Skip to main content
Back

NUS Discovers AI Models Already Know How They'll Be Judged

A new paper shows base LLMs already know how a judge will score them — and a 160-example training cycle is enough to surface that latent ability

  • SEE (Self-Evaluation Elicitation) is a new method showing base LLMs already predict judge scores before any training, reframing self-evaluation as elicitation, not acquisition.
  • Qwen3-4B-Base prompted few-shot achieves 0.50–0.70 calibration against an external judge across three benchmarks with zero targeted training.
  • The two-phase cycle (Calibration-Coupled RL + Masked Judge Distillation) uses only 160 unique examples — 31x fewer than the RL baseline — while outperforming it on all metrics.
  • Self-evaluation gains transfer to held-out judges (Claude Sonnet 4.6, Gemini 3.1 Flash-Lite), suggesting the model learns a general notion of quality, not one judge's quirks.
  • Practical applications include judge-free inference-time reranking, uncertainty-based routing, and lightweight quality monitoring in production pipelines.
  • Code is available on GitHub; full paper at arXiv:2606.05122.

Every time you use an LLM judge to score a model's outputs, you're paying for an external API call, waiting on latency, and trusting that the judge's preferences are stable. A new paper from researchers at the National University of Singapore and Beijing University of Technology asks a disarmingly simple question: what if the model being judged already knows what score it's going to get?

The answer, according to the paper introducing Self-Evaluation Elicitation (SEE), is that it largely does. Before any specialized training, a base language model prompted with a few examples can already predict an external judge's multi-attribute quality scores well above chance. The research reframes the problem of judge-aligned self-evaluation from one of acquisition (teaching the model something new) to one of elicitation (surfacing what it already knows).

Why this matters now

LLM-as-a-judge now agrees with human reviewers about 85% of the time, and has become the default method for evaluating LLM applications at scale. But the judge is always an external party. A model that predicts how it will be judged can rerank its own samples, defer when it expects a low score, or escalate a difficult prompt to a stronger model , none of which requires querying the judge at inference time.

Prior work on self-evaluation has focused on verifiable tasks like mathematics and coding, where the target is a scalar measure of correctness relative to a known answer. It remained unclear whether a model could predict an external judge's multi-attribute quality scores in the open-ended setting, where no verifiable answer exists. SEE is the first method to tackle this directly.

The latent ability hiding in plain sight

The researchers discovered that the base model already approximates the judge to a substantial degree. Prompted few-shot in the scoring format, Qwen3-4B-Base predicts an external judge's five attributes with a nonlinear calibration of 0.50 to 0.70 across three benchmarks, well above random guessing, despite never having been trained to do so specifically.