# Research Papers — 2026-05 · bundle page 43 of 326

> Full-text markdown of papers #2101–2150 of 16266 published in 2026-05, newest first. AI-readable mirrors of the full 16266-paper 2026-05 archive (50 papers per file).

Canonical page: https://aipod.dev/papers/archive/2026-05/page/43
[← Previous bundle](https://aipod.dev/papers/archive/2026-05/page/42.md)
[Next bundle →](https://aipod.dev/papers/archive/2026-05/page/44.md)

### Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet

Authors: Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, Adam Pearce, Craig Citro, Emmanuel Ameisen, Andy Jones, Hoagy Cunningham, Nicholas L Turner et al.
Published: 2026-05-28
Categories: cs.AI
arXiv: https://arxiv.org/abs/2605.29358v1

We demonstrate that sparse autoencoders can extract interpretable features from Claude 3 Sonnet, a production-scale language model, addressing the open question of whether dictionary learning methods scale beyond small transformers. We trained sparse autoencoders with up to 34 million features on the model's middle layer residual stream, using scaling laws to guide hyperparameter selection. The resulting features are multilingual and multimodal (generalizing to images despite text-only training), respond to both concrete instances and abstract discussions of concepts, and can be used to steer model behavior in ways consistent with their interpretations. We find features corresponding to famous entities and locations, as well as more abstract concepts like sarcasm or errors in code. We also identify features relevant to ways in which language models might cause harm--including features representing deception, power-seeking, sycophancy, and bias--and show that these causally influence model outputs when manipulated. Additionally, we conduct analyses of feature interpretability, geometry, and computational function. However, significant limitations remain: our suite of features is incomplete, and we lack rigorous methods for evaluating whether our features faithfully capture model computations.

---

### PassNet: Scaling Large Language Models for Graph Compiler Pass Generation

Authors: Yiqun Liu, Yingsheng Wu, Ruqi Yang, Enrong Zheng, Honglei Qiu, Sijun He, Tai Liang, Jingjing Wu, Yuhan Zhou, Yiwei Zhang, Dongyan Chen, Weihan Yi et al.
Published: 2026-05-28
Categories: cs.AI, cs.LG, cs.PL
arXiv: https://arxiv.org/abs/2605.29357v1

Modern tensor compilers such as TorchInductor deliver substantial speedups on mainstream models, yet face a systematic performance ceiling on long-tail workloads -- our profiling shows that 43% of real-world subgraphs experience end-to-end slowdowns under default compilation. While LLMs offer a path toward automated optimization, existing efforts focus on standalone kernel generation. We argue that pass generation -- where LLMs author structured graph transformations that integrate directly into compiler pipelines -- is the more appropriate abstraction. We propose PassNet, the first large-scale ecosystem for LLM-based compiler pass generation, comprising: (1) PassNet-Dataset, over 18K unique computational graphs from 100K real-world models; and (2) PassBench, 200 curated long-tail fusible tasks (comprising 2,060 subgraphs in total) evaluated under the Error-aware Speedup Score (ES_t) -- a metric unifying correctness, stability, and performance -- with layered integrity defenses against systematic LLM exploitation. Experiments reveal that PassBench is both highly discriminative and genuinely unsaturated: the best frontier model trails TorchInductor by 37% in aggregate, yet on individual subgraphs LLMs achieve up to 3x speedup over the same compiler -- indicating that the bottleneck is consistency, not capability. Fine-tuning a small model on merely ~4K PassNet trajectories yields a 2.67x improvement approaching frontier-model performance, demonstrating substantial headroom and validating PassNet as live training infrastructure for advancing LLM-driven compiler optimization. All data, benchmarks, and tooling are publicly available.

---

### Neural-Behavioral Representation of Natural Whole-body Movement in Monkeys

Authors: Jieshi He, Puzhe Li, Yanan Sui, Mu-ming Poo
Published: 2026-05-28
Categories: cs.LG, q-bio.NC
arXiv: https://arxiv.org/abs/2605.29355v1

Understanding how cortical activity represents natural whole-body behaviors in primates remains challenging. Limited by the diversity of movements and inaccessibility of large-scale neural representation of whole-body kinematics, previous motor decoding studies focused on constrained tasks and limited limb movements. Here, we present a neural-behavioral recording and modeling framework for freely moving monkeys, combining large-scale epidural cortical signals from distributed sensory- and motor-related areas with synchronized multi-view motion capture through a custom-made data collection platform. We reconstructed whole-body monkey kinematics and learned a compact behavior prior using an autoregressive encoder-decoder model. Conditioned on neural signals, the model decoded accurate and realistic whole-body movement without explicit physical constraints. Our results provide a novel proof-of-concept approach for decoding natural whole-body movements in primates using large-scale intracranial neural activity.

---

### Harmless Yet Harmful: Neutral Prompting Attacks for Stealthy Hallucination Steering in Agent Skills

Authors: Chia-Yi Hsu, Chia-Mu Yu, Chun-Ying Huang, Jun Sakuma
Published: 2026-05-28
Categories: cs.CR, cs.LG
arXiv: https://arxiv.org/abs/2605.29354v2

LLM-powered coding agents increasingly participate in software development workflows by generating code, selecting dependencies, and producing package installation commands. This creates a new software supply chain risk: when an agent hallucinates a non-existent package, an attacker may register the hallucinated name and later compromise users who install it. Existing package hallucination attacks and defenses primarily focus on naturally occurring hallucinations, targeted dependency steering, or post-hoc package validation. In this paper, we introduce \emph{Neutral Prompting Attack} (NPA), a highly stealthy attack paradigm in which semantically benign instructions, such as encouraging imagination and exhaustiveness, increase package hallucination propensity without containing explicit malicious intent. Unlike targeted dependency steering, NPA does not specify an attacker-chosen package. Instead, it shifts the model's dependency generation behavior toward more speculative package names. We evaluate NPA across multiple coding-oriented LLMs and package hallucination benchmarks. Our results show that NPA increases both \emph{Hallucination ASR} and \emph{Pip Install ASR}, changes the distribution of hallucinated package names, and evades existing static-analysis, LLM-based, and agent-based Skill defenses. These findings reveal that harmless-looking prompts can covertly manipulate hallucination behavior and create downstream software supply chain risks.

---

### Harmless Yet Harmful: Neutral Prompting Attacks for Stealthy Hallucination Steering in Agent Skills

Authors: Chia-Yi Hsu, Chia-Mu Yu, Chun-Ying Huang, Jun Sakuma
Published: 2026-05-28
Categories: cs.CR, cs.LG
arXiv: https://arxiv.org/abs/2605.29354v1

LLM-powered coding agents increasingly participate in software development workflows by generating code, selecting dependencies, and producing package installation commands. This creates a new software supply chain risk: when an agent hallucinates a non-existent package, an attacker may register the hallucinated name and later compromise users who install it. Existing package hallucination attacks and defenses primarily focus on naturally occurring hallucinations, targeted dependency steering, or post-hoc package validation. In this paper, we introduce \emph{Neutral Prompting Attack} (NPA), a highly stealthy attack paradigm in which semantically benign instructions, such as encouraging imagination and exhaustiveness, increase package hallucination propensity without containing explicit malicious intent. Unlike targeted dependency steering, NPA does not specify an attacker-chosen package. Instead, it shifts the model's dependency generation behavior toward more speculative package names. We evaluate NPA across multiple coding-oriented LLMs and package hallucination benchmarks. Our results show that NPA increases both \emph{Hallucination ASR} and \emph{Pip Install ASR}, changes the distribution of hallucinated package names, and evades existing static-analysis, LLM-based, and agent-based Skill defenses. These findings reveal that harmless-looking prompts can covertly manipulate hallucination behavior and create downstream software supply chain risks.

---

### DeepFake Forensics AI: A Multi-Modal Detection and Blockchain-Anchored Evidence Management Platform

Authors: Naisha Minnah
Published: 2026-05-28
Categories: cs.CR, cs.CV
arXiv: https://arxiv.org/abs/2605.29353v1

The proliferation of AI-generated synthetic media poses a critical threat to the integrity of digital evidence in legal and forensic contexts. Existing deepfake detection systems typically address a single modality and provide no mechanism for tamper-proof evidence preservation. We present DeepFake Forensics AI, a unified platform that detects synthetic media across image, video, and audio modalities, identifies generative architecture fingerprints, and anchors forensic evidence immutably on the Ethereum blockchain. Our system trains four independent neural networks from scratch: an EfficientNet-B4 image detector (AUC = 0.9868), a Bidirectional LSTM video detector (AUC= 0.9628), an ECAPA-TDNN audio detector (EER = 18.63%), and a novel GAN fingerprinting module (accuracy = 99.88%) that identifies the generative architecture behind a fake image. Evidence files are hashed with SHA-256, stored on IPFS via Pinata, and registered on-chain via a Solidity smart contract with role-based access control. The platform provides a React frontend and FastAPI backend suitable for deployment in forensic and legal workflows. To our knowledge, this is the first system to unify multi-modal deepfake detection with blockchain-based chain-of custody management.

