Large language models can memorize and reproduce verbatim training data, raising privacy and IP concerns. Recent theory links attention heads to this memorisation, while empirical work shows many heads are redundant and can be pruned with little performance loss. This study shows that zeroing out a large fraction of redundant heads before inference dramatically reduces memorisation (91.5% drop) and even improves language‑modeling performance on GPT‑2‑XL.
By Aryan Chowdhury, Riza Batista-Navarro
The paper introduces DAPS, a dependency‑aware premise selector for large language model theorem proving in Lean 4. By curating underused dependency resources from 53 Lean Blueprint projects and extracting typed, multi‑level Mathlib4 dependencies, DAPS employs a structural neighborhood encoder and a group‑level contrastive objective to overcome frequency skew, isolated ranking, and out‑of‑distribution degradation. It achieves a mean Recall@32 of 89.25 on Mathlib4‑Heldout, outperforming the strong selector LeanHammer, and improves theorem proving on miniF2F and IMO‑ProofBench.
By Zixuan Chen, Wenyuan Jiang, Junling Wang, Mrinmaya Sachan, Yinya Huang
The paper introduces BULLPEN, a framework that builds compact embeddings of large language models (LLMs) by jointly analyzing model answers, correctness, and question text. BULLPEN expands on prior work by adding 20 new benchmarks focused on behavior, style, and provenance, and demonstrates that embeddings derived from only 400 questions retain most of the predictive power of full‑pool methods. Across 260 open‑weight models, BULLPEN’s answer‑text encoder achieves higher accuracy on new tasks and preserves over 91% of full‑pool skill with minimal inference cost.
By Russell Ro, Qiming Cui, Reza Abbasi-Asl
The paper presents three lightweight modules to improve medical lesion segmentation, especially in data‑scarce and computationally constrained settings. The modules include a DT‑Morph Combined Loss for better topological and boundary optimization, a 3D Group Prompt that captures domain‑universal knowledge through volumetric image‑mask and text prompts, and a Many‑to‑One Conditional Segmentation Head that supports fast incremental learning with minimal computational overhead. Ablation studies on abdominal and extra‑abdominal datasets confirm the effectiveness of these components for both organ and lesion segmentation tasks.
By Ruichen Li, Zhao Yao
The paper investigates how standard evaluations of detectors for hijacked large language model (LLM) agents may conceal important weaknesses. By auditing eleven evaluations on the AgentDrift and AgentDojo benchmarks, the authors show that detectors often perform poorly when faced with rare attacks, unseen tools, or data from a different benchmark, with precision dropping dramatically at low attack rates and rankings reversing as tool coverage falls. They also find that a released LLM action monitor, despite high AUROC on its own data, would still generate many false alerts when applied to a new benchmark.
"whyItMatters":"The study highlights that current evaluation practices can give a misleading sense of detector reliability, underscoring the need for more robust, cross‑benchmark testing to ensure real‑world safety of LLM agents."
By Mahmoud Abbasi, Juan M. N\'u\~nez V., Amin Shahraki
The paper introduces a causal framework that distinguishes between measurement errors and mechanism shifts as two distinct sources of anomalies. By modeling outliers as latent interventions on both true and observed variables, the authors derive conditions under which these anomaly types can be differentiated. They present an efficient inference algorithm that localizes root causes and classifies anomaly types, achieving state‑of‑the‑art performance on synthetic and real‑world datasets.
By Hendrik Suhr, David Kaltenpoth, Jilles Vreeken
Deep Neural Sheaf Diffusion (DNSD) is a new graph neural network architecture that replaces the sheaf Laplacian with a sheaf adjacency operator to keep informative signals across many layers. It also uses normalization, odd nonlinearities, and gating, and replaces scalar attention with matrix-valued edge functions. Experiments show DNSD outperforms existing GNNs and Neural Sheaf Diffusion on both synthetic long‑range and real‑world graph benchmarks, achieving up to 30 percentage‑point gains in accuracy.
By R\'emi Bourgerie, \v{S}ar\=unas Girdzijauskas, Viktoria Fodor
The paper introduces the concept of adaptive personalization under preference epistemic uncertainty, arguing that conversational AI must adjust its response strategy as human preferences evolve. A study shows people prefer state‑contingent strategies when uncertainty changes, and the authors build a multilingual benchmark to test language models on this task. Results reveal a recognition‑generation gap: models can choose the right strategy from options but fail to generate it correctly, defaulting to exploit‑first behavior; targeted training on internal strategy representations improves performance on out‑of‑distribution personas.
By Priyanka Dey, Rajarshi Das, Vikas Bahirwani, Qi Zeng, Julia Kruk, Qi Qi
EDU-Jev is a family of single‑pass decision models (0.8B, 2B, and 4B parameters) designed for educational tasks that require selecting from a fixed set of options. The models output a probability distribution over predefined choices without generating text, using a pointer head to score each option. In twelve education benchmarks, EDU‑Jev 4B achieved 67.6% accuracy, outperforming five commercial decision APIs by 9.4 percentage points and the best open 4B model by 18.7 points, while the 2B version can process about 150 decisions per second on a single H100 GPU.
By Jiazheng Li, Yulan He
MedResearcher‑R1 is a medical deep‑research agent that uses a Knowledge‑Informed Trajectory Synthesis (KISA) framework to build medical knowledge graphs and generate complex multi‑hop question‑answer pairs, especially for rare medical entities. It also incorporates a medical retrieval engine and general‑purpose tools to reliably synthesize medical evidence. The resulting 2,100+ trajectories across 12 specialties enable supervised fine‑tuning and reinforcement learning, producing a 32B model that outperforms larger proprietary systems on MedBrowseComp and shows strong performance on GAIA and xBench benchmarks.
