The paper introduces a probe‑free method called the Neuron Separability Index (NSI) to assess how individual neurons in Large Language Models (LLMs) distinguish grammatical from ungrammatical constructions. Using linguistic minimal pairs across 68 paradigms and seven checkpoints, the study finds that raw separability peaks earlier for morphological and syntactic distinctions, but after permutation normalization, single‑unit selectivity is sparse, weak, and narrowly tuned, with rare strongly selective "grandmother neurons". Additionally, whole‑vector linear separability, single‑neuron selectivity, and behavioral competence are largely dissociated, and targeted ablations further separate activation selectivity from causal reliance.
By Linyang He, Nima Mesgarani
The paper investigates where exactly‑once semantics should be enforced for tool‑using agents—within the model, the agent harness, or the tool contract—by evaluating 25,930 episodes across nine models, three harnesses, two contract variants, and fifteen recovery conditions. Using the LIMBO sandbox, the study shows that when an immediate read‑back is available, frontier models rarely duplicate lost acknowledgements, whereas weaker models do; when read‑back is unavailable, the contract’s idempotency keys explain most duplicate behavior. The authors prove that verification‑only policies cannot guarantee exactly‑once under late commits without bounded in‑flight time, and that waiting only helps when delays are short and predictable.
whyItMatters":"The findings clarify that enforcing exactly‑once semantics largely depends on the tool contract and fault type, guiding designers on where to focus reliability mechanisms for LLM agents."
By Jiapeng Li
The paper explores using large language models (LLMs) as AI respondents to convert policy documents into structured survey responses. It introduces a long-context in‑context learning pipeline that maps policy text to predefined survey categories such as policy instruments, target groups, and thematic areas, and includes a secondary LLM validation step. Evaluation on a multi‑country dataset shows high agreement (84‑95%) with human responses for structured indicators, though free‑text fields differ, indicating that hybrid human‑AI workflows can enhance policy monitoring efficiency while still requiring human oversight.
By Carolyn Cole, Matthias Deschryvere, Toqeer Ehsan, Arash Hajikhani
IterSynth introduces a role-decoupled, iterative synthesis framework for deep search agents, separating planning and synthesis into distinct Planner and Synthesizer modules that maintain a persistent summary state. This design mitigates role coupling and context noise, while the new Role-Decoupled Policy Optimization (RDPO) enhances training by combining outcome rewards with turn-level rubric evaluations. Experiments on five long-horizon benchmarks show IterSynth-8B outperforming prior ≤8B agents by 4.2% and delivering significant zero-shot gains over ReAct on proprietary models.
By Xingyu Wu, Yuchen Yan, Zhengxi Lu, Siqi Chen, Xin ZHANG, Aiting Liu, Chao Deng, Jie Liu, Jin Ma, Jian Shao, Jun Xiao, Yongliang Shen
The paper investigates whether large language model–based coding agents can automatically synthesize programs that solve generalized task and motion planning (TAMP) problems across diverse instances. Using Claude Code and Codex, the authors evaluate 980 generated programs on 100 held‑out environments from KinDER and PDDLStream, achieving mean success rates between 56 % and 95 %—higher than hand‑engineered planners and other baselines—while requiring an order of magnitude less computation per instance. The study demonstrates that coding agents can calibrate physical models, test edge cases, and refine strategies, suggesting they are a strong baseline for generalized TAMP.
By Matteo Merler, Bowen Li, Josh Roy, Yichao Liang, Qianwei Wang, Yixuan Huang, Tom Silver
KernelOPT is a multi‑agent system that optimizes GPU kernels generated by compilers like PyTorch Inductor by treating compiled models as structured artifacts. It preserves vendor library calls and focuses on Triton sub‑kernels, using five profiling‑guided LLM agents and a four‑gate verification cascade to ensure correctness and performance before re‑stitching the model. On 250 KernelBench problems, KernelOPT achieves geometric mean speedups of 1.40×, 1.15×, and 1.07× over torch.compile at three optimization levels.
