LLM Tool Use

LLM: Large Language Model

Large language model tool use examines how models decide when tools are needed, choose among them, provide valid arguments, and use returned results to complete tasks. Common approaches include prompting frameworks such as ReAct, supervised fine-tuning and reinforcement learning on tool-call trajectories, retrieval and filtering for large tool catalogs, and planning with execution feedback or memory. Benchmarks such as BFCL, AppWorld, and MCP-based evaluations test structured calls, multi-step workflows, and real-world tool interactions. A central challenge is making tool use reliable and efficient: tools can improve computation and access to current information, but unnecessary calls, bad arguments, misleading results, and failures in long-horizon reasoning can make performance worse.

Latest papers

295 papers47 papers in the last four weeks, up 38% on the four weeks before. 0.3% of all new papers.

Context Is a Pipeline, Not a Window: Information Attrition in LLM Agent Harnesses
Oct 26, 2026Yuvraj Sehgal, Mahsa Panahandeh, Naser Ezzati-JivanLLM Context Management LLM Agent Harnesses
Department of Computer Science Brock University St. Catharines, Ontario, Canada · School of Electrical Engineering and Computer Science University of Ottawa Ottawa, Ontario, Canada
Matched Capabilities, Different Behaviors: A Controlled Study of MCP and CLI Interfaces for Language Agents
Oct 25, 2026Zhenning Shi, Liang Pang, Yongqi LiTool-Augmented Language Model Agents LLM Tool Use
The Hong Kong Polytechnic University · Institute of Computing Technology, Chinese Academy of Sciences
Prompt Sensitivity in Tool-Augmented Model Comparisons
Oct 22, 2026Mateusz Modrzejewski, Tomasz RadzikowskiLLM Evaluation Prompt Sensitivity
Institute of Computer Science, Warsaw University of Technology
Before the Ticket: Grounding HPC User Support in Documentation and Resolved Tickets with an Agentic LLM Assistant
Oct 21, 2026Elham J. Barezi, Primus C Kabuo, Niveditha NerellaQuestion Answering Document QA
Rosen Center for Advanced Computing, Purdue University, West Lafayette, IN, USA
Measure First, Then Allocate: Calibrated Rollout Budgets for MCP Agents in Reinforcement Learning
Oct 20, 2026Hieu Tran, Zonghai Yao, Sunjae Kwon +1RL for LLM Agents RL for Language Models
University of Massachusetts Amherst · University of Massachusetts Lowell
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Oct 19, 2026Ke Zhang, Sahchit Chundur, Jerry Peprah Owusu +2LLM Tool Use LLM Agent Reliability
University of California, Riverside · University of California, Irvine
Learning Which Device to Call: Online Tool Grounding for LLM Agents in IoT
Oct 18, 2026Yichen Liu, Xiaochong Jiang, Haoran Yu +1LLM Tool Use Tool Selection for Language Model Agents
Independent Researcher, Seattle, WA, USA · Independent Researcher, Boston, MA, USA
Muse Glimmer 30B: An On-Device Agentic Model
Oct 17, 2026Arash Ebrahim Abtahizadeh, Muna Aghamelu, Adil Ahmed +326LLM Agents LLM Tool Use
Explore Separately, Adopt Selectively: Execution-Guided Tool Use over Large Tool Libraries
Oct 15, 2026Anna Mazhar, Seiji Maekawa, Nikita BhutaniTool Discovery for LLM Agents Tool-Using Agents
Cornell University · Megagon Labs
Rethinking the Role of Tool Graph Memory in Multi-Turn Tool Use
Oct 15, 2026Uiji Hwang, Richeng Xuan, Taeuk KimLLM Tool Use Agent Memory
Hanyang University · Hunyuan, Tencent
Tool-Augmented Large Language Models for Reliable Medical Calculation from Patient Notes
Oct 15, 2026Yang Zhong, Ling Luo, Xuanlyu Zhu +4LLM Tool Use Clinical Decision Support
College of Computer Science and Technology, Dalian University of Technology, Dalian, China · Air Force Communications NCO Academy, Dalian, China
Incentivizing Tool-Use Reinforcement Learning for Post-Arrival Emergency Care Prediction
Oct 15, 2026Pengfei Hu, Chenwei Wu, Qiao Jin +4Clinical Outcome Prediction Clinical Risk Prediction
Stevens Institute of Technology · University of Michigan, Ann Arbor · National Institutes of Health +1
TRACER: Trajectory-Aware Tool Retrieval for LLM Agents
Oct 15, 2026Tanish Kher, Sarvesh DK, Abedelkadir Asi +2Tool Selection for Language Model Agents LLM Tool Use
Microsoft
From Relevance to Validity: Zero-Shot Tool Selection via Functional and Contextual Validation
Oct 14, 2026Xiaoxuan Dong, Jin Zhang, Zichen Yuan +11Tool Selection LLM Tool Use
UESTC · LZU · CityUHK +5
Compiling Expert Procedures into Verified Programs: When a Small Model Is Enough
Oct 13, 2026Vishvesh Bhat, Asad Aali, Muhammad Ahmed MohsinLLM Tool Use LLM-Based Program Synthesis
CoreThink AI · Stanford University
Plan or Retrieve First? A Scale Study of Tool Discovery Architectures
Oct 13, 2026Mingfei Zhang, Amit Surana, Ziauddin Majid +2Tool Discovery for LLM Agents LLM Tool Use
Apple
Agaram 2.5B: The First Tamil-First Language Model Trained from Scratch with Tool Calling. A Playbook for Under-Represented Languages
Oct 13, 2026Ramchand Kumaresan, Madhusudhanan G, Kaushik KezendiraneLanguage Modeling Low-Resource Language Processing
Murai Labs · Independent researcher
FRACTION: Fraud Explanation Generation via Agentic Graph Context Navigation
Oct 13, 2026Cecelia Soh, Jun Hu, Yuan Li +6Faithfulness of Language Model Explanations LLMs for Graph Learning
National University of Singapore · Hefei University of Technology · ByteDance Inc. +1
ToolHACE: A Dataset for Detecting Hallucinations in Tool-Based LLM Responses
Oct 12, 2026Nikolay Ivanov, Sergey Pletenev, Elisei Rykov +8Hallucination Detection LLM Hallucination Detection
Skoltech · AIRI · T-Tech +1
Davinci-Tool: Scaling Agentic Tool-Use Supervision with Verifiable Executable Environments
Oct 12, 2026Xiaoyang Yuan, Ji Zhao, Xingyu Hu +7LLM Tool Use LLM Agent Training
Tongji University · Shanghai Innovation Institute · University of Science and Technology of China +1
Sensitivity or Criterion? Measuring and Moving When Language Models Act
Oct 11, 2026Sixun Dong, Yebowen Hu, Ming Yin +3Agent Evaluation LLM Tool Use
University of Central Florida · Zoom Communications · Duke University +1
SAT-RL: Training Language Models for Logical Reasoning with Symbolic Solvers
Oct 9, 2026Jialiang Sun, Kuldeep MeelRL for Language Model Reasoning Symbolic Reasoning
University of Toronto
Evaluating LLMs as Agentic Workflow Executors in Real-World Customer Support Skills
Oct 9, 2026Md Tahmid Rahman Laskar, Gundeep Singh, Xue-Yong Fu +11LLM Agent Evaluation Agentic Workflows
Dialpad Inc.
Which Metadata Matters for LLM Tool Selection in MCP? Evidence from Controlled Factor-Level Interventions
Oct 9, 2026Xiaofan Li, Fengchun Qiao, Xing GaoLLM Tool Use Model Context Protocol
University of Delaware, Newark, DE, USA · University of South Florida, Tampa, FL, USA
Masked Diffusion Language Agents for Tool-Integrated Chemical Reasoning
Oct 9, 2026Mengdi Liu, Chenghao Jia, Hong Chang +1Chemical Reasoning in LLMs LLM Tool Use
State Key Laboratory of AI Safety, Institute of Computing Technology, CAS, China · University of Chinese Academy of Sciences (CAS), China
How Do Agentic LLMs Decide to Call Tools? A Tool-Call Vector Shaped by Suppression
Oct 7, 2026Xijie Gong, Tingxu Han, Jiahao Zhang +5LLM Tool Use LLM Decision-Making
Mohamed bin Zayed University of Artificial Intelligence · University of Electronic Science and Technology of China · Nanjing University +3
Structured but Silent: Probing Capability Requirements in LLM Hidden States
Oct 6, 2026Kyojun Choo, Minsoo Song, Yunju Kang +1LLM Tool Use Language Model Probing
Soongsil University
Mind the Accent Gap: British Accent Robustness in Speech-Driven Financial Voice Assistants
Oct 5, 2026Aadam Haq, Oggi Rudovic, Malcolm Chadwick +5ASR Evaluation Voice Agent Evaluation
NatWest, UK
SchemaFill: Efficient LLM Tool Calling via Slot-Parallel Speculative Decoding
Oct 5, 2026Zhi-Kai Chen, Song-Yan Li, De-Chuan Zhan +1Speculative Decoding LLM Inference Acceleration
School of Artificial Intelligence, Nanjing University, China · National Key Laboratory for Novel Software Technology, Nanjing University, China · Nanjing University, China
When LLMs Sit Above Diagnostic Tools: Unrealized Complementarity in Industrial Fault Diagnosis
Oct 4, 2026Donghwan KimLLM Reliability LLM Tool Use
Independent Researcher

