The paper introduces Tool Primitives, a design that replaces rigid API schemas with natural language interfaces for tool calling, enabling seamless inter-tool communication. It builds ToolFace, a repository of over 25,000 functions that LLMs can dynamically retrieve, and HEART, a harness engineering framework that orchestrates tool use with planning, routing, and verification. Experiments show HEART outperforms fine‑tuned models and leading commercial LLMs while cutting API costs by up to 85%.
By Haibo Jin, Suijin Wang, Xucheng Yu, Haojing Luo, Haohan Wang
The paper introduces OBJECTION, an inference-time pipeline that adds an Adversarial Lawyer Agent to each of the three reasoning steps—offense, unlawfulness, and culpability—in legal judgment prediction models. By actively injecting defense arguments, the agent challenges the model’s default assumption of guilt, which is common in datasets biased toward guilty outcomes. Using a new Natural Innocent dataset of 3.4k real cases, OBJECTION reduces the False Guilty Rate from 82.93% to 16.69%, demonstrating significant improvement in substantive legal reasoning.
By Jaehoon Jeong, Jay-Yoon Lee
LookStep is a new end‑to‑end framework for Vision‑Language Navigation that integrates Language‑Centric Future State Modeling with an Event‑Driven Rolling Memory. It uses language labels to predict coarse navigation progress and future states for candidate actions, and autonomously decides which observations to store in a bounded memory with semantic roles. Empirical results show that LookStep outperforms existing methods on VLN‑CE tasks, achieving a 49.7% success rate on R2R‑CE Val‑Unseen while improving memory efficiency and reducing data requirements.
By Kun-Yang Yu, Yingzhe Li, Hongyu Xu, Shi-Yu Tian, Zhi Zhou, Yang Chen, Ming Yang, Sheng Wang, Qing Yu, Lan-Zhe Guo, Yu-Feng Li
The paper introduces TUX, a Tacit Understanding Index that measures how similarly humans and large language models (LLMs) place concepts along subjective spectra in a task inspired by the game Wavelength. Using 241 human participants and 200 profile-conditioned LLM agents across four models, the study finds that human–agent pairs with similar traits achieve higher TUX scores, indicating that tacit alignment is linked to person-level characteristics. Regression analyses show that richer predictor sets—including individual traits, decision-making styles, and confidence—improve the explainability of TUX beyond simple trait-distance baselines.
By Yueshen Li, Hanyi Min, Vedant Das Swain, Koustuv Saha
The paper introduces SPACE, a method for enabling large language model agents to emit variable-length action chunks in long-horizon tasks. By distilling chunk-boundary supervision from programmatic skills derived from successful trajectories, SPACE overcomes the tendency of agents to either act one step at a time or commit to overly long sequences. Experiments on ALFWorld and ScienceWorld demonstrate that SPACE raises success rates by 7.0%–31.3% and cuts LLM decision rounds by up to 78.9%.
By Yanting Yang, Can Jin, Jinman Zhao, Jiahao Wu, Yang Zhou, Zhepeng Wang, Zhendong Wang, Mu Zhou, Dimitris N. Metaxas
The paper introduces Document-Mediated Reinforcement Learning (DMRL), a framework that treats skill document optimization in advertising recommendation as a sequence of structured editing actions. An upper-level agent edits the skill documents while a frozen lower-level task agent evaluates the edits via A/B testing. DMRL incorporates Dual-Relative Policy Optimization for robust advantage estimation and a Long-term Reward Predictor that models population heterogeneity to estimate long-term outcomes, achieving superior performance on a large-scale short‑video ads platform.
By Wei Zhang, Hongji Li, Song Sun, Peng Yu, Xue Yang, Lei Zhao, Peng Jiang
The study couples the Met Office Unified Model with distributed reinforcement learning agents, using a DDPG actor that applies bounded potential‑temperature corrections across 70 vertical levels. Training is performed on ten nudged forecasts, after which the frozen policy is evaluated in a non‑nudged forecast, demonstrating numerical stability. The learned policy reduces Z₅₀₀ MAE in four of six latitude bands—up to 45.8% in the northern tropics—and decreases MSLP error by up to 27.3% in certain bands, indicating promising bias‑correction potential.