---

### Attention as In-Context Empirical Bayes: A Two-Stage View via Particle Dynamics

Authors: Matthew Smart, Soumya Ganguly, Nilava Metya, Alexandre V. Morozov, Anirvan M. Sengupta
Published: 2026-05-28
Categories: cs.LG, math.DS, stat.ML
arXiv: https://arxiv.org/abs/2605.29351v1

We study minimal attention-only transformers under all-token corruption and show they admit a two-stage empirical Bayes interpretation. A single attention step computes a kernel-weighted posterior mean with respect to the empirical distribution defined by the context. Depth refines this distribution through particle dynamics (Stage 1), while a long-range skip-connection carries the noisy input as a query for posterior inference (Stage 2), revealing distinct statistical roles for depth and attention residuals. The framework isolates a minimal setting in which the context itself induces a depth-dependent energy landscape governing in-context inference. We show that effective denoising can emerge without an explicit noise schedule: a fixed kernel bandwidth and finite integration horizon suffice, yielding a principled depth-noise relationship. We further establish a posterior-mean recovery guarantee for a class of well-behaved priors, where the empirical estimator converges to the Bayes-optimal predictor under asymptotic conditions. Connecting these dynamics to reverse-diffusion limits, our results provide a statistical interpretation of attention as in-context inference via sample-based posterior estimation, without explicit density modeling.

---

### ConMoE: Expert-Pool Consolidation via Prototype Reassignment for MoE Compression

Authors: Yilun Yao, Jiaming Pan, Elsie Dai, Peizhuang Cong, Yaoming Li, Tong Yang
Published: 2026-05-28
Categories: cs.AI
arXiv: https://arxiv.org/abs/2605.29350v1

Mixture-of-Experts (MoE) language models reduce per-token computation but still require storing and serving all experts, making deployment memory-intensive. Existing post-training compression methods mainly shrink this cost by pruning experts or merging their weights. We formulate post-training MoE compression as expert-pool consolidation: retaining a smaller set of pretrained experts as reusable prototypes and deterministically remapping each original expert reference to one selected prototype. This view separates the reduced expert pool from the reuse structure that represents the original expert slots, and allows prototype sharing within local layer scopes while preserving the original router interface. We propose ConMoE, a train-free prototype remapping framework that selects retained experts using calibration-based contribution and replaceability signals, then redirects original expert calls to the selected prototypes without weight updates or post-compression fine-tuning. Experiments on three pretrained MoE language models show that ConMoE matches or outperforms strong pruning and merging baselines in several settings, achieving the best average score on deepseek-moe-16b-base at both 25% and 50% routed-expert reduction, while remaining competitive on Qwen3-30B-A3B and OLMoE-1B-7B-0125. Ablations indicate that deterministic reassignment is the most stable component, whereas broader cross-layer sharing and post-hoc weight fusion are model-dependent.

---

### Draft-OPD: On-Policy Distillation for Speculative Draft Models

Authors: Haodi Lei, Yafu Li, Haoran Zhang, Shunkai Zhang, Qianjia Cheng, Xiaoye Qu, Ganqu Cui, Bowen Zhou, Ning Ding, Yun Luo, Yu Cheng
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29343v3

Speculative decoding accelerates large language model inference by pairing a target model with a lightweight draft model whose proposed tokens are verified in parallel. A common way to build draft models, like EAGLE3 or DFlash is supervised fine-tuning (SFT) on target-generated trajectories. However, we observe that SFT quickly plateaus: the draft model's acceptance length on test data stops improving. The reason is an offline-to-inference mismatch: In SFT, the drafter learns from fixed target-generated trajectories, whereas during speculative decoding it is evaluated on blocks proposed under its own policy. This motivates on-policy distillation (OPD), where the target model supervises the drafter on draft-induced states. Yet OPD remains difficult for draft models, as they cannot reliably roll out complete sequences independently, whereas target-assisted generation makes the collected sequences follow the target distribution and thus eliminates the on-policy signal. We therefore propose Draft-OPD, which uses target-assisted rollout for stable continuations and replays drafting from the verification-exposed error positions. This allows the drafter to learn from target feedback on both accepted and rejected proposals, focusing training on the draft-induced errors that limit speculative acceptance. Experiments show that Draft-OPD achieves over $5\times$ lossless acceleration for thinking models across diverse tasks, improving over EAGLE-3 and DFlash by 23\% and 13\%.

---

### Draft-OPD: On-Policy Distillation for Speculative Draft Models

Authors: Haodi Lei, Yafu Li, Haoran Zhang, Shunkai Zhang, Qianjia Cheng, Xiaoye Qu, Ganqu Cui, Bowen Zhou, Ning Ding, Yun Luo, Yu Cheng
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29343v2

Speculative decoding accelerates large language model inference by pairing a target model with a lightweight draft model whose proposed tokens are verified in parallel. A common way to build draft models, like EAGLE3 or DFlash is supervised fine-tuning (SFT) on target-generated trajectories. However, we observe that SFT quickly plateaus: the draft model's acceptance length on test data stops improving. The reason is an offline-to-inference mismatch: In SFT, the drafter learns from fixed target-generated trajectories, whereas during speculative decoding it is evaluated on blocks proposed under its own policy. This motivates on-policy distillation (OPD), where the target model supervises the drafter on draft-induced states. Yet OPD remains difficult for draft models, as they cannot reliably roll out complete sequences independently, whereas target-assisted generation makes the collected sequences follow the target distribution and thus eliminates the on-policy signal. We therefore propose Draft-OPD, which uses target-assisted rollout for stable continuations and replays drafting from the verification-exposed error positions. This allows the drafter to learn from target feedback on both accepted and rejected proposals, focusing training on the draft-induced errors that limit speculative acceptance. Experiments show that Draft-OPD achieves over $5\times$ lossless acceleration for thinking models across diverse tasks, improving over EAGLE-3 and DFlash by 23\% and 13\%.

---

### WorldMemArena: Evaluating Multimodal Agent Memory Through Action-World Interaction

Authors: Chengzhi Liu, Yuzhe Yang, Sophia Xiao Pu, Yepeng Liu, Lin Long, Yichen Guo, Nuo Chen, Zhaotian Weng, Elena Kochkina, Simerjot Kaur, Charese Smiley, Xiaomo Liu et al.
Published: 2026-05-28
Categories: cs.CV, cs.CL
arXiv: https://arxiv.org/abs/2605.29341v2

Multimodal large language models are increasingly deployed as long-horizon agents, where memory must do more than recall: it must track an evolving world, revise what has gone stale, and surface the right evidence at decision time. Existing benchmarks measure recall over static dialogue, collapse memory into a single end-of-task accuracy, and reduce visual observations to captions, leaving us unable to localize failures to writing, maintenance, retrieval, or use. The rise of agent harnesses that author their own memory sharpens this gap, since we have no principled way to compare hand-designed pipelines with self-managing alternatives. To close these gaps, we formulate multimodal agent memory as an Action-World Interaction Loop with an observable four-stage lifecycle, and instantiate it in WorldMemArena: 400 multi-session multimodal tasks spanning Lifelong Evolution (evolving personal and task states) and Agentic Execution (memory from real observations, actions, and feedback), annotated with gold memory points, updates, distractors, and evidence chains for stage-level diagnosis. This enables the first head-to-head comparison of long-context, manually designed (RAG and external memory systems), and harness-based memory agents. Results show that: (1) better memory writing and storage do not guarantee better performance; (2) multimodal memory still struggles to fully use visual evidence; (3) systems are unstable across domains and degrade on realistic agentic trajectories; and (4) harness memory is more flexible but remains costly and less reliable.

---

### A Study on Question-Answer Dataset for LLM Safety Evaluation with a Focus on Illegal Activities

Authors: Kenji Imamura, Masao Ideuchi, Atsushi Fujita
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29340v1

In this paper, we discuss question-answer dataset for LLM safety evaluation, with a focus on illegal activities. Specifically, on the basis of manual analysis of AnswerCarefully, we introduce several additional information, methods for creating question-answer examples, and a rubric for evaluating LLM-generated responses. The outcomes of this study are intended to be shared with the "JAI-Trust" project.

