Invited talk: Generative Adversarial Harnesses -- Kangwook Lee

Meta-harness optimization treats the executable setup around an LLM (prompts, tools, retrieval, memory, state updates, and decision procedures) as an object of optimization, rather than a manually designed component. It is effective when task success can be measured directly, such as through unit tests, exact answers, or executable checks. Many open-ended LLM tasks, however, lack such metrics: essay quality, research synthesis, and long-form argumentation are difficult to score directly. We propose Generative Adversarial Harnesses, a framework for optimizing LLM harnesses in settings where direct evaluation is unavailable but high-quality reference outputs exist. Our framework casts harness optimization as a game between executable LLM agents: a discriminator harness that learns to distinguish expert outputs from generator outputs, and a generator harness that is optimized to produce outputs matching the reference distribution. The contribution is to lift adversarial optimization from model weights or prompts to full executable LLM harnesses, making meta-harness optimization applicable to open-ended domains where correctness is hard to measure but strong examples are available.

This work is joint with Beomhan Baek, Junsoo Oh, Seonho Lee, and Junhyuck Kim.