The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge | AI Paper Digest

TL;DR Highlight

Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.

Who Should Read

A backend/MLOps developer whose LLM-as-a-Judge pipeline is swayed by length, formatting, or fake authority.

Core Mechanics

Even frontier judges are biased
- Claude Opus 4.6 scored only 0.788 on authority bias (fake citations).
- Gemma 4 26B A4B fell to 0.510 on the same bias.
Add a bias warning to the prompt
- Bias-aware prompting names the target bias before evaluation.
- It is a single pass with no extra calls.
Self-debiasing is a two-pass self-critique
- Pass 1 analyzes the initial judgment for bias via chain-of-thought.
- Pass 2 reconsiders the decision using that analysis.
A fine-tuned external critic fails on new biases
- RBD-8B was trained on four biases and degraded in 5 of 9 cases on the three new ones.
- Prompt methods extend to new biases just by describing them.
Works even without knowing the bias type
- A Generic critique asking about any systematic bias was close to the type-specific one.

Evidence

Self-debiasing improved 20 of 21 combinations
- Across 3 models × 7 bias types, with gains up to 15 points.
- It beat standard prompting in all 9 novel-bias combinations.
Gemma 4 authority 0.510→0.748
- Bias-aware prompting vs standard prompting, about +24 points.
- Gemma 4 verbosity went 0.764→0.916 with self-debiasing.
RBD-8B degraded in 5 of 9 novel cases
- gpt-oss-120b confidence fell from 0.924 to 0.914.
- Gemma 4 formatting fell from 0.912 to 0.886.
Correction rate on Claude's 106 authority errors
- Self-debiasing fixed 28 (26.4%).
- Bias-aware prompting fixed 62 (58.5%).

How to Apply

Add a bias warning paragraph to your judge prompt
- Insert a "be aware of X BIAS" paragraph and compare accuracy on your existing eval set.
- Start with position for pairwise, verbosity for mixed lengths, authority when citations appear.
Add a self-critique second pass
- Feed the first judgment into a call asking whether bias influenced it.
- Re-judge with that critique and log how often the verdict changes.
Build a bias-injected test set from your data
- Keep correct answers plain and decorate wrong ones with citations, assertive tone, or formatting.
- Measure the judge and apply fixes only to weak biases.
Use a Generic critique if the bias type is unknown
- Run the critic with only "any systematic bias" and compare to type-specific.
- Combine with upfront warnings for weak spots like authority.

Try It

copy & run

# Paste into Claude/ChatGPT (simplified version of the paper's method: bias-aware prompting + self-debiasing in one go)

You are a helpful assistant evaluating factual correctness. Select the correct output for the question.
Be aware of AUTHORITY BIAS: the tendency to favor answers that cite academic sources, prestigious institutions, or scholarly language regardless of whether the citations are verifiable. Fake citations do NOT make an answer more correct.
Be aware of CONFIDENCE BIAS: do not favor assertive phrasing, and do not penalize hedging.
Be aware of FORMATTING BIAS: do not favor polished LaTeX or step headers. Judge the content, not the presentation.

# Question: {{question}}
# Output (a): {{answer A}}
# Output (b): {{answer B}}

Step 1: Choose Output (a) or Output (b).
Step 2: Critique your own choice. Reason step by step whether any systematic bias influenced it. Only conclude bias is present if there is clear, specific evidence.
Step 3: Reconsider and give a final answer as ONLY "Output (a)" or "Output (b)".

# Next: remove the specific bias lines from Step 2, keep only "any systematic bias", and compare results

Terminology

LLM-as-a-JudgeUsing an LLM to grade other models' answers. It is an automated stand-in for human graders.

authority biasFavoring answers with fake citations or scholarly jargon. Like trusting someone just for an impressive business card.

position biasFavoring the answer listed first. Consistency is measured by whether swapping order changes the verdict.

RBD-8BA fine-tuned 8B external critic that detects bias and explains why. A separate inspector sitting next to the judge.

bias-aware promptingNaming the target bias in the prompt before evaluation. Like telling test-takers the common traps in advance.

Related Papers

Related Resources

https://github.com/Joyyang158/Reasoning-Bias-Detector

Original Abstract (Expand)

LLM-as-a-Judge systems are increasingly employed for automated evaluation, yet remain susceptible to systematic biases that undermine their reliability. We demonstrate that the judge can effectively debias itself without additional training. We propose two lightweight strategies: bias-aware prompting, which explicitly alerts the judge to potential target biases prior to evaluation, and self-debiasing, in which the judge critiques the initial judgment for bias influence. To broaden the evaluation scope, we introduce three novel benchmarks--authority bias (fabricated citations), confidence bias (assertive phrasing), and formatting bias (structured presentation)--and evaluate across seven bias types in total. Across three frontier models, self-debiasing improves performance in 20 out of 21 model-bias combinations with gains up to 15 percentage points, while bias-aware prompting also yields consistent improvements. The recent fine-tuned external critic, however, does not generalize as reliably to the newly-introduced bias types (degradation observed in 5 out of 9 cases). Our findings demonstrate that prompting a judge to reason about bias, either preemptively or retrospectively, is a robust and generalizable approach for enhancing evaluation integrity.