---

### DMC-CF: Dynamic Multimodal CounterFactual QA benchmark for Causal Reasoning

Authors: Junzhe Zhang, Huixuan Zhang, Guirong Wang, Xingyao Zhang, Pei Liu, Lin Qu, Hu Wei, Xiaojun Wan
Published: 2026-05-28
Categories: cs.CV
arXiv: https://arxiv.org/abs/2605.29339v1

With the rapid advancement of multimodal large language models (MLLMs), models have demonstrated increasingly powerful multimodal capabilities. However, whether MLLMs trained through statistical learning can truly understand the causal relationships underlying the real world remains a key research question. In recent years, numerous multimodal causal reasoning datasets have been proposed. Nevertheless, these datasets are either limited in scale or constructed from synthetic images and videos, cartoon-based content, or other non-realistic multimodal sources. To address these limitations, we collect real-world videos and construct DMC-CF-Static, a large-scale benchmark for multimodal causal counterfactual reasoning. Furthermore, to mitigate issues such as data contamination in traditional static evaluation, we represent causal events using causal graphs and propose the Dynamic Graph Intervention (DGI) framework to build the dynamic evaluation benchmark DMC-CF-Dynamic from DMC-CF-Static. Experimental results on the overall DMC-CF, which includes both static and dynamic evaluation benchmarks, demonstrate that the multimodal causal reasoning capabilities of current multimodal large language models in real-world scenarios still require substantial improvement.

---

### Enhancing Factuality through Consensus and Consistency in Summarization Using Minimum Bayes Risk Decoding

Authors: Riza Setiawan Soetedjo, Yusuke Sakai, Hidetaka Kamigaito, Jingun Kwon, Manabu Okumura, Taro Watanabe
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29336v1

Improving the quality of model-generated summaries, especially factuality, the accuracy of a summary with respect to its source content, remains a challenge. While reranking could select the optimal output from multiple generated candidates, it is limited to only using the source as guidance, resulting in unreliable summaries. To address this limitation, we propose ConSUM that reranks candidate summaries by considering two factors: consistency to the source document and consensus among the other candidates. Consensus is established using Minimum Bayes Risk (MBR) decoding over the set of generated summaries, while ensuring consistency by employing factuality-aware metrics that compare the summary against the source. Rigorous testing demonstrates that our system is competitive with existing methods, with human evaluations further confirming that its generated summaries are preferred over those from other systems. Our code is available at https://github.com/naist-nlp/ConSUM .

---

### Rethinking FID Through the Geometry of the Reference Dataset

Authors: Yunghee Lee, Byeonghyun Pak
Published: 2026-05-28
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2605.29335v1

Fréchet Inception Distance (FID) is widely used to evaluate image generators, yet lower FID does not always correspond to better sample quality. We show that this mismatch depends in part on the geometry of the reference dataset. In a controlled study across six datasets, distributional density and effective rank significantly explain how FID changes as sample quality improves. Concentrated datasets tend to yield more favorable FID trends, whereas more dispersed datasets can make FID worsen despite better samples. Attribution to precision and recall and ablations with alternative feature spaces and distances support the same conclusion. These results suggest that distributional metrics should be interpreted together with the geometry of the reference dataset for more reliable benchmarking.

---

### EarthShift: a benchmark for measuring robustness to real-world distribution shifts in Earth observation

Authors: Kelsey Doerksen, Hannah Kerner
Published: 2026-05-28
Categories: cs.CV
arXiv: https://arxiv.org/abs/2605.29330v1

Current Earth observation benchmarks focus on measuring performance on diverse tasks and applications, typically measuring generalization in-distribution. But when models are deployed, they must generalize to myriad out-of-distribution scenarios, such as new time periods, geographies, scales, and sensors. We introduce EarthShift: the first public testbed for benchmarking robustness across multiple realistic distribution shifts encountered in remote sensing. EarthShift enables users to measure distributional robustness by comparing performance in- and out-of-distribution using datasets from paired datasets from different sources, temporal windows, geographic locations, and sensors. Our experiments on 8 geospatial foundation models (GFMs) and 11 tasks covering 5 shift types show that GFMs consistently perform 15-20% worse out-of-distribution on average regardless of model architecture, size, pre-training or fine-tuning strategy. We show that GFM robustness is similar to that of generic vision foundation models, and even fully-supervised models. This highlights a need for future research to strive for improvements in distributional robustness, not just performance, which can be benchmarked using EarthShift. We release our code and datasets to provide a testbed to guide future work to create foundation models that are robust and reliable in real-world applications. Code and data for EarthShift are available at: https://earthshift.github.io

---

### Mixing Vector Model for Copolymer Inference via Mixed Integer Linear Programming

Authors: Jianshen Zhu, Raveena Rai, Taiyo Sohkawa, Naveed Ahmed Azam, Kazuya Haraguchi, Liang Zhao, Tatsuya Akutsu
Published: 2026-05-28
Categories: q-bio.QM, cs.LG
arXiv: https://arxiv.org/abs/2605.29329v1

A novel two-phase molecule inference framework, mol-infer, has recently been developed to infer chemical graphs with prescribed abstract structures and desired property values through mixed integer linear programming (MILP) under the two-layered model, with guaranteed optimality and exactness relative to the given learned prediction function and structural constraints. In this study, we extend this framework to copolymers by introducing a simple feature representation, called the mixing vector (MV) model. In the proposed model, a copolymer feature vector is represented as a convex combination of MILP-tractable monomer descriptors weighted by the mixing ratio of the constituent monomers. This representation does not require explicit sequence-class information and is therefore naturally compatible with MILP-based inverse design. Under this model, we construct prediction functions for several copolymer property datasets using artificial neural networks, reduced quadratic multiple linear regression, and random forests. The proposed representation achieves practically useful predictive performance across multiple physicochemical property datasets; in particular, the best test R^2 score exceeds 0.7 for nine of the ten datasets and exceeds 0.9 for six datasets. We also formulate a multi-monomer inverse-design problem under the MV representation with a prescribed mixing ratio and show that the resulting MILP instances remain tractable, even for three-monomer settings. Finally, we perform an external consistency check by re-evaluating the inferred candidates and comparing the re-computed property values with those predicted by the learned model. Overall, the proposed framework gives a tractable first step toward model-level exact inverse design of copolymers under the two-layered model.

---

### Reasoning-preserved Efficient Distillation of Large Language Models via Activation-aware Initialization

Authors: Junlin He, Yihong Tang, Tong Nie, Guilong Li, Binyu Yang, Jinxiao Du, Lijun Sun, Wei Ma
Published: 2026-05-28
Categories: cs.CL, cs.LG
arXiv: https://arxiv.org/abs/2605.29327v1

Efficient Distillation (EDistill) compresses large language models (LLMs) by structured pruning parameters and tuning lightweight modules with high training efficiency. Although these EDistilled LLMs achieve state-of-the-art (SOTA) performance on general ability benchmarks relative to similarly sized LLMs, we identify a severe degradation in their multi-step reasoning ability, which we term reasoning collapse. We systematically analyze the geometric origins of reasoning collapse and show that the SOTA EDistill method based on width-reducing projection matrices suffers from eRank collapse, in which the effective rank (eRank) of hidden representations drops. We theoretically explain how singular values of randomly initialized projection matrices become unevenly distributed, leading to eRank collapse and thus token indistinguishability. To address this issue, we propose RED (Reasoning-preserved Efficient Distillation) for LLMs, which introduces activation-aware initialization to initialize projection matrices as channel-selection matrices, thus theoretically mitigating eRank collapse. Experiments on Llama and Qwen series demonstrate that RED substantially recovers reasoning while maintaining high training efficiency and SOTA general ability.

