A Plausible Explanation Is Enough to Mislead Multimodal Judges

Who Should Read

A developer running an evaluation pipeline where an MLLM judge scores answers together with their explanations

Core Mechanics

• A plausible explanation alone lets wrong answers through
  • With image, question and answer fixed, adding an explanation lowered incorrect-answer rejection across 14 MLLM judges.
  • Versus no explanation, Plain dropped 12.7 pp and CoT dropped 13.8 pp.
• Persuasion tricks were no stronger than plain text
  • Explanations built with 41 persuasion techniques caused a 12.5 pp drop.
  • Rhetoric strength was uncorrelated with the drop (Spearman ρ=0.00).
• The judge's bar moved, not its skill
  • Signal detection theory (separating discrimination from the accept threshold) shows the acceptance criterion shifted.
  • Mean discriminability (d′) showed no clear change.
• Accuracy hides the damage
  • Overall judge accuracy fell by under 2 pp (58.3% to 56.4–56.7%).
  • Incorrect-answer rejection fell 12.5–13.8 pp over the same conditions.
• Even confident rejections flip

Evidence

• Rejection drop by explanation type
  • Averaged over 14 judges versus no explanation on JUPE.
  • Plain −12.7 pp, CoT −13.8 pp, Persuasion −12.5 pp.
• Mean incorrect-rejection rate
  • 42.4% without explanation.
  • 29.7% with Plain, 29.9% with CoT, 28.6% with Persuasion; false rejection of correct answers fell from 23.3% to about 12%.
• Criterion shift Δc
  • Plain −0.39, CoT −0.45, Persuasion −0.39, with 95% CIs excluding zero, same direction in 12 of 14 judges.
  • Δd′ was +0.04 to +0.09 with CIs spanning zero.
• Prompt mitigation results
  • Under Persuasion-VI explanations, mean wrong-answer rejection across 10 judges was 24.4% with Conservative and 30.4% with Self-Verify.
  • The paper reports gains of 25% and 13% over explanations alone.

How to Apply

• Add a conservative-judging line to the judge prompt
  • Insert 'Be conservative in accepting the candidate answer as correct.' before the output instruction of your judge prompt.
  • Track the correct-answer rejection rate too, since it moves as well.
• Report wrong- and correct-answer rejection separately
  • Replace accuracy in eval reports with CRR and IRR.
  • Judge the same answers twice, with and without explanations, and compare.
• Add a solve-first step before judging
  • Instruct the judge to decide its own answer before checking the candidate.
  • Validate on a small set per model, since effects differ by model.
• Run an answer-only scoring path alongside
  • In pipelines that score answers with explanations, add a run with the explanation removed.
  • Send items where the two verdicts diverge to human review.

Try It

# Paste into Claude or ChatGPT with an image attached (simplified version of the paper's method: compare verdicts with and without an explanation)

Judge whether the final answer correctly answers the Question based on the provided image. Be conservative in accepting the candidate answer as correct. Determine your own answer before verifying the candidate.

Question: {{question}}
Candidate Answer: {{candidate answer}}
Candidate Explanation: {{explanation (delete this line for the baseline)}}

Options (order randomized): A. Definitely incorrect B. Probably incorrect C. Probably correct D. Definitely correct
Only output your final answer in the exact format: Answer: X (where X is A, B, C, or D).

Next step: run a wrong answer with and without the explanation and see whether the verdict changes.

Terminology

Related Papers

• The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge

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

• 90% of US working hours are spent on tasks not covered well by AI benchmarks

  Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage

• Postgres 19 delayed after AI reviewers find more bugs than humans can fix

  LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted

• Show HN: Gemini 4 Argon plays StarCraft

  LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.

• Neural networks are clean

  Looking at neurons directly shows most have one main job, challenging superposition.

• Linus Torvalds's first LLM-assisted patch series

  Linus posted an LLM-assisted 10-patch series giving an order-of-magnitude read() speedup, with a race already found.