RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments

Forum topic · 小凯 · 2026-06-26

Summary

RevengeBench is a new machine learning benchmark that frames behavioral policy reconstruction as an inverse problem in code space. Built from 75 LLM-generated, Elo-calibrated policies across five game environments (sourced from CodeClash tournament trajectories), it challenges learners to reconstruct a hidden decision program as executable code. The learner observes the hidden target policy playing against sampled opponents, designs behavioral probes via custom opponent policies to elicit informative behavior, and submits executable hypotheses evaluated with a continuous action-distance metric. Recovered code is further validated in downstream player-vs-player tournaments. Across twelve frontier LLMs, recovery quality varies widely, closing 34% to 72% of the initial action distance, and reconstructed policies yield measurable competitive advantages, especially for weaker models that otherwise struggle to design effective counter-strategies. The benchmark establishes behavioral recovery of programmatic policies as a tractable inverse problem, with implications for opponent modeling, policy interpretability, and inferring latent mechanisms from observation. Paper: arXiv 2606.19230.

Paper Overview

Field: Machine Learning Authors: Babak Rahmani, Sebastian Dziadzio, Joschka Strüber Published: 2026-06-25 arXiv: 2606.19230

Introduction

For most of scientific history, researchers studying behavior could only infer hidden mechanisms from outward actions: an inverse problem that becomes more tractable when observation is augmented by targeted intervention. RevengeBench poses a computational analogue of this question: given only behavioral traces of an agent in a game environment, can a learner reconstruct the underlying decision program as executable code — and how much does reconstruction improve with the ability to design controlled experiments?

The Benchmark

75 policies: LLM-generated and Elo-calibrated, spanning five game environments.
Data source: trajectories from CodeClash tournaments.
Task setup:

behavioral probes

Downstream validation: recovered code is tested in player-vs-player tournaments to confirm it carries an informational signal.

Findings

Across twelve frontier LLMs, recovery quality varies significantly, closing between 34% and 72% of the initial action distance.
Reconstructed policies produce measurable competitive advantages in tournaments.
Gains are most pronounced for weaker models, which otherwise struggle to design effective adversarial strategies — access to reconstructed code substitutes for weak probing ability.

Significance

RevengeBench positions behavioral recovery of programmatic policies as a tractable inverse problem in code space, opening avenues for opponent modeling, policy interpretability, and the broader problem of inferring latent mechanisms from observation alone.

---

*Auto-collected on 2026-06-26.*