From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking

• URL: http://arxiv.org/abs/2506.19724v1
• Date: Tue, 24 Jun 2025 15:39:20 GMT
• Title: From Reproduction to Replication: Evaluating Research Agents with Progressive Code Masking
• Authors: Gyeongwon James Kim, Alex Wilf, Louis-Philippe Morency, Daniel Fried,
• Abstract summary: AutoExperiment is a benchmark that evaluates AI agents' ability to implement and run machine learning experiments.<n>We evaluate state-of-the-art agents and find that performance degrades rapidly as $n$ increases.<n>Our findings highlight critical challenges in long-horizon code generation, context retrieval, and autonomous experiment execution.
• Score: 48.90371827091671
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Recent progress in autonomous code generation has fueled excitement around AI agents capable of accelerating scientific discovery by running experiments. However, there is currently no benchmark that evaluates whether such agents can implement scientific ideas when given varied amounts of code as a starting point, interpolating between reproduction (running code) and from-scratch replication (fully re-implementing and running code). We introduce AutoExperiment, a benchmark that evaluates AI agents' ability to implement and run machine learning experiments based on natural language descriptions in research papers. In each task, agents are given a research paper, a codebase with key functions masked out, and a command to run the experiment. The goal is to generate the missing code, execute the experiment in a sandboxed environment, and reproduce the results. AutoExperiment scales in difficulty by varying the number of missing functions $n$, ranging from partial reproduction to full replication. We evaluate state-of-the-art agents and find that performance degrades rapidly as $n$ increases. Agents that can dynamically interact with the environment (e.g. to debug their code) can outperform agents in fixed "agentless" harnesses, and there exists a significant gap between single-shot and multi-trial success rates (Pass@1 vs. Pass@5), motivating verifier approaches to our benchmark. Our findings highlight critical challenges in long-horizon code generation, context retrieval, and autonomous experiment execution, establishing AutoExperiment as a new benchmark for evaluating progress in AI-driven scientific experimentation. Our data and code are open-sourced at https://github.com/j1mk1m/AutoExperiment .

Related papers

• Code with Me or for Me? How Increasing AI Automation Transforms Developer Workflows [66.1850490474361]
  We conduct the first academic study to explore developer interactions with coding agents.<n>We evaluate two leading copilot and agentic coding assistants, GitHub Copilot and OpenHands.<n>Our results show agents have the potential to assist developers in ways that surpass copilots.
  arXiv  Detail & Related papers  (2025-07-10T20:12:54Z)
• EXP-Bench: Can AI Conduct AI Research Experiments? [38.30861763360086]
  We introduce EXP-Bench, a novel benchmark designed to evaluate AI agents on complete research experiments.<n>Given a research question and incomplete starter code, EXP-Bench challenges AI agents to formulate hypotheses, design and implement experimental procedures, execute them, and analyze results.<n>With the pipeline, EXP-Bench curated 461 AI research tasks from 51 top-tier AI research papers.
  arXiv  Detail & Related papers  (2025-05-30T16:46:29Z)
• AutoReproduce: Automatic AI Experiment Reproduction with Paper Lineage [62.049868205196425]
  AutoReproduce is a framework capable of automatically reproducing experiments described in research papers in an end-to-end manner.<n>Results show that AutoReproduce achieves an average performance gap of $22.1%$ on $89.74%$ of the executable experiment runs.
  arXiv  Detail & Related papers  (2025-05-27T03:15:21Z)
• R&D-Agent: Automating Data-Driven AI Solution Building Through LLM-Powered Automated Research, Development, and Evolution [60.80016554091364]
  R&D-Agent is a dual-agent framework for iterative exploration.<n>The Researcher agent uses performance feedback to generate ideas, while the Developer agent refines code based on error feedback.<n>R&D-Agent is evaluated on MLE-Bench and emerges as the top-performing machine learning engineering agent.
  arXiv  Detail & Related papers  (2025-05-20T06:07:00Z)
• CodeScientist: End-to-End Semi-Automated Scientific Discovery with Code-based Experimentation [48.12054700748627]
  We introduce CodeScientist, a novel ASD system that frames ideation and experiment construction as a form of genetic search jointly.<n>We use this paradigm to conduct hundreds of automated experiments on machine-generated ideas broadly in the domain of agents and virtual environments.
  arXiv  Detail & Related papers  (2025-03-20T22:37:17Z)
• Commit0: Library Generation from Scratch [77.38414688148006]
  Commit0 is a benchmark that challenges AI agents to write libraries from scratch.<n>Agents are provided with a specification document outlining the library's API as well as a suite of interactive unit tests.<n> Commit0 also offers an interactive environment where models receive static analysis and execution feedback on the code they generate.
  arXiv  Detail & Related papers  (2024-12-02T18:11:30Z)
• MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation [96.71370747681078]
  We introduce MLAgentBench, a suite of 13 tasks ranging from improving model performance on CIFAR-10 to recent research problems like BabyLM. For each task, an agent can perform actions like reading/writing files, executing code, and inspecting outputs. We benchmark agents based on Claude v1.0, Claude v2.1, Claude v3 Opus, GPT-4, GPT-4-turbo, Gemini-Pro, and Mixtral and find that a Claude v3 Opus agent is the best in terms of success rate.
  arXiv  Detail & Related papers  (2023-10-05T04:06:12Z)
• SIERRA: A Modular Framework for Research Automation and Reproducibility [6.1678491628787455]
  We present SIERRA, a novel framework for accelerating research development and improving results. SIERRA accelerates research by automating the process of generating executable experiments from queries over independent variables. It employs a modular architecture enabling easy customization and extension for the needs of individual researchers.
  arXiv  Detail & Related papers  (2022-08-16T15:36:34Z)
• SIERRA: A Modular Framework for Research Automation [5.220940151628734]
  We present SIERRA, a novel framework for accelerating research developments and improving results. SIERRA makes it easy to quickly specify the independent variable(s) for an experiment, generate experimental inputs, automatically run the experiment, and process the results to generate deliverables such as graphs and videos. It employs a deeply modular approach that allows easy customization and extension of automation for the needs of individual researchers.
  arXiv  Detail & Related papers  (2022-03-03T23:45:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.