MLRC-Bench

Tests whether a language agent can propose and implement a genuinely novel ML method across 7 real research-competition tasks, scored against each competition's own baseline and top human result.

Also known as: MLRC-Bench: Can Language Agents Solve Machine Learning Research Challenges?

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryagentic
Subcategorymachine-learning research-competition agent benchmark (novel methods vs. baseline and top human)
Page statusactive
Metricrelative_improvement_to_human (primary), with absolute_improvement_to_baseline reported alongside
Directionhigher_is_better
Unit%
Dataset size7
Dataset licenceCC-BY-4.0 (Hugging Face dataset card and the companion Space's cardData both state this)
PublisherUniversity of Michigan; LG AI Research; University of Illinois at Chicago

What it measures

MLRC-Bench gives an agent an open machine-learning research problem adapted from a real competition at a recent ML conference (NeurIPS, KDD, ECCV and similar venues, spanning areas such as LLM safety, computer vision and AI for science): a baseline code repository, background literature, and dev-set data to iterate against. The agent must propose a new method, implement it in code, and run it, with its final version evaluated once on a held test split. Unlike end-to-end agent benchmarks that grade a full pipeline with an LLM-as-judge, MLRC-Bench isolates the specific steps of proposing and implementing a novel method and grades the result with each competition's own objective metric, explicitly to avoid rewarding an agent for sounding innovative rather than being effective.

Task format

The agent works inside a per-task sandboxed environment (its own conda/Docker setup, described in a README the agent can read) with shell and code-execution access, iterating on a development split for a time or message budget. Its method is then run once via the task's own `main.py -m <method_name> -p test` command on the held test split; some tasks (llm-merging, machine-unlearning, product-rec) require submitting a run to Kaggle or AIcrowd to obtain the final score rather than scoring locally.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub