Prior Scores Bias LLM Evaluation Results
Listen to the summary
Uses a voice available on your device
Audio options
On this page 4 sections
Related concepts 2 concepts
Key Takeaways
- Seven out of eight tested LLMs exhibited statistically significant anchoring effects when provided with prior-score metadata.
- Categorical judgments are compromised, with anchored metadata causing a 10.18 percent rate of wrong labels on previously correct evaluations.
- The bias manifests as a block on 48 percent of potential error corrections, effectively silencing the judge's ability to fix mistakes.
- Standardized effect sizes measured by Cohen’s d reached an absolute value of 0.71, indicating a strong influence from the provided anchors.
Summary & Methodology Analysis
The study evaluates anchoring bias in LLM-as-a-judge systems by testing eight distinct models against twenty fixed tasks. The researchers utilized a combination of API-served models, including GPT-4.1, Claude-4.5-Sonnet, and DeepSeek-R1 accessed via a unified LiteLLM proxy, and five locally hosted open-weight models, specifically the Llama 3 family (3B, 8B, and 70B), Qwen2.5-7B, and Gemma-2-9B. To quantify the bias, the team established memory conditions that included metadata containing revision, attempt, and prior-score data, comparing these to a baseline condition without prior metadata. The methodology mirrors standard evaluation frameworks like MT-Bench and G-Eval, which frequently rely on structured prompt inputs to derive scores.
The technical investigation involved a targeted probe using token-level log-probabilities, the likelihood assigned by a model to a specific token during generation, to identify threshold-like response patterns. This allowed the researchers to measure the influence of anchors restricted to the 0 to 3.99 range. By applying this to both numerical scoring and categorical industry data with human-labeled ground truth, the authors confirmed that the bias persists across different decision types. The quantitative impact was significant, with a Cohen’s d effect size of 0.71 and a 10.18 percent error rate on categorical tasks, demonstrating that the presence of prior metadata systematically steers the model away from independent evaluation.
Despite these findings, the study is subject to specific constraints. The reliance on twenty fixed texts limits the breadth of the current findings, and the transferability to unseen tasks or models has yet to be established. Furthermore, the industry validation was conducted on a single 441-sample domain, meaning the observed effect magnitudes may not generalize to broader, more diverse datasets. Finally, because the study restricted anchors to the 0 to 3.99 range, other anchor regimes remain outside the scope of the researchers' claims, and the paper does not specify how different ranges might alter the outcome.
Interactive System Flowchart
Cross-Examination & FAQs
A deeper dive clarifying mechanics, constraints, and baseline evaluations.
Q1. What is the primary finding regarding LLM-as-a-judge systems?
The paper finds that providing prior evaluation scores as metadata introduces anchoring bias, which systematically compromises the independence of the judge's output.
Q2. Does this bias affect more than just numerical scores?
Yes, the bias extends to categorical decisions, where anchored metadata blocked 48 percent of error corrections and induced wrong labels in 10.18 percent of previously correct judgments.
Q3. Are all models equally affected by this bias?
Seven out of the eight models evaluated showed statistically significant anchoring effects, indicated by 95 percent task-stratified bootstrap intervals below zero.
Q4. Which models were included in the evaluation?
The evaluation included GPT-4.1, Claude-4.5-Sonnet, DeepSeek-R1, the Llama 3 family (3B, 8B, 70B), Qwen2.5-7B, and Gemma-2-9B.
Q5. What is the magnitude of the observed anchoring effect?
The standardized effect size, measured by Cohen’s d, reached an absolute value of 0.71.
Q6. What range of scores was used for the anchor tests?
The anchors were restricted to the 0 to 3.99 range.
Q7. Is the observed bias transferable to all tasks?
The paper states that transferability to unseen tasks or models remains to be established and the industry validation was limited to one 441-sample domain.
Q8. How did the researchers conduct the tests across different models?
The API models were accessed via a unified LiteLLM proxy, and the study utilized twenty fixed evaluation tasks across four categories.
Q9. Does the paper suggest any mitigation strategies for these biases?
The paper investigated mitigation strategies including Chain-of-Thought reasoning and explicit disregard warnings, though the findings focused on the presence of bias.