imperfect data, methods, and unmodelled chance, while the ceiling itself shows that chance in reproduction alone sets a non-trivial limit on predicting individual lives. --- ### PersuaRL: Reinforcement Learning-Driven Multi-Expert Selection for Persuasive Dialogue Generation in Insurance Authors: Rohan Kirti, Akash Ghosh, Aryan Vats, Niladri Ghosh, Shipra Shriparn, Roshni Ramnani, Anutosh Maitra, Sriparna Saha Published: 2026-09-01 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.01188v1 Large Language Models (LLMs) are revolutionizing digital communication by powering conversational agents deployed across domains such as customer service, digital sales, and insurance. These agents, built on LLMs, can understand user input, retrieve relevant information, and generate coherent responses. However, while they excel at factual communication, they often lack the ability to engage in truly persuasive, context-sensitive dialogue, especially in domains like insurance, where trust and clarity are critical. Building on this need within the insurance domain, our work focuses on improving the persuasiveness of digital agents, aka LLMs. To support this, we introduce InsureDial, a Persuasive Insurance Dialogue dataset, designed to capture the nuances of persuasive communication specific to motor insurance interactions. We introduce PersuaRL, a reinforcement learning-based framework that equips LLM-driven dialogue agents with the ability to adaptively explore, select, and coordinate strategies across multiple expert modules, guided by the evolving dialogue context, to achieve more effective persuasion. We conduct extensive automatic human and qualitative evaluations on two benchmark persuasion dialogue datasets, including our InsureDial. Our evaluations consistently demonstrate that PersuaRL outperforms baseline, generating contextually appropriate and highly persuasive responses. --- ### Athena: Vulnerability-Affected Library Identification via Knowledge Graph Completion Authors: Phong Trinh Duy, Trang Dang Yen, Hung Nguyen-Huu, Bach Le, Quyet-Thang Huynh, Dieu Hoang Vu, David Lo, Thanh Le-Cong Published: 2026-09-01 Categories: cs.SE, cs.AI, cs.CR arXiv: https://arxiv.org/abs/2609.01187v1 A single vulnerability in a widely used library can cascade through millions of dependent applications, yet more than half of vulnerability database entries contain missing or incorrect affected-library information. Existing automated approaches neglect the relational structure of vulnerability databases, treating identification as an isolated text retrieval problem. In this paper, we propose Athena, the first graph-based approach for vulnerability affected library identification. Athena models vulnerability databases as a knowledge graph and reformulates the identification problem as knowledge graph completion (KGC). It comprises three key modules: a Modeling module that constructs a security knowledge graph integrating CVEs, libraries, CWE weakness types, CPE products, and software ecosystems; a Completion module that applies a modular KGC backbone to predict missing affected libraries for a given CVE via link prediction; and a Re-ranking module that retrieves KGC candidates and rescores them using a fine-tuned LLM augmented with knowledge graph embeddings, jointly leveraging structural and textual information. Our experiments on VulLib demonstrate that Athena significantly outperforms four state-of-the-art baselines, achieving a 32% improvement in Avg. F1 over the best baseline (i.e., VulLibGen). Notably, our KGC backbone with only 110M parameters already surpasses VulLibGen's best configuration at 7B parameters, demonstrating the effectiveness of graph-based modeling; the re-ranking module then provides substantial further gains, consistently outperforming the best baseline across all evaluated LLM backbones. --- ### LLMPEDIA: Browsing, Verifying, and Comparing the Parametric Encyclopedic Knowledge of LLMs Authors: Muhammed Saeed, Simon Razniewski Published: 2026-09-01 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.01182v1 Flagship language models appear saturated on benchmarks like MMLU (Hendrycks et al., 2021), scoring above 90% - yet benchmarks test only what the experimenter thought to ask, the availability bias of fixed question sets. LLMPEDIA makes this bias measurable and browsable. We recursively materialized ~1.3M articles from three model families' parametric memory (GPT-5-mini, DeepSeek-V3.2, Llama-3.3-70B) without retrieval, then audited a stratified sample of atomic claims against Wikipedia and a curated web stack, coloring every claim supported, refuted, or insufficient (Saeed and Razniewski, 2026). On a uniform random sample the true rate is 68.4% - more than 21 pp below MMLU - with 30.5% of claims insufficient: assertions no benchmark probes and the world's largest encyclopedia cannot adjudicate - long-tail knowledge or plausible hallucination, the evidence cannot tell - extending to free text the coverage gap GPTKB established for triples (Hu et al., 2025). The resulting live, open encyclopedia lets visitors inspect this frontier one claim at a time through five one-click views - link-traversal exploration, claim-level