By Pritthijit Nath, Sebastian Schemm, Peter Haynes, Emily Shuckburgh, Mark Webb
NS-Copilot is a large‑language‑model driven multi‑agent system designed to automate neuroscience data analysis. It integrates domain‑specific pre‑trained models for modalities such as EEG and extracellular spike data, and uses a natural‑language interface to orchestrate agents that plan, generate code, and synthesize results. In benchmarks on Alzheimer’s, Parkinson’s, and working‑memory spike decoding, the system consistently outperformed strong baselines across multiple trials.
By Wuche Liu, Yiran Qiao, Linlin Hou, Rui Yang, Shusen Pu, Song Wang, Jing Ma
PRISM (Proactive Risk Intelligence and Safety Management) is an agentic multi-model architecture designed to shift autonomous transportation safety from reactive crash avoidance to proactive, continuous risk management. It uses inverse crash‑probability modeling to transform binary crash classifiers into dynamic safety scores, and runs three specialized models—trajectory kinematics, environmental risk, and VRU interaction—coordinated by a reinforcement‑learning reasoning layer. Across 1,296 naturalistic driving scenarios, PRISM achieved a mean safety score of 68/100, classified 77.6% of situations as advisory, and flagged 3.8% as near‑misses, with 11% requiring intervention or emergency response, highlighting trajectory risk and VRU proximity as key safety factors.
By Joyjit Roy, Samaresh Kumar Singh, Sushanta Das
ExecRetrieval is a new benchmark for code‑embedding retrieval that contains 939 Python tasks, each with a verified correct implementation and up to four single‑edit buggy distractors generated mechanically. The dataset allows direct testing of a retriever’s ability to functionally discriminate correct code from near‑clone incorrect code, rather than relying on lexical similarity. Experiments on 23 dense embeddings and BM25 show that while the best system can retrieve the correct code within the top 10 results, it often fails to rank the correct implementation first, with rank‑1 errors dominated by paired buggy variants.
By Aaryan Kapoor, Md Abdullah Al Hafiz Khan
GEM is a Generative LiDAR world model that uses a deformable Mamba architecture to better handle the disorder of LiDAR point clouds and distinguish dynamic objects from static structures. The model tokenizes LiDAR sweeps, unsupervisedly disentangles dynamic and static features, and applies a tri‑path deformable Mamba for selective scanning and adaptive gating fusion, improving spatial‑temporal understanding. Experiments show GEM outperforms existing methods across multiple benchmarks, and it can be paired with a planner and BEV controller for autonomous rollout and "what‑if" scenario generation.
By Yang Wu, Zhaojiang Liu, Qiang Meng, Youquan Liu, Renliang Weng, Jianjun Qian, Jian Yang, Jin Xie
ShallowStream is a framework for streaming video understanding that uses the shallow layers of a multimodal large language model (MLLM) to encode frames and build a lightweight index. During streaming, it maintains an always‑on index via the KV cache of shallow layers, and at query time it scores context frames using shallow‑layer attention and selects diverse evidence for answering. The approach matches the performance of leading streaming methods while cutting per‑frame prefill latency and 10‑second end‑to‑end latency by up to 52.1× and 11.9×, respectively.
By Jitai Hao, Ke Yang, Qiang Huang, Jun Yu
The paper investigates when information sharing enhances decentralized discovery by separating its effects on pooled estimation and independent rescue actions in finite discovery models. It shows that a registered incremental-sharing protocol improves discovery only when pooled residual error decreases faster than an independent rescue attempt, and that equilibrium selection can determine whether sharing is beneficial. The study uses synthetic, finite models without human or organizational data.
By Yohei Nakajima
The paper investigates how persistent memory in AI agents can lead to over‑trust in stale facts, creating a "Memory Trust Gap" that worsens as model capability increases. Using a benchmark with Benefit and Safety suites across Qwen3 models of varying sizes, the authors show that larger models are more prone to harmful over‑trust, especially when metadata is absent or misleading. They also demonstrate that mitigation strategies such as exposing metadata or pre‑resolving conflicts improve accuracy, but the effectiveness depends on model size and dataset.
