Decoding Multi-Turn LLMs: The Game of Extraction vs. Containment

In the nuanced world of multi-turn large language model (LLM) evaluation, reporting a simple win-rate has been the norm. However, the Adversarial Information Deduction Game (AIDG) introduces a rigorous framework that breaks down this evaluation into a two-player partially observable stochastic game (POSG), focusing on two distinct roles: Seeker and Holder. This decomposition shines a light on three critical failure modes that LLMs face.

The Three Modes of Failure

First, there's the issue of cooperative-prior leakage, where models inadvertently reveal more than intended. Then, we encounter constraint-reasoning interference, which disrupts the flow of logical deduction. Lastly, inefficient hypothesis-space traversal hampers the ability of the models to explore possibilities effectively. In a study of 439 games across six new LLMs, these failure modes played a significant role.

Offensive performance varied widely with a sigma of 53.3 ELO, contrasted sharply by the tightly clustered defensive performance at just 1.9 ELO. This emphasizes a glaring gap in the models' offensive strategies. What's more, confirmation framing, a technique that primes responses, boosted extraction success by 7.75 times compared to uninformed deduction. Yet, despite these tactics, constraint violations led to 41.3% of deductive failures, with no correlation to model scale.

Offense vs. Defense: A Surprising Gap?

Color me skeptical, but the narrative of surprise around the containment-over-extraction gap doesn't hold water. It's a predictable outcome when defensive actions are locally resolvable while offensive maneuvers require globally coupled planning. The AIDG framework offers a clear attribution of this gap, revealing it as a consequence of how models are structured and trained.

What they're not telling you: the meticulous design choices, like turn-decay weighting and the Bradley-Terry rating model, aren't just arbitrary selections. They're based on explicit assumptions that shape the evaluation process. This approach allows for more nuanced insights into model capabilities, but it also raises a question about industry standards: Why has it taken so long to adopt such detailed methodologies?

Why This Matters

So, why should we care? The AIDG's structured evaluation offers a lens into understanding LLMs' real-world applicability. In scenarios where both extraction and containment are critical, like legal document analysis or sensitive information retrieval, knowing these strengths and weaknesses is key.

I've seen this pattern before where claims of breakthroughs overshadow the complexities underneath. The difference here's the quantifiable breakdown of performance into its fundamental components. As researchers and developers push the boundaries of LLMs, adopting frameworks like AIDG isn't just beneficial, it's necessary. Without it, we're left with overfitting solutions that tell us very little about genuine capabilities.