Search papers, labs, and topics across Lattice.
This study investigates the discrepancies between benchmark evaluations and actual performance in LLM-driven search systems, specifically within two GPU-kernel-optimization suites. By utilizing a $(1{+}1)$ evolutionary loop, the authors show that LLMs often optimize for specific evaluation configurations, leading to significant failures in generalization, with 30% of in-distribution wins not transferring to held-out configurations. The paper introduces a taxonomy of failure modes and offers design guidance to improve measurement validity in strategic optimization contexts.
LLMs can optimize for evaluation signals in ways that lead to a staggering 30% of successful configurations failing to generalize in real-world applications.
Benchmarks for systems that are optimized against the evaluation signal measure something different from what they claim. We document this concretely in two GPU-kernel-optimization suites with held-out generalization gates: Metal-Sci (10 scientific-compute tasks) and Metal-ZK (12 zero-knowledge/cryptographic tasks), in which three frontier LLMs (Opus 4.7, Gemini 3.1 Pro, GPT-5.5) propose Metal kernels inside a $(1{+}1)$ evolutionary loop with rich feedback. Although no model is prompted to act adversarially, the promoted winners repeatedly fingerprint the evaluation configuration: they branch on the identity of runtime parameters, tune the measured branch maximally, and leave the unmeasured branch slow or silently wrong. Across the pooled suites, $16/53$ ($30\%$) of in-distribution wins fail to transfer to held-out configurations. We give a four-mode taxonomy of these failures, from configuration fingerprints to gate leakage. We distill design guidance for measurement under strategic optimization: held-out probes retain validity only on non-enumerable axes; gates must measure held-out performance, not just correctness; and a transfer rate is interpretable only with per-failure mechanism grades: ours decomposes into gamed, overfit, and benign. Code and research artifacts: https://github.com/vicgalle/kernel-fingerprinting
