RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments

A team of researchers has introduced RevengeBench, a benchmark designed to test whether large language models can reconstruct hidden decision-making code by observing behavior alone, and found that recovery quality varies sharply across frontier models, from 34 to 72 percent of the behavioral gap closed [1]. The work, led by Sebastian Dziadzio and colleagues, frames the task as a computational analogue of a classic scientific inverse problem: inferring latent mechanisms from outward actions, made more tractable through targeted intervention [1]. The benchmark comprises 75 LLM-generated, Elo-calibrated policies drawn from CodeClash tournament trajectories, spanning five distinct game environments [1]. A learner agent observes a hidden target policy playing against sampled opponents, then designs custom opponent policies as behavioral probes to elicit informative responses before submitting an executable hypothesis [1]. Evaluation uses continuous action-distance metrics, measuring how much of the initial distance from a random policy is closed by the reconstructed code [4]. Across twelve frontier LLMs, the strongest model closed 72 percent of that gap, while the weakest managed only 34 percent [1]. The researchers note that active probing helped some models and did not degrade performance for others, suggesting that intervention yields benefits when the learner is capable enough to exploit it [4]. To assess whether the recovered code carries strategic value, the team fed each reconstructed policy back to the same LLM and asked it to write a counter-policy against the original target [4]. The resulting counter-policies produced a measurable competitive advantage, with the ordering blind < recovered < oracle holding consistently across all models [4]. The effect was especially pronounced for weaker models that otherwise failed to design effective counter-strategies on their own [1]. The paper, accepted at the ICML 2026 AIWILD workshop, positions behavioral recovery of programmatic policies as a tractable inverse problem in code-space [3]. The authors argue the benchmark opens a path to opponent modeling, policy interpretability, and the broader question of inferring latent mechanisms from observations [1]. The submission is a regular paper of nine pages, with authors authorizing public release of names and data upon acceptance [3].