Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.13303

Agreement Without Alignment: Label-Free Rubric Clarification Increases LLM-Judge Consensus but Not Agreement with Humans

Sangdo Han·Moonsoo Kim
View original source ↗Read source document ↗Export metadata ↓

Abstract

LLM judges are increasingly deployed as panels, and disagreement among panel members is often read as a symptom of an unclear rubric. We test a label-free procedure that turns this disagreement into rubric repairs. The procedure decomposes each rubric into sub-criteria, ranks them by cross-judge disagreement on a development split, and has an editor model rewrite only the most contested ones. An additional stage that moves text-checkable sub-criteria to deterministic checkers transferred none in our evaluation. Human labels are never used to select or edit clauses; they are used only to evaluate the outcome. We pre-specified four hypotheses in a version-controlled protocol and evaluated six rubric conditions with a panel of five judges from three providers on SummEval, the HealthBench meta-evaluation, and USR Topical-Chat, for a total of 129,770 judge requests, excluding retries. On SummEval, the procedure raised inter-judge Krippendorff's $α$ by 0.069 (95% CI 0.046 to 0.092) over the original rubric and by 0.074 over a control that applied the same edits to an equal number of randomly chosen clauses that were not selected. Agreement with expert ratings showed no detectable change (Kendall $τ$ difference -0.007, 95% CI -0.037 to 0.021). On HealthBench and USR, neither inter-judge agreement nor human agreement met the preregistered effect criteria. In a preregistered secondary comparison, human agreement after editing the contested clauses was lower than after editing the control clauses on all three datasets in point estimate, although no interval excluded zero. Selecting clauses by disagreement makes judges agree with each other more than editing arbitrary clauses does, but in our setting the added consensus did not bring the judges closer to human annotators. We will release code, frozen rubrics, item identifiers, and judge scores.

Explore related subjects

cs.CLcs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Sangdo Han, Moonsoo Kim. 2026-10-09. Agreement Without Alignment: Label-Free Rubric Clarification Increases LLM-Judge Consensus but Not Agreement with Humans. https://arxiv.org/abs/2610.13303

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

OARelatedWork: A Large-Scale Dataset of Related Work Sections with Full-texts from Open Access Sources

This paper introduces OARelatedWork: a dataset for related work generation from open-access sources. It is the first large-scale multi-document summarization dataset for related work generation, containing whole related work sections and full texts of cited papers. Its validation and test splits are constructed so that every source paper is available in full text, enabling controlled evaluation of full-text related work generation. The dataset includes 94 450 papers from multiple domains. With OARelatedWork, we aim to shift the field from generating parts of related work sections from abstracts only to generating entire related work sections from all available content. We (i) benchmark a wide spectrum of models, highlighting that synthesizing massive full-text contexts remains challenge for tested cost-effective Large Language Models (LLMs): under our statement-level judge, GPT-4o-mini's evidence-grounded True rate drops from 92.9% with abstracts to 83.8% with full texts. We (ii) empirically analyze human writing behavior through a human evaluation over 40 papers and 408 factual statements, revealing that authors frequently introduce abstractive claims ungrounded in localized source texts; consequently, tested LLMs actually surpass human baselines in strict, evidence-grounded factuality. Finally, we (iii) conduct a fine-grained meta-evaluation, revealing that standard reference-based metrics are inadequate for evaluating such long-form structured outputs, and introduce a robust statement-level evaluation framework to address this gap.

cs.CL↗
arXiv2026

Knowing When to Stop: Efficient Context Processing via Latent Sufficiency Signals

Large language models (LLMs) process entire input contexts indiscriminately, which is inefficient when the information required to answer a query is localized within the context. We present dynamic context cutoff, a novel method enabling LLMs to self-terminate processing upon acquiring sufficient task-relevant information. Through analysis of model internals, we discover that specific attention heads inherently encode "sufficiency signals" -- detectable through lightweight classifiers -- that predict when critical information has been processed. This reveals a new efficiency paradigm: models' internal understanding naturally dictates processing needs rather than external compression heuristics. Comprehensive experiments across six QA datasets (up to 40K tokens) with three model families (LLaMA/Qwen/Mistral, 1B-70B) demonstrate 3.4% accuracy improvement while achieving 1.33x token reduction on average. Furthermore, our method demonstrates superior performance compared to other context efficiency methods at equivalent token reduction rates. Additionally, we observe an emergent scaling phenomenon: while smaller models require probing for sufficiency detection, larger models exhibit intrinsic self-assessment capabilities through prompting.

cs.CL↗
arXiv2026

ReasoningShield: Safety Detection over Reasoning Traces of Large Reasoning Models

Large Reasoning Models (LRMs) leverage explicit reasoning traces, known as Chain-of-Thought (CoT) reasoning, to decompose complex problems into intermediate steps before deriving final answers. However, these reasoning traces introduce unique safety challenges: harmful content can be embedded in intermediate steps even when final answers appear benign. Our study reveals that existing state-of-the-art moderation tools experience significant performance degradation on CoT moderation, with F1 scores dropping by up to 35.2% compared to traditional answer moderation. To address these challenges, we present ReasoningShield, a comprehensive benchmark and a suite of strong lightweight models for CoT safety moderation. Our benchmark includes 9.2K annotated (Query, CoT) pairs with structured stepwise risk analysis, consisting of a 7K-sample training set and a 2.2K human-annotated test set, spanning 10 risk categories, 3 safety levels, and 8 LRMs with diverse reasoning paradigms. To further demonstrate the utility of the benchmark, we develop lightweight models with 1B/3B parameters using a two-stage training strategy. Our models achieve 91.8% F1 on our benchmark, substantially outperforming leading tools such as LlamaGuard-4 by 35.6% and commercial models such as GPT-4o by 15.8%, while generalizing effectively across diverse reasoning paradigms and unseen scenarios. Resources are publicly available: Project Page: https://cosmosyi.github.io/ReasoningShield/, Code: https://github.com/CosmosYi/ReasoningShield, Models: https://huggingface.co/ReasoningShield/models, and Dataset: https://huggingface.co/datasets/ReasoningShield/ReasoningShield-Dataset.

cs.CL↗