Neither Blinding Nor a Jury, Neither Capability Nor Strategy: What Brings a Verifier to Reliability
Across 64,800 controlled verifications, a model approves its own wrong answers 11.9 percentage points more often than other models do, and telling it who wrote the answer changes almost nothing. Only execution grounding restores reliability: a grounded judge catches 93.3% of errors against 76.7% for the best ungrounded jury.
My contribution: mathematics domain; prompt-parity test that keeps verifier strategy a clean independent variable.
Which Verifier Catches What? Mapping Error Severity to Verifier Strategy in Multi-Agent LLM Pipelines
Controlled fault injection into a plan/solve/verify pipeline. Removing the verifier raises execution-stage error propagation to 27.3%; chain-of-thought and rubric verifiers repair 95 to 98% of the faults they catch, against 78.5% for direct prompting.
My contribution: First author. Design, experiments, analysis.