---

### NeuroEdge: Real-Time Hand Gesture Recognition with High-Density EMG Using Deep Learning at the Edge

Authors: Peter Chudinov, Zhenyu Lin, Jay Motamarry, Srihita Panati, Xiaorong Zhang, Zhuwei Qin
Published: 2026-05-28
Categories: cs.LG
arXiv: https://arxiv.org/abs/2605.29326v1

High-density electromyography (HD-EMG) has emerged as a powerful modality for decoding fine-grained neuromuscular activity, enabling real-time neural-machine interfaces (NMIs) for applications such as prosthetic control, rehabilitation, and augmented interaction. While deep learning approaches such as convolutional neural networks (CNNs)have demonstrated high classification accuracy for EMG-based gesture recognition, their deployment on embedded hardware remains a major challenge due to computational and memory constraints. This paper presents NeuroEdge, a real-time HD EMG-based NMI system that performs gesture recognition entirely on resource-constrained microcontrollers. The system features two custom-designed modules: the HD-EMG StreamBridge, a wireless communication interface that streams raw HD-EMG data from a Quattrocento amplifier to an ESP32 microcontroller; and the EdgeDL Inference Engine, a lightweight deep learning framework executing on a Sony Spresense microcontroller. A compact 1-dimensional CNN optimized for embedded inference processes, sliding windows of EMG data in real time. Data streaming and inference are pipelined and synchronized through an architecture that utilizes Direct Memory Access (DMA) for data transfer and Serial Peripheral Interface (SPI) burst communication between the ESP32 and Spresense, ensuring low-latency performance. Experimental results show that NeuroEdge achieves a real-time classification accuracy of 90% across seven hand gestures, with a total average latency of 83 ms using 192 channels of HD-EMG recorded from the forearm. Our system demonstrates the feasibility of deploying complex HD-EMG-based gesture recognition on microcontroller-based edge devices, bridging the gap between high-resolution biosignal acquisition and deep learning-based embedded inference for next-generation NMIs.

---

### Multi-Stage VLM Pipeline for Zero-Shot Traffic Accident Understanding

Authors: Fumiya Tatematsu, Fumihiko Takahashi
Published: 2026-05-28
Categories: cs.CV
arXiv: https://arxiv.org/abs/2605.29325v1

We present the 1st-place solution to the ACCIDENT challenge at the CVPR 2026 AUTOPILOT Workshop, which asks for zero-shot prediction of accident timing, impact centroid, and collision type from CCTV footage. On a frozen Qwen3-VL-32B-Instruct checkpoint we build a three-stage pipeline (full-video joint prediction, time refinement, and single-frame grounding of the impact centroid), run the same pipeline a second time on a 235B Mixture-of-Experts sibling, blend the two outputs 9:1, and finally snap each predicted point onto the nearest vehicle detection. The final system reaches Public LB 0.55469 / Private LB 0.57080, roughly +0.21 over the strongest host baseline (Molmo-7B, 0.358) and wins the challenge. We ablate each component, report the negative results that shaped the final design, and release the code at https://github.com/fuumin621/cvpr2026-accident-1st-place-solution.

---

### STAMP: Training Explicit Memory for Mobile GUI Agents in Controllable and Scalable Virtual Environments

Authors: Junyang Wang, Haiyang Xu, Xi Zhang, Zhaoqing Zhu, Ming Yan, Jieping Ye, Jitao Sang
Published: 2026-05-28
Categories: cs.CL, cs.CV
arXiv: https://arxiv.org/abs/2605.29324v1

Mobile GUI agents excel at immediate reactive control but frequently fail in realistic, long-horizon tasks that require memory. This failure stems from a fundamental conflict between limited context windows and token-heavy screenshots. To save the limited context, agents must progressively discard older visual history, permanently losing crucial transient information. Furthermore, existing action-centric datasets fail to teach agents what or when to explicitly memorize, and augmenting static real-world data is prohibitively expensive and lacks interactive verification. To resolve this, we present STAMP, a framework that trains explicit memory in mobile agents through controllable virtual environments, where deterministic memory variables are programmatically injected into synthesized tasks to control what must be memorized, when it should be encoded, and when it must later be retrieved, thereby producing verifiable supervised data at scale and enabling online reinforcement learning through environment-driven reward feedback. Evaluated on our newly introduced Memory-World benchmark, the resulting Stamp-GUI agent achieves state-of-the-art performance among GUI-specialized models and sets a new high watermark on our Memory-World benchmark, demonstrating exceptional memory accuracy and task resilience while maintaining strong general mobile navigation capabilities.

---

### Rethinking Stepwise Model Routing: A Cost-Efficient Table Reasoning Perspective

Authors: Shenghao Ye, Yuxiang Wang, Yu Guo, Dong Jin, Shuangwu Chen, Jian Yang
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29319v1

Large Reasoning Models (LRMs) achieve strong performance on table reasoning tasks but incur substantial inference cost due to long reasoning traces. Stepwise model routing mitigates this issue by dynamically assigning reasoning steps to smaller or larger models. However, stepwise model routing for table reasoning remains underexplored. Through empirical analysis, we find that reasoning steps involving tables contain two types of tokens with distinct uncertainty distributions: table tokens grounded in table structure, such as cell values and headers, and text tokens representing surrounding natural-language reasoning. The uncertainty of both token types is correlated with the risk that the model makes an error in the next reasoning step. However, existing methods fail to model them separately, leading to suboptimal routing decisions. To address this, we propose EcoTab, a table-aware stepwise routing framework for efficient table reasoning. At each reasoning step, EcoTab separately estimates the uncertainties of table tokens and text tokens, maps them to next-step failure risks for the small model, and combines the two risks for routing. Experiments on multiple table reasoning benchmarks show that EcoTab consistently outperforms strong baselines and achieves a better balance between accuracy and efficiency.

---

### FreeForm: Reduced-Order Deformable Simulation from Particle-Based Skinning Eigenmodes

Authors: Donglai Xiang, Vismay Modi, Rishit Dagli, Ty Trusty, Gilles Daviet, Anka He Chen, Nicholas Sharp, David I. W. Levin
Published: 2026-05-28
Categories: cs.GR, cs.CV
arXiv: https://arxiv.org/abs/2605.29318v1

We present a novel formulation for mesh-free, reduced-order simulation of deformable hyperelastic objects. Existing work in reduced-order elastodynamic simulation represents the input geometry by either meshes, which can be difficult to obtain due to challenges in scanning and triangulating complex shapes, or by neural fields that require per-shape optimization. We propose to adopt a Reproducing Kernel Particle Method (RKPM) representation, which enables the construction of reduced-order skinning weights by solving a generalized eigensystem on the Hessian matrix of the elastic energy. We demonstrate that this formulation not only leads to a 40x training speedup compared with the per-shape optimization of neural fields, but also achieves lower simulation error when evaluated against the converged results of finite element method. We show our simulation results on a wide variety of objects in different representations including meshes and Gaussian splats, as well as the application of our method in the downstream task of robot simulation.

---

### FoRA: Fisher-orthogonal Rank Adaptation for Parameter-Efficient Fine-Tuning

Authors: Juneyoung Park, Seongbae Lee, Han-Sang Lee, Kyuho Lee, Minjae Kim, Seungheon Hyeon, Kiduk Kwon, Seongwan Kim, Jaeho Lee
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29317v2

Parameter-efficient fine-tuning(PEFT) has largely focused on LoRA and its accuracy-oriented variants, leaving the original goal of reducing trainable parameters has receivedcomparatively little attention. We introduce FoRA, which revisits this goal by reducing the number of adapted layers rather than adapter rank. FoRA selects task-informative layers via a single-pass diagonal Fisher score (under 1% of training cost) and trains the LoRA down-projection at selected layers on the Stiefel manifold, preserving column orthonormality and effective rank. FoRA consistently outperforms LoRA and DoRA at half their parameter budget, and falls within 0.7-0.8 accuracy points of AdaLoRA at one-quarter its parameter count, across five LLaMA-family backbones. Cross-architecture experiments on twelve backbones from the LLaMA, Qwen3, and Gemma families confirm consistent gains from 270M to 32B parameters. The two components combine super-additively: Fisher selection alone matches rank reduction at the same budget, while the Stiefel constraint provides the decisive additional gain.

---

### CapTalk: Text-Guided Stylization and Speech-Driven 3D Head Animation

Authors: Xuangeng Chu, Yuan Gan, Ziteng Cui, Shuhong Liu, Jian Wang, Bing Zhou, Tatsuya Harada
Published: 2026-05-28
Categories: cs.CV
arXiv: https://arxiv.org/abs/2605.29316v1

Audio-driven 3D facial animation aims to generate synchronized lip movements and vivid facial expressions from arbitrary audio clips. While existing methods can produce synchronized lip motions, they often rely on predefined identity or style latent features, which limits users' ability to freely control speaking styles. Moreover, applying a fixed style or identity to an entire audio segment typically results in facial animation styles that do not adapt to the emotional content of the audio. To address these challenges, we revisit the entanglement between style and emotion, construct a large-scale dataset with textual descriptions of both style and emotion, and propose a novel talking head generation framework that enables separate control over style and emotion. Our model takes as input both textual descriptions of speaking style and character emotion, as well as the driving audio stream, enabling real-time generation of highly synchronized lip movements and facial expressions that match the provided descriptions. Furthermore, our model supports dynamic emotion control during inference, allowing it to handle scenarios where the target emotion changes throughout the speech.

