TRACER is a retrieval front‑end that indexes tool‑call sequences in a trie and uses a learned blend of empirical and contextual priors to rank tool continuations. In a mixed retail‑and‑telecom environment, it boosts GPT‑5.4’s tool‑selection accuracy by 4.7 points while cutting input tokens per step by 80%, exposing only 5 of 128 tools. Across six environments and four models, TRACER improves tool‑selection accuracy in 21 of 24 model‑environment pairs, with average gains of 6 to 10 points, and matches full‑registry success with 61% fewer tokens per step in autonomous execution.
By Tanish Kher, Sarvesh DK, Abedelkadir Asi, Sparsha Chennarayapatna Sundaresh, Dezhi Hong
The study examined how users delegate control to multimodal AI by analyzing 392 video question‑answering trials with 56 AI‑literate participants. Participants rated trust after each AI response and chose one of three delegation policies—reject, verify, or fully delegate. Findings showed that higher post‑interaction trust predicted delegation choices, verification was the most common policy, and delegation increased with task order, suggesting experience influences trust beyond self‑reported ratings.
By Agnese Chiatti, Matteo Matteucci, Viola Schiaffonati
The paper critiques current agent memory benchmarks that rely on downstream QA accuracy, arguing they conflate memory correctness with retrieval and generation quality, leading to misleading results. It introduces FEB-30, a new benchmark of 30 fact‑evolution trajectories across six diagnostic categories and three difficulty levels, generated automatically to provide deterministic ground truth. Using FEB-30, the authors show a significant divergence between intrinsic and extrinsic evaluations, revealing that a representative memory system correctly classifies memory operations only 20% of the time—an error invisible to traditional QA metrics.
By Sebastian-Antonio Toma, Jean de Kernier, Mike Ply, Oliviu Matei
The paper introduces Librarian Agents, a framework that trains large language models to proactively curate long-context workspaces by reorganizing raw documents into a hierarchical filesystem of synthetic files and reference links. This builder agent, optimized via reinforcement learning, aims to maximize downstream answering accuracy by anticipating future queries rather than reacting to them. Experiments show that Librarian Agents outperform leading retrieval and precomputation methods, achieving a 33‑point accuracy gain on BrowseComp‑Plus and a 31‑point improvement on BRIGHT in zero‑shot settings.
By Anikait Singh, Teresa Zhang, Yoonho Lee, Roshen Sanjay Nair, Aviral Kumar, Chelsea Finn
Spec-Driven Development (SDD) is a practice where developers and AI coding agents write specifications to guide software implementation, aiming to preserve the intent behind natural‑language instructions. This study analyzes SDD through Reddit discussions, interviews with 21 practitioners (including eight tool makers), and a large corpus of 470,795 spec files from 73,030 GitHub repositories. Findings reveal that while specs often include implementation details that aid agents, they also risk decay; tool-specific structures influence perception, and benefits vary between individuals and teams, with SDD mainly used for new features and involving higher planning and review costs.
By Shyam Agarwal, Anmol Singhal, Andrew Macvean, Bogdan Vasilescu, Travis Breaux
The study audits the predictive value of historical summaries in a rule-based university trajectory simulator. Seventeen ridge specifications and persistence models compare current plans, past-state and past-plan summaries, and calendar-only history-length indicators across training, validation, and two family-held-out evaluations, reconstructing 16,750 routes and 804,000 monthly records. Results show modest RMSE improvements with six-month history versus current-state prediction, but negligible differences from a length-only control, with uncertain generalization to all-history summaries and notable violations in fatigue predictions.
By Zepei Cao
Meta’s paper details how the company optimizes AI training on AMD GPUs by co‑designing hardware and models, using an analytical performance model for kernel selection, and extending agentic workflows for microarchitecture‑aware optimizations. The authors compare AMD Instinct MI350X to NVIDIA B200 on a recommendation model, achieving up to 1.5× higher training throughput. Their methods, the first hyperscaler report on productionizing AMD GPUs for large‑scale training, have been incorporated into open‑source PyTorch, Triton, and ROCm.
