2026 · Preprint
RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments
Tests whether agents can recover executable decision policies from behavioral traces and self-designed opponent probes, using 75 synthetic policies across five game arenas and measuring held-out action agreement and downstream competitive value.
Scope and limitations
The targets are fixed synthetic policies, not full games or adaptive adversaries. Agreement on finitely evaluated states does not uniquely identify source code, and some arenas have substantial variance. The public package omits some paper-only plotting and ablation suites; reproduction also needs CodeClash, a container runtime, and model access.
Why it fits this collection
Provides a game-domain benchmark for agent-assisted behavioral reverse engineering and active experimental design.
Publication and review details
- Venue
- arXiv
- Status
- Preprint
- Publication / posting date
- June 24, 2026
- Date source
- arXiv v1 submission history; the non-archival AIWILD workshop acceptance does not create a separate publication date.
- First arXiv submission
- June 24, 2026
- arXiv
- 2606.26094
- Reviewed arXiv version
- v1
- DOI
- 10.48550/arXiv.2606.26094
Accepted to the non-archival Second Workshop on Agents in the Wild (AIWILD) at ICML 2026: its OpenReview record was published on 2026-05-23 and is linked from the ICML page for the July 11, 2026 workshop. The workshop call states a non-archival policy, so this entry keeps the arXiv preprint and first-submission date. The repository and an author’s publication page also report the acceptance. The artifact is official, but its documented omitted experiments are not treated as released.
Sources reviewed
Source review only; experiments were not rerun for this collection. Cite the original authors using the linked paper or BibTeX.