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

Accuracy is Not All You Need

NeurIPS 2024 · Large language models › Language, safety and interpretability · approach: Uncertainty quantification for LLMs · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that LLM compression (quantization, layer dropping, pruning) can keep benchmark accuracy nearly unchanged while many individual answers flip correct/incorrect, and proposes flips and KL-divergence as distance metrics for evaluating compression.

Problem

Accuracy on benchmarks is the standard way to validate LLM compression, but similar accuracy can hide large changes in per-question model behavior that end users see, especially in free-form generation.

Approach

Introduces "flips" (percentage of answers changing correct to incorrect or incorrect to correct between baseline and compressed model) and KL-divergence as distance metrics, showing they correlate with each other and with MT-Bench degradation while accuracy stays flat.

Key results

• Across six quantization schemes on seven tasks, accuracy changes are within 0-2% while flips reach up to 13.6% (e.g., Llama2-7b SQ W8A8 on MMLU/ARC-style tasks).
• On MMLU, Spearman correlation between KL-divergence and flips is 0.981.
• Llama2-70b chat with BnB W4A4 and W8A8 keeps MMLU accuracy within about 1% but scores lower on MT-Bench; GPT-4 judge gives roughly 0.4 point drop on a 10-point scale.

Measured effects reported in the paper

Surprising findings

• Correct-to-incorrect and incorrect-to-correct transitions are roughly balanced, so accuracy stays flat despite many flips.
• Adding symmetric Gaussian noise to log-likelihoods leaves perplexity unchanged while generation quality degrades.

Limitations

• Flips only signal behavioral divergence; they may not translate into visible degradation in every downstream task.
• Qualitative MT-Bench evaluation is subjective and may not be representative.
• MT-Bench scoring relies on GPT-4 as judge, which may be noisy; small score differences may be judge noise.

Experimental setup

Datasets: MMLU PIQA HellaSwag ARC Easy ARC Challenge LAMBADA Winogrande GSM8K TriviaQA WikiText-2

Benchmarks: MMLU PIQA HellaSwag ARC Easy ARC Challenge LAMBADA Winogrande GSM8K TriviaQA MT-Bench

Metrics: accuracy flips AllFlips KL divergence perplexity MT-Bench score Spearman correlation

Baselines: GPTQ AWQ SmoothQuant LLM.int8() (BnB) Wanda SliceGPT layer dropping

Builds on: LLM.int8() GPTQ AWQ SmoothQuant Wanda

Closest papers by meaning

• Compressing LLMs: The Truth is Rarely Pure and Never Simple (ICLR 2024 · similarity 0.92)
• Language Modeling Is Compression (ICLR 2024 · similarity 0.90)
• LLMCBench: Benchmarking Large Language Model Compression for Efficient Deployment (NeurIPS 2024 · similarity 0.90)
• Compressing Large Language Models by Joint Sparsification and Quantization (ICML 2024 · similarity 0.90)
• LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM Compression (NeurIPS 2026 · similarity 0.90)
• Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression (ICML 2025 · similarity 0.89)