🤖 AI Summary
This study addresses the "cherry-picked coverage omission" (CCO) problem in large language model (LLM) judges, wherein models erroneously output directional verdicts (e.g., SUPPORTS/REFUTES) instead of the appropriate CONFLICTING label when presented with mixed evidence, leading to unsafe commitments. The work formally characterizes CCO for the first time and introduces a task-contract-based diagnostic protocol. To mitigate this issue, the authors propose a dual-channel reference probe coupled with an external commitment control layer that orthogonally leverages evidence structure and model confidence, along with a NO-COMMIT routing mechanism to enhance safety. Experiments reveal that over 84% of three-way LLM judges exhibit CCO on the AVeriTeC conflict subset, and majority voting exacerbates the problem. In contrast, the dual-channel probe significantly improves CONFLICTING recognition (p < 1/2001) and demonstrates consistent effectiveness on VitaminC-Mixed.
📝 Abstract
LLM judges increasingly turn verdicts into system commitments. Under mixed evidence (claims with both supporting and refuting sources) this is unsafe: when the schema exposes CONFLICTING as the authorized non-directional verdict, returning SUPPORTS/REFUTES is an unauthorized directional commitment, a failure we name Cherry-pick Override (CCO). We define CCO under an explicit task contract and report it with a same-denominator diagnostic protocol paired with matched-coverage bootstrap and an apples-to-apples random-veto null. On AVeriTeC's Conflicting subset (N_C = 150), three-option judges return a directional verdict on more than 84% of mixed-evidence claims; under the typed schema, three-judge majority voting amplifies direction-on-conflict on AVeriTeC (0.887 vs. 0.840; 95% CI [+0.013, +0.080]) but does not replicate on VitaminC-Mixed. Walking an intervention ladder of common single-channel fixes (typed vocabulary, panel aggregation, confidence thresholding, validator-only filtering), each leaves a distinct residual failure: panel aggregation suppresses single-judge CONFLICTING dissent in 48% of CCO cases; the panel is well-calibrated for direction (ECE = 0.07 on pure-S/R) so confidence cannot operationally separate CCO from correct directional commits; validator-as-classifier nearly halves pure-evidence accuracy. A minimal two-channel reference probe reaches operating points neither single channel reaches; under the random-veto null its promotion to CONFLICTING is structurally targeted on AVeriTeC (empirical p < 1/2001) and weaker but in the same direction on VitaminC-Mixed, a selectivity result rather than a magnitude one. We argue for an external commitment-control layer that separates verdict generation from commitment authorization, using structural evidence and confidence as orthogonal channels and NO-COMMIT as a routed controller state.