My taste and the automated benchmark disagreed almost completely

2026-07-05 · Source: How I AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, quick

Summary

The analysis reveals a significant divergence between human "taste" and automated benchmark results when evaluating large language models (LLMs) acting as judges. Specifically, Opus 4A and GPT 5.5 were used to judge other models, including themselves. The author observed GPT 5.5 consistently acting as the toughest judge, even rating itself lower than another judge. While the LLM judges generally agreed, they were perceived as overall generous, necessitating a "double bench" approach to balance their evaluations. This discrepancy highlights that model effectiveness is highly dependent on the specific job and task fit, suggesting that traditional metrics may not fully capture nuanced performance aspects. The author intends to integrate more subjective "taste" into future judgment processes.

Key takeaway

For Machine Learning Engineers evaluating LLM performance, recognize that automated benchmarks may not align with nuanced human judgment. If you are relying solely on LLM judges, consider integrating your own "vibe checks" or subjective criteria into the evaluation process. This can help encode more qualitative "taste" into model assessments, especially for tasks where models are already highly proficient, ensuring a more robust and context-aware performance evaluation.

Key insights

The author's subjective evaluation ("taste") diverged significantly from automated LLM benchmark results, highlighting limitations in current judging methods.

Principles

• LLM judges can exhibit inherent biases.
• Model performance is task-dependent.
• Saturated tasks make poor benchmarks.

Method

A "double bench" approach was used to balance evaluations from two LLM judges (Opus 4A and GPT 5.5), aiming to mitigate individual judge biases and generosity.

In practice

• Incorporate human "taste" into LLM judging.
• Avoid agentic tasks for benchmarking.
• Use multiple LLM judges for balance.

Topics

• LLM Evaluation
• Model Judging
• Benchmark Bias
• GPT 5.5
• Opus 4A
• Agentic Tasks

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Sonnet 5 review: I ran 64 generations to find out if it's worth it — Benchmarking frontier LLMs requires combining automated metrics with human "taste" to capture nuanced performance.
• From Feelings to Metrics: Understanding and Formalizing How Users Vibe-Test LLMs — Formalized "vibe-testing" bridges the gap between LLM benchmark scores and real-world user utility.
• # I Got Tired of Saying “This Response Is Better” Without Being Able to Prove It — Combining human domain expertise with automated pattern detection improves LLM evaluation accuracy and provides objective vocabulary for subjective judgments.
• Evaluation: Model Benchmarks and LLM Application Assessment — LLM evaluation requires a multi-faceted approach, combining intrinsic, deterministic, and subjective methods with standardized benchmarks.
• LLM Judges Have Dark Current: A Psychometric Datasheet for LLM-as-a-Judge Evaluation — LLM-as-a-judge systems require psychometric evaluation as measurement instruments to understand their true biases and sensitivities.
• What Do Models Still Suck At? - Peter Gostev, Arena.ai, BullshitBench — LLMs still struggle with identifying nonsense and consistently meeting user expectations in complex, real-world tasks.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by How I AI.