# Agentic Reverse Engineering: From I/O to Code with Discovery Agent

> Synthesizes executable programs from input/output observations using LLM-guided evolutionary search, progressively revealed examples, simple-to-complex mutations and execution-error feedback. An active variant chooses inputs and queries a hidden reference program.

- Authors: Yihong Dong; Jiaru Qian; Haoran Zhang; Peixu Wang; Binhua Li; Zhi Jin; Yongbin Li; Ge Li; Xiaokang Yang; Xue Jiang
- Venue: NeurIPS 2026
- Status: Conference accepted
- Publication / posting date: 2026 (exact date not verified)
- Date source: Official NeurIPS 2026 poster listing confirms acceptance, not a proceedings publication date. First arXiv submission was 2026-05-14; reviewed v2 was revised on 2026-10-18.
- First arXiv submission: May 14, 2026
- arXiv: 2605.15334
- Reviewed arXiv version: v2
- DOI: 10.48550/arxiv.2605.15334
- Sources reviewed: 2026-10-20
- Added to the catalog: 2026-10-20
- Original paper: <https://arxiv.org/abs/2605.15334>
- PDF: <https://arxiv.org/pdf/2605.15334v2>
- Code: <https://github.com/JiaruQian/IO2Code>
- BibTeX: <https://awesomerea.org/papers/citations/paper-dio-agent.bib>
- Catalog page: <https://awesomerea.org/papers/paper-dio-agent/>
- Record JSON: <https://awesomerea.org/papers/paper-dio-agent/record.json>
- Chinese version: [简体中文](https://awesomerea.org/zh/papers/paper-dio-agent/index.md)

## Scope and limitations

IO2CodeBench has 125 generated tasks spanning basic operations, algorithms, geometry and image-to-value mappings, normally with eight visible and fifteen held-out examples. Active exploration covers a 25% subset from each level. Multimodal results average sample pass ratios, while other levels count tasks passing every held-out example. Finite test success does not establish original-source recovery or complete equivalence; arbitrary applications and stateful services are not evaluated.

## Why it fits this collection

Provides a concrete observation-to-program reconstruction workflow and separates visible-example fitting from held-out behavior checks.

## Publication and review notes

The official poster and pinned repository retain the earlier title From I/O to Code with Discovery Agent; all ten authors match v2 and DataCite. The registered DOI identifies the arXiv work. The paper uses the arXiv non-exclusive license. Author-linked code was checked at ef15827ab7090bf2d227fb1eef51d91b0163339e: its README documents Python 3.10+, provider credentials, benchmark/oracle tasks and experiment runners. No repository-wide code/data license was verified in the reviewed README and root tree.

## Primary sources

1. <https://arxiv.org/abs/2605.15334>
2. <https://arxiv.org/html/2605.15334v2>
3. <https://api.datacite.org/dois/10.48550/arxiv.2605.15334>
4. <https://neurips.cc/virtual/2026/loc/sydney/poster/154711>
5. <https://github.com/JiaruQian/IO2Code/blob/ef15827ab7090bf2d227fb1eef51d91b0163339e/README.md>
6. <https://github.com/JiaruQian/IO2Code/tree/ef15827ab7090bf2d227fb1eef51d91b0163339e>

## About this entry

Reviewed against the paper and its primary sources. Cite the original authors using the linked paper or BibTeX.
Original editorial content: MIT. Linked projects, papers and assets retain their own rights and licenses.