Benchmarking at the Edge of Comprehension
Samuele Marro, Jialin Yu, Emanuele La Malfa, Oishi Deb, Jiawei Li, Yibo Yang, Ebey Abraham, Sunando Sengupta, Eric Sommerlade, Michael Wooldridge, Philip Torr
TL;DR
The paper tackles the problem of evaluating frontier LLMs when full human comprehension and ground-truth verification are no longer feasible, a regime they call post-comprehension. It proposes Critique-Resilient Benchmarking (CRB), an adversarial, critique-based evaluation framework where correctness is defined as resistance to verified critiques within a bounded budget, with humans acting as bounded verifiers. The protocol uses an itemized bipartite Bradley-Terry model to jointly rank answerers and question authors, incorporating feasibility gating to filter ill-posed tasks and providing separate scores for answering strength ($eta_b$) and benchmarking strength ($oldsymbol{ ext{ extalpha}}_a$). Experimental validation in mathematics across eight frontier LLMs shows stable rankings, meaningful correlations with external benchmarks, and robustness to adjudication by weaker models, suggesting CRB can meaningfully measure progress where ground-truth is hard to obtain. Overall, the work reframes benchmarking as an adversarial, evidence-based game and offers a scalable path to evaluate AI progress in high-complexity domains.
Abstract
As frontier Large Language Models (LLMs) increasingly saturate new benchmarks shortly after they are published, benchmarking itself is at a juncture: if frontier models keep improving, it will become increasingly hard for humans to generate discriminative tasks, provide accurate ground-truth answers, or evaluate complex solutions. If benchmarking becomes infeasible, our ability to measure any progress in AI is at stake. We refer to this scenario as the post-comprehension regime. In this work, we propose Critique-Resilient Benchmarking, an adversarial framework designed to compare models even when full human understanding is infeasible. Our technique relies on the notion of critique-resilient correctness: an answer is deemed correct if no adversary has convincingly proved otherwise. Unlike standard benchmarking, humans serve as bounded verifiers and focus on localized claims, which preserves evaluation integrity beyond full comprehension of the task. Using an itemized bipartite Bradley-Terry model, we jointly rank LLMs by their ability to solve challenging tasks and to generate difficult yet solvable questions. We showcase the effectiveness of our method in the mathematical domain across eight frontier LLMs, showing that the resulting scores are stable and correlate with external capability measures. Our framework reformulates benchmarking as an adversarial generation-evaluation game in which humans serve as final adjudicators.