AgentGAN
An adversarial multi-agent framework drawing on the ideas of generative adversarial networks, in which generator agents rewrite problem statements to remove memorization cues and the discriminator consists of frozen code models, used to assess a model's genuine reasoning ability.
Timeline (last 90 days)
AgentGAN is an adversarial multi-agent framework for automatically rewriting problem statements to remove memorization cues, in which the generator agent repeatedly rewrites the problem until a set of frozen code models collectively answer incorrectly, while executable legality gating ensures the answer and complexity class remain unchanged.
Adversarial hardening narrows the class gap between NPC and P twins from 24.8 percentage points to 14.6 percentage points, suggesting that part of the surface complexity gap stems from models' familiarity with textbook wording.
All Facts (2)
AgentGAN is an adversarial multi-agent framework for automatically rewriting problem statements to remove memorization cues, in which the generator agent repeatedly rewrites the problem until a set of frozen code models collectively answer incorrectly, while executable legality gating ensures the answer and complexity class remain unchanged.
50%UnverifiedAdversarial hardening narrows the class gap between NPC and P twins from 24.8 percentage points to 14.6 percentage points, suggesting that part of the surface complexity gap stems from models' familiarity with textbook wording.
50%