factuality, cross-model and political-persona comparison, and a guided topic drill-down - each page, claim, and verdict at a stable URL. LLMPEDIA is live at https://llmpedia.net --- ### Monocular Depth Estimation from a Single Image: Progress and Opportunities Authors: Muxin Liu, Xiaoyang Lyu, Yang-Tian Sun, Yi-Hua Huang, Ziyi Yang, Peng Dai, Xiaojuan Qi Published: 2026-09-01 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.01172v1 Monocular depth estimation has long stood as a fundamental challenge in computer vision, enabling a wide range of applications including 3D reconstruction, robotics, autonomous driving, and augmented reality. This survey traces the field's evolution from early learning-based methods to the emergence of transformative foundation models. We begin by framing the problem, distinguishing between relative and metric depth estimation, and highlighting the key challenges that have shaped a decade of research. We then present common problem formulations and introduce the most widely used datasets, covering indoor, outdoor, and synthetic data. Following this, we review major advances prior to the foundation model era, distilling core insights from influential methods that contributed to improvements in accuracy, efficiency, and robustness. The survey then turns to the recent surge of foundation-model-based approaches, categorizing them into discriminative and generative paradigms and emphasizing the critical roles of large-scale pretraining (e.g., DINOv3) and synthetic data. We compare representative models using both quantitative benchmarks and qualitative examples, and discuss natural extensions to video-based depth estimation. Further, to illustrate real-world impact, we highlight the integration of depth estimation into applications such as visual SLAM, content generation, and robot perception. Finally, we outline open challenges and promising research directions as the field advances further into the era of foundation models. --- ### Pre-carved Niches: The Formation Dynamics of Modular Task Partitions in Early LLM Training Authors: Guangqi Li, Yongxin Li Published: 2026-09-01 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.01170v1 Large language models exhibit a modular internal organization that mirrors well-studied functional networks of the human brain, but how this organization forms during training is unknown: prior work has characterized finished models, not the formation process. We track formation step by step: we train a Pythia-410M model from scratch (two trajectories, bf16 and fp32) and run attribution patching at every step, alongside probes for gradient norms, effective updates, weight norms, and first-order loss decomposition across 14 tasks in four cognitive domains. Three findings. First, the modular map is pre-carved: before any learning, the dominant task pair already overlaps at ~3.6x the attribution substrate (a task-independent baseline), and its layer-0 concentration is an architecture-level constant on this model family. Second, the partition locks in through two sharp jumps whose amplitudes do not track the learning-rate schedule (the second reaching 20.4 sigma quiet-window / 6.2 sigma global), accompanied by gradient-level relative deprivation--winners receive 2.25->2.73x the loser's gradient supply, 9.5-11.5 standard deviations below a random control--that does not propagate to updates or weights. Third, deviation from the substrate appears only in the domain being learned, consistent with the hypothesis that modularity tracks learning. We close by separating the feature-level account we can defend from the mechanistic questions we cannot, and we pre-register the scale-threshold hypothesis behind our ongoing 2.8B experiments. --- ### Jailbreaking Text-to-Image Models Through Cracks: Navigating Heterogeneous Safety Filters via Multi-Agent Debate Authors: Kaiyan Wen, Shijie Zhang, Lu Yu, Guangdong Bai Published: 2026-09-01 Categories: cs.AI, cs.MM arXiv: https://arxiv.org/abs/2609.01168v2 Text-to-image (T2I) models remain vulnerable to jailbreak attacks that elicit Not-Safe-For-Work (NSFW) content, despite increasingly being guarded by heterogeneous, multi-layer safety stacks combining text filters, image classifiers, and cross-modal detectors. Existing jailbreak studies either optimize against individual filters or query the complete pipeline with aggregate feedback, making it difficult to identify the active constraint and adapt to conflicts across safety layers. In this paper, we introduce the Detection Surface, a unified geometric framework that characterizes the decision boundaries induced by heterogeneous T2I safety filters and their joint effect on the jailbreak search space. This formulation reveals that successful evasion is governed by a sparse and non-convex region shaped by cross-layer conflicts, where mutations that bypass one filter may increase exposure to another. Motivated by this analysis, we propose CRACK, a multi-agent debate framework for adaptive jailbreak search that decomposes jailbreak search into exploration, diagnosis, and arbitration. CRACK coordinates an Attack Agent, a Defense Agent, and a Judge Agent to iteratively