By Aheli Poddar, Sanskar Prasad, Arindam Samanta, Subha Chakraborty, Vishal Goyal, Rohit Singh Rathaur
SheetMind is a Manager‑Action‑Reflection framework that evaluates how much spreadsheet agent performance derives from the agents themselves versus the shared action interface. In a controlled study on all 221 tasks of the SheetCopilot Benchmark, replacing the high‑level action API with primitive cell operations drops accuracy by 47.1 points, while adding a Reflection Agent improves performance by 4.5 points and a Manager by 1.4 points. The framework also shows that decomposition changes failure modes, reducing silent wrong outputs from 33% to 25%, and that GPT‑5 and GPT‑5‑mini achieve similar performance, whereas GPT‑3.5 underperforms significantly.
By Lyuhao Chen, Xi Cheng, Yanming Kang, Ruiyan Zhu, Ke Liu, Rakesh Chowdary Machineni, Yulang Fei, Brian Zhu, Daniel Jin, Binze Cai, Zheng Qi, Neeraj Parihar, Zhoutian Xu, Oliver Gao
SPARQL-LLM is an open‑source, triplestore‑agnostic system that generates SPARQL queries from natural language using lightweight metadata and dedicated components for indexing, prompt building, and execution. It achieves up to 59 % higher F1 scores than the next best system on a multilingual challenge and on bioinformatics knowledge graphs, while being up to 27 × faster and costing no more than $0.01 per question. The project is publicly available on GitHub and is already deployed on real‑world decentralized knowledge graphs such as expasy.org/chat.
By Panayiotis Smeros, Vincent Emonet, Ruijie Wang, Ana-Claudia Sima, Tarcisio Mendes de Farias
The paper proposes that two architectural assumptions—(1) attention and MLPs share a key‑value form <phi(S)>U, and (2) components read from an additive residual stream—are sufficient to answer three interpretability questions: component interaction, information routing, and token attribution. By treating these selections as a computational graph, the authors develop Unpack, a backward attribution method that validates interaction scores, recovered routes, and token attribution against established tests across models ranging from 160M to 6.9B parameters. The study also shows that contribution and causal effect can differ, with a recognizable signature in how components change when a task is removed.
By Po-Kai Chen, Aske Plaat, Niki van Stein
The paper introduces a neuro‑symbolic architecture that combines symbolic planning, reinforcement learning, and a large language model (LLM) to address novelties in dynamic open‑world environments. The LLM is used to identify missing operators, generate symbolic plans, and write reward functions, enabling the reinforcement learning agent to learn control policies for newly identified operators. The proposed method outperforms state‑of‑the‑art approaches in both operator discovery and learning within continuous robotic domains.
By Hong Lu, Pierrick Lorang, Timothy R. Duggan, Jivko Sinapov, Matthias Scheutz
Task-Aware Spectral Pruning (TASP) is a post‑training framework that tailors sparse masks to specific tasks by calibrating module‑level spectral descriptors against task‑specific ablation effects. It constructs masks that close grouped‑query‑attention and SwiGLU dependencies, routing each user turn to a single compiled mask that remains fixed during prefill and decoding. In experiments, TASP achieves a 43% active‑FLOP reduction while preserving 97.7% of the dense BF16 performance on Llama‑3‑70B, and delivers a 1.44× speedup on an A100 80GB with INT8‑weight/BF16‑compute, reducing decode latency from 45.2 to 31.3 ms/token.
By Ibne Farabi Shihab, Fariya Afrin, Sanjeda Akter, Anuj Sharma
TraceGuard is an adaptive, rank‑based filtering method designed to detect and remove poisoned image‑text pairs from multimodal training corpora. It evaluates six corpus‑level features—such as cross‑modal neighborhoods, recurring text, and changes after text‑span erasure—to identify suspicious examples without training the victim model. In experiments across 19 attack scenarios, TraceGuard successfully removes an average of 98.4% of poisoned data while discarding only 5.4% of clean data, reducing residual attack impact to at most 1% in most configurations.
By Haoyang Li, Yaxin Xiao, Linyan Dai, Jiawen Fu, Zi Liang, Jason Xue, Qingqing Ye, Haibo Hu
The paper introduces Stochastic Semantic Evidence Graphs (SSEGs), a hierarchical stochastic directed acyclic graph that models uncertainty in AI-agent workflows, from evidence and retrieval to generation and decision mapping. SSEGs expand language nodes into autoregressive token subgraphs, optionally apply semantic reduction and calibration, and preserve uncertain claim–passage relations while propagating Fréchet bounds. The authors derive pathwise error bounds, use nodewise terms to trigger governance checks, and demonstrate through experiments that SSEGs can detect and quantify where uncertainty enters and propagates in AI outputs.