By Ailing Yu, HaoYang Hong, Lan Yao, Jingnan Liu, Zhe Chen, Jiajun Yin, Yuan Wang, Xinhao Liao, Zhiling Ye, Ji Li, Yun Yue, Hansong Xiao, Hualei Zhou, Chunxiao Guo, Peng Wei, Junwei Liu, Jinjie Gu
Ctrl‑G is a knowledge‑graph augmented retrieval‑augmented generation (KG‑RAG) framework that jointly limits structural and interaction exposure when sending graph evidence to a remote large language model (LLM). It constructs query‑conditioned sanitized views via structural aggregation and pruning, and uses a hierarchical controller to prune exploration paths while verifying evidence against the original KG. By reusing abstract, verified reasoning traces in local memory, Ctrl‑G reduces repeated disclosure across questions and achieves the highest accuracy on five of six benchmarks, outperforming the strongest baseline by up to 13.8 points.
By Xingyu Tan, Xiaoyang Wang, Qing Liu, Xiwei Xu, Xin Yuan, Liming Zhu, Wenjie Zhang
arXiv:2610. 41030v1 Announce Type: new Abstract: Robot perception is increasingly data driven, yet field data is slow, costly, and sometimes disruptive to collect.
By Rugved Katole, Vedant Patil, Jenna Kline, Christopher Stewart
RoboArchitect automates the design of agent harnesses that orchestrate frozen robot policies, vision tools, and direct arm commands. By observing rollouts, diagnosing failures, and iteratively revising the harness, it improves success rates from 9.7% to 43.3% on unseen scenes and from 12.7% to 40.2% on unseen tasks without updating model weights. The evolved harness also develops persistent memory and intuitive control strategies, outperforming finetuned policies especially when the benchmark diverges from the policy’s training data.
By Shengguang Wu, Suning Huang, Cherie Ho, Ria Doshi, Hao Zhu, Serena Yeung-Levy
IR275K is a curated benchmark for infrared multi‑frame super‑resolution, comprising 594 video sequences and 275,196 frames with train/validation/test splits and a reproducible x4 evaluation protocol. It evaluates seven learned baselines, the CGMamba reference model, and bicubic interpolation under both bicubic degradation and synthetic mixed sensor noise, revealing that EDVR performs best under bicubic degradation while MSDTGP excels under mixed sensor noise. Ablations with CGMamba demonstrate that using real neighboring frames improves PSNR by about 1 dB compared to center‑frame repetition, and that center‑frame conditioning also benefits reconstruction across conditions.
By Jie Deng, Heyang Wang, Changxin Wang, Junkai Shen, Hongyi Chen, Zhiping He, Hongxing Qi, Xudong Zhang, Jianyu Wang
VEvo-Harness is a causally-guarded, continual self‑evolving system designed to improve agentic image generation. It introduces a visual traceback mechanism, visual continual governance, a lifecycle‑managed harness state, and a heterogeneous sandbox policy to ensure only causally valid skill updates are incorporated. The system outperforms existing agentic methods and rivals proprietary generators while preserving continual evolution and generalization.
By Yuan Wang, Yuhang Yu, Mengting Chen, Tianrun Xu, Zihao Pan, Xuetao Feng, Wei Li
The paper introduces SPLICE, a differentially private synthetic data algorithm that uses noisy sliced Chebyshev polynomial moments to fit a distribution. SPLICE achieves a Sliced Wasserstein-1 distance of ∼O(1/√n) from the target distribution, a bound that does not depend on the data dimension. The authors also prove a new moment‑matching result and provide a practical implementation that outperforms existing DP synthetic data methods on a large benchmark.
By Lucas Rosenblatt, R. Teal Witter, Christopher Musco
Stepwise-Querying is a method that extracts step-by-step explanations from a black-box teacher model without needing its internal chain-of-thought. By independently querying the teacher for intermediate derivations, validity checks, and corrections for each step, the gathered explanations are used to fine-tune smaller student models. Experiments on mathematics and code benchmarks show that this approach outperforms three baselines in 14 of 16 comparisons, and the introduced Stepwise Explanation Information Gain metric helps diagnose the usefulness of intermediate reasoning.
By Dicong Yu, Wenhua Wang, Tian Wang
The paper introduces Step-wise Rubrics as Rewards (SRaR), a reinforcement learning framework that assigns rubric-based scores to individual reasoning steps in large language models. By normalizing step scores and combining them with outcome rewards through a decoupled advantage estimator, SRaR preserves criterion-level structure and mitigates issues like rewarding flawed steps or encouraging endless self‑correction. Experiments on 16K problems and six math benchmarks show that SRaR outperforms prior rubric methods, improving accuracy, increasing faithful reasoning rates, and reducing self‑correction loops.
By Weichu Xie, Haozhe Zhao, Wenpu Liu, Yongfu Zhu, Liang Chen, Minghao Ye, Zirong Chen, Yuqi Xu, Shuai Dong, Ziyue Wang, Xinbo Xu, Kean Shi, Ruoyu Wu, Xiaoying Zhang, Wenqi Shao, Baobao Chang, Nan Duan, Jiaqi Wang
arXiv:2602. 14020v2 Announce Type: replace-cross Abstract: Heavy-tailed data make covariance estimation sensitive to the clipping level: stronger clipping reduces variance but increases bias.
By Even He, Zaizai Yan
OpenVIO is an open‑source visual‑inertial odometry system released under the MIT license, designed for both research and practical use. It supports pinhole, fisheye, and stitched 360° camera models and includes a detailed description of its components and design choices. Benchmark results comparing OpenVIO to other open‑source VIO systems on public datasets are also provided.
By Seungwon Choi