TraceML: An Empirical Analysis of Human-Agent Planning in Machine Learning Development
TraceML introduces a version-level schema pairing human and agent work on the same Kaggle competitions. It includes 4,465 human trajectories across 134 competitions, with seven competitions also worked by two agent scaffolds, yielding 430 paired human and 207 agent trajectories. Each code version records score, timestamp, action, intent, edit size, and score effect. The analysis shows experts alternate data work, validation, model changes, and ensembling, while agent scaffolds collapse into narrow loops: Codex re-weights ensembles and tunes submissions, MLEvolve mutates its model in place, and neither pivots.
Large language models write correct code for isolated problems but remain far weaker at autonomous machine-learning development, where an agent must revise data pipelines, models, and validation over hours of feedback, and on most competitions still finishes below strong human competitors. Outcome-based benchmarks record this gap but not its cause, because they grade the final submission and discard the development process behind it. TraceML pairs human and agent work on the same competitions under one version-level schema: 4,465 human Kaggle trajectories across 134 competitions, seven of which are also worked by two agent scaffolds, giving 430 paired human and 207 agent trajectories. Every code version carries its score, its timestamp, and labels for the action taken, its intent, the edit size, and the score effect. Read this way, the gap becomes concrete. Experts alternate data work, validation, model changes, and ensembling, and return to approaches they had set aside. Each agent scaffold instead collapses into a narrow loop: Codex spends its steps re-weighting ensembles and tuning submissions, MLEvolve mutates its model in place, and neither pivots at the h…
TraceML provides a process-level dataset that captures version-level actions, intents, edit sizes, and score effects for both human and agent trajectories. This enables analysis of planning and adaptation strategies beyond final scores. The observed agent behavior—narrow loops without pivoting—suggests current scaffolds lack mechanisms for strategic exploration and recovery from dead ends.
The gap between human and agent performance in ML development highlights a limitation in autonomous coding agents for complex, iterative tasks. This may slow adoption of fully autonomous ML development tools and increase demand for human-in-the-loop systems that combine agent efficiency with human strategic oversight.
TraceML offers a dataset and methodology for evaluating and improving autonomous ML development agents. Companies developing such agents could use it to identify weaknesses and guide product development. The research may also inform the design of hybrid human-agent workflows that leverage the strengths of both.
Future work may focus on improving agent planning capabilities, such as incorporating memory of past approaches, explicit exploration strategies, and better validation feedback loops. The TraceML dataset could serve as a benchmark for evaluating such improvements. Observable next signals include new agent architectures that demonstrate pivoting behavior and improved competition rankings.