Post-hoc Alignment of LLM-judges to Human Judgment Distribution
Sebastian Steindl, Nikos Voskarides, Alberto Gasparin, Diego Marcheggiani
cs.CL
Quick verdict
Claims LLM-judges match humans on hard labels but not soft; NAPHA entropy-routed post-hoc alignment helps, but gains hinge on imperfect entropy routing.
Summary
The paper asks whether LLM-as-a-judge systems can predict not just aggregated hard labels but the full distribution of human annotations (Human Judgment Distributions, HJD). Across five datasets (SummEval, TopicalChat, ChaosNLI, DynaSent round 2, Anecdotes), the authors find that LLM judges — mainly Claude-4-Sonnet, with GPT-OSS-120B and Qwen3-32B in appendices — approach human-level hard-label accuracy (e.g., ChaosNLI macro-F1 0.99/0.81/0.61 by entropy stratum vs human 0.93/0.74/0.53) but predict soft-labels poorly. They propose NAPHA: compute the entropy of predicted soft-labels, bin instances into low/medium/high tercile classes, and route each to a class-specific one-hidden-layer MLP trained with KL loss on a 20% data split. NAPHA consistently reduces DistCE/JSD across three base models and five datasets — e.g., Anecdotes SLP-SE DistCE 0.299→0.272 with predicted routing, 0.172 with oracle routing. Gains concentrate on high-entropy instances; low-entropy alignment sometimes degrades. Oracle experiments show far larger gains if entropy routing were accurate, but entropy-class prediction F1 is only 0.35–0.72. Limitations: sparse soft targets (3–5 annotators on three datasets), the assumption that all disagreement is valid, routing that conflates model uncertainty with human disagreement, no significance testing, and no stated code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- LLM judges achieve near human-level hard-label accuracy on most of five datasets but predict the full human judgment distribution (soft-labels) poorly.
- NAPHA, an entropy-class-routed post-hoc alignment method, consistently improves soft-label prediction across three base models and five datasets, with the largest gains on high-entropy instances.
- Oracle entropy routing would substantially increase NAPHA's effectiveness, making entropy-class prediction the key practical bottleneck.
- A per-class lightweight MLP matches or beats standard post-hoc calibration baselines (temperature scaling, Dirichlet calibration, linear/parameterized temperature scaling).
Strengths
Systematic hard-vs-soft-label evaluation with sound human baselines
Five HLV datasets, three base models, three backbones (Section 5, Tables 1–3); human hard-label performance estimated via leave-one-out bootstrapping (Appendix G), correctly avoiding the self-comparison bias that inflates human agreement with only 3 annotators.
Honest oracle/predicted decomposition of NAPHA's gains
Tables 3 and 4 separate predicted routing (Anecdotes SLP-SE DistCE 0.299→0.272) from oracle routing (0.172), explicitly exposing routing quality — not the aligner — as the bottleneck, corroborated by entropy-class F1 of only 0.35–0.72 (Table 10).
Lightweight, deployment-realistic method
Useful ablations including alternatives that tie the MLP
Weaknesses
No statistical significance testing for the central consistency claim
Predicted-routing gains are modest and low-entropy alignment degrades
Post-alignment scores are nearly identical across base models, suggesting dataset-prior learning
Slightly overstated human-parity claim
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Preliminary Review: Post-hoc Alignment of LLM-judges to Human Judgment Distribution
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is NAPHA and how does it work?
NAPHA (eNtropy-Aware Post-Hoc Alignment) is a lightweight post-processing method for LLM judges: it computes the Shannon entropy of the model's predicted soft-labels, assigns each instance to a low/medium/high entropy class using terciles from the training data, and routes it to a small per-class MLP trained with KL-divergence loss to map predictions toward the human annotation distribution. It needs only output tokens, so it works with closed-weight APIs like Claude.
Are LLM judges as accurate as human annotators?
According to this paper, on aggregated hard labels LLM judges reach near-human agreement on most of the five studied datasets — for example, on ChaosNLI they beat the average annotator (macro-F1 0.99 vs 0.93 at low entropy) — but they are substantially worse at predicting the full distribution of human judgments (soft-labels), which human label variation research argues is the more informative target.
How much training data does NAPHA need?
The paper's ablation (Table 13) shows performance stabilizes at 10% of the training split, so only a few hundred human-labeled instances per dataset are needed to train the alignment models — the authors use a 20/80 train/test split throughout.
Does NAPHA work with closed-source models, or does it need model internals?
It works with closed-source models: NAPHA operates only on the model's text outputs (predicted label distributions), requiring no logits or internal states. The main experiments use Claude-4-Sonnet via its API, with open-weight GPT-OSS-120B and Qwen3-32B confirming the same conclusions.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 3, 4, 7 and 8 are images; only their captions were extracted, so plotted values (entropy correlations, prompt screenshots) could not be verified against the text. All tables referenced in the review were present in the extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.