The paper introduces the Post‑Audit Mirage, a framework showing that offline tests may not detect whether an agent update remains safe after deployment. It defines a benchmark that uses matched safe and harmful lifecycles to expose identical verifier‑visible evidence while an isolated scorer records opposite deployment truth, and presents a theorem on the error‑abstention tradeoff using observational equivalence. The IdentifiedRangeMonitor tool applies fresh group‑labelled harm outcomes to decide whether to deploy, hold, or remain uncertain, and experiments demonstrate that monitoring can recover missing distinctions in favorable settings but may stay undecided for rare groups or small effects.
The paper presents a proof‑of‑concept workflow that uses an AI agent to translate natural‑language prompts into early‑stage structural models. By connecting the agent to a COMPAS‑based structural modelling API via a Model Context Protocol server, the system generates and edits a graph‑based structural data model with explicit building elements. The resulting artifacts can be inspected, modified, serialized, and integrated into downstream AEC workflows, demonstrating the feasibility of combining natural language, agentic workflows, and domain‑specific structural data structures for AI‑assisted structural co‑design.
The paper introduces Active Preference Acquisition (APA) for long‑term personalized web agents, defining the Prospective Personalization Value (PPV) of a missing preference as its expected cumulative utility over future tasks minus the cost of asking. APA selects questions within a per‑interaction budget, prioritizing preferences that may not benefit the current task but can be reused across many future tasks. Experiments on the ActiveWebPref benchmark show APA acquiring 80.3% of amortization‑dependent preferences and achieving higher net utility than baseline methods, while limiting the number of questions asked.
ReflectPO introduces a reinforcement learning framework that enables large language model agents to recover from errors during long-horizon tasks involving external tools. The method consists of Reflection-Guided Sampling, which uses branch differences and repeated visits in interaction trajectories to generate recovery hints, and Tree-Structured Credit Assignment, which organizes trajectories into trees and assigns action-level credit based on parent–child reward differences and recovery gains. Experiments on two long-horizon tool-use benchmarks demonstrate that ReflectPO outperforms existing baselines and improves agents’ ability to detect deviations, recover from mistakes, and continue progressing toward task completion.
The paper introduces two lightweight inference harnesses, OPTICYCLE and OPTISTATE, designed to enable large language models to translate natural-language optimization problems into solver programs without additional training. Evaluated across nine benchmarks and 11 Qwen variants (0.6B–27B parameters), the harnesses consistently improve macro accuracy, with execution-guided revision boosting performance by up to 49.46 percentage points. Despite their simplicity, both harnesses outperform six adapted agentic workflows and four post-trained optimization models of comparable size, though the benefit of explicit formulation state varies by model.
Lean Finder V2 is a semantic search framework for Mathlib that tackles the problem of short, ambiguous user queries by using adversarial training and query rewriting to disambiguate them. It achieves a new state‑of‑the‑art performance on the MathlibQR benchmark, improving retrieval by at least 11% over previous engines, and further boosts proof success when integrated with Claude Code.
Trace2World is a framework that reconstructs app worlds from real interaction traces, separating global behavior reconstruction from implementation. It reconciles traces into a shared, source‑linked app model and compiles an executable environment, allowing independent replay of source workflows to localize mismatches for bounded implementation repair. Using 1,570 traces, the system builds 50 environments and achieves 91.1% end‑to‑end reproduction on 200 workflows across ten apps, outperforming a strong baseline and improving benchmark success and external GUI benchmark scores.
PhyEvolve is a multi-agent framework that improves physics-aware video generation by enhancing natural-language prompts through structured state-evolution planning. It employs three agents—a Director, a Physical Evolver, and a Composer—to ground scenes, identify physical mechanisms, and construct causally linked state transitions that guide video generation. The method works with existing generators via prompt replacement and achieves state‑of‑the‑art performance, outperforming the prior agentic baseline by an average of 7.9 percentage points on four major benchmarks.
ULME introduces a post‑training framework that unifies text embedding and generation within a single causal decoder‑only LLM by inserting learnable slots for semantic and reasoning embeddings. It trains three objectives—contrastive, language‑modeling, and stop‑gradient predictive—so that generation actively shapes the embedding space. Across Qwen2.5 and Qwen3 models (0.5B–8B), ULME improves MTEB scores by 3.3–4.1 points, tightens BRIGHT and AdvBench‑IR, and matches generation‑only fine‑tunes on standard accuracy benchmarks.
The paper investigates whether scaling data, diversity, and model capacity benefits embodied agents in virtual worlds. Using 120 training setups across four model sizes, six data budgets, and five data mixtures, the authors find that more data improves in‑corpus performance, diversity aids transfer to related worlds, and larger models boost data efficiency but may generalize less to distant worlds. They also identify distinct training phases and fit a scaling law, concluding that continued pre‑training alone is insufficient for fully generalist agents.
By Kuan Zhang, Yukun Chen, Zhihao Yang, Yue Su, Xiangnan Wu, Zirong Chen, Run Luo, Jinkun Hou, Tao Tan, Yinhe Zheng, Yiming Li
FogTTA is an online test-time adaptation framework that enhances Transformer-based object detectors, specifically RF-DETR, for foggy weather conditions. It operates without source data or retraining by using a teacher-student architecture: the deployed model generates pseudo labels from weakly augmented inputs, which are refined and used to train a student model on strongly augmented samples with Varifocal loss. The teacher model is updated via exponential moving averaging, enabling real-time, stable adaptation and improved detection accuracy in foggy environments.