Oct 26, 2026J/K move · Enter open · S save

Context Is a Pipeline, Not a Window: Information Attrition in LLM Agent Harnesses

Yuvraj Sehgal, Mahsa Panahandeh, Naser Ezzati-Jivan

Department of Computer Science Brock University St. Catharines, Ontario, Canada · School of Electrical Engineering and Computer Science University of Ottawa Ottawa, Ontario, Canada

LLM Context Management LLM Agent Harnesses LLM Tool Use LLM Agent Evaluation

Open paper AskPDF arXiv

LLM agents are judged by two numbers: the size of the context window, and how often the final answer is correct. Both numbers miss a third thing: how much of a tool's output is lost before the model ever sees it. We measured this loss in a real tool-using agent. The agent finds the cause of injected faults by reading Linux kernel traces. The context window was almost empty: across 720 runs, the single largest prompt used only 5.75% of it. Yet a third of all tool results (3,827 of 11,189) were cut before they reached the model, and every cut result became broken JSON. The harness cut each result by slicing its JSON text at a fixed length. This does not shorten a result evenly. It deletes whatever data sits near the end. One field reached the model in only 4.7% of calls. A second field, in the very same results, reached the model in 99.3% of calls. Which field held the fault's evidence depended on when the faulty process started: early or late in the trace. The agent's diagnosis accuracy matched this same split. So we ran controlled experiments to test one question: did the data loss cause the drop in accuracy? This paper makes four contributions. First, we model an agent's context as a five-stage pipeline, with a measured failure at each stage. Second, we replay 864 real tool results, including one from a production trace where the shipped cut deletes all 128 records of the fault's evidence. Third, a controlled experiment shows that the cut's real cost appears first as extra work, not as wrong answers: 43% more tool calls, twice the computation, and yield falling from 0.61 to 0.31. The agent still finds the right answer when it has another way to reach it. Fourth, a survey of seven agent frameworks finds two that cut tool output the same unsafe way. Whether the data cut truly causes the accuracy drop is still an open question.