Research map / Multi-agent systems and games / LLM agents

DeepTRACE: Auditing Deep Research AI Systems for Tracking Reliability Across Citations and Evidence

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: Software engineering agent benchmarks

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.

Introduces DeepTRACE, an audit framework with eight metrics that evaluates generative search engines and deep research agents on answer text, source, and citation reliability, using a 303-query corpus and an LLM judge validated against human raters.

Problem

Generative search engines and deep research agents often produce one-sided, overconfident answers with unsupported statements and inaccurate citations, and existing benchmarks evaluate isolated components rather than end-to-end, community-relevant reliability.

Approach

Turns 16 community-identified failure cases from a usability study into eight measurable metrics (one-sided, overconfident, relevant statements, uncited sources, unsupported statements, source necessity, citation accuracy, citation thoroughness) computed from statement decomposition, a citation matrix and an LLM-judged factual support matrix.

Key results

• GPT-5(DR) achieves 87.5% citation thoroughness, 79.1% citation accuracy, and 12.5% unsupported statements, the strongest deep research configuration.
• Deep research agents reduce overconfidence (<20% for all DR engines) but remain highly one-sided on debate queries (54.7–94.8%).
• Perplexity deep research has 97.5% unsupported statements, and GSEs show 40–68% citation accuracy.

Surprising findings

• Deep research agents with more listed sources (e.g., YouChat DR with 57 sources) show high unsupported rates, indicating more sources does not imply better grounding.
• Perplexity GSE gives overconfident answers in most cases while being one-sided, despite its high volume of statements.

Limitations

• Factual support judged by an LLM with only moderate agreement with humans (Pearson 0.62), and confidence scoring at 0.72 correlation.
• Excludes UI-level and multimodal factors and does not assess whether answers are correct.
• Evaluation snapshot as of August 2025 on closed commercial systems that may change over time.

Experimental setup

Datasets: DeepTRACE corpus (303 queries from usability study sessions and ProCon.org)

Benchmarks: DeepTRACE

Metrics: one-sided answer rate overconfident answer rate relevant statement rate uncited sources unsupported statements source necessity citation accuracy citation thoroughness

Baselines: GPT-5 web search YouChat Perplexity Copilot Gemini Deep Research

Builds on: RAGAS ClashEval DeepResearch Bench

Closest papers by meaning

• From Intent to Evidence: A Categorical Approach for Structural Evaluation of Deep Research Agents (NeurIPS 2026 · similarity 0.88)
• WebWeaver: Structuring Web-Scale Evidence with Dynamic Outlines for Open-Ended Deep Research (ICLR 2026 · similarity 0.87)
• MiroEval: Benchmarking Multimodal Deep Research Agents in Process and Outcome (NeurIPS 2026 · similarity 0.85)
• Deep Research Arena: The First Exam of LLMs’ Research Abilities via Seminar-Grounded Tasks (AAAI 2026 · similarity 0.84)
• Auditing Sabotage Bench: A Benchmark for Detecting and Fixing Research Sabotage in ML Codebases (NeurIPS 2026 · similarity 0.84)
• Trust, but Don’t Verify: Epistemic Blind Spots in LLM Source Evaluation (NeurIPS 2026 · similarity 0.84)