generate prompt mutations, obtain layer-specific diagnostic feedback, and optimize mutation strategies through reward-guided refinement. Through repeated rounds of debate, CRACK adapts its search direction to the evolving cross-layer constraints while preserving the original harmful intent. Extensive experiments across multiple T2I models, datasets, and safety configurations show that CRACK achieves Attack Success Rates (ASR) of up to 99.63% under composite defenses, while requiring fewer queries than existing methods and maintaining semantic fidelity. --- ### Nonparametric inference for density-dependent McKean--Vlasov diffusions Authors: Denis Belomestny, Ekaterina Morozova Published: 2026-09-01 Categories: math.ST, math.PR, stat.ML arXiv: https://arxiv.org/abs/2609.01166v1 The present research is devoted to the nonparametric estimation of a density-dependent drift coefficient in a multivariate McKean--Vlasov diffusion from independent observations at a common time, as well as the stationary density. Under certain assumptions on the (known) potential, we reduce the problem to the one-dimensional one and construct a sieve maximum-likelihood estimator based on sparse ReQU neural networks subject to structural and Hölder constraints. Using the endpoint-adapted graded approximation, we achieve the rate of $\left(b_n\log n/n\right)^{2(β+1)/(2β+3)}$ for the Kullback-Leibler divergence between the true and estimated stationary densities, with $b_n$ being at most a logarithmic factor. Similarly, it is shown that the constructed estimator for the drift coefficient converges to the true one at the rate of $\left(b_n\log n/n\right)^{β/(2β+3)}$ in the $L^2$-metric. A matching Assouad lower bound proves minimax optimality of this bound up to logarithmic factors. --- ### CopyShield: A Cross-Level Benchmark of Copyright Defenses in LLMs Authors: Maryam Alshehyari, Dushyant Singh Chauhan, Samuele Poppi, Martin Takac, Salem Lahlou, Nils Lukas Published: 2026-09-01 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.01161v1 Large language models can reproduce memorized text verbatim, yet copyright defenses are usually evaluated under incompatible protocols. We introduce CopyShield, a controlled benchmark comparing three representative defenses at distinct intervention levels: contrastive decoding (output), Direct Preference Optimization (behavioral), and activation intervention (representation). We evaluate CopyShield on two model families, LLaMA-3.1-8B and Mistral-7B-v0.3, using controlled memorization over five public-domain books and a shared protocol measuring literal leakage, calibrated non-literal leakage, utility, and degeneracy. Across these methods, intervention level is associated with distinct compliance-utility trade-offs. On LLaMA-3.1-8B, contrastive decoding remains near-degeneracy-free (0-2%) but reaches a literal-suppression floor at NV-Recall 0.192-0.203. DPO nearly eliminates literal leakage (0.263 to 0.002) but induces paraphrase-loop degeneracy in 58% of QA outputs, with no utility gain over the SFT baseline. Activation intervention attains the lowest non-literal flagging rate (1/200) by blocking 84% of non-literal queries before generation. Human evaluation confirms that DPO has low coherence, whereas activation lowers perceived copyright risk through broad refusal. On Mistral-7B-v0.3, the output- and representation-level patterns persist, while DPO degeneracy falls to 10-14%, showing that its severity is model-dependent. Together, CopyShield provides cross-level reference baselines and identifies targeted non-literal suppression as an open challenge. The code is available at https://github.com/spotai-mbzuai/CopyShield.git. --- ### Superposed Latent Autoencoder Authors: Quanling Zhao, Jiaying Yang, Tianqi Zhang, Ziyang Hao, Fatemeh Asgarinejad, Flavio Ponzina, Tajana Rosing Published: 2026-09-01 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2609.01158v1 Autoencoders typically meet tight latent-memory budgets by making each latent representation smaller, sacrificing representational capacity. We ask a different question: can multiple wider latents be stored together instead? We introduce the Superposed Latent Autoencoder (SLAE), which preserves high-capacity latent representations while sharing storage through learned superposition. SLAE transforms latents into storage-friendly codes, binds them with randomized keys, superposes multiple codes into a single memory tensor, and learns to recover each latent before decoding. Under the same storage budget, SLAE replaces irreversible dimensional bottlenecks with structured interference that can be suppressed. Across CIFAR-10/100, SVHN, STL-10, Tiny ImageNet, and a wide range of memory budgets, SLAE substantially improves the reconstruction--memory tradeoff, reducing reconstruction error by up to 56% over conventional autoencoders at matched storage. Further analysis shows that SLAE's advantage comes from making wider representations usable under the same storage budget. These gains also extend beyond reconstruction: the information preserved by SLAE improves downstream classification by up to 16.79 percentage points under the same memory budget. Our results suggest