The paper introduces two new kernel‑based methods, GAIM and Q‑DAIM, for compressing the key‑value cache in attention mechanisms by formulating the problem in a Hilbert space and applying empirical interpolation techniques. To reduce the quadratic cost of these methods, the authors propose surrogate models—including Taylor series, Performer, and Random Fourier Features—that enable linear‑time compression. The work also extends to streaming attention, offering node‑selection penalties and interpolation updates, and demonstrates competitive performance on the LongBench‑E benchmark.
By Michel Fabrice Serret, Laura Grigori
The paper presents a machine‑learning based method for finding optimal transport protocols in underdamped stochastic systems with nonlinear Landauer boundary conditions. It focuses on steering the system from a single‑peak to a double‑peak position distribution using only an external potential. The authors compare their numerical results to a perturbative expansion valid in the overdamped limit.
By Julia Sanders, Dario Lucente, Marco Baldovin, Paolo Muratore-Ginanneschi
The paper demonstrates that Matryoshka Representation Learning (MRL) embeddings contain a natural coarse‑to‑fine hierarchy, with prefix length aligning to gold taxonomy granularity. By training a simple order‑embedding probe called PEEL on these frozen embeddings, the authors can recover high‑fidelity taxonomic DAGs from a concept vocabulary alone, achieving strong results on two OLLM benchmarks and transferring to new domains without target‑edge supervision. The study shows that pre‑trained retrieval embeddings encode a decodable granularity gradient, enabling a minimal probe to extract explicit parent‑child (is‑a) structure.
By Eunchan Kim, Inwhee Joe
The paper introduces a specification-guided decoding framework that blends large language model generation with symbolic reasoning to enforce formal specifications during output generation. It offers a forward-only variant that guarantees compliance by construction and a backtracking variant that revises earlier choices when later constraints become unsatisfiable. The approach is model-agnostic, requires no retraining, and improves compliance on configuration and constraint-solving benchmarks while reducing repeated inference compared to rejection sampling.
By Alessandro Previti, Konstantinos Vandikas, Mario Alviano, Mats Carlsson, Albin Larsson Forsberg, Alexandros Nikou, Athanasios Karapantelakis, Aneta Vulgarakis
BiHashFormer is a dual‑branch transformer that uses hashing to efficiently detect objects in high‑resolution wide (HRW) shots. It employs a RoI Selector to pick target‑containing windows, a compression branch that refines attention via HashAttention by discarding half the key‑value pairs, and a compensation branch that recovers missed objects with HashMiner. Experiments on the PANDA gigapixel benchmark show a 37.5% reduction in backbone FLOPs and an AP50 of 81.0%, while the parallel structure avoids inference delays and improves stability compared to single‑branch models.
By Jingchen Huang, Wenxi Li, Chenyang Lyu, Moran Liu, Haozhe Lin, Yuchen Guo
Grounding Image Composition with Label-Conditioned Visual Evidence introduces the KU-PCPE benchmark, which provides image-level composition labels and class-specific geometric elements but lacks explicit visual evidence for each label. The dataset contains 4,093 training masks and a 1,327-task expert test split, with annotations that include segment-versus-line ratings to indicate whether region or semantic-guideline evidence better explains a label. Performance metrics show a gap between expert agreement (.867 Dice) and model predictions (.664 for the strongest deployable model, .672 for a proposal oracle, and .395 for a VLM proposal selector).
By Armin Dadras, Tingyu Lin, Robert Sablatnig, Franziska Proksa, Markus Seidl
The paper investigates the faithfulness of visual information in multimodal chain-of-thought (MCoT) reasoning by large vision-language models (LVLMs). It shows that visual cues in MCoT are often inaccurate and largely ignored, as predictions change little when visual input is altered but significantly when textual input is changed. To improve visual faithfulness, the authors propose an annotation‑free Sufficient‑Component Cause Model (SCCM) that encourages the generation of minimal yet sufficient visual components, and demonstrate its effectiveness across fine‑grained perception and reasoning benchmarks.
