liveBook · Manning

Self-enhancement bias

As explained in LLM Evaluation and Alignment, The Foundational Ideas

Self-enhancement bias is a tendency for a language-model judge to favor outputs generated by itself or by models from the same model family, including when those outputs are not objectively better.

Self-enhancement bias is a language-model judge’s tendency to favor answers produced by itself or by models from the same family, even when those answers are not objectively better. section 4.3.3

Diagnose the bias through controlled comparisons:

  • GPT-4 judges favored GPT-4 responses by about 10% more than human evaluators, while Claude-v1 favored Claude responses by about 25%; GPT-3.5 showed no significant self-preference. section 4.3.3
  • Self-preference remains detectable after controlling for response quality, but its magnitude varies by task type and model family. section 4.3.3
  • Similar training data, reinforcement-learning procedures, diction, sentence structures, and reasoning styles may make familiar outputs appear better. section 4.3.3

Evaluation and RLHF risk

When GPT-4 compares a GPT-4 response with a Claude response, it may rate its own model family more favorably. In an RLHF pipeline, repeated preference for the judge’s style can optimize stylistic mimicry instead of genuine quality improvement. section 4.3.3 section 4.3.5

Figure 4.8 Self-enhancement bias varies across model families. Some judges (GPT-4, Claude) show elevated win rates for their own outputs, while others (GPT-3.5) do not exhibit significant self-preference.

Compare self-preference across GPT-4, Claude, and GPT-3.5 model families. section 4.3.3

Figure 4.9 Self-enhancement bias can create feedback loops in RLHF pipelines. If the judge prefers its own style, training may optimize for stylistic mimicry rather than genuine quality improvement.

Trace how judge preferences can create RLHF feedback loops that reward stylistic similarity. section 4.3.3

How LLM Evaluation and Alignment, The Foundational Ideas covers it

How deep it goes
Explains definition, comparative evidence, stylistic mechanism, caveats, and ensemble mitigation, but does not provide a measurement procedure.
Examples
Uses GPT-4, Claude-v1, and GPT-3.5 win-rate comparisons plus an RLHF feedback-loop scenario.
What it assumes
Assumes familiarity with LLM-as-a-judge evaluation, model comparisons, and human preference judgments.
What you can do after
Diagnose self-preference and select diverse judge ensembles instead of treating one model’s judgment as ground truth.
How it connects
Connects model evaluation with human preferences, model families, stylistic familiarity, reinforcement learning from human feedback, and judge ensembles.

Also explained in

Compare how 2 books explain self-enhancement bias →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of LLM Evaluation and Alignment, The Foundational Ideas. The section links point to the parts of the book it draws on.