Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data

Hugging Face Daily Papers Papers
self-evaluation calibration reinforcement-learning distillation llm model-evaluation elicitation

Summary

This paper introduces Self-Evaluation Elicitation (SEE), which uses calibration-coupled reinforcement learning and masked distillation to elicit latent judge calibration in base LLMs with minimal data, improving calibration across benchmarks while preserving answer quality.

Large language models are increasingly evaluated by other models, raising a natural question: can a model predict how a judge will score its own output? We find that the ability is largely present before any targeted training: prompted few-shot, a base model already predicts an external judge's multi-attribute quality scores on open-ended responses well above chance across three benchmarks. We introduce Self-Evaluation Elicitation (SEE), a method that surfaces this latent ability through a short cycle comprising a calibration-coupled reinforcement learning phase that improves the answer and predicts the judge, followed by a masked distillation phase that sharpens the prediction while leaving the answer untouched. From 160 unique examples, roughly 31x fewer than a reinforcement learning baseline, SEE improves held-out calibration across three benchmarks while preserving answer quality. The elicited self-evaluation is sharply localized within the model's own token distribution and stable across judges it was never trained against, indicating a transferable notion of quality rather than a single judge's preference. These results reframe judge-aligned self-evaluation as a problem of elicitation rather than acquisition.
Original Article
View Cached Full Text

Cached at: 06/09/26, 08:41 AM

Paper page - Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data

Source: https://huggingface.co/papers/2606.05122

Abstract

Self-Evaluation Elicitation (SEE) method improves model calibration for quality assessment through calibration-coupled reinforcement learning and masked distillation, demonstrating transferable quality evaluation beyond specific judge preferences.

Large language models are increasingly evaluated by other models, raising a natural question: can a model predict how a judge will score its own output? We find that the ability is largely present before any targeted training: prompted few-shot, a base model already predicts an external judge’smulti-attribute quality scoreson open-ended responses well above chance across three benchmarks. We introduceSelf-EvaluationElicitation (SEE), a method that surfaces this latent ability through a short cycle comprising acalibration-coupledreinforcement learningphase that improves the answer and predicts the judge, followed by amasked distillationphase that sharpens the prediction while leaving the answer untouched. From 160 unique examples, roughly 31x fewer than areinforcement learningbaseline, SEE improves held-outcalibrationacross three benchmarks while preserving answer quality. The elicitedself-evaluationis sharply localized within the model’s owntoken distributionand stable across judges it was never trained against, indicating a transferable notion of quality rather than a single judge’s preference. These results reframe judge-alignedself-evaluationas a problem of elicitation rather than acquisition.

View arXiv pageView PDFGitHub1Add to collection

Get this paper in your agent:

hf papers read 2606\.05122

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.05122 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.05122 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.05122 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Calibrating LLMs with Semantic-level Reward

arXiv cs.CL

Proposes CSR, a framework that calibrates LLMs directly in semantic space using a novel semantic calibration reward, reducing ECE by up to 40% and improving AUROC by up to 31% over verbalized-confidence baselines across multiple datasets.

Confidence Calibration in Large Language Models

arXiv cs.AI

This paper analyzes the confidence calibration of 11 popular LLMs, finding that they are generally overconfident, especially on hard tasks, and underconfident on easy tasks. It introduces LifeEval, a test for evaluating calibration across difficulty levels.

Generating and Refining Dynamic Evaluation Rubrics for LLM-as-a-Judge

arXiv cs.CL

This paper proposes a training-free method to automatically generate fine-grained evaluation rubrics for LLM-as-a-judge without human annotation, and further introduces an iterative fine-tuning strategy for a rubric generator that outperforms larger proprietary models.

Retrieval-Augmented Linguistic Calibration

arXiv cs.CL

This paper proposes Retrieval-Augmented Linguistic Calibration (RALC), a post-hoc pipeline for calibrating confidence signals in LLMs by modeling linguistic confidence as a distribution and using retrieval-augmented rewriting. It introduces Faithfulness Divergence metric and shows significant improvements across benchmarks.