PulseAugur
EN
LIVE 21:54:47

LLM judge accuracy skyrockets after prompt correction

By PulseAugur Editorial · [1 sources] ·

An LLM judge using Gemini 3.1 Flash-Lite initially failed to accurately assess answers, scoring only 60% accuracy. This was due to a system prompt designed for a support bot, which incorrectly prioritized checking for a "next step" rather than factual correctness. After modifying the prompt to focus solely on accuracy and relevance, the model's agreement with human labels improved to 95%. The author also noted that both Gemini and Claude Haiku 4.5 incorrectly passed an answer indicating a lack of information, highlighting a potential ambiguity in labeling or judge instructions. AI

IMPACT Highlights the critical importance of precise system prompts for accurate LLM evaluation and performance assessment.

RANK_REASON The item discusses the author's personal experience and findings with LLM evaluation, rather than a new release or significant industry event.

Read on dev.to — LLM tag →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

LLM judge accuracy skyrockets after prompt correction

How we ranked this

Signal score
13 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Commentary
The item discusses the author's personal experience and findings with LLM evaluation, rather than a new release or significant industry event.
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
product, other
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

Full methodology in our editorial standards.

COVERAGE [1]

  1. dev.to — LLM tag TIER_1 English(EN) · sarthak bhardwaj ·

    My LLM judge failed "Paris." as the capital of France. The rubric was grading the wrong thing.

    <p>My LLM judge agreed with my hand labels on 12 of 20 answers: 0.60 accuracy, Cohen's kappa 0.20. Same model, same 20 answers, one system prompt changed, and it agreed on 19 of 20: 0.95, kappa 0.90. The model was fine. I had handed it a rubric that graded something my labels nev…

RELATED ENTITIES

RELATED TOPICS