acceptodds
Under review as a conference paper at ICLR 2027

TRUST: LLM Evaluation with Imperfect Judges via Reliability Transfer under Sparse Target Supervision

Abstract

Modern large-scale LLM evaluation increasingly relies on imperfect LLM judges, yet reliable estimation of model performance remains difficult when human annotations are available for only a small fraction of target examples. Current approaches either treat the judge as a black-box predictor or model the judge's error behavior, but the former suffers from bias whereas the latter is ineffective in low-supervision regimes. We address these limitations by studying how learnt judge behavior on auxiliary benchmarks can be transferred to a target benchmark to improve LLM evaluations. Concretely, we introduce TRUST, a two-stage framework that 1) aligns source and target examples in semantic-confidence space and filters poorly matched source samples, and 2) transfers judge behavior in each semantic-confidence region through an adaptively weighted likelihood to aid model performance evaluation. Our analysis characterizes how judge reliability transfer improves model evaluation, showing it depends on the tradeoff between variance reduction from additional auxiliary benchmarks supervision and bias from residual source benchmarks–target benchmark judge behavior mismatch. Across diverse LLM evaluation benchmarks, TRUST substantially reduces LLM failure-rate estimation error by up to approximately 20% relative to target-only and prediction-powered inference (PPI) baselines, with the largest gains experienced when target human supervision is extremely limited.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Maximum likelihood estimation, LLM-as-a-Judge, LLM certification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AeKJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TRUST, a framework for improving the estimation of target LLM failure rates by transferring judge reliability information from auxiliary benchmarks to a target benchmark. The framework leverages semantic clustering, confidence stratification, and target-guided filtering to identify local regions where judge behavior can be reliably transferred. The authors also provide a theoretical analysis and validate the proposed method across multiple benchmarks.

Critical strengths
  1. The paper addresses an important problem, as accurately estimating LLM error rates is challenging given the high cost of human annotation.

  2. The idea of transferring TPR/FPR rather than directly transferring error rates is novel.

Critical weaknesses
  1. The proposed method focuses on binary correct/incorrect labels. However, LLM evaluation can also involve multi-level scoring tasks. It is unclear how the proposed method and its corresponding theoretical analysis could be extended to these broader settings.

  2. The proposed pipeline is relatively complex, yet some of its components may not be essential. For example, Figure 12 shows that the performance is relatively insensitive to the filtering threshold.

  3. The paper analyzes the mean squared error (MSE), which can be decomposed into squared bias and variance. A separate analysis of these two components would provide additional insights into the proposed method.

Questions for authors

See Weaknesses above.

Reviewer_xmrVOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the estimation of an LLM’s failure rate using limited human-labeled target data, abundant judge-labeled target data, and labeled auxiliary benchmarks. The proposed framework, TRUST, aligns source and target examples through semantic–confidence partitioning and source filtering, then incorporates source judge reliability into an adaptively weighted likelihood. It transfers local true-positive and false-positive rates rather than source model failure rates. The theoretical analysis characterizes the tradeoff between variance reduction and bias from reliability mismatch. Experiments across four benchmarks and two judges, together with component ablations, report improved failure-rate estimation in sparse-target-supervision settings.

Critical strengths

A meaningful evaluation problem with a concrete methodological contribution. Reliable aggregate performance estimation under limited human supervision is practically important. TRUST provides a structured way to reuse auxiliary annotations by transferring judge reliability while retaining target-specific failure-rate estimation. Its bias–variance analysis helps clarify when such transfer can be beneficial.

Substantial empirical support for the proposed approach. The evaluation includes multiple target-only estimators, within- and cross-task transfer settings, and two judges. Ablations examine partitioning, filtering, adaptive weighting, and a simple source-pooling alternative. Together, these provide useful evidence that the gains depend on how source information is incorporated, rather than merely on access to additional annotations.

Critical weaknesses

The evaluation does not isolate a potentially important failure mode: source and target examples may be semantically similar yet judged under different rubrics. In this setting, semantic similarity and judge confidence need not imply similar conditional error rates, so the alignment and filtering stages may retain misleading source information. Adaptive weighting offers a safeguard, but it relies on the sparse target labels whose insufficiency motivates the method. The Jigsaw–HSO transfers provide partial evidence under differing criteria, but the aggregate results do not establish how reliably the method detects and handles rubric-induced mismatch. This leaves some uncertainty about its applicability when evaluation criteria change.

Questions for authors

How does TRUST distinguish transferable judge reliability from misleading similarity when source and target examples have similar content but different evaluation rubrics? In particular, is the judge’s instruction or rubric explicitly represented in the semantic alignment, or is alignment based only on the evaluated prompt–response pair?

Could you provide a controlled analysis that holds the underlying examples fixed or closely matched while varying the source and target judging criteria, with ground-truth labels consistent with each criterion? Reporting the resulting TPR/FPR mismatch, transfer weights, and estimation error relative to target-only methods would help establish whether sparse target supervision can reliably suppress harmful transfer. An analysis of rubric-disagreement cases within the existing Jigsaw–HSO experiments would also be informative.

Reviewer_Khz8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses a problem that when the target data has very few labels, how to better estimate the performance of a model on the target data. The proposed method using semantic-based clustering and confidence stratification to align source and target samples, then transfers the judge's reliability from source data to target data for more reliable performance estimation.

Critical strengths

The paper is well-motivated, the addressed problem has practical meaning. The idea of using semantic similarity to align the target and test data is intuitive and reasonable.

The paper provided comprehensive math analysis.

Critical weaknesses

Limited experiment setting. The experiments are limited to multi-choice and classification tasks, and the judge model is asked to judge the correctness of the target model based solely on its own understanding. It is similar to measure the agreement rate between the judge model and target model, and it’s unclear whether the method generalizes to more practical open-ended evaluation tasks. For example, how well the does clustering method work when the response are longer.

The target model and judge models used in experiments are restricted to 7b-8b models and both Qwen models. It’s unclear how well the model works with models from different model family and stronger judge models.

Questions for authors

How does the method perform in a standard reward bench scenario, where the answer is open-ended?

2 comments

Please indicate whether your concern has been addressed, or whether there are remaining concerns that should be discussed.

Dear Reviewer Khz8,

As the author–reviewer discussion period closes today, we wanted to briefly check whether our response, including the additional results, has addressed your concerns. We would be happy to clarify anything further.

Best regards,


Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.