LLMs, GenAI, RAG, agents, evaluation · insight-0558
Control for answer length when an LLM judge or reward model scores outputs
- Evidence type
- empirical finding
- Difficulty
- medium
- Roles
- Applied AI, Data scientist
- Industries
- cross-industry
Summary: situation, action, reason, boundary
Situation
A new prompt makes answers 40 percent longer, and the LLM judge's win rate jumps.
Action
Report length next to every judged score, compare at similar lengths or fit a length-adjusted win rate, and add a rubric item that penalises unnecessary content.
Reason
LLM judges and automatic evaluators have been shown to favour longer outputs, and controlling for length made one benchmark harder to game and better aligned with human rankings.
Boundary
LLM-as-judge evaluations, preference data collection, reward-model training.
Symptom or risk
The team optimises for verbosity, customers get longer and slower answers, and quality may not have improved.
Exceptions
When completeness is the goal, longer may be genuinely better; the rubric should say so explicitly.
Trade-offs
Length control adds analysis and can penalise legitimately thorough answers.
Pitfalls
Tuning prompts against the judge until outputs grow without anyone noticing.
Lessons and concepts
Claims and their sources
Zheng et al. document verbosity bias in LLM judges, which favour longer answers.
Dubois et al. show LLM auto-evaluators prefer longer outputs and that a regression-based length control reduces gameability and raises correlation with human-preference rankings.