---

### PatchBoard: Schema-Grounded State Mutation for Reliable and Auditable LLM Multi-Agent Collaboration

Authors: Shuyu Zhang, Yaqi Shi, Jiarui Zhang, Yanxiao Zhao, Lu Wang
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29313v2

LLM multi-agent systems often coordinate through natural-language dialogue or loosely structured shared memory, making intermediate state difficult to validate, attribute, and audit. We introduce PatchBoard, a schema-grounded collaboration architecture that replaces inter-agent dialogue with validated JSON Patch mutations over a shared structured state. An Architect agent constructs a task-specific schema and workflow rules, while a deterministic kernel validates each proposed state mutation against schema constraints, role-specific write contracts, and runtime invariants before committing it transactionally. On 630 matched ALFWorld episodes, PatchBoard achieves an 84.6% success rate, compared with 30.8% for LangGraph and 61.6% for Flock, while reducing tokens per successful task to 45.5k, compared with 368.3k and 64.2k, respectively.

---

### PatchBoard: Schema-Grounded State Mutation for Reliable and Auditable LLM Multi-Agent Collaboration

Authors: Shuyu Zhang, Yaqi Shi, Lu Wang
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29313v1

LLM multi-agent systems often coordinate through natural-language dialogue or loosely structured shared memory, making intermediate state difficult to validate, attribute, and audit. We introduce PatchBoard, a schema-grounded collaboration architecture that replaces inter-agent dialogue with validated JSON Patch mutations over a shared structured state. An Architect agent constructs a task-specific schema and workflow rules, while a deterministic kernel validates each proposed state mutation against schema constraints, role-specific write contracts, and runtime invariants before committing it transactionally. On 630 matched ALFWorld episodes, PatchBoard achieves an 84.6% success rate, compared with 30.8% for LangGraph and 61.6% for Flock, while reducing tokens per successful task to 45.5k, compared with 368.3k and 64.2k, respectively.

---

### Rubric-Guided Process Reward for Stepwise Model Routing

Authors: Shenghao Ye, Yu Guo, Zhengheng Li, Shuangwu Chen, Jian Yang
Published: 2026-05-28
Categories: cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2605.29310v2

Stepwise model routing improves the efficiency of Large Reasoning Models (LRMs) by assigning each reasoning step to a suitable model. Recent methods formulate routing as a sequential decision process and train the router with reinforcement learning. However, although they model routing as a process, they still supervise the router with outcome rewards. Such rewards only reflect final answer correctness and fail to evaluate intermediate routing decisions, which can weaken performance and generalization. To address this gap, we propose RoRo, a rubric-guided process reward framework for stepwise model routing. RoRo first collects diverse routing trajectories and constructs preference pairs based on outcome, cost, and process quality. It then trains a Rubricor to generate a query-specific evaluation rubric and a Judge to score routing trajectories under this rubric through alternating optimization. The resulting process rewards are combined with outcome rewards to optimize the routing policy via GRPO. Experiments on five reasoning benchmarks under both same-family and cross-family settings show that RoRo consistently outperforms strong baselines and achieves better accuracy and cost trade-offs.

---

### Rubric-Guided Process Reward for Stepwise Model Routing

Authors: Shenghao Ye, Yu Guo, Zhengheng Li, Shuangwu Chen, Jian Yang
Published: 2026-05-28
Categories: cs.AI, cs.CL
arXiv: https://arxiv.org/abs/2605.29310v1

Stepwise model routing improves the efficiency of Large Reasoning Models (LRMs) by assigning each reasoning step to a suitable model. Recent methods formulate routing as a sequential decision process and train the router with reinforcement learning. However, although they model routing as a process, they still supervise the router with outcome rewards. Such rewards only reflect final answer correctness and fail to evaluate intermediate routing decisions, which can weaken performance and generalization. To address this gap, we propose RoRo, a rubric-guided process reward framework for stepwise model routing. RoRo first collects diverse routing trajectories and constructs preference pairs based on outcome, cost, and process quality. It then trains a Rubricor to generate a query-specific evaluation rubric and a Judge to score routing trajectories under this rubric through alternating optimization. The resulting process rewards are combined with outcome rewards to optimize the routing policy via GRPO. Experiments on five reasoning benchmarks under both same-family and cross-family settings show that RoRo consistently outperforms strong baselines and achieves better accuracy and cost trade-offs.

---

### GrepSeek: Training Search Agents for Direct Corpus Interaction

Authors: Alireza Salemi, Chang Zeng, Atharva Nijasure, Jui-Hui Chung, Razieh Rahimi, Fernando Diaz, Hamed Zamani
Published: 2026-05-28
Categories: cs.CL, cs.AI, cs.IR, cs.LG
arXiv: https://arxiv.org/abs/2605.29307v1

Large Language Model (LLM) search agents have shown strong promise for knowledge-intensive language tasks through multiple rounds of reasoning and information retrieval. Most existing systems access information using a retriever that takes a keyword or natural language query and returns a ranked list of documents using an index of pre-computed document representations. In this work, we explore a complementary perspective in which the search agent treats the corpus itself as the search environment and finds evidence by issuing executable shell commands. We introduce GrepSeek, an optimized direct corpus interaction (DCI) search agent that trains a compact search agent to find, filter, and compose evidence from large text corpora. To address the instability of learning behavior directly with reinforcement learning on large corpora, we propose a two-stage training pipeline. First, we construct a cold-start dataset using an answer-aware Tutor and answer-blind Planner to generate verified, causally grounded search trajectories. Second, we refine the initialized policy with Group Relative Policy Optimization (GRPO), allowing the agent to improve its task-oriented search behavior through direct interaction with the corpus. To make DCI practical at scale, we further use a semantics-preserving sharded-parallel execution engine that accelerates shell-based retrieval by up to $7.6\times$ while preserving byte-exact equivalence with sequential execution of the shell command. Experiments across seven open-domain question answering benchmarks show that GrepSeek achieves the strongest overall token-level $F_1$ and Exact Match. Our analysis also highlights the limitations of purely lexical interaction on queries with substantial surface-form variation, suggesting DCI as a practical and competitive method for search agents that can complement existing retrieval paradigms in the real world.

---

### Entropy-KL Divergence-based Token Masking: A Novel Approach for Selective Fine-tuning of Large Language Models

Authors: Qi Liu, Mingdi Sun, Yongyi He, Zhi Zheng, Tong Xu, Yi Zheng, Zhefeng Wang, Enhong Chen
Published: 2026-05-28
Categories: cs.AI
arXiv: https://arxiv.org/abs/2605.29303v1

Supervised fine-tuning (SFT) followed by reinforcement learning (RL) has become a standard post-training paradigm for large language models. This paradigm provides a cold-start for RL exploration, avoiding the inefficiency of pure RL where on-policy sampling yields insufficient positive samples. However, in practice, existing approaches often use a small amount of data for SFT initialization compared to the RL phase, which can cause the model to fit the limited samples and shift away from its pre-trained distribution. This distribution shift impedes the model's ability to effectively explore during subsequent RL training. To address this challenge, we propose that in low-data regimes, SFT should prioritize activating task-relevant capabilities rather than memorizing specific content. Along this line, we propose EKSFT (Entropy-KL Selective Fine-Tuning), which selectively masks tokens that exhibit either high entropy or high KL divergence from a reference model. By excluding these high-uncertainty, distribution-shifting tokens from imitation, EKSFT injects task-specific knowledge while preserving the integrity of the model's pre-trained distribution. Empirical evaluations on mathematical reasoning benchmarks demonstrate that EKSFT consistently outperforms standard SFT. Further RL fine-tuning from the EKSFT model yields consistently better post-RL performance, indicating improved exploration for the RL stage. Our codes and datasets are available at https://github.com/MINE-USTC/EKSFT.