a new principle for representation compression: instead of making every latent smaller, keep representations wide and let them share memory. --- ### Subword Segmental BabyLMs: Learning to Tokenise for Sample-Efficient Pretraining Authors: Francois Meyer Published: 2026-09-01 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.01151v1 In the standard LM training pipeline, subword tokenisation is applied as a preprocessing step. Subword segmental language modelling is an alternative paradigm in which tokenisation is learned during training, allowing the model to discover subword units that optimise its training objective. In this paper, we present our submission to the 2026 BabyLM Challenge, for which we develop two new subword segmental LMs: SubSegGPT and SubSegDeBERTa. SubSegGPT is a decoder-only model that learns tokenisation during autoregressive pretraining. SubSegDeBERTa is an encoder-based model that jointly learns to generate and tokenise masked words. We train both for the Strict and Strict-small tracks. Our top submission to Strict is SubSegDeBERTa, which achieves notable gains in zero-shot evaluation. Our top submission to Strict-small is SubSegGPT, which outperforms tokenisation-based baselines. Our results show that learnable subword tokenisation can improve sample-efficiency for BabyLM pretraining. We analyse the subword learning dynamics of our models and find that tokenisation gradually converges on subword units that balance morphological alignment and fine-grained segmentation. --- ### Dotting the Eye: An Intent-Driven Image Retouching Agent for Visual Focus Enhancement Authors: Chujie Qin, Zilong Zhang, Zewei Chang, Chunle Guo, Ruixing Wang, Tao Hu, Ming-Ming Cheng, Chongyi Li Published: 2026-09-01 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.01148v1 Image retouching is commonly formulated as enhancing overall visual quality through color adjustment, but in practice, it also serves to emphasize visual focus by guiding viewers' attention toward a specific subject or region. Achieving such focus-oriented retouching is inherently challenging, as it requires well-coordinated global and local adjustments to manipulate perceptual saliency while maintaining visual naturalness. This intricate process typically demands substantial professional expertise. In this study, we propose EyeControl, a MLLM-driven agent with a diffusion-based retouching executor that enables visual focus enhancement under weak user intent. With only a few clicks or coarse strokes, EyeControl directs visual attention to the intended region, effectively "dotting the eye" of the image. The core idea is to explicitly link the weak user intention with the target editing region and the corresponding tonal adjustment operations during retouching. To achieve this, the system first interprets the intent and image content to infer the visual focus and generate structured intent guidance for the retouching executor. Second, the retouching executor is encouraged to respond more strongly to the target region, explicitly aligning its attention map with a designed pseudo-intent map. We also introduce an operation-consistency constraint to improve coordination between global and local adjustments, achieving more natural and coherent retouching. Additionally, we contribute ControlArt-Bench, a high-quality evaluation dataset for visual focus enhancement. Extensive evaluations demonstrate that EyeControl yields perceptually appealing results with stronger intent alignment. Code can be found in https://github.com/DragonisCV/EyeControl. --- ### On the Design Fundamentals of Pixel Text Representation Learning Authors: Chaohao Yuan, Ruifeng Yuan, Zhuoxu Huang, Yu Rong, Hong Cheng, Hou Pong Chan, Chenghao Xiao Published: 2026-09-01 Categories: cs.CV, cs.CL arXiv: https://arxiv.org/abs/2609.01147v1 Text-rich visual inputs require models that can read, retrieve, and compress language directly in pixel space, yet existing pixel-text encoders struggle with fixed resolution pretraining, visual shortcut learning, weak visual grounding, and multilingual visual text understanding. In this work, we investigate the fundamental design principles required for robust visual text representation learning. Through systematic controlled ablations, we identify four critical components: variable image resolutions and rendered font sizes provide spatial proxies for high-resolution document generalization; natural image-text pairs are indispensable for grounding and prevent text-only collapse; layout-aware rendering helps prevent pixel-level shortcuts; and a two-stage multilingual curriculum enables effective cross-lingual alignment. By integrating these principles into a scalable training recipe, we train Pixel Linguist II, a native-resolution vision encoder trained with on-the-fly rendering, unified contrastive grounding, and a multilingual curriculum over 280M training examples. Pixel Linguist II sets new state-of-the-art results on English, cross-lingual, and multilingual Visual STS and ViDoRe, while also enabling better MLLM downstream evaluation. Notably, Pixel Linguist II remains robust under 80\% visual token compression, showing great promise for optical