aiqbench analysis
StepCOPS introduces a rigorous statistical framework for selecting language model policies by certifying lower-tail performance floors, addressing the risk of mean scores masking rare failures. By combining independent proposal splits with exact binomial tests and Holm's step-down procedure, it ensures high-probability guarantees that selected models meet safety thresholds without being overly conservative. This method represents a significant advancement in post-training pipelines for large language models, offering a more reliable approach to policy selection than existing confidence bound techniques.
StepCOPS: Closed-Testing Lower-Tail Certificates for Language-Model Policy Selection
arXiv:2609.29549v1 Announce Type: new Abstract: Post-training pipelines must select one language-model policy from many checkpoints, prompts, and decoding rules. Mean evaluator scores can conceal rare failures, whereas simultaneous candidate-wise confidence bounds can be unnecessarily conservative. We introduce StepCOPS, which uses an independent proposal split to nominate one lower-tail floor per candidate, exact binomial tests on a fresh certification split, and Holm's step-down procedure to certify a set of floors. With probability at least $1-\delta$, every certified floor, including the largest floor used for policy selection, is below its candidate's population lower $\alpha$-quantile. This guarantee assumes i.i.d. evaluation units while allowing arbitrary within-unit dependence across candidates. Across 24 predeclared configurations and 11 benchmarks, StepCOPS obtains 96.4% selected-policy coverage over 500 paired trials, raises the certified floor by 1.5 points over both proposal-Bonferroni and exact COPS, remains 0.6 points below the large-reference jury oracle, and abstains in 2.4% of trials. Shadow-judge, benchmark-native, artifact, and leave-one-judge-out audits characterize the proxy boundary: the guarantee applies to the fixed jury score, not directly to human safety.
Abstract:Post-training pipelines must select one language-model policy from many checkpoints, prompts, and decoding rules. Mean evaluator scores can conceal rare failures, whereas simultaneous candidate-wise confidence bounds can be unnecessarily conservative. We introduce StepCOPS, which uses an independent proposal split to nominate one lower-tail floor per candidate, exact binomial tests on a fresh certification split, and Holm's step-down procedure to certify a set of floors. With probability at least $1-\delta$, every certified floor, including the largest floor used for policy selection, is below its candidate's population lower $\alpha$-quantile. This guarantee assumes i.i.d. evaluation units while allowing arbitrary within-unit dependence across candidates. Across 24 predeclared configurations and 11 benchmarks, StepCOPS obtains 96.4% selected-policy coverage over 500 paired trials, raises the certified floor by 1.5 points over both proposal-Bonferroni and exact COPS, remains 0.6 points below the large-reference jury oracle, and abstains in 2.4% of trials. Shadow-judge, benchmark-native, artifact, and leave-one-judge-out audits characterize the proxy boundary: the guarantee applies to the fixed jury score, not directly to human safety.
Computation and Language (cs.CL)
arXiv:2609.29549 [cs.CL]
(or arXiv:2609.29549v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2609.29549 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 261 words