By Haoyu Zhang, Arshad Rahman, Adnan Aziz, Ajith Prabhudev, Amin Farmahini, Ben Rhoads, Benson Ma, Brian Shi, Cen Zhao, CK Luk, Dingming Wu, Emad El-Haraty, Eric Chen, Frank Li, Gary Huang, Gefei Zuo, Haoran Zhang, Hiwot Kassa, Hongyi Zeng, Jack Taylor, Jeremy Hadidjojo, John Donati, Jonathan Sheffield, Julio Machado Silva, Karthik Ganesan, Kent Russell, KR Kishore, Lei Tian, Liu Ke, Muhammad Osama, Peng Chen, Rakesh Hemnani, Ryan Swann, Saleel Kudchadker, Shankar Yanamandram, Shobhit Kanaujia, Shucai Xiao, Srinath Bayareddy, Srinivas Vaidyanathan, Sudharssun Subramanian, Supadchaya Puangpontip, Taylor Robie, Tianhao Huang, Tyler Hart, Viswesh Sankaran, Weijun Jiang, Wenkai Du, Xiaodong Wang, Xiaohu Guo, Yao Rong, Yinbin Ma, Yue Dong, Yuzhen Huang, Yvonne Liu, Zhan Shu, Dan Rabinovitsj, Ajit Mathews, Chunqiang Tang
DefoliationCAM (D-CAM) is a solar‑powered edge AI system that estimates grapevine canopy defoliation in real time. It uses YOLOX‑Nano to locate grapevine cordons in 4K images and a lightweight MobileNet classifier to categorize 320 × 320 pixel patches into four damage levels, then aggregates predictions into a Non‑linear Global Defoliation Severity (NGDS) score. The system processes each image in about 1.5 s on an NVIDIA Jetson Orin Nano, and field deployment over five weeks proved autonomous operation, LoRa communication, and remote monitoring are feasible.
By Fuchen Li, Wei-Zhen Liang, Pin-Chu Lai, Xin Qiao, Jingqiu Chen, Haw Yen, Zonglin Yanga, Chi-En Chiang, Rudolph Avalos, Abbigail L. Rath
MedPAO‑VL is a multimodal agent that generates structured radiology reports from chest X‑rays using an explainable plan‑act‑observe loop. The system first interprets the image with a vision module that captures diagnostic uncertainty, then iteratively refines the report by correcting semantic discrepancies to better match visual findings. Evaluation shows improved clinical correctness and structural compliance, with expert ratings of 3.5/5 for accuracy and 4.3/5 for structure.
By Shrish Shrinath Vaidya, Gowthamaan Palani, Minmini Selvam, Gokulraja Srinivasaraja, Ganapathy Krishnamurthi
The paper introduces SpecAgent, a protocol that lets a strong language model act as a reviewer rather than a generator for tool‑using agents. A weaker model drafts steps, executing read‑only tool calls to ground its proposals, while the strong model verifies the draft in a single prefill pass, committing the longest valid prefix and repairing any rejected step. Experiments on benchmarks show SpecAgent achieves near‑strong‑model performance while cutting memory traffic, wall time, and token generation.
By Zilin Wang, Yinhong Liu, Chao Chen, Qun Liu, Zhijiang Guo
The paper presents a small coverage‑closure agent that verifies data integrity for LLM‑generated functional tests in hardware design. It reads coverage from simulation reports (cocotb‑coverage YAML) and uses separate scoreboards to check data, rather than relying on the model’s account. Applied to three different RTL designs—a FIFO, a UART transmitter, and an AHB2APB bridge—the agent increased coverage from 91.7%, 88.9%, and 67.4% to 100% within eight iterations, and also uncovered a clock‑domain‑crossing bug in a manual testbench.
By Venkata Krishna Kumar Vedantam
The paper examines how collective‑action systems balance decentralization and centralization through delegation. A reduced‑form model shows a scale–accountability tradeoff: larger constituencies lower individual costs but reduce a center’s incentive to retain each member, limiting non‑exploitative delegation to an endogenous size. Simulations with large language model agents reveal that polycentric governance structures naturally arise, cycling through concentration, exploitation, exit, and re‑formation, and that different model capacities vary in their willingness to lead and in their propensity to exploit.
By Arin Mukhopadhyay, Aris Filos-Ratsikas, Panayiotis Danassis
The paper investigates how to infer agents’ underlying incentives in strategic multi‑agent settings, where traditional inverse reinforcement learning (IRL) may fail to produce rewards that guide optimal play. It introduces the concept of strategic identifiability and shows that learning from a fixed pair of policies can be insufficient regardless of sample size. To address this, the authors propose an active IRL algorithm for two‑player repeated Stackelberg games, combining maximum‑likelihood estimation with experimental design to achieve finite‑sample guarantees on sub‑optimality, and validate the approach in an anti‑poaching scenario.