---

### ViASNet: A Video Ad Saliency Network for Predicting Dynamic Saliency and Viewer Engagement

Authors: Jianping Ye, Michel Wedel
Published: 2026-05-28
Categories: cs.CV
arXiv: https://arxiv.org/abs/2605.29302v1

The digital media landscape has seen a pervasive shift toward short-form video advertising on TV, social media and e-commerce platforms. The present study focuses on deep saliency prediction for short-form video advertising. Deep saliency models have been used to generate predictions of human eye fixation patterns with the purpose of enhancing user interaction with digital technology and optimizing its design. For video ads, dynamic saliency maps capture where and when viewers are looking, revealing why video ads are effective, and how their content should be optimized. We develop and test a new deep dynamic saliency prediction model called ViASNet (Video Ad Saliency Network), which has an architecture founded on the 3D U-Net, and accommodates the influence of audio and the semantic meaning of scenes. We assess the model's performance on 151 video ads, each seen by about 20 viewers wile their eye movements were tracked, and explore the critical factors influencing model performance through ablation experiments. We calculate the entropy of the predicted saliency maps frame-by-frame as a diagnostic tool to identify ads and scenes that fail to engage viewers, and illustrate its use on test data of 15 unseen ads. Our study reveals that ad design and testing can be sped up considerably through automated systems built on deep saliency models such as ViASNet.

---

### MusTBench: Benchmarking and Advancing Temporal Grounding in Music LLMs

Authors: Daeyong Kwon, Qiyu Wu, Shinobu Kuriya, Junghyun Koo, Shuyang Cui, Zhi Zhong, Wei-Hsiang Liao, Hiromi Wakaki, Yuki Mitsufuji
Published: 2026-05-28
Categories: cs.CL, cs.AI, cs.SD
arXiv: https://arxiv.org/abs/2605.29300v2

Recent Large Audio-Language Models (LALMs) have demonstrated promising abilities in understanding musical content. However, whether their responses are grounded in the correct temporal regions of the audio remains underexplored. This limitation is particularly critical for music understanding, where key information often occurs as temporally localized events, such as instrument entries and rhythmic transitions. To address this gap, we introduce MusTBench, a music-expert-validated benchmark designed to evaluate temporal grounding in LALMs through five temporally grounded question-answering tasks. To further improve temporal grounding in existing models, we propose MusT, a novel four-stage temporal optimization recipe spanning music encoder adaptation, LLM adaptation, LLM supervised fine-tuning, and RL-based optimization. Experiments on MusTBench show that existing LALMs struggle with precise temporal grounding, while MusT brings significant improvements over strong baselines. These results establish temporal grounding as a key missing capability in current LALMs and position MusTBench as a challenging benchmark for future research in temporally grounded music understanding.

---

### MusTBENCH: Benchmarking and Advancing Temporal Grounding in Music LLMs

Authors: Daeyong Kwon, Qiyu Wu, Shinobu Kuriya, Junghyun Koo, Shuyang Cui, Zhi Zhong, Wei-Hsiang Liao, Hiromi Wakaki, Yuki Mitsufuji
Published: 2026-05-28
Categories: cs.CL, cs.AI, cs.SD
arXiv: https://arxiv.org/abs/2605.29300v1

Recent Large Audio-Language Models (LALMs) have demonstrated promising abilities in understanding musical content. However, whether their responses are grounded in the correct temporal regions of the audio remains underexplored. This limitation is particularly critical for music understanding, where key information often occurs as temporally localized events, such as instrument entries and rhythmic transitions. To address this gap, we introduce MusTBENCH, a music-expert-validated benchmark designed to evaluate temporal grounding in LALMs through five temporally grounded question-answering tasks. To further improve temporal grounding in existing models, we propose MusT, a novel four-stage temporal optimization recipe spanning music encoder adaptation, LLM adaptation, LLM supervised fine-tuning, and RL-based optimization. Experiments on MusTBENCH show that existing LALMs struggle with precise temporal grounding, while MusT brings significant improvements over strong baselines. These results establish temporal grounding as a key missing capability in current LALMs and position MusTBENCH as a challenging benchmark for future research in temporally grounded music understanding.

---

### Pocket-Dentist: On-Device Dental Image Understanding via Efficient Multimodal Large Language Models

Authors: Kai Bian, Xucheng Guo, Bin Chen, Lingyan Ruan, Yiran Shen, Ting Dang, Hong Jia
Published: 2026-05-28
Categories: cs.CV, cs.AI
arXiv: https://arxiv.org/abs/2605.29299v3

Evaluations of dental vision-language models remain fragmented across datasets, task definitions and metrics, and often ignore their computational cost. This limits their widespread deployment for dental screening outside specialist centres, where timely inference, limited hardware, and local handling of patient images are vital for practical, privacy-preserving clinical prescreening. Here we present Pocket-Dentist, an efficiency-aware benchmark for dental multimodal question answering that brings together three datasets spanning approximately 1,159 patients from BRAR and MetaDent, five task types and seven metrics. Across 14 typical VLMs, our results reveal an interesting observation: compact VLMs, such as 2B-parameter models, become competitive with much larger VLMs on most metrics after lightweight adaptation while requiring substantially lower computational costs in dental image understanding. Deployed locally on an iPhone 17 Pro, our finetuned compact VLM Pocket-Dentist-2B processed each sample in 4.31 s, reducing latency by 4.9x and memory use by 2.3x compared with a 7B baseline. Our project page is available at https://2026-icml.github.io/pocket-dentist-icml.

---

### EvoGM: Learning to Merge LLMs via Evolutionary Generative Optimization

Authors: Tao Jiang, Xinmeng Yu, Chenhao Yi, Yiling Wu, Yan Li, Ran Cheng, Dongmei Jiang, Jianguo Zhang
Published: 2026-05-28
Categories: cs.NE
arXiv: https://arxiv.org/abs/2605.29295v1

Evolutionary model merging provides a powerful framework for the automated, training-free composition of LLMs through parameter-space search. However, existing methods predominantly rely on stochastic, hand-crafted operators that overlook the underlying performance landscape of the coefficient space. We propose Evolutionary Generative Merging (EvoGM), a framework that transcends manual heuristics by employing learnable generative modeling to optimize merging coefficients. Specifically, EvoGM features a dual-generator architecture with cycle-consistent learning to adaptively sample and refine promising merging candidates. By constructing winner-loser pairs from historical search trajectories, our framework effectively captures high-performance parameter distributions and maximizes data efficiency. This generative process is seamlessly integrated into a multi-round evolutionary pipeline, where elite merged models iteratively serve as new expert foundations. Extensive experiments across diverse benchmarks demonstrate that EvoGM significantly outperforms state-of-the-art baselines, exhibiting robust performance on both seen and unseen tasks. Code and data are available at https://github.com/JiangTao97/evogm.

---

### Turbulence-Robust Dynamic Object Segmentation with Multi-Signal Priors and SAM2 Refinement

Authors: Bolian Peng, Ying Tang, Xu Liu, Long Sun, Xiaoqiang Lu
Published: 2026-05-28
Categories: cs.CV
arXiv: https://arxiv.org/abs/2605.29292v1

This technical report presents our solution for the CVPR 2026 UG2+ Challenge Track 3: Dynamic Object Segmentation in Turbulence (DOST). We design a training-free multi-signal segmentation pipeline that combines pretrained motion estimation, self-supervised semantic priors, background anomaly modeling, manually calibrated proposal fusion, and SAM2-based mask refinement. The method uses RAFT for dense motion responses, DINOv2 for semantic objectness priors, ViBe for training-free background modeling, and pretrained SAM2 for box-prompt mask refinement. Instead of optimizing an end-to-end segmentation network, our system operates entirely in inference mode. This design is suitable for the DOST setting, where severe atmospheric turbulence produces pseudo-motion, blur, and intermittent target visibility, making a single motion cue unreliable. The final submitted masks are evaluated by the official leaderboard, which reports 0.425041 mIoU and 0.457206 mDice. Since no task-specific model training or fine-tuning is performed, stronger learned temporal association, adaptive proposal selection, or task-specific adaptation may further improve the system.

---

### Diagnosing Harmful Continuation in Answer-Correct Long-CoT Training Traces

Authors: Chen He, Yuhao Wu, Lei Wang, Wenxuan Zhang, Fumin Shen
Published: 2026-05-28
Categories: cs.AI
arXiv: https://arxiv.org/abs/2605.29288v1

Long chain-of-thought (CoT) traces are widely used as supervision for reasoning-oriented LLM SFT, yet answer-correct traces can still lead to markedly different fine-tuning outcomes. We study post-conclusion continuation in answer-correct long-CoT data: a continuation where the answer appears sufficiently supported, but the trace continues with additional reasoning that remains in the supervised target. To test its training effect, we use a delete-only editor to construct answer-preserving suffix removal and compare CoT-based SFT on the original and processed traces. We observe improved SFT outcomes after removing the editor-identified post-conclusion continuation, suggesting that this continuation is harmful to training in our setting. We therefore refer to this empirically supported phenomenon as harmful continuation. Beyond this intervention, we further characterize the removed post-conclusion continuation through uncertainty and hidden-state progress. We observe persistent local uncertainty together with weakened terminal-directional progress, forming an uncertainty--geometry mismatch. Finally, we instantiate Harmful Continuation Cut (HCC), a lightweight boundary proxy that approximates the editor-identified post-conclusion continuation boundary.