By Matthew Francis Dixon
The paper evaluates two large language models, Claude Sonnet 4.5 and Claude Opus 5, on the bidirectional English Resource Grammar (ERG) tasks of generating English from Minimal Recursion Semantics (MRS) and parsing English into MRS. In generation, Opus achieves 76.3 BLEU—surpassing a 72k‑pair trained system and matching a million‑pair system—while Sonnet scores 65.7 BLEU, rising to 69.6 when selecting from ACE’s candidates. In parsing, both models lag behind ACE, attaining only 57.2 and 65.5 F₁ respectively, with exact‑match on about 1 % of sentences, highlighting that high generation scores do not guarantee accurate semantic parsing.
By Soham Dan
ARGUS is a pipeline that builds document-level Event Knowledge Graphs (EKGs) from U.S. employment‑discrimination complaints. It uses a 5W1H-inspired schema, legal-domain models, and LLM‑based structured generation to extract fact‑bearing statements, create chunk‑level event graphs with participant, temporal, and causal structure, and merge them into comprehensive document representations. Evaluation shows that graph‑structured classifiers outperform raw and linearized baselines on claim classification, and EKG‑only retrieval improves document‑scoped QA, though overall open‑retrieval gains are limited by low first‑stage candidate recall.
By Sriram Kannan, Swetha Saseendran, Vishnu Vardhan Reddy Kandi, Leslie Barrett, Madhavan Seshadri, Enrico Santus
The paper introduces Privileged Self-Practice (PSP), a method that retains privileged information (PI) in the prompt rather than the loss during on‑policy self‑distillation for multi‑turn agents. PSP injects short per‑task instructions from an analyzer model when rollouts fail, sampling again with the instruction in context and training with the unchanged GRPO objective. Experiments on AppWorld and SWE‑bench Verified show PSP consistently outperforms plain GRPO, boosting task‑goal completion by up to 65% and resolved rate by up to 61% across three student models.
By Xingyu Su, Abhishek Kumar, Qing Ping, Youzhi Luo, Jonathan Buck, Zach Zhang, Subramanian Chidambaram, Vinayak Arannil
Cross-Modal Emotion Understanding: A Transformer-GAT Approach for Dialogue Emotion Recognition proposes a hybrid framework that combines a Transformer and a Graph Attention Network to capture both global semantic information and fine-grained relationships between modalities. The model is evaluated on the IEMOCAP and MELD datasets, achieving weighted F1 scores of 72.45% and 77.37%, respectively, and surpasses state‑of‑the‑art methods. These results suggest that integrating multimodal features with balanced global and local context modeling can provide deeper emotional insights for dialogue emotion recognition.
By Jiaqi Qiao, Yifan Lyu, Xiujuan Xu
HarnessPAI is a model‑ and embodiment‑agnostic framework that treats code as an executable, evolvable interface for Physical AI. It separates short‑term open‑loop program execution from long‑term closed‑loop evolution, using feedback to refine programs and distill reusable skills. Across various robots, HarnessPAI outperforms pure action models and code‑as‑policy baselines, achieving significant gains on tasks like LIBERO‑PRO and RoboCasa without retraining the underlying model.
By Xin Wang, Wenhao Wu, Menghao Zhang, Zhi Wang, Kun Shao, Jian Luan, Yang Li, Qing Li, Shangding Gu, Huichi Zhou, Shuqing Shi, Fei Ni, Shuo Lu, Weicheng Meng, Kang Li, Jin Wu, Kang Zhao, Shangmin Guo, Gen Li, Yongqiang Tang, Zhizhong Zhang, Yuan Xie, Heng Qu
The paper reports on a large‑scale verified search experiment using a 30B language model on a laptop, evaluating three operator packages—schematic notebooks, named obstacles, and behavioural repulsion—in a factorial design across nine construction problems. Results show that the full composition of operators closes the seed‑to‑record gap more effectively than any single component, increases construction‑hash diversity, and that memory plus repulsion consistently avoids collapse. A frontier proposer achieves similar gains in far fewer samples, but the search ultimately stalls near a plateau where the reference family is adopted and optimized only when provided as code.
By Roberto I. Ono Filho
A useful visual representation is a statistic of the observed past that retains the latent factors shared with the future and discards patch-private noise. We formalize this requirement as a condition...