By Zujing Liu, Junwen Pan, Qi She, Yuan Gao, Guisong Xia
Sparkle introduces a new dataset of approximately 140,000 video pairs for instruction‑guided background replacement, covering five common themes. The authors present a scalable pipeline that decouples foreground and background guidance with strict quality filtering, addressing the lack of precise background cues in existing data. Experiments show that models trained on Sparkle outperform all current baselines on both the OpenVE-Bench and the newly released Sparkle-Bench.
By Ziyun Zeng, Yiqi Lin, Guoqiang Liang, Mike Zheng Shou
The study evaluates quantum kernel methods for classification against tuned classical kernels, addressing inconsistencies in prior benchmarks by implementing protocol controls such as balanced bandwidth tuning, fold‑specific kernel recomputation, and hardware‑based cost measurement. A factorial design across 324 experimental cells—varying dimensionality, training size, class separability, and data structure—reveals that optimizing the quantum feature map’s bandwidth yields the largest accuracy gains, surpassing any quantum‑classical performance differences observed. Consequently, tuned quantum kernels are statistically indistinguishable from tuned RBF kernels on four real datasets, and the apparent advantages on synthetic data diminish with increasing dimensionality and are largely attributable to protocol variance rather than inherent quantum benefits.
By Ronak Bhalgami
The paper critiques current LLM safety evaluations that rely solely on severity taxonomies, arguing they ignore the probability that a threat will actually be executed. It introduces Expected Harm, a metric that multiplies severity by execution likelihood—modeled via execution cost—to better gauge real-world risk. Empirical studies show that models over‑refuse low‑likelihood, high‑cost threats while remaining vulnerable to high‑likelihood, low‑cost ones, a miscalibration traced to the distribution of safety examples in supervised fine‑tuning. This misalignment can be exploited to double the success rate of jailbreak attacks.
By Yen-Shan Chen, Zhi Rui Tam, Cheng-Kuang Wu, Yun-Nung Chen
FrontierMat is a new benchmark comprising 330 real-world problems from cutting-edge materials research in areas such as aerospace, semiconductors, embodied intelligence, and nuclear fusion. Each problem is validated by experts and requires a language model to infer a withheld outcome or draw a conclusion from experimental evidence, with 130 problems demanding cross‑scale reasoning. The strongest general‑purpose model scores only 53.7%, and domain‑adaptation techniques can raise performance to 55.4%, matching the best proprietary models while improving on public materials benchmarks.
By Feng Chen, Jianguo Huang, Yushan Xiao, Yue Li, Bijun Tang, Xuyu Dong, Xingguo Wang, Feiyang Xie, Jianhui Gu, Dongfeng Cao, Hongda Chen, Yongji Gong, Ruiqi Li, Shuxin Li, Xingji Li, Ruitao Lv, Li Song, Lei Wang, Ye Zhang, Xiaoxu Zhao, Xitao Zheng, Bo An, Zheng Liu
The paper investigates model merging as a scalable alternative to multi-task training for multi-domain automatic speech recognition (ASR). It benchmarks 11 merging algorithms across 10 European Portuguese domains, assessing in‑domain accuracy, robustness to distribution shift, and performance on English and multilingual data. The authors introduce BoostedTSV‑M, a new merging method that mitigates rank collapse through singular‑value boosting, achieving better performance than full fine‑tuning while maintaining out‑of‑distribution generalisation in a single model.
By Carlos Carvalho, Francisco Teixeira, Thomas Rolland, Alberto Abad
The paper introduces DUGO, a Dual-Scale Gaussian Memory framework for Online Temporal Action Localization in egocentric procedural videos. DUGO combines a Dual-Scale Temporal Encoder—pairing depth‑wise local representations with global self‑attention—and a Gaussian Memory Bank that stores action memory as stochastic, input‑conditioned Gaussian prototypes. Experiments on EGTEA, EPIC‑Kitchens‑100, and Muses show that DUGO achieves state‑of‑the‑art performance while improving on non‑egocentric benchmarks.
By Mohammed Aman Bhuiyan, Md. Sazzad Hossain Adib, Samiul Basir Bhuiyan, Mir Faiyaz Hossain, Syed Mohammed Shamsul Islam, Md Zakir Hossain, Shafin Rahman
ReVEAL is a new framework for reasoning video object segmentation that decouples multi‑modal large language model (MLLM) reasoning from downstream segmentation. It introduces Visual Evidence Tokens (VETs) to link visual regions with target, background, or distractor roles, and uses a Dynamic Attention Layer Router to aggregate role‑specific grounding evidence. Trained on the ReVEAL‑47K dataset with a Contrastive Attention Grounding Loss, ReVEAL guides a frozen SAM2 model to predict target masks and achieves state‑of‑the‑art performance on five benchmarks without requiring costly cross‑model alignment.