---

### UniNote: A Unified Embedding Model for Multimodal Representation and Ranking

Authors: Jinghan Zhao, Wenwei Jin, Anqi Li, Jintao Tong, Luya Mo, Jiawei Li, Bin Li, Yao Hu
Published: 2026-05-28
Categories: cs.IR, cs.CV
arXiv: https://arxiv.org/abs/2605.29287v2

Item-to-Item (I2I) retrieval is a fundamental part of modern content platforms, supporting critical industrial workflows from recommendation engines to content auditing. While multimodal embedding methods have advanced general retrieval, they often falter in I2I scenarios due to the challenges of balancing global content representation with fine-grained local retrieval, the systemic inefficiency of decoupled embedding-and-ranking pipelines, and the inherent trade-offs between model precision and serving latency. To solve these issues, we propose \textbf{UniNote}, a unified embedding model designed for industrial I2I retrieval. Tailored retrieval strategies are introduced to support representation learning over complex, multimodal content at varying granularities. To operationalize these strategies, UniNote employs a two-stage training paradigm: the first stage leverages contrastive SFT to establish robust base embeddings, while the second stage refines ranking quality through a reinforcement learning (RL) process that aligns the model with content relevance. Our results show that UniNote achieves SOTA performance across diverse I2I tasks. Deployed at Xiaohongshu and integrated with Matryoshka Representation Learning (MRL), UniNote achieved significant improvements in retrieval quality and cost efficiency in large-scale applications.

---

### Do Physics Foundation Models Learn Generalizable Physics? A Bias-Aware Benchmark Across Physical Regimes and Distribution Shifts

Authors: Mengdi Chu, Yang Liu, Ayan Biswas, Han-Wei Shen
Published: 2026-05-28
Categories: cs.LG, cs.AI
arXiv: https://arxiv.org/abs/2605.29283v2

Recent physics foundation models claim general spatiotemporal forecasting ability, yet their evaluations often collapse performance into a single average score under a fixed training distribution. This makes it difficult to determine whether a model has learned generalizable physical dynamics or only performs well under particular settings. We construct a benchmark with 8 physical dynamics, 3 training-data mixtures, and 25 test regimes induced by dynamic-scale and initial-condition complexity shifts, covering in-distribution, distribution-shift, and out-of-distribution settings. We evaluate five physics foundation model architectures and four model variants per architecture (scratch and three pretrained sizes), resulting in 60,000 measurements. Our results show that current physics foundation models behave as conditional rather than universal generalists: their generality depends on the physical regime, temporal scale, initial-condition setting, pretraining, model size, and architecture. Improving the training data distribution only partially mitigates this limitation. Pretraining and scaling are also unable to reliably remove their ability biases. We argue that improving physics foundation models requires moving beyond scaling models or expanding data, toward learning mechanisms that better capture transferable physical knowledge across regimes, temporal scales, and distribution shifts.

---

### LoopFM: Learning frOm HistOrical RePresentations of Foundation Model for Recommendation

Authors: Shali Jiang, Hua Zheng, Boyang Liu, Laming Chen, Kenny Lov, Chuanqi Xu, Lisang Ding, Qinghai Zhou, Can Cui, Xiaolong Liu, Xiaoyi Liu, Yasmine Badr et al.
Published: 2026-05-28
Categories: cs.LG, cs.AI, cs.IR
arXiv: https://arxiv.org/abs/2605.29280v2

Knowledge distillation (KD) transfers a single scalar prediction from a large foundation model (FM) to compact vertical models (VMs), suffering from diminishing transfer ratio -- the fraction of FM improvement captured by the VM -- as a single scalar cannot convey the rich intermediate knowledge that larger FMs learn. To address this bottleneck, we propose LoopFM (Learning frOm HistOrical RePresentations of FM), a framework that opens a high-bandwidth transfer channel by structuring FM intermediate embeddings as input features (e.g., user history sequence) for downstream VMs, without requiring real-time FM inference at serving and architectural coupling between FM and VM. We provide a theoretical framework for LoopFM with a gain decomposition and transfer-ratio analysis. On three public benchmarks, LoopFM demonstrates strong AUC improvements (e.g., 6%+ on TaobaoAd) and complementary knowledge transfer capability with KD. On industrial-scale systems (billions of examples, trillion-parameter FMs), LoopFM approximately doubles the knowledge transfer ratio on top of KD, delivering a +0.5% conversion improvement in the first half after its initial launch, and +1.03% and +1.22% conversion improvement from two individual launches in the subsequent half.

---

### Accommodation Goes Both Ways: Studying Linguistic Convergence Between Humans and Language Models

Authors: Terra Blevins
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29278v1

As LLMs become increasingly integrated into daily life, understanding how their presence will shape human linguistic behavior is an open question. We present a large-scale study of linguistic convergence in human-LLM dialogue, examining how humans and LLMs accommodate each other's linguistic style during multi-turn conversations. Using an asymmetric convergence metric on WildChat, a corpus of real-world ChatGPT transcripts, we find that while LLMs significantly overconverge toward their users on both function word and open-class features across eight languages, human convergence rates in this setting are broadly consistent with human-human baselines. These findings suggest that accommodation in human-LLM dialogue is asymmetric: while LLMs dramatically overfit to their users' style, humans linguistically accommodate LLMs no differently than they would another person.

---

### Code-QA-Bench: Separating Code Reasoning from Documentation Memorization in Repository-Level QA

Authors: Jun Zhang, JianYing Qu, Hanwen Du, Zhongkai Sun, Yehua Yang, Qiao Zhao
Published: 2026-05-28
Categories: cs.SE, cs.AI
arXiv: https://arxiv.org/abs/2605.29277v1

We present Code-QA-Bench, a fully automated framework for synthesizing repository-level code understanding benchmarks that separates genuine code comprehension from documentation recall and pretraining memorization. The framework makes two methodological contributions: (1) an answer-first generation pipeline where a tool-equipped agent explores source code to produce verified gold answers before deriving questions, ensuring every task is grounded in real code structure; and (2) a three-condition experimental design evaluating agents under closed-book (no repository), code-only (documentation removed), and documented (full repository) conditions, with deltas directly quantifying documentation utility and memorization. We generate 528 code-derivable and 100 doc-dependent tasks across 10 Python repositories from SWE-Bench, scored by an LLM judge on accuracy, completeness, and specificity. Experiments on four frontier models reveal that code access is the dominant factor (+0.23 mean gain over closed-book), documentation provides modest additional benefit (+0.071 on doc-dependent tasks), and code-only $\approx$ documented on code-derivable tasks, validating the design. The framework is open-source and applicable to any well-documented Python repository.

---

### Prompt-Level Reward Specifications for Open-Ended Post-Training

Authors: Zijun Weng, Xiaohui Hu, Shuangyong Song, Yongxiang Li, Kaidong Yu, Xuanjing Huang
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29275v1

Open-ended post-training benefits from rewards that make prompt-specific success conditions explicit, rather than relying only on post-hoc scalar scores. In instruction following, writing, and decision-support tasks, response quality depends on local requirements, holistic preferences, and explicit constraints, but existing reward methods often leave these criteria implicit or cover only narrowly verifiable cases. We propose a prompt-level reward specification framework that separates reward specification from reward computation. Given only prompts, our framework constructs reusable task-adaptive rubrics and executable hard-constraint checkers offline, making reward criteria explicit before training and reusable across rollouts. At scoring time, artifact-anchored rubric and code scores are combined with an independent global score for residual holistic quality, yielding a normalized hybrid reward over requirement satisfaction, holistic quality, and deterministic constraints. The framework requires no human preference annotations, reference answers, or a separately trained reward model. Experiments show that the resulting reward improves offline RM-style response ranking and supports online reinforcement learning across multiple open-ended benchmarks. Ablations further show that rubrics, global scoring, and executable verification provide complementary supervision.

