Research map / Large language models / Language, safety and interpretability
Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation
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.
Proposes signal (relative dispersion across models) and noise (relative std over final checkpoints) metrics, shows their ratio predicts small-to-large decision accuracy and scaling-law error, and tests interventions (subtask filtering, checkpoint averaging, bits-per-byte) on 30 benchmarks and 465 open-weight models.
Problem
Small-scale language model experiments are used to make development decisions (dataset choice) and to fit scaling laws, but many benchmarks give unreliable small-to-large predictions; there is no cheap intrinsic measure of whether a benchmark is useful for this.
Approach
Define signal as the relative dispersion of final-checkpoint scores across models and noise as the relative std of the final n intermediate checkpoints of one run; their ratio (SNR) predicts decision accuracy, and noise predicts scaling-law prediction error. Interventions targeting SNR (subtask filtering, checkpoint averaging, bits-per-byte metric) improve both prediction settings.
Key results
- SNR correlates with decision accuracy (R=0.791, R^2=0.626) across OLMES benchmarks.
- Noise correlates with scaling-law prediction error for 13B target (R=0.653, R^2=0.426) across 30 tasks.
- Switching to bits-per-byte improved SNR for most benchmarks (e.g., GSM8K 1.2 to 7.0, MBPP 2.0 to 41.8) and lowered scaling-law error for 73.3% of benchmarks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| average final checkpoints of both small prediction and large target models (ablation) | DataDecide models · All 30-task average | decision accuracy: 68.9 -> 71.3 | Helped |
| average checkpoints of the small models used to fit the scaling law (ablation) | OLMo 2 scaling-law models predicting 13B · All 30-task average | prediction error: 1.03 -> 0.86 | Helped |
| average checkpoints of the small models used to fit the scaling law (ablation) | OLMo 2 scaling-law models predicting 13B · Math 6-task average | prediction error: 11.33 -> 2.30 | Helped |
| average checkpoints of the small models used to fit the scaling law (ablation) | OLMo 2 scaling-law models predicting 13B · Code 4-task average | prediction error: 3.15 -> 2.75 | Helped |
| average checkpoints of both small prediction and large target models (ablation) | DataDecide models · Knowledge 19-task average | decision accuracy: 71.3 -> 71.7 | Helped |
| average final checkpoints of the small prediction models only (ablation) | DataDecide models · All 30-task average | decision accuracy: 68.9 -> 70.7 | Helped |
| average final checkpoints of the large target models only (ablation) | DataDecide models · All 30-task average | decision accuracy: 68.9 -> 69.5 | Helped |
| filter subtasks by selecting the highest-SNR subset | MMLU and AutoBencher · MMLU and AutoBencher | decision accuracy: +2.6% for MMLU; +5% for AutoBencher | Helped |
| use bits-per-byte instead of the primary task metric | evaluation benchmarks · GSM8K | signal-to-noise ratio: 1.2 to 7.0 | Helped |
| use bits-per-byte instead of the primary task metric | evaluation benchmarks · MBPP | signal-to-noise ratio: 2.0 to 41.8 | Helped |
| use bits-per-byte instead of the primary task metric | evaluation benchmarks · all benchmarks | decision accuracy: 90.0% of all benchmarks | Helped |
| use bits-per-byte instead of the primary task metric | evaluation benchmarks · all benchmarks | scaling-law prediction error: 73.3% of all benchmarks | Helped |
Surprising findings
- For MMLU, a 16-subtask subset had higher SNR than the full test set, and a 6-subtask AutoBencher subset likewise, with fewer instances.
- A 300-instance ARC-Easy subset showed lower noise than the full 30K-instance AutoBencher.
Limitations
- Scope limited to decision accuracy and scaling-law error settings and to noise from training checkpoints; other noise sources (seed, data order, evaluation config) not fully covered.
- Noise estimated only over final n checkpoints (n=30 for scaling law target), with guidance that n=20 is needed for tight bounds.
- Relies on one particular SNR instantiation (relative dispersion over models).
Experimental setup
Datasets: OLMoE mix DataDecide pretraining corpora MMLU ARC HellaSwag GSM8K MBPP AutoBencher
Benchmarks: MMLU ARC Challenge ARC Easy HellaSwag GSM8K HumanEval MBPP AutoBencher OLMES Core 9 TriviaQA
Metrics: signal-to-noise ratio decision accuracy scaling law prediction error relative standard deviation Spearman correlation R^2
Baselines: final checkpoint evaluation random subtask ordering primary task metric (accuracy / exact match / pass@1) seed noise SNR (single-model SNR)
Builds on: OLMo 2 DataDecide OLMES AutoBencher Kendall's tau
Closest papers by meaning
- Measuring what Matters: Construct Validity in Large Language Model Benchmarks (NeurIPS 2025 · similarity 0.90)
- BenchmarkCards: Standardized Documentation for Large Language Model Benchmarks (NeurIPS 2025 · similarity 0.90)
- A Probabilistic Perspective on Unlearning and Alignment for Large Language Models (ICLR 2025 · similarity 0.90)
- An Empirical Study on Noisy Data and LLM Pretraining Loss Divergence (NeurIPS 2026 · similarity 0.90)
- Language models scale reliably with over-training and on downstream tasks (ICLR 2025 · similarity 0.89)
- An Empirical Analysis of Uncertainty in Large Language Model Evaluations (ICLR 2025 · similarity 0.89)
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.