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

arXiv: https://arxiv.org/abs/2602.10144 · HTML: https://mlsanity.com/paper/2602.10144/

## Key facts

- Cited by 7 papers in the map (data through 14 Oct 2026).
- Builds on 29 of its 45 references that are in the map.
- First posted on arXiv 9 Feb 2026.

## Builds on (most cited first, 29 of 29)

- [The Llama 3 Herd of Models](https://mlsanity.com/paper/2407.21783/) (2024)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Measuring Mathematical Problem Solving With the MATH Dataset](https://mlsanity.com/paper/2103.03874/) (2021)
- [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://mlsanity.com/paper/2205.14135/) (2022)
- [TruthfulQA: Measuring How Models Mimic Human Falsehoods](https://mlsanity.com/paper/2109.07958/) (2021)
- [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://mlsanity.com/paper/2311.12022/) (2023)
- [GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers](https://mlsanity.com/paper/2210.17323/) (2022)
- [MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark](https://mlsanity.com/paper/2406.01574/) (2024)
- [Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them](https://mlsanity.com/paper/2210.09261/) (2022)
- [Solving Quantitative Reasoning Problems with Language Models](https://mlsanity.com/paper/2206.14858/) (2022)
- [gpt-oss-120b & gpt-oss-20b Model Card](https://mlsanity.com/paper/2508.10925/) (2025)
- [A Survey on Evaluation of Large Language Models](https://mlsanity.com/paper/2307.03109/) (2023)
- [Instruction-Following Evaluation for Large Language Models](https://mlsanity.com/paper/2311.07911/) (2023)
- [SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot](https://mlsanity.com/paper/2301.00774/) (2023)
- [Accelerating Large Language Model Decoding with Speculative Sampling](https://mlsanity.com/paper/2302.01318/) (2023)
- [Abstractive Text Summarization Using Sequence-to-Sequence RNNs and Beyond](https://mlsanity.com/paper/1602.06023/) (2016)
- [A Simple and Effective Pruning Approach for Large Language Models](https://mlsanity.com/paper/2306.11695/) (2023)
- [FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision](https://mlsanity.com/paper/2407.08608/) (2024)
- [A Survey on Efficient Inference for Large Language Models](https://mlsanity.com/paper/2404.14294/) (2024)
- [MuSR: Testing the Limits of Chain-of-thought with Multistep Soft Reasoning](https://mlsanity.com/paper/2310.16049/) (2023)
- [Adding Error Bars to Evals: A Statistical Approach to Language Model Evaluations](https://mlsanity.com/paper/2411.00640/) (2024)
- [Accelerated Sparse Neural Training: A Provable and Efficient Method to Find N:M Transposable Masks](https://mlsanity.com/paper/2102.08124/) (2021)
- [Understanding and Mitigating Numerical Sources of Nondeterminism in LLM Inference](https://mlsanity.com/paper/2506.09501/) (2025)
- ["Give Me BF16 or Give Me Death"? Accuracy-Performance Trade-Offs in LLM Quantization](https://mlsanity.com/paper/2411.02355/) (2024)
- [Accuracy is Not All You Need](https://mlsanity.com/paper/2407.09141/) (2024)
- [Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation](https://mlsanity.com/paper/2508.13144/) (2025)
- [Inference Optimization of Foundation Models on AI Accelerators](https://mlsanity.com/paper/2407.09111/) (2024)
- [A Proximal Operator for Inducing 2:4-Sparsity](https://mlsanity.com/paper/2501.18015/) (2025)
- [Block-regularized 5×2 Cross-validated McNemar's Test for Comparing Two Classification Algorithms](https://mlsanity.com/paper/2304.03990/) (2023)

## Built on by (most cited first, 7 of 7)

- [Statistically-Lossless Quantization of Large Language Models](https://mlsanity.com/paper/2605.02404/) (2026)
- [Certifying Compressed Language Models: An Audit and a Statistical Toolkit](https://mlsanity.com/paper/2608.15046/) (2026)
- [Accuracy is Not Enough: A Divergence-Based Approach to Evaluate Fidelity Loss in Quantized LLMs](https://mlsanity.com/paper/2609.07664/) (2026)
- [Attention Quantization for Tabular Foundation Models](https://mlsanity.com/paper/2609.13031/) (2026)
- [A Calibrated Instrument for Measuring How Inference Optimizations Affect Output Quality](https://mlsanity.com/paper/2609.18005/) (2026)
- [Disaggregated Quantization: Specializing LLM Prefill and Decode](https://mlsanity.com/paper/2609.26333/) (2026)
- [What Does a Harness Repair? A Preregistered Study of Visibility, Baseline Adequacy and Evaluation Defects](https://mlsanity.com/paper/2610.05533/) (2026)

## Abstract (arXiv)

Minimizing the inference cost and latency of foundation models has become a crucial area of research. Optimization approaches include theoretically lossless methods and others without accuracy guarantees like quantization. In all of these cases it is crucial to ensure that the model quality has not degraded. However, even at temperature zero, model generations are not necessarily robust even to theoretically lossless model optimizations due to numerical errors. We thus require statistical tools to decide whether a finite-sample accuracy deviation is an evidence of a model's degradation or whether it can be attributed to (harmless) noise in the evaluation. We propose a statistically sound hypothesis testing framework based on McNemar's test allowing to efficiently detect model degradations, while guaranteeing a controlled rate of false positives. The crucial insight is that we have to confront the model scores on each sample, rather than aggregated on the task level. Furthermore, we propose three approaches to aggregate accuracy estimates across multiple benchmarks into a single decision. We provide an implementation on top of the largely adopted open source LM Evaluation Harness and provide a case study illustrating that the method correctly flags degraded models, while not flagging model optimizations that are provably lossless. We find that with our tests even empirical accuracy degradations of 0.3% can be confidently attributed to actual degradations rather than noise.