---

### Learnable Assessment Skills for LLM-based Automated Scoring: Rubric Construction via Iterative Optimization

Authors: Yun Wang, Xin Xia, Xuansheng Wu, Xiaoming Zhai, Ninghao Liu
Published: 2026-05-28
Categories: cs.CL
arXiv: https://arxiv.org/abs/2605.29274v1

LLM-based automated scoring approaches near-human performance, but scaling to new tasks remains bottlenecked by the per-item human configuration of upstream stages such as rubric construction. Human experts bypass this bottleneck through evaluation heuristics developed over extensive practice. We ask whether LLMs can learn similar heuristics directly from scoring experience, and formalize this as the concept of assessment skills: item-independent natural-language procedural knowledge that guides LLMs through specific stages of the scoring workflow. Focusing on rubric construction as a first instantiation, we propose an iterative framework that decomposes a skill into a fixed scaffold and learnable item-agnostic rules, refining the rules through LLM-driven diagnosis of scoring errors and validation-gated selection. The framework requires no expert-written rubric. On all ten ASAP-SAS items, optimized skills substantially improve LLM-based scoring and frequently surpass the dataset-provided expert rubric. Cross-item transfer experiments further reveal that learned skills capture both generalizable and item-specific patterns.

---

### A Theoretical and Experimental Study of a Novel Adaptive Learning Algorithm

Authors: Sakshi Kumari, Shyam Kumar M, Sushmitha P
Published: 2026-05-28
Categories: cs.LG, math.OC
arXiv: https://arxiv.org/abs/2605.29273v1

A crucial component of machine learning algorithms is minimizing loss functions with less computational cost and less oscillations. While adaptive learning rate-based optimizers have been widely used for real-world tasks, they do not guarantee convergence, which is why AMSGrad was later introduced to investigate the non-convergence behaviour of Adam. In this paper, popular adaptive optimization methods like Adam and AMSGrad are critically reviewed with an emphasis on their fundamental design concepts. To address limitations of the above mentioned optimizers, a new optimizer variant, C-Adam, is proposed based on the line of sight approach. A theoretical proof for convergence is also provided and the optimizer is validated through a number of real-life based numerical experiments.

---

### Causal Label Recovery in Payment Networks

Authors: Gaurav Dhama
Published: 2026-05-28
Categories: cs.LG, cs.AI, stat.ML
arXiv: https://arxiv.org/abs/2605.29272v1

Fraud detection models in payment networks train on chargeback labels that are systematically biased. Every label must survive three sequential gates: authorization (declined transactions generate no labels), issuer reporting (unreported fraud is invisible), and delay (pending chargebacks are missing at training time). Labels that do arrive may be corrupted by first-party misuse or issuer misclassification. A companion paper (arXiv:2605.27557) proved that these four impairments impose a minimax lower bound on detection performance. This paper asks: can that bound be achieved? We formalize the observation pipeline as a sequential missing-data problem with three propensity stages and a corruption layer, and construct the Sequential Triply Robust (STR) estimator. The STR corrects for all four impairments simultaneously and achieves the semiparametric efficiency bound -- no estimator can have lower asymptotic variance. It is sequentially triply robust: at each gate, consistency requires only that either the propensity model or the outcome regression is correctly specified, not both. We provide corruption correction via noise-rate-adjusted pseudo-labels, empirical Bayes shrinkage to stabilize inverse-propensity weights for small issuers, a plug-in variance estimator yielding valid confidence intervals, and a Bernstein concentration inequality for finite-sample guarantees. On the operational side, we derive the optimal training delay -- the maturity window that minimizes the sum of label-quality loss and model staleness -- and prove that the STR permits training on data that is days old rather than months old, decoupling model freshness from the chargeback maturity cycle. The STR provably dominates naive chargeback-based training in mean squared error for any sample size.

---

### CoHyDE: Iterative Co-Training of LLM Rewriter & Dense Encoder for Tool Retrieval

Authors: Vaishali Senthil, Ashutosh Hathidara, Sebastian Schreiber
Published: 2026-05-28
Categories: cs.AI, cs.IR, cs.LG
arXiv: https://arxiv.org/abs/2605.29271v2

Tool retrieval over large API catalogs is a core bottleneck for LLM agents: user queries arrive in colloquial, often underspecified language, while the catalog uses technical API vocabulary that no fixed encoder can bridge on its own. The two dominant training approaches, contrastive encoder fine-tuning and HyDE-style query expansion with a frozen LLM, address this problem from opposite ends and fail in complementary directions: the fine-tuned encoder excels when the query's surface form already matches the catalog but collapses when it does not, while zero-shot HyDE is more robust to underspecified queries yet generates catalog-unaware hypothetical descriptions that degrade retrieval when queries are well-formed. We introduce CoHyDE, an iterative procedure that trains the dense encoder and the LLM rewriter as a single co-evolving system: the encoder is retrained with InfoNCE on catalog-style hypothetical descriptions produced by the rewriter, and the rewriter is preference-aligned via DPO against the encoder's retrieval scores, with both sides warm-started on the tool catalog before the loop begins. On a ~10k tool subset of the ToolBench catalog, three rounds of CoHyDE improve over the strongest single-component baseline by +2.5 pp NDCG@5 on standard queries and +6.3 pp on held-out vague queries, with gains as large as +8 pp on the hardest vague tier. Ablations confirm that co-training is the key ingredient: using either component in isolation fails to match CoHyDE on both well-formed and vague queries, with losses of up to -8 pp on vague queries.

---

### CoHyDE: Iterative Co-Training of LLM Rewriter & Dense Encoder for Tool Retrieval

Authors: Vaishali Senthil, Ashutosh Hathidara, Sebastian Schreiber
Published: 2026-05-28
Categories: cs.AI, cs.IR, cs.LG
arXiv: https://arxiv.org/abs/2605.29271v1

Tool retrieval over large API catalogs is a core bottleneck for LLM agents: user queries arrive in colloquial, often underspecified language, while the catalog uses technical API vocabulary that no fixed encoder can bridge on its own. The two dominant training approaches, contrastive encoder fine-tuning and HyDE-style query expansion with a frozen LLM, address this problem from opposite ends and fail in complementary directions: the fine-tuned encoder excels when the query's surface form already matches the catalog but collapses when it does not, while zero-shot HyDE is more robust to underspecified queries yet generates catalog-unaware hypothetical descriptions that degrade retrieval when queries are well-formed. We introduce CoHyDE, an iterative procedure that trains the dense encoder and the LLM rewriter as a single co-evolving system: the encoder is retrained with InfoNCE on catalog-style hypothetical descriptions produced by the rewriter, and the rewriter is preference-aligned via DPO against the encoder's retrieval scores, with both sides warm-started on the tool catalog before the loop begins. On a ~10k tool subset of the ToolBench catalog, three rounds of CoHyDE improve over the strongest single-component baseline by +2.5 pp NDCG@5 on standard queries and +6.3 pp on held-out vague queries, with gains as large as +8 pp on the hardest vague tier. Ablations confirm that co-training is the key ingredient: using either component in isolation fails to match CoHyDE on both well-formed and vague queries, with losses of up to -8 pp on vague queries.

---

### Indexing the Unreadable: LLM-Native Recursive Construction and Search of Service Taxonomies

Authors: Wei Zheng, Yang Yan, Yiyang Shao, Jinyang Li, Zeze Chang, Yukuang Jia, Qiming Mao, Chihyung Wang, Jingbin Zhou
Published: 2026-05-28
Categories: cs.AI
arXiv: https://arxiv.org/abs/2605.29270v1

The era of the Internet of Agents (IoA) is taking shape: LLM agents are expected to fulfill user goals by orchestrating fast-growing populations of Model Context Protocol (MCP) servers, Agent-to-Agent (A2A) endpoints, reusable skills, and other LLM-callable services. Yet LLMs face a structural mismatch with this regime: effective context is a scarce resource that does not scale with the number of services. Concatenating thousands of service descriptions into a prompt overflows the context window, and even when the window is large enough, models systematically under-attend to information in the middle of long inputs, the well-documented Lost-in-the-Middle phenomenon. This is fundamentally a question of context management for service discovery. To address this, we propose an LLM-native progressive-disclosure scheme and its concrete instantiation, A2X (Agent-to-Anything service discovery): an LLM-driven pipeline that automatically organizes the registered services into a hierarchical taxonomy and walks it layer by layer at query time, so that every LLM call sees only a small candidate set highly relevant to the user query. This decouples effective-context scarcity from registry size and significantly reduces token consumption while improving retrieval accuracy. Compared to full-context dumping, A2X achieves a 6.2-point Hit Rate gain at one-ninth the prompt-token cost; compared to the state-of-the-art open-source embedding-based baseline, A2X improves Hit Rate by more than 20 points.

---