By Zhongyuan Bao, Cunjie He, Lejun Zhang, Zenglin Xu, Yanwei Fu
TinyAstral introduces a two‑stage method to transfer improvements from a large teacher robot agent to a compact student agent. First, the teacher’s harness—its tools, feedback, memory, and recovery logic—is recursively self‑improved by LLM agents, then the enhanced harness is frozen and used to distill a small 4B student that matches the teacher’s performance. The distilled student achieves near‑teacher success rates, outperforms larger open‑source agents, and runs with roughly one‑fifth the computational cost.
By Kehong Gong, Xiaoxu Zheng, Guanli Hou, Seunghun Lee, Mengqi Guo, Runjia Tan, Zhengyu Wen, Qi Wang, Xiaoyu He, Mingyuan Zhang, Michael Bi Mi
The paper introduces h1, a simple agentic loop that interfaces a language model with Lean 4, avoiding complex human-designed workflows. h1 achieves state‑of‑the‑art performance on benchmarks such as PutnamBench, FATE‑X, and FormalConjectures100‑Solved, solving eight open problems and matching or surpassing specialized systems like Goedel‑Architect and AlphaProof Nexus. Experiments show that an efficient Lean interface is crucial for cost and success, suggesting that future evaluations should prioritize execution feedback over elaborate workflows.
By Skander Moalla, Ahmad Rammal, Niket Patel, Remi Munos, Julia Kempe
arXiv:2610.41169v1 Announce Type: new
Abstract: The ability to integrate new information and deduce its consequences is central to learning, understanding, and reasoning in large language models (LLM...
By Alexander Detkov, Dashiel Karish, Matthew Oleynikov, Matt Thomson
RECAST is a reconstruction-based framework that generates chain-of-thought (CoT) explanations from closed‑source reasoning models that only expose final answers, such as Claude‑Sonnet‑5.5, GPT‑6.1, and Gemini 3. Unlike prior extraction methods that rely on visible reasoning fields or scratchpads, RECAST conditions a decoder on the question and the model’s final answer to produce a first‑person structured rationale, cleans it, and outputs data ready for supervised fine‑tuning. The system is open‑source, packaged as an agent skill, and demonstrates successful qualitative reconstructions on the mentioned models using an answer‑only protocol.
By Qi Wang, Kang Xu, Chengcheng Wan, Jiangtao Wang
ROTrans is a benchmark that evaluates large language models (LLMs) on the task of migrating code from ROS1 to ROS2. It contains 108 scaffolded completion tasks derived from over 70 open‑source ROS1 repositories, organized into a hierarchical taxonomy that tests interface‑level, behavioral‑level, and system‑level correctness. Each task comes with a verified ROS2 reference implementation, a runtime test suite that exercises topics, services, actions, and parameters, and a container for reproducibility, allowing candidates to be scored by inserting only their target files into a fixed reference package.
"whyItMatters":"The study reveals that current frontier LLMs achieve less than 10.9% pass@1 on these tasks, highlighting significant gaps in LLM performance for complex robotics software development and underscoring the need for execution‑guided repair techniques to improve migration success."
By Yichen Zhao, Minghao Shao, Zhuoyan Yu, Yijun Shen, Boyuan Chen, Yik-Cheung Tam, Muhammad Shafique
SERIAL-7 is a Korean‑English benchmark that evaluates large language models on their ability to generate machine‑verifiable structured outputs across seven representations. It introduces an executable verification framework that separates semantic derivation (Derive), constrained artifact generation (Realize), and cross‑representation preservation (Preserve), using 177 validated artifacts and 43 atomic representation constraints. The benchmark shows that reliable structured generation requires explicit, executable verification of how artifacts are realized and preserved, as evidenced by varying compliance rates across ten models.
By Youngkyung Seo, MiHyeon Kim, Jisoo Lee, Hwijung Ryu