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 recovery as an inverse problem in code space. Introduced by Babak Rahmani, Sebastian Dziadzio, and Joschka Strüber (arXiv:2606.19230), the benchmark contains 75 LLM-generated, Elo-calibrated policies across five game environments, drawn from CodeClash tournament trajectories. A learner observes a hidden target policy playing against sampled opponents, designs behavioral probes as custom opponent policies to elicit informative behavior, and then submits an executable hypothesis, scored by a continuous action-distance metric. Recovered code is further validated in downstream player-versus-player tournaments. Across twelve frontier LLMs, reconstruction quality varies widely, closing 34% to 72% of the initial action distance, and recovered policies yield measurable competitive advantages, especially for weaker models that otherwise struggle to design effective counter-strategies. The work positions behavioral recovery of programmatic policies as a tractable inverse problem, with implications for opponent modeling and policy interpretability.

Overview

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

Abstract

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. We pose a computational analogue: 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 this reconstruction improve with the ability to design controlled experiments?

We introduce RevengeBench, a benchmark of 75 LLM-generated, Elo-calibrated policies across five game environments, drawn from CodeClash tournament trajectories. The learner observes the hidden target policy play against sampled opponents and designs behavioral probes in the form of custom opponent policies that elicit informative behavior. The learner then submits an executable hypothesis, evaluated with a continuous action-distance metric. Recovered code is further validated to carry an informative signal in downstream player-versus-player tournaments.

Key Findings

Among twelve frontier LLMs, reconstruction quality varies widely: 34% to 72% of the initial action distance is closed.
Reconstructed policies yield measurable competitive advantages in tournaments, particularly for weaker models that otherwise struggle to design effective counter-strategies.

Significance

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

--- *Collected automatically on 2026-06-26*