OpenDiscoveryTrace: The First Dataset That Watches How AI Scientists Actually Think

When we evaluate AI scientists today, we mostly ask one question: did it get the right answer? A new dataset presented at the ICML 2026 AI for Science Workshop argues that this is fundamentally the wrong question - and proves the point with numbers that are hard to ignore.

OpenDiscoveryTrace is the first public, structured dataset of complete AI scientific agent process traces. It won the Best Dataset Award at the workshop's Dataset Proposal Competition. The core finding is both elegant and unsettling: two frontier models can arrive at the same answer, at the same rate, using reasoning processes that are almost nothing alike.

The Problem With Outcome-Only Evaluation

1 Existing benchmarks for autonomous AI scientists evaluate only final outputs - generated code, hypotheses, or papers - yet discard the reasoning process by which those outputs were obtained. 1 This makes it impossible to audit scientific methodology, diagnose failure modes, or distinguish systematic reasoning from fortunate guessing.

The authors identify three concrete harms from this blind spot 1:

1. Conflated quality and luck. A model that stumbles into a correct answer scores identically to one that reasons its way there.
2. Ungovernable systems. Regulatory and institutional governance increasingly requires complete decision traces for auditing - something output-only benchmarks cannot supply.
3. Stuck improvement loops. Fixing an agent requires understanding where it fails, and failures are trajectory properties, not output properties.

What the Dataset Contains

2 OpenDiscoveryTrace comprises 558 trajectories across seven models: 372 from three frontier models (GPT-5.4, Claude Opus 4.6, and Gemini 3.1 Pro, 124 each, fully balanced across domains and difficulty levels) and four open-weight models (Qwen2.5-7B, Mistral-7B-v0.3, Phi-3.5-mini, and Qwen2.5-1.5B, 30 each), plus 60 live-retrieval variant trajectories.

3 Each trajectory records a structured 9-field-per-step trace - including 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.

The trace schema is deliberately rich. Capturing revision_triggers and self_reported_confidence at every step means researchers can study not just what a model did, but why it changed its mind and how certain it claimed to be along the way. These fields are invisible in any output-only evaluation.

The Headline Finding: Same Score, Completely Different Behavior

This is where the dataset earns its keep. 4 Pilot analysis on 363 LLM-judged trajectories reveals that all three frontier models achieve comparable success rates (84 - 89%). That looks like a tie - until you examine the traces.

4 Claude Opus 4.6 produces 30× more errors per trajectory than GPT-5.4 (2.5 vs. 0.08, p < 0.0001, Cliff's δ = 0.613), while arriving at the same answers. And the error profiles are qualitatively different: 5 Claude's errors are 66.7% tool misuse, while GPT-5.4's errors are 83.6% reasoning errors.

In plain terms: Claude fumbles with its tools constantly but reasons past those fumbles; GPT-5.4 uses tools cleanly but miscalculates more often. Both produce correct answers, but they are doing fundamentally different things. 6 Process traces expose this dimension; output-only benchmarks cannot.

This has major implications for which model you'd want in a production scientific pipeline. An agent that misuses tools at high frequency is a liability in any environment where tool calls have real-world side effects - API costs, database writes, irreversible lab instructions. Knowing the error type matters as much as knowing the error rate.

Five Benchmark Tasks and Open Infrastructure

Beyond the dataset itself, the authors define five benchmark prediction tasks with baselines ranging from logistic regression and random forests to LSTMs and Transformers. This gives the community concrete targets: can a model trained on process traces predict where an agent will fail, or when its self-reported confidence is miscalibrated?

The dataset, trace schema, agent harness, and benchmark definitions are all publicly available on Hugging Face, making OpenDiscoveryTrace immediately usable for reproducible research.

Why It Matters

The AI-scientist space is crowded and moving fast. Systems like The AI Scientist, AgentLaboratory, and a growing roster of specialized research agents are moving from demos to real deployment. As that happens, the evaluation infrastructure needs to keep pace.

OpenDiscoveryTrace makes a principled case that the next generation of AI scientist benchmarks must be process-aware, not just outcome-aware. For practitioners, this means:

• Auditing becomes possible. A structured 9-field trace is something a compliance team can review; a final paper or hypothesis is not.
• Debugging gets specific. "The model failed" is actionable only if you know whether it failed because of a bad tool call, a reasoning error, or a confidence miscalibration.
• Model selection changes. Two models with the same accuracy score may have very different deployment risk profiles once you look inside the trajectory.

The dataset is a research artifact, but the argument it makes is an engineering argument: you cannot responsibly deploy what you cannot inspect. In a world where AI agents are running real experiments on real data, that argument is hard to dismiss.