By Ali Abedi, Q. M. Jonathan Wu, Ning Zhang, Shiqi Tian
The paper introduces a new benchmark for personalized shopping that focuses on determining how far an autonomous agent should commit to purchasing an item. It defines a three-valued, evidence‑conditioned boundary—‘don’t buy’, ‘confirm first’, or ‘buy automatically’—based on product information sufficiency, reversibility, price norm, and urgency. Experiments with instruction‑tuned models show that these agents often overreach, committing even when fit is low or evidence is missing, and that a shallow learner trained on a person’s own labels can predict their auto‑purchase line more accurately than any prompt‑model combination.
By Ryotaro Shimizu, Sai HtaungKham, Mo Fan
The paper presents a method that derives BI‑RADS descriptors from breast MRI segmentation masks and uses a large language model to generate natural‑language rationales grounded in those descriptors. A BI‑RADS decision tree with eight quantitative radiomic measurements, lesion size, and component count informs the descriptors, which are then verbalised by the model. Applying this framework to 1,506 cases in the MAMA‑MIA dataset produced 3,012 reproducible BI‑RADS rationales with high faithfulness (99.98%) and counterfactual accuracy (99.3%).
By Hrishikesh Singh, Aarti Sharma, Divyanshu Kumar, Millie Pant
AgentPolicyForge is a framework that leverages executable environments to build policy benchmarks and iteratively train policy guards for task agents. It introduces PolicyGym, a dataset covering ten policy families with 698 training scenarios and 254 human‑reviewed test scenarios, where tool calls are executed and outcomes recorded to assess compliance. The framework supports an attacker‑generated request loop that fine‑tunes guards on balanced, environment‑verified examples, leading to a 40‑44% reduction in attack success rates compared to policy prompting alone.
By Chao Shang, Wanyu Du, Yi Zhang, Sandesh Swamy
TrainBridge is an agentic approach that transfers optimization guidance from GPU training to Ascend NPU training. It extracts optimization intents from highly optimized GPU execution and uses them to rewrite NPU programs, then refines these rewrites through iterative profiling. Experiments show that TrainBridge yields an average 1.7× speedup across four workloads and nearly 2.0× speedup on 27 Modded‑NanoGPT speedrun records.
By Yupeng Chen, Kunxiang Zhao, Xingjian Wang, Yue Zhao, Yijing Wu, Ye Liu, Kai Li, Congliang Chen, Ruoyu Sun
ASTEROID is a student–teacher supervised learning framework that learns efficient test‑time exploration in partially observable sequential decision problems. By training a history‑conditioned student on on‑policy contexts while a privileged expert provides action labels conditioned on that history, the student learns to search while uncertain and exploit once the hidden state is revealed. Across eleven diverse environments, ASTEROID achieves roughly twice the returns under fixed exploration budgets, uses about 100× fewer interactions than reinforcement learning baselines, and transfers well to vision‑denied real‑world robot manipulation, improving success by +40% over standard distillation.
By Sriyash Poddar, Yanda Bao, Jacob Krantz, Matthew Chang, Xavier Puig, Roozbeh Mottaghi, Natasha Jaques, Abhishek Gupta
The paper reviews and categorizes large language models that combine step‑by‑step reasoning with interactive search, treating retrieval as an action within an iterative agentic loop. It analyzes these systems along four architectural axes—retrieval coupling, controller design, knowledge source, and grounding strictness—while also considering the learning signal as a cross‑cutting factor. The authors conduct controlled ablations and benchmark analyses to clarify how design choices affect search, stopping, and grounding behavior, aiming to provide a unified perspective for future research.
By Biswas Poudel, Nilson Chapagain, Amit Kumar, Xianshun Jiang
FlashNext‑RL is a model‑specific reinforcement learning training framework designed for the Qwen3.8‑Flash‑Next language model. It optimizes BF16 training of the full 48‑layer model on a single 8×B200 node by freezing early layers, offloading host memory, and tuning kernels for mixture‑of‑experts, sparse attention, and gated residual streams. The optimizations raise non‑padding training throughput from ~395 tokens/s/GPU to 4,438 tokens/s/GPU and enable a live software‑engineering RL run with 1,024 rollouts and a 131,072‑token context, completing iterations in about 22.4 minutes.
By Zhewei Yao, Yuxiong He
SANE is an agent‑based framework that automatically generates scorer‑specific perturbation recipes and expected score relations for large‑language‑model judges, using behavioral requirements encoded as reusable agent instructions. It validates each recipe against the scorer’s intent and the instantiated test case, producing a growing inventory of aligned recipes as more scorers are added. In experiments across five domains, SANE yields 3.7× more validated test families than the Judge Reliability Harness, detects deficiencies in all domains—including professionalism and toxicity—and achieves strong human alignment in judge ranking on faithfulness (Spearman ρ = 0.852).
By Ethan Ludwick, Yu-Chung Hsiao
The paper introduces OACap, a caption-guided active observation system for detailed video captioning that requests targeted observations to fill missing information and iteratively refines its captions. OACap consistently improves performance on general video-understanding benchmarks while using only a text‑only reader. The authors also present FSCap‑Bench, a new evaluation suite with Core‑Support Factuality Score and Cross‑Modal Semantic Round‑Trip Score to assess factual and semantic preservation in captions.
By Hao Shi, Hong Yan, Xiaoyang Zhang, Qing Cui, Chao Huang