context compression. Our code and resources are available at https://github.com/Pixel-Linguist/Pixel-Linguist-II. --- ### StainPresetNet: Stain Preset Network for Fast Multi-to-Multi Stain Normalization Authors: Hongtao Kang, Die Luo, Li Chen, Jing Cai, Junbo Hu, Xiuli Liu, Shenghua Cheng Published: 2026-09-01 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.01146v1 Stain normalization reduces color variations caused by variations in staining protocols and imaging conditions, thereby enhancing computer-aided diagnostic system performance. Traditional methods derive mapping relationships from individual or limited reference images through pixel-wise transformation, offering style flexibility but suffering from inaccurate color mapping extraction. While existing deep-learning-based approaches achieve accurate dataset-wide color mapping through complex neural networks, they face challenges including computational inefficiency, artifact generation, and fixed normalization directions requiring model retraining for directional changes. To address these limitations, we propose StainPresetNet - a novel framework that combines structural preservation with dataset-level color mapping while maintaining computational efficiency. Our method implements pixel-wise normalization guided by preset reference images, enabling multi-directional adaptability without retraining. Evaluations on cytopathology and histopathology datasets demonstrate that StainPresetNet achieves superior color mapping accuracy compared to conventional methods, effectively improves classifier generalization in diagnostic tasks, and reduces computational overhead by 90\% versus existing deep learning approaches. The proposed preset-guided mechanism facilitates flexible adjustment of normalization directions through simple reference image replacement, overcoming the directional rigidity of current deep-learning-based solutions. --- ### Revisiting Face Recognition for Monozygotic Twins: The Celeb Twins Test Set Authors: Michael Zang, Haiyu Wu, Mrinal Sharma, Kevin W. Bowyer Published: 2026-09-01 Categories: cs.CV, cs.AI arXiv: https://arxiv.org/abs/2609.01141v1 Past literature on face recognition for monozygotic (("identical") twins points to facial marks and mirror asymmetry as possible directions for improved accuracy of twins recognition. The Celeb Twins Test Set (CTTS) contains web-scraped image pairs for 80 sets of celebrity twins. It is the only twins test set with meta-data for twins with distinguishing skin marks and possible mirror asymmetry. CTTS is organized in the manner of face verification test sets such as LFW, CALFW, CPLFW, CFP-FP, and AgeDB-30. Current deep CNN matchers can achieve over 76% accuracy in classifying CTTS same-person / different-person image pairs. We show that current matchers do not make use of skin marks, or asymmetry, and discuss reasons for this. Finally, we discuss the feasibility of using generative AI tools such as Grok, ChatGPT and Gemini to create images of imagined monozygotic twins as a means to increase representation of twins in face recognition training sets. --- ### Does task decomposition improve automatic NLG evaluation? Authors: Sebastian Steindl, Nikos Voskarides, Alberto Gasparin, Diego Marcheggiani Published: 2026-09-01 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.01139v1 The LLM-as-a-judge (LLMaJ) framework has emerged as a promising solution for cheap, reproducible, reference-free Natural Language Generation (NLG) evaluation. Prior work seeks to improve LLMaJ by decomposing evaluation tasks into simpler sub-tasks. In this work, we systematically compare LLMaJ methods with and without decomposition on multiple NLG datasets. We find no evidence that LLMaJ with task decomposition leads to performance gains over a fair baseline that does not use decomposition. Instead, we find that previously reported performance gains in decomposition-based LLMaJ stem from using human labels as training data, and not task decomposition itself. Also, we find that, when human labels are available, LLMaJ without using task decomposition can perform comparably to human annotators. --- ### Different Changes Require Different Reasoning: Change-Type-Specialized Experts for Robust Change Captioning Authors: Jiyoung Park, InJae Oh, Jung Uk Kim Published: 2026-09-01 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.01136v1 Change captioning is the task of generating natural language descriptions that explain the changes between a pair of images. Although different change types (e.g., color shifts, object additions) exhibit distinct visual cues and require specialized reasoning processes, existing methods often overlook these distinctions. To address this limitation, we propose Multi-Expert Diagnosis for Image Change (MEDIC), a novel framework that introduces change-type awareness by explicitly modeling change categories. MEDIC employs type-specialized memory experts that dynamically retrieve type-relevant visual patterns conditioned on the input. This design enables each expert to capture diverse variations within its change type while focusing on the most informative visual cues. By softly routing inputs across