By Jundong Hu, Shekar Ramachandran
The paper introduces Belief-Calibrated Optimization (BCO), a method that records and updates a persistent in‑context document representing an agent’s belief about how the environment responds to edits. By continually revising this world model as new candidates are evaluated, BCO improves the performance of frozen LLM agents across five benchmarks, outperforming a control lacking the world model. An offline ablation shows that the document’s content provides reusable, accurate predictions of environmental responses, beyond mere form.
By Yuhan Chen, Zhihua Tian, Mahavir Dabas, Charith Peris, Rahul Gupta, Ming Jin, Feiyang Kang, Siyuan Zhang, Nan Wang, Ruoxi Jia
The paper reports a case study of a large language model (LLM) coding agent tasked with building a multi‑component data system from a detailed specification. During a single session the agent introduced five defects, which were categorized by violated constraints and detection methods. The study also evaluates the agent’s retrieval‑filtering strategy on the HotpotQA benchmark, showing that filtering to a graph‑identified entity set yields higher recall than unfiltered search, with a statistically significant gap across all tested budgets.
By Phanindra Reddy Madduru
Zeta‑Lite is a WebAssembly‑based in‑browser SQL database that brings concurrent, snapshot‑isolated transactions and copy‑on‑write database branching to the client side. It is a compact 2.87 MB gzipped build of the Zeta engine, offering a full PostgreSQL‑compatible feature set—including joins, CTEs, window functions, JSONB, full‑text search, HNSW vector search, and graph queries—while maintaining high read/write throughput across major browsers. The engine’s log‑centric asynchronous MVCC core enables overlapping transactions on a single thread and unique branching capabilities rarely seen even in server‑side databases.
whyItMatters":"Zeta‑Lite’s concurrent, branchable design provides a lightweight, privacy‑preserving, and offline‑ready memory layer for in‑browser AI agents, enabling them to explore, test, and commit speculative changes efficiently."
By Gene Zhang
READY or Not: Reliable Enterprise Agent Deployment introduces a framework for qualifying AI agents for enterprise workflows. It measures reliability and operating cost under various oversight policies, selects the minimum‑cost policy that meets a specified reliability target, and statistically qualifies it on held‑out cases. In a clinical audit study, READY revealed that two agents with nearly identical autonomous accuracy required markedly different levels of human review to achieve the same reliability target.
By Veronica Chatrath (Christy), Bryan Zhu (Christy), Jingxuan Fan (Christy), George Pu (Christy), Soham Dinesh Tiwari (Christy), Soham Dan (Christy), Ryan Young (Christy), Yuan (Christy), Li, Yuang Yao, Apaar Shanker, Minglai Yang, Daniel Yue Zhang, Yunzhong He, Ying Liu, Chenguang Wang, Zhijun Yin, Yuan Xue
SkillGLoW introduces a new way for large language model agents to self‑improve by consolidating procedural skills shared across related tasks. Instead of storing all skills in a single global document or a flat per‑task pool, SkillGLoW aggregates local skills into procedural families, compresses them into de‑instantiated global priors, and regenerates instance‑specific details on demand. Experiments on four diverse benchmarks show that these priors improve performance by an average of 17.2 points over a no‑skill baseline, are more compact than per‑task pools, and enable better transfer to unseen tasks.
By Ao Yan, Xin Zhang, Jiawei Du, Joey Tianyi Zhou
PhoenixNest-Video is an evidence‑grounded multimodal agent designed for automated video interview assessment. It constructs a semantic video graph as working memory, retrieves information conditioned on rubrics across visual, audio, and textual streams, and outputs per‑criterion scores tied to the candidate’s materials. Trained with rubric‑based reinforcement learning, the system achieves 91.50% grade‑level accuracy on VInterview‑2025, outperforming larger proprietary models while providing traceable evidence for each score.
By Fan Yuxuan, Huang Miaojun, Zhang Haimei, Wu Jingshen, Liu Hao