Source-linked AI summary
Guiding LLM Peer Reviewers: The Impact of Score Anchors on Review Evidence and Accuracy
Judita Preiss, Yunhan Yang
TL;DR
The paper examines whether external score guidance changes only LLM evaluation scores or also the evidence and framing of review rationales. Using paired no-guidance and oracle-guided reviews from 98 internal REF-style outputs, it finds improved scoring accuracy without simple score copying, while rationale shifts and evidence coverage are uneven by direction.
Problem
It is unclear whether external score guidance changes the evidence in LLM review rationales as well as their final scores.
Method
The study compares no-guidance and oracle-guided reviews of 98 outputs, testing score following and analysing review frames and evaluation-point coverage against human references.
Results
Oracle guidance improves scoring accuracy without simple score copying; corrected mismatches coincide with review-frame changes, while human upgrade points receive more consistent coverage than downgrade points.
Takeaways & Limitations
Score-guided review generation should be evaluated by changes in rationale evidence and framing, not only by final-score correctness.
Takeaways & Limitations
The small, domain-specific, self-selected dataset and oracle guidance limit generalisation beyond REF-candidate-like outputs and known reference scores.
Abstract
from arXiv · showhide
Large language models (LLMs) are increasingly used for research quality evaluation, with prior work exploring their scoring accuracy and the plausibility of review rationales. However, less is known about whether external score guidance changes the evidence presented in the generated review as well as the final score. This study uses 98 Allied Health Professions research outputs submitted for internal REF-style assessment, with specialist human review reports and adjudicated 1-4 reference scores. No-guidance baseline reviews are compared with oracle-guided reviews, where the supplied score is set to the rounded human reference score; extracted evaluation points are used to compare human and LLM evidence use. Using this design, oracle guidance improves scoring accuracy, with score-following checks showing that models do not simply copy the supplied score. Corrected score mismatches are associated with changes in the generated review frame, showing that the score signal can steer review rationales. This effect is direction-dependent: LLM reviews cover human strength or upgrade points more reliably than human weakness or downgrade points, with the weakest alignment for expert downgrade evidence. The results show that score-guided review generation can be evaluated at the level of review evidence, as well as the final score.
1 Introduction
The paper asks whether external score guidance changes only LLM review scores or also the evidence and framing used to justify them. It studies this question in an internal REF-style evaluation and introduces analyses linking score correction to rationale changes.
- Prior correlations between LLM scores and expert quality proxies do not establish accurate scoring on the target evaluation scale.
- The study evaluates whether an external score signal changes both the final score and the evidence in the accompanying rationale.
- The dataset contains 98 anonymised Allied Health Professions outputs with private specialist reviews and rounded human reference scores on a 1–4 scale.
- A full guide-score sweep tests direct score-following, while oracle guidance supplies the human reference score for comparison with no guidance.
- The paper proposes protocols for separating score copying from substantive rationale changes and for comparing human and LLM score-supporting evidence.
- Successful score correction is associated with directional frame shifts: stronger framing for underestimates and reduced novelty or world-leading framing for overestimates.
2 Methods
The methods analyse LLM outputs at both score and review-content levels. They compare guidance conditions against human reference scores and annotate review framing and evaluation-point coverage.
- The analysis examines both assigned scores and generated review content.
- Score-level evaluation measures direct score-following, exact-match accuracy, and mean absolute error against human reference scores.
- Scores below, equal to, or above the human reference are labelled underestimates, correct matches, or overestimates.
- Review-content analysis pairs no-guidance and oracle-guided reviews for the same paper and generation setting, grouping pairs by score transition.
- Sentence-level labels estimate broad evaluative framing, while point-level analysis identifies claims that support, lower, or qualify a score.
- Human-point coverage records whether each human-review point is covered, partially covered, or absent in the corresponding LLM review.
3 Experiments
Experiments tested oracle score guidance across multiple LLM, prompt, input, and generation settings, then examined score mismatches, evidence themes, review-frame shifts, and human-point coverage. Guidance improved score accuracy without simple copying, but its effects on review content were uneven across strength and weakness evidence.
- Experimental setup: The experiments used 98 anonymised outputs with three paper-input variants and multiple LLM, prompt, hyperparameter, and guide-score settings.The primary text-level analysis used Llama-3.1-8B prompt v4 with full-text input and three generation settings.
- Score-level results: Oracle guidance improved exact-match accuracy across all selected full-text configurations, while MAE decreased in every setting.The guided scores were therefore closer to human reference scores even when they were not exact matches.
- Score-level results: 31.89% of final LLM scores exactly matched supplied guide scores, indicating guidance acted as an anchor rather than a deterministic output constraint.The supplied and final scores correlated at 0.32, and no paper-configuration pair copied all four guide-score values.
- Mismatch analysis: Guidance corrected 31 of 68 baseline underestimates and 38 of 82 baseline overestimates, while many mismatches persisted in the same direction.The mismatch groups motivated paired text-level comparisons of corrected and persistent underestimates and overestimates.
- Evidence frames: Human and LLM reviews differed in score-raising and score-lowering evidence themes, especially in international significance, practical-impact limits, generalisability, methodology, and evidence scale.Human higher-score points gave more emphasis to international/global significance, while human downgrade evidence more often included practical-impact and international-significance limits.