type-specialized experts and learning dedicated representations for each change category, MEDIC generates more precise and type-aware change descriptions. Extensive experiments demonstrate that the proposed MEDIC consistently outperforms existing methods across diverse and challenging datasets. The code is available at \href{https://github.com/VisualAIKHU/MEDIC}{GitHub}. --- ### Overfitting Mitigation via Singular Value Decomposition in Minimum Bayes Risk Decoding Authors: Riza Setiawan Soetedjo, Yusuke Sakai, Hidetaka Kamigaito, Katsuhiko Hayashi, Taro Watanabe Published: 2026-09-01 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.01135v1 Minimum Bayes Risk (MBR) decoding enables high-quality text generation by selecting the hypothesis that maximizes a utility metric over sampled pseudo-references. However, it is highly susceptible to metric overfitting: it can irregularly inflate the chosen utility metric at the direct expense of other unoptimized evaluation metrics. To mitigate this, we introduce SVD-MBR, which frames the pairwise utility matrix as a noisy information signal. By computing a low-rank approximation via Singular Value Decomposition (SVD) and retaining only the top-$k$ components, we effectively decouple true consensus from metric noise. Experiments demonstrate that SVD-MBR successfully regularizes decoding, yielding substantial gains across a range of generalized metrics. Furthermore, we reveal that this denoising is metric-dependent: neural metrics encode a robust low-rank consensus ideal for SVD, whereas surface-level metrics struggle to separate signal from metric noise. --- ### Scaled Idempotence in Transformer Attention: Paired OV Geometry and Shared-Value Algebras Authors: Jiming Feng, Junliang Li Published: 2026-09-01 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.01129v1 We identify a recurrent algebraic regularity in Transformer attention: a sparse subset of effective OV operators $T=OV^\top$ nearly closes under composition, $T^2\approxαT$. Across six pretrained endpoints spanning 2.8B--235B parameters, 3.98--8.00% of heads reach squared closure alignment $\mathcal{P}\geq0.9$, while no matched within-layer O/V mismatch does. An exact principal-coordinate factorization, $T=Q_OKQ_V^\top$ and $T^2=Q_O(KDK)Q_V^\top$, separates within-support transport from read--write return geometry. Across all 7,304 heads in nine MHA/GQA models, scrambling only the orientation of $K$ while preserving singular values, norms, factor spans, and principal angles reduces median closure from 0.336 to $1.04\times10^{-4}$; trained orientation wins for 98.64% of heads and in every layer. Constructive searches show that high closure is feasible in every surveyed layer, but usually not attained. Retrospective trajectories in three independently trained lineages further separate broadly available capacity from the orientations attained by final strong heads. Under exact value sharing, headwise closure extends to a right-action algebra, $T_iT_j=α_jT_i$. Seven-model experiments verify the approximate law and reveal distinct oblique projections with a shared value-defined kernel. These results characterize scaled idempotence as a sparse trained orientation within broadly available geometric capacity and show how value sharing extends a headwise relation into a local operator algebra. --- ### When Does Online Adaptation Pay on the Edge? A Leakage-Free Evaluation of Warmup, Learning-Rate Selection, and Resource Trade-offs for Time-Series Forecasting Authors: Takumi Fujimoto, Hiroaki Nishi Published: 2026-09-01 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.01126v1 Online adaptation can help edge time-series forecasting under distribution drift, but its measured benefit is sensitive to evaluation choices. We study six public multivariate streams, including building-sensor and smart-meter data, under a leakage-free streaming protocol. We identify two additional sources of comparison bias. First, the warmup budget of the static baseline has a two-sided effect: insufficient warmup undertrains the baseline, whereas excessive warmup can degrade its pre-drift generalization. Across six dataset-backbone settings, the estimated adaptation benefit changes by 3.0 to 18.8 percentage points (pp) over the 1,000-20,000-step warmup range. Second, comparing SGD with momentum (SGD+m) and Adam at a shared default learning rate conflates optimizer quality with rate sensitivity. We select both the warmup budget and each optimizer's online rate using a held-out pre-drift validation slice without accessing test data. Under this validation-only procedure, Adam outperforms SGD+m in 310 of 360 evaluated cells, while 4 Adam cells remain below the static baseline. We further characterize accuracy against adaptation-state memory and A100-measured per-update latency for full, head-only, and calibration-based adaptation. In the evaluated PatchTST frontier settings, several parameter-efficient variants are nondominated on the adaptation-state-memory axis. Smart-meter analyses also show that reported gains depend on meter-selection rules. These findings support a validation-only commissioning procedure, while target-device latency and energy