Breaking

Paper2Code Released to Auto-Generate Code From ML Papers, Accepted at ICLR 2026

June 30, 2026 at 16:30 EDT

AI Agents
Software Dev & Coding
Open Source

An open-source project called Paper2Code that automatically generates code repositories from machine learning papers has been released. Its core multi-agent LLM system, PaperCoder, processes a paper in three stages—planning, analysis and code generation—to output a working repository.

ICLR 2026 · PaperCoder / Paper2Code

A multi-agent LLM that reads ML papers and writes the whole code repo

PaperCoder turns a scientific paper into a working, paper-faithful code repository — splitting the job across specialized agents to attack the ML reproducibility crisis. Open source under Apache-2.0.

specialized agent stages in the pipeline

$0.50–0.70

estimated cost per paper (o3-mini)

1–5

correctness score, o3-mini-high over 8 samples

The pipeline: paper in → repo out

INPUT

PDF / LaTeX

→

STAGE 1

Planning

→

STAGE 2

Analysis

→

STAGE 3

Code Generation

→

OUTPUT

Working repo

Demo: regenerates an implementation repo from "Attention Is All You Need." Runs on o3-mini, or open-source models like DeepSeek-Coder-V2-Lite-Instruct via vLLM.

Why it matters

Reported to outperform existing baselines on PaperBench and its own Paper2Code Benchmark — a step toward solving the "reproducibility crisis," where shipped code is often incomplete and snippet tools can't assemble a full repo.

Caveats

PDFs need preprocessing (s2orc-doc2json); output completeness tracks the eval score, so human review and debugging stay recommended. Some ask why the strong Claude Sonnet series isn't used. Real-world usage remains early-stage.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.