cs.IRNov 18, 2026

Can Programs Replace LLM Judges? Executable Program Judges for Scalable and Efficient Relevance Labeling

Authors: Eugene Agichtein, Charles L. A. Clarke, Kaustubh Dhole, William Stern

Organizations: Emory University, Atlanta, GA, USA · University of Waterloo, Waterloo, ON, Canada · Stanford University, Stanford, CA, USA

Abstract

Large language model (LLM) judges have become a common way to create large-scale relevance judgments for retrieval evaluation. However, each judgment requires a potentially expensive call to a frontier model, and the judgments are hard to interpret and may not be reproducible. In this paper, we ask whether an LLM can write a readable program for each topic that replaces its per-document judgments. To address these issues, we propose ExecJudges, a method for generating fast, inspectable and scalable judges for automatic relevance labeling. Specifically, ExecJudges uses a frontier LLM to turn assessment guidelines into executable programs, which rely only on lexical matching, standard NLP libraries like parsers, and pool statistics. An ExecJudges-L variant also uses examples of human labels if available. We evaluate ExecJudges on seven TREC collections by comparing the produced labels with trained NIST assessors, frontier LLM judges, and recently proposed per-topic classifiers. Our results show that ExecJudges agrees with NIST assessors at the level of the Claude Sonnet 5.5 LLM judges on TREC DL 2019-2021. With a small number of human labels per topic, the ExecJudges-L variant agrees with NIST more than per-topic classifiers with the same labels on all seven collections, and either exceeds the agreement of LLM judges provided the same labels or is within about 0.05 of it in Krippendorff's alpha. The labels rank the official TREC DL runs with Kendall's tau of 0.87 to 0.94, and writing the programs incurs a one-time cost of 0.17to0.17 to 0.49 per topic, with no additional costs for grading. Our results suggest that ExecJudges offers a reproducible, inspectable, efficient and cost-effective alternative to LLM judges for labeling deep result pools for system evaluation or ranking improvements.

Figures & tables

Explore similar work

Feb 6, 2026

FairJudge: An Adaptive, Debiased, and Consistent LLM-as-a-Judge

Existing LLM-as-a-Judge systems suffer from three fundamental limitations: limited adaptivity to task- and domain-specific evaluation criteria, systematic biases driven by non-semantic cues such as position, length, format, and model provenance, and evaluation inconsistency that leads to contradictory judgments across different evaluation modes (e.g., pointwise versus pairwise). To address these issues, we propose FairJudge, an adaptive, debiased, and consistent LLM-as-a-Judge. Unlike prior approaches that treat the judge as a static evaluator, FairJudge models judging behavior itself as a learnable and regularized policy. From a data-centric perspective, we construct a high-information-density judging dataset that explicitly injects supervision signals aligned with evaluation behavior. Building on this dataset, we adopt a curriculum-style SFT-DPO-GRPO training paradigm that progressively aligns rubric adherence, bias mitigation, and cross-mode consistency, while avoiding catastrophic forgetting. Experimental results on multiple internal and public benchmarks show that FairJudge consistently improves agreement and F1, reduces non-semantic biases, and outperforms substantially larger instruction-tuned LLMs. All resources will be publicly released after acceptance to facilitate future research.
Sep 26, 2026

Opening LLM Judges: Recovering Preference Signals Beyond the Final Verdict

LLM judges are widely used to evaluate model outputs, but their verdicts can be unreliable: a judge may favor the worse answer for its position, length, or other surface features. When a judge is wrong, is the information needed to judge correctly absent from the model, or present in its internal representations but not reflected in the output? We study this across 64 open-weight evaluators and 14 datasets, including causal interventions on 41 judges (editing activations mid-run to see whether the verdict changes). On LLMBar, built so the superficially better answer is the worse one, the verdicts of 50 judges agree with human labels only 0.456 of the time, even after averaging both answer orders. Yet a small probe on the same judges' activations, with no weight updates, reaches 0.846, and 0.686 once surface features such as length and position are residualized out (0.507 with shuffled labels). The gap holds across eight benchmarks and model families, but is not universal: a score of how well surface features alone predict the human label, computed before any probe is trained, predicts the size of the gain (Spearman rho = 0.90). On rubric tasks that score one answer at a time, leaving no surface cue to exploit, reading the internals gives no advantage. The interventions also show that editing activations mid-network already changes the verdict, before it can be read off directly, and locate the pathways carrying position and length bias. At the same label budget, the recovered signal lets a judge flag cases where it is likely wrong and yields better labels for preference learning. A wrong verdict, then, does not mean the judge lacks the information, and a simple diagnostic shows when it is worth recovering.
Sep 24, 2026

Decision Models vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places

LLM judges now score outputs against rubrics, in benchmarks and as training rewards. Decision models return probabilities over permitted answers with a confidence score but no free text. Two such models are Jev and an open reproduction, Perplexity's Decider. We compare each with three flash-tier LLM judges on nine panels from seven benchmarks with human judgments, giving every judge identical criterion texts. The LLM judges run in two setups, reading a whole rubric at once or one criterion at a time. Our findings suggest both decision models can often stand in for them. The LLM judges cost 16 to 357 times as much and take 5 to 350 times as long, yet in both setups each decision model trails them significantly in at most 6 of 27 comparisons. The two kinds of judge err alike, agreeing more with one another than with the human-annotated judgment labels on ordinal criteria. Over 85% of flash-tier LLM verdicts on either decision model's most confident errors repeat them, where independent errors would give about half. Each decision model's confidence ranks its errors on most ordinal panels, so each is a natural first stage for a cascade deferring uncertain verdicts to an LLM judge. Yet with flash-tier fallbacks, such cascades gained a mere 2.7 points at best over the strongest single judge, even with oracle thresholds, suggesting they mainly save cost. A cascade needs a fallback that is right where its first stage errs, but the LLM judges are mostly wrong in the same places as Jev, and the Decider is already the strongest single judge on most panels. Even each vendor's higher-tier model repeats those errors, and as a fallback it lifted no cascade more than 1.7 points above both tiers' strongest single judge. Holding across setups, effort levels, and tiers, these findings suggest designing future decision models afresh to get right what LLM judges get wrong, and vice versa.