AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

When LLMs get significantly worse: A statistical approach to detect model degradations

Jonas M. Kübler, Kailash Budhathoki, Matthäus Kleindessner, Xiong Zhou, Junming Yin, Ashish Khetan, George Karypis

ICLR 2026 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks · 5 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Compare per-sample scores with exact one-sided McNemar test instead of treating accuracies as independentdegradation detection · Llama-3.1 8B Instruct KV-FP8detection of 0.79% drop: p = 1.69e-05 detects dropImproved
Use the McNemar-based tests on KV-FP8 for the 70B modelexact McNemar tests · Llama-3.3 70B Instruct KV-FP8detection: 0.29% drop, p_pool 2.51e-02Improved
Apply tests to theoretically lossless variants (rerun, transformers, TP1, A100)exact McNemar tests · Llama-3.1 8B Instructfalse positives: p_pool 5.64e-01 to 9.21e-01Improved
Use hard accuracy-difference threshold (>2%) as degradation rulethreshold baseline · Llama-3.1 8B Instruct KV-FP8detection: 0.79 vs 2% thresholdHurt
Use hard flip-probability threshold (>=5%) as degradation ruleflip-rate baseline · Llama-3.1 8B Instruct FP8false positives: 8.71% flips vs 9.03% for KV-FP8Hurt
Remove examples that never flip from MMLU-Pro (ablation)degradation testing · Llama-3.1 8B Instruct MMLU-Proflip rate KV-FP8: 15.49% -> 24.42%Improved
Use w4a16 INT4 quantizationLlama-3.1 8B Instruct · Open LLM Leaderboard v2pooled accuracy: 42.43 -> 40.70Hurt
Use FP8 weights/activationsLlama-3.1 8B Instruct · Open LLM Leaderboard v2pooled accuracy: 42.43 -> 42.48No clear effect
Use 2:4 sparsity (pruned and distilled)Llama-3.1 8B base · Open LLM Leaderboard v2pooled accuracy: 32.72 -> 30.13Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy