OpenDiscoveryTrace: A New Benchmark for Auditing the Reasoning Process of AI Scientists

Authors: Aayam Bansal, Keertan Balaji

arXiv: 2609.09203 [cs.AI]

Submitted: 5 September 2026

Venue: Best Dataset Award, ICML 2026 Workshop on AI for Science (AI Scientists: Tools, Co-authors, or Founders?)

Overview

Current benchmarks for autonomous AI scientists judge only the final output—generated code, hypotheses, or papers—while discarding the reasoning process behind those outputs. This blind spot makes it impossible to audit scientific methodology, diagnose failure modes, or distinguish systematic reasoning from lucky guessing.

OpenDiscoveryTrace addresses this gap with a public dataset of 558 complete AI scientific agent trajectories that capture how models reason, not merely what they produce. Each trajectory records a structured nine-field-per-step trace—thoughts, tool calls, observations, errors, revision triggers, and self-reported confidence—as models execute 124 scientific tasks spanning drug discovery, materials science, genomics, and scientific literature analysis.

Dataset Composition

The dataset covers seven models:

• Three frontier models (GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro)—124 trajectories each, fully balanced across domains and difficulty levels.
• Four open-weight models (Qwen2.5-7B, Mistral-7B-v0.3, Phi-3.5-mini, and Qwen2.5-1.5B)—30 trajectories each.
• Plus 60 live-retrieval variant trajectories.

Key Finding: Process Traces Reveal What Output-Only Evaluation Misses

Pilot analysis on 363 LLM-judged trajectories shows that process traces expose behavioral differences invisible to output-only evaluation. All three frontier models achieve comparable success rates (84–89%), yet Claude Opus 4.6 produces 30× more errors than GPT-5.4 (2.5 vs. 0.08 per trajectory, p < 0.0001, Cliff's δ = 0.613). Their error profiles are also qualitatively distinct:

• Claude Opus 4.6: 66.7% tool misuse
• GPT-5.4: 83.6% reasoning errors

This gap underscores the value of auditing how agents fail, not just whether they succeed.

Benchmark Tasks and Baselines

The authors define five benchmark tasks with baselines from logistic regression, random forests, LSTMs, and Transformer models, providing a foundation for standardized process-level evaluation.

Availability

The dataset, trace schema, agent harness, and benchmark definitions are publicly available under CC BY 4.0 to support research on process-level evaluation, scientific agent auditing, and AI governance.

• Code: github.com/aayambansal/OpenDiscoveryTrace
• Dataset: huggingface.co/datasets/aayambansall/OpenDiscoveryTrace

Why It Matters in 2026

As autonomous research agents move from demos into real scientific pipelines—writing grant proposals, running wet-lab simulations, and drafting peer-reviewed manuscripts—evaluation standards have lagged behind capability. Output-only benchmarks such as SWE-bench and paper-generation leaderboards cannot tell a regulator, journal editor, or collaborating lab why a model reached a conclusion. OpenDiscoveryTrace pushes the field toward trace-based evaluation, a prerequisite for reproducibility, meaningful auditing, and the emerging AI-governance frameworks now being drafted for scientific AI systems.

Subjects: Artificial Intelligence (cs.AI)

Cite as: arXiv:2609.09203 [cs.AI]

License: CC BY 4.0

via ArXiv AI