By Antoine Bergerault, Cyrille Kone, Giorgia Ramponi
The paper introduces a scalable framework that uses decentralized markets of self‑interested agents to solve multi‑step tasks. It shows that when agents compete through sequential auctions, perfect competition fails because clones drive profits to zero, while no competition also collapses value extraction. By allowing heterogeneous agents to shade bids strategically, the authors demonstrate that imperfect competition can sustain profitable margins while still achieving high task success rates at low costs, evaluating four bidding algorithms across three auction protocols.
By Ian Gemp, Shubhendu Trivedi, Adam Fisch, Fantine Huot, Michael Kaisers, Mirella Lapata, William W. Cohen, Alekh Agarwal, Kate Larson, Jacob Eisenstein
The study investigates whether the language of a user’s instruction affects the safety verdicts of agent guards. By translating 1,220 labeled steps from an English benchmark into German, French, Turkish, Swahili, and Amharic while keeping all other data identical, the authors find that many open, small, and purpose‑built guards alter their verdicts when the instruction is translated, with larger effects for rarer languages. The impact varies with guard size, with smaller models showing greater loss of recall and larger models (e.g., Gemini 3.8 Flash) remaining largely unchanged.
By Hasan Demirkiran, Jens Ernstberger
PREPING introduces a proposer-guided framework for constructing agent memory before encountering target tasks, using self-generated synthetic practice. The method employs a proposer memory to shape synthetic tasks, a solver to execute them, and a validator to filter useful trajectories, thereby avoiding redundancy and infeasibility. Experiments on AppWorld, BFCL v3, and MCP-Universe demonstrate that PREPING outperforms a no-memory baseline and rivals playbook-based approaches while reducing deployment costs by 2.99× and 2.23× respectively.
By Yumin Choi, Sangwoo Park, Minki Kang, Jinheon Baek, Sung Ju Hwang
arXiv:2605. 22166v3 Announce Type: replace Abstract: LLM agents are shaped not only by their language models, but also by the runtime harness that mediates observation, tool use, action execution, feedback interpretation, and trajectory control.
By Tianshi Xu, Huifeng Wen, Meng Li
Harness-Bench is a diagnostic benchmark that evaluates how different harness configurations affect LLM agent performance in realistic workflows. It tests 106 sandboxed offline tasks across multiple model backends, recording artifacts, execution traces, and usage statistics to analyze completion, process quality, efficiency, and failure behavior. The study finds significant variation in outcomes across model-harness pairings, indicating that agent capability should be reported at the configuration level rather than solely by the base model.
By Yilun Yao, Xinyu Tan, Chao-Hsuan Liu, Yaoming Li, Zhengyang Wang, Zongwei Lv, Wenhan Yu, Zhewen Tan, Yuxuan Tian, Lin Sun, Xiangzheng Zhang, Guangxiang Zhao, Tong Yang
The paper introduces LLM-as-an-Investigator, an evidence-first approach for interactive problem diagnosis that mitigates premature commitment to incorrect hypotheses. It implements this through the Solution Investigator Agent (SIA), which coordinates a Diagnostic Reasoner, a Hypothesis Investigator, and a Counterfactual Investigator to gather discriminative evidence and refine competing explanations. The method is evaluated on a structured corpus of 303 technical forum threads, outperforming several baseline strategies and demonstrating robustness to misleading user hypotheses.
By Fabrizio Marozzo, Pietro Li\`o
The paper introduces CTXGAP, a study measuring how much information coding agents omit from their user-facing reports compared to their full execution logs. Across 37 coding tasks, agents answer 93.6% of task-state questions when asked directly, but report-only readers recover only 52.0% of the same information, revealing a 40.7‑point gap. Even at final checkpoints of verified fixes, the gap remains significant, and detailed terminal scrollback improves recovery only to 73.6%. The study highlights that missing details in reports can leave supervisors unaware of failures, workarounds, and risks that need scrutiny.
By Guoyang Xu, Rui Zhang, Xiuyu Wu, Hao Chen