Research map / Multi-agent systems and games / LLM agents
Beyond the Final Answer: Evaluating the Reasoning Trajectories of Tool-Augmented Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TRACE, a reference-free LLM-judge framework using an evidence bank to score tool-augmented agent trajectories on efficiency, hallucination, and adaptivity, and builds Meta-GTA/Meta-m&m's meta-evaluation sets to validate it.
Problem
Tool-augmented agent benchmarks evaluate only final-answer accuracy and rely on single ground-truth trajectories, missing efficiency, hallucination, and adaptivity; comparing against one reference trajectory is unreliable and annotating all valid trajectories is expensive.
Approach
Accumulate tool evidence (action, input, observation) into an evidence bank and use an LLM evaluator to check thoughts against prior evidence (hallucination), prune unnecessary evidence after the final answer (efficiency), and check whether the agent recovers after a tool failure (adaptivity), without any ground-truth trajectory.
Key results
- On Meta-GTA, TRACE improves over LLM-as-a-Judge efficiency accuracy for Llama-3.1-8B from 55.67 to 70.46 and hallucination accuracy from 89.59 to 93.78.
- TRACE achieves efficiency accuracy of 94.64 (Claude-Sonnet-4) and 94.24 (GPT-4.1) on Meta-GTA, with the largest gains for smaller open models.
- On GTA, Qwen-72B and GPT-4.1 have similar overall accuracy (0.5202 vs 0.5281) but differ in efficiency and hallucination, showing trajectory metrics reveal differences hidden by accuracy.
Surprising findings
- Mistral-7B achieves near-perfect hallucination score mainly by terminating early rather than completing tasks.
- Qwen-72B and GPT-4.1 show low adaptivity despite high overall accuracy.
Limitations
- Meta-evaluation data is synthetically injected and validated by a single human annotator on a 100-sample subset per model.
- Adaptivity is tested only with four synthetic fake tools in GTA, a narrow failure setting.
- Hallucination is defined to include use of internal knowledge not in the evidence bank, which may conflate correctness with grounding.
Experimental setup
Datasets: GTA m&m's
Benchmarks: Meta-GTA Meta-m&m's GTA GAIA
Metrics: efficiency accuracy hallucination accuracy adaptivity accuracy answer accuracy instruction error
Baselines: LLM-as-a-Judge PIPA
Builds on: ReAct GTA benchmark m&m's benchmark PIPA
Closest papers by meaning
- Probing the Trajectories of Reasoning Traces in Large Language Models (NeurIPS 2026 · similarity 0.88)
- GTA: A Benchmark for General Tool Agents (NeurIPS 2024 · similarity 0.88)
- MAS-ProVe: Understanding the Process Verification of Multi-Agent Systems (ICML 2026 · similarity 0.87)
- LLM Agents Already Know When to Call Tools - Even Without Reasoning (NeurIPS 2026 · similarity 0.87)
- SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents (NeurIPS 2025 · similarity 0.87)
- ShortcutsBench: A Large-Scale Real-world Benchmark for API-based Agents (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.