Verbosity & Self-Enhancement Bias

Verbosity & Self-Enhancement Bias Mitigation addresses two pervasive cognitive distortions in LLM judges: 1) Verbosity Bias (the tendency for judge LLMs to favor lengthy, overly verbose, formatting-heavy responses over concise accurate answers by a 30%+ margin), resolved via length-penalty normalization or informational density rubrics; 2) Self-Enhancement Bias (the systemic tendency of a model family like GPT or Claude to assign higher scores to its own outputs), mitigated by deploying neutral third-party cross-vendor judges (e.g., Claude judging GPT, GPT judging Llama) or ensemble juries.

Objective cross-vendor benchmarking, preventing RLHF/DPO alignment from reward-hacking via fluff generation, and factual density audits.

Without mitigation, models optimize for length gaming in RLHF, bloating user responses with verbose boilerplate; debiasing anchors evaluations strictly to information density and substantive reasoning.

Derive the logistic regression calibration model with length covariates to mathematically normalize Verbosity Bias?

Design prompt strategies that penalize verbosity by explicitly prioritizing factual density over length?

Quantify the empirical severity of Self-Enhancement bias when models judge outputs from their own architectural lineage?

Explain the application of Cross-Judge Evaluation Matrices where competing models cross-evaluate peers to reach equilibrium?

How does style normalization and verbal quirk removal prevent judge models from recognizing their own stylistic fingerprints?