metabench -- A Sparse Benchmark of Reasoning and Knowledge in