- Review-frame shifts: Corrected underestimates shifted toward stronger positive framing, whereas corrected overestimates reduced novelty/world-leading framing and slightly increased methodological-weakness framing.Persistent mismatches showed weaker or opposite movement, and human upgrade points were covered more consistently than downgrade points across robustness checks.
4 Discussion
Score guidance can steer both final scores and review rationales rather than merely causing score copying. Its effect is direction-dependent: positive framing is easier to elicit than reliable incorporation of expert downgrade evidence.
- Score guidance influences both final scores and review rationales in cases where it corrects baseline mismatches.
- Overestimates are more likely to persist when rationales retain strong positive framing or omit downgrade evidence used by human reviewers.
- Score guidance is useful but incomplete because expert downgrade evidence remains harder to incorporate reliably.
Limitations
The study’s evidence is constrained by a small, self-selected, domain-specific dataset, oracle guidance, and automatic text annotation. Its findings therefore do not generalize directly across all research evaluation settings or deployed predicted-score systems.
- The dataset contains 98 self-selected health-domain outputs from an internal REF-style process, with scores concentrated around 3* and few very low-quality outputs.
- The findings are evidence from a controlled case study rather than general estimates for all research evaluation settings or the full 1–4 score range.
- Oracle guidance tests the effect of a correct score signal but does not capture uncertainty from predicted scores in deployment.
- Text-level analyses rely on automatic annotation, including zero-shot sentence labels and Qwen-based extraction and coverage judgements.
A.1.1 Dataset and preprocessing
The dataset comprises private internal assessment materials from a non-random REF-style process, prepared as anonymised PDFs for text-based analysis. Outputs were cleaned and structurally processed before evaluation.
- Dataset: The dataset contains private assessment reports and scores for outputs considered plausible candidates rather than a random sample of publications.
- Dataset: Each output received qualitative comments and recommended scores from two to four reviewers, producing an adjudicated 1–4 reference score.
- Dataset: The study used 98 anonymised health-domain outputs with human reports and adjudicated scores, removing identifying and publication-related information where possible.
- Preprocessing: PDF text was extracted with PDFMiner, manually cleaned, and organized into top-level sections using GROBID and custom Python scripts.
A.1.2 Prompt templates
Review generation used a shared system prompt with either no score guidance or an external oracle-score block. Three variants changed only the ordering of prompt blocks.
- Prompt conditions: The no-guidance prompt generated baseline reviews, while the oracle-guidance prompt added an external score block.
- Prompt conditions: All conditions used the same system prompt.
- Prompt variants: Variants v4, v7, and v9 differed only in the order of the rule, format, and paper-text blocks.
A.1.3 Model, decoding, and configuration screening
The study evaluates three LLMs across varied inputs, prompts, and decoding settings, then screens configurations to select a compact set for main text analysis.
- Three models were evaluated: Llama-3.1-8B-Instruct, Falcon3-10B-Instruct, and Llama-3.1-405B-Instruct.
- The broader output set crossed three input-length conditions with 11 decoding settings, producing 33 input–decoding configurations per model/prompt setting.
- Configuration screening compared signed score-difference vectors using normalized mutual information to reduce redundancy among baseline outputs.
- The shared system prompt directed models to assess originality, significance, and rigour while providing scores and detailed reasons.
- The primary text-analysis setting used Llama-3.1-8B, prompt v4, full-text input, and three retained temperature/top-p configurations.The retained settings were 0.1/0.9, 0.3/0.8, and 0.5/0.9.
A.1.4 Annotation labels, prompts, and text features
The appendix describes the prompts, annotation layers, computational pipeline, and supplementary checks used to analyse review evidence and score transitions across settings.
- A.1.4 Annotation labels, prompts, and text features: Three text-derived feature groups captured sentence-level review frames, extracted evaluation points, and coverage of human evaluation points.
- A.1.4 Annotation labels, prompts, and text features: The task prompt required evaluations of Originality, Significance, and Rigour, each paired with a qualitative report and quantitative score.
- A.1.4 Annotation labels, prompts, and text features: Sentence-level labels classified review stance and themes such as strength, weakness, significance, and novelty.
- A.1.4 Annotation labels, prompts, and text features: Qwen3-32B extracted human evaluation points and judged whether corresponding points were covered in LLM reviews.
- A.1.5 Compute resources and runtime: The pipeline comprised LLM review generation, sentence-level zero-shot labelling, and Qwen-based point-level annotation.
- A.1.5 Compute resources and runtime: For 8B and 10B models, each model/prompt/input setting produced 55 output files, while the 405B setting produced 165 files overall.The reported estimates were approximately 18.3 sequential hours per 8B or 10B model/prompt/input setting and 412.5 GPU-job hours for the 405B setting.
- A.2.1 Supplementary score-guidance results: Oracle-guidance score results aggregated three input lengths and 11 decoding settings, with negative ∆MAE indicating reduced distance from the human reference.
- A.2.1 Supplementary score-guidance results: Score-following checks tested whether final scores were fully determined by supplied guide scores, while transition tables counted corrections, persistent mismatches, and new errors.