Beyond Solution-Centric Search: Adaptive Inquiry and Knowledge Revision for Autonomous ML Engineering

TL;DR - An arXiv preprint proposing an "information paradigm" for autonomous ML engineering agents, replacing solution-centric tree/graph/chain search with an evolving information state; its instantiation, Iris, reaches a 64.9% any-medal rate on MLE-Bench under a 12-hour budget.

• Reframes long-horizon agent design: instead of the search structure dictating information flow, an explicit information state represents task understanding and drives solution improvement.
• Iris runs an inquiry-revision loop — it derives local action plans from the current state and takes "epistemic actions" that probe decision-critical unknowns without altering the retained solution.
• Information management synthesizes cross-experiment observations into task knowledge made of revisable claims with explicit scope and status, updated as evidence arrives; decision contexts are assembled from raw evidence, structured summaries, or task knowledge at the needed granularity.
• Reported results: highest any-medal rate (64.9%) among compared systems on MLE-Bench at a 12-hour budget, plus cross-domain generalization across four tasks in harness engineering and model post-training.