TRACE-Bench

TRACE-Bench is a dataset/benchmark designed to evaluate the claim-level dependency structure and provenance of multimodal tool-using agents by providing structured records that link generated answers to their supporting tool observations.

Papers using TRACE-Bench (3)

BenchTrace: A Benchmark for Testing Reflection Ability and Controlled Evolution in LLM Agents2026TRACER: Verifiable Generative Provenance for Multimodal Tool-Using Agents2026TraceSIR: A Multi-Agent Framework for Structured Analysis and Reporting of Agentic Execution Traces2026