remain to be measured. Code, data, and all reported numbers: https://github.com/keiotakmin/tsf-edge-adaptation. --- ### P-PatchDiff: Progressive Patch Diffusion Models for Low-light Image Enhancement Authors: Ruoyu Guo, Haonan Zhong, Maurice Pagnucco, Yang Song Published: 2026-09-01 Categories: cs.CV arXiv: https://arxiv.org/abs/2609.01123v1 Recent advancements in low-light image enhancement have leveraged diffusion models for their strong ability to generate perceptually realistic, detailed images. Patch diffusion models further offer a promising solution to size-agnostic image restoration while improving efficiency. However, existing methods typically rely on small, fixed patches (e.g., 64$\times$64) that cannot capture image-level brightness context, whereas enlarging the receptive field improves brightness and colour estimation but substantially increases computational cost. Moreover, low-light images often exhibit uneven brightness across regions, making it necessary to ensure that locally enhanced patches remain visually coherent when combined into the full image. To address these limitations, we propose P-PatchDiff, a scalable progressive patch diffusion framework for low-light image enhancement that dynamically adjusts patch size throughout the denoising process, enabling a gradual shift from local to global views. A Multi-Patch Alignment strategy is also introduced to normalise features across varying patch scales using an estimated global brightness proxy. Rather than pursuing pixel-level reconstruction accuracy, P-PatchDiff focuses on scalability and coherent brightness across the whole image, allowing the model to perceive multi-scale information and better enhance regions with varying brightness. We empirically demonstrate that P-PatchDiff effectively enhances images ranging from 400 $\times$ 600 to 4K and is 80$\times$ faster than existing patch diffusion models while using less than 9GB of memory. The code is available at https://github.com/RuoyuGuo/P-PatchDiff. --- ### DNC-IMM: Early Lane-Change Intention Recognition via Neural Calibration Based on Driving Context Information Authors: Woong-Chan Byun, Seung-Hyun Kong Published: 2026-09-01 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2609.01120v1 Early recognition of lane-change intention is essential for proactive decision-making in autonomous driving and advanced driver assistance systems. This paper proposes a Dual Neural-Calibrated Interacting Multiple Model (DNC-IMM) that improves adaptability to driving context while preserving the probabilistic structure and interpretability of a conventional IMM. The proposed method encodes driving-context information, including target-vehicle motion, gaps to surrounding vehicles, and relative velocities, with a neural network that calibrates both the transition-probability matrix and measurement likelihoods. The final intention is determined from the calibrated IMM mode posterior rather than from a separate direct classifier. Experiments on the highD dataset demonstrate that the proposed method reliably recognizes lane-change intentions before lane crossing and provides particularly strong performance at the earlier 2-3 s prediction horizons. --- ### Latent Recurrent Thoughts: Recurrent Refinement of Proposed Latents for Reasoning with Frozen LLMs Authors: Zhaoliang Chen, Jie Fu Published: 2026-09-01 Categories: cs.AI, cs.CL arXiv: https://arxiv.org/abs/2609.01117v1 Chain-of-thought reasoning unfolds in discrete token space: each step is committed as text, errors propagate, and eliciting good traces presupposes traces to imitate. Reasoning instead in a model's continuous representation space - where intermediate states are vectors rather than words - sidesteps these constraints, but leaves open how those latent states should be computed. We approach this along two axes. First, we keep a large language model (LLM) frozen and use it for what it is already good at - modeling and decoding sequences - while a small auxiliary network supplies continuous latent thoughts as input. Second, we produce those latents by recurrence: a tiny recurrent reasoner refines them over many steps, decoupling the depth of computation from the size of the model, so that the latents are a product of iterative processing rather than a single forward pass. We instantiate this as Latent Recurrent Thoughts (LRT): a task-dedicated proposer supplies base latents, a recurrent reasoner refines them through bounded residual corrections, and the frozen LLM decodes the answer. On symbolic reasoning with answer supervision but no reasoning traces (Countdown-4, Sudoku) and on natural-language reasoning (HumanEval, MBPP, StrategyQA), LRT substantially outperforms prior frozen-decoder continuous-space reasoning methods under an identical decoder, prompt, data, and training budget, and outperforms non-thinking-mode chain-of-thought prompting on the same backbone at a small fraction of its inference compute. --- ### EDRAC: Benchmarking Arabic Dialect Reading Comprehension Authors: Noor Abo Mokh, Kirill Chirkunov, Teresa Lynn, Nizar Habash, Reham Marzouk, Malik H. Altakrori, Younes Samih, Muhammed Abu Odeh, Nour Rabih, Rahaf Alshahrani, Hamad Alshehhi, Hamdan Al-Ali et al. Published: 2026-09-01 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2609.01113v1 Dialectal Arabic (DA) remains under-resourced compared to Modern Standard Arabic (MSA), particularly for machine reading comprehension (MRC) and question answering (QA). Existing Arabic QA benchmarks primarily focus on formal written MSA or multiple-choice QA, with limited coverage of naturally spoken dialects. Here, we aim to bridge this gap. We introduce EDRAC, the first large-scale benchmark for dialectal Arabic machine reading comprehension (MRC) and generative QA, covering five major dialects: Egyptian, Moroccan, Emirati, Syrian, and Saudi Arabic. EDRAC contains 499 passages derived from naturally occurring spoken interactions and 4,977 corresponding QA pairs generated through a human--LLM collaborative pipeline combining iterative generation, LLM-as-a-judge evaluation, and human verification. We benchmark Arabic-centric and multilingual LLMs on EDRAC using lexical and semantic metrics. Our results reveal substantial gaps between semantic answer quality and dialectal fidelity, highlighting the limitations of existing evaluation metrics for dialectal Arabic generation. EDRAC provides a realistic and challenging MRC benchmark for future research on dialectal Arabic NLP. --- ### ClinTraceBench: Source-Verifiable Longitudinal Clinical Reasoning over EHR-Derived Dialogues Authors: Huimin Wang, Zhengyi Zhao, Yutian Zhao Published: 2026-09-01 Categories: cs.CL arXiv: https://arxiv.org/abs/2609.01111v1 Clinical LLM assistants must reason over multi-visit patient trajectories, yet whether the compact history representations used to scale them---retrieval, structured timelines, LLM summaries, agentic memory---preserve the longitudinal signal clinical reasoning needs has not been measured. We introduce ClinTraceBench: 385 MIMIC-IV-derived verified dialogues with event-ID provenance, a nine-task taxonomy (T1--T9), and L0--L4 deterministic + L5 human-audit validation (98.92\% agreement). We evaluate eight history representation strategies---a no-context floor, \textit{last-visit-only}, \textit{full-context}, BGE-M3 \textit{dense-retrieval}, two compression schemes, and two agentic-memory systems (\textit{Mem0}, \textit{A-Mem})---across four backbones (DeepSeek-V3, GPT-4o-mini, Haiku~4.5, Sonnet~4.6) on 6{,}271 questions: 32 cells, 200{,}672 predictions. Four findings: (SP4) a controlled T3 injection probe isolates compression-induced \textit{relation} loss---with the attribution sentence present \textit{before} construction, \textit{Mem0}, \textit{A-Mem} and \textit{llm-summary} still recover only 0--5.3\% of the injected positives; (SP1) compressed strategies pay an aggregation tax on multi-visit trends and cross-patient comparisons; (SP2) the blind-to-full gap spans $+29.8$~pp (GPT-4o-mini) to $+62.7$~pp (Haiku); (SP3) abstention scales non-monotonically with context length. On the Pareto frontier Haiku dominates Sonnet under \textit{full-context} (\$25.76 vs.\ \$106.21), inverting the ``biggest backbone wins'' heuristic. --- ### Replicating TRACE: A Practitioner's Guide to Its Threshold and Particle Budget Authors: Alex Chadyuk, Alicia Zhang, Roy Kucukates Published: 2026-09-01 Categories: cs.LG arXiv: https://arxiv.org/abs/2609.01108v1 TRACE (Math & Lienhart, arXiv:2602.01135) reads causal graphs over event types out of a pretrained autoregressive sequence model by thresholding a per-position conditional-mutual-information estimate at a fixed tau. We independently replicate its headline synthetic result: with tau selected on a validation split, mean per-sequence F1 against exact interventional truth reaches 0.90-0.91 at vocabulary size 1000 (paper: 0.91) and 0.86-0.91 from 100 to 2000. First, the optimal threshold is pinned to the truth margin, not to any constant: at every size the errors at tau* straddle the delta = 0.05 margin defining ground truth (missed true edges lie just above it, accepted false ones just below), and the blind optimum lands near delta/2 times the estimator's calibration, confirmed out of sample at 5000. Second, at a single global threshold TRACE mostly recovers a direct, adjacent-influence graph: lag-1 true edges are recalled at 0.97-0.99, while true edges at lag 2 or more read orders of magnitude lower---the reading-scale price of randomizing mediating positions, which an exact test of direct causal effect requires when the truth is unknown. A per-lag threshold family recovers a third to a half of lag-2 truth; on lag-uniform data one validated threshold recalls every lag at 0.40-0.87, 8-26 pp below an atomic-intervention control at lags 3-6. Third, the default lag decay of the paper's synthetic benchmark concentrates about 85% of interventional truth at lag 1 and pushes the rest below the estimator's noise floor, so headline F1 there certifies lag-1 recovery only and conflates the benchmark's skew with the algorithm's own limit; a flatter decay separates the two. Fourth, F1 saturates from N = 2 particles at the selected threshold---a property of the threshold's margin over the noise floor, not of the estimator, which converges as N^(-1/2). We distill five practitioner rules. ---