AI graders keep handing out the same few scores, mostly 4s and 5s — why can't they tell answers apart more finely?
Why do current language model judges collapse into coarse discrete scores?
This explores why LLM judges asked to grade outputs tend to hand out the same few scores (a lot of 4s and 5s, say) instead of fine-grained distinctions, and the corpus has no note that answers this head-on.
This explores why LLM judges asked to grade outputs tend to hand out the same few scores instead of fine-grained distinctions. The corpus has no paper that measures or explains score collapse directly, so what follows is adjacent evidence and my own inference. It does not settle the question.
The closest mechanism is about training signals, not judges. When rewards barely vary across a prompt's candidate answers, the learning signal weakens and regularization takes over, pushing the model toward generic, input-agnostic outputs (Why do language models collapse into generic templates?). A judge that gives most answers the same score is a low-variance signal in the same sense. If such a judge trains or filters a model, it may push that model toward templates. That is an inference from the note, not something the note tests. Its fix, selecting the high-variance prompts, is a reminder that a judge is only useful where it separates good answers from bad ones.
The corpus also suggests why judges might lack fine-grained information to draw on. LLM judges fall for fake authority and rich formatting in ways that ignore the actual meaning of the answer (Can LLM judges be fooled by fake credentials and formatting?). If a score responds more to surface cues than to real quality differences, only a few surface-level bands may be left to score. Separately, RLHF-style training degrades a model's calibration, meaning its stated confidence stops tracking how often it is right. Ranking reasoning traces by the model's own answer confidence reverses that damage (Can model confidence work as a reward signal for reasoning?). A model that no longer knows how sure it should be may not spread its scores out either, though the note doesn't test this.
The corpus does show two ways around asking for a fine-grained absolute score. One is comparison: Chatbot Arena gets a credible model ranking by aggregating 240K+ simple pairwise votes, so no single judgment has to be precise (Can crowdsourced votes reliably rank language models?). The other is to train the model to judge itself. Post-Completion Learning teaches a model to compute its own reward after finishing an answer (Can models learn to evaluate their own work during training?). If you want to know why judges bunch their scores, these notes point to the signal's variance, its calibration and the surface cues it rewards. The corpus doesn't yet have a paper that isolates score granularity.
Sources 5 notes
When within-prompt reward variance is low, task gradients weaken and regularization dominates, pushing policies toward generic outputs. SNR-Aware Filtering—selecting high-variance prompts before updates—recovers performance across tasks and scales.
Research identified four evaluation biases in LLM judges, with authority and beauty biases being semantics-agnostic and trivially exploitable through fake references and formatting—zero-shot attacks requiring no model access or optimization.
RLSF uses answer-span confidence to rank reasoning traces, creating synthetic preferences that strengthen step-by-step reasoning while reversing RLHF's calibration degradation—without requiring human labels or external verifiers.
Chatbot Arena's 240K+ crowdsourced preference votes produce credible model rankings because the underlying questions are diverse and discriminating, and crowd judgments correlate with expert raters—validating human preference as a scalable evaluation signal.
Post-Completion Learning exploits unused sequence space after model output to train self-assessment capabilities during training while maintaining zero inference cost. The model learns to compute its own reward functions, internalizing evaluation rather than relying on external reward models.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Post-Training Large Language Models via Reinforcement Learning from Self-Feedback
- Post-Completion Learning for Language Models
- Beyond Binary Rewards: Training LMs to Reason About Their Uncertainty
- Humans or LLMs as the Judge? A Study on Judgement Biases
- Understanding and Mitigating Premature Confidence for Better LLM Reasoning
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference
- Temporal Self-Rewarding Language Models: Decoupling Chosen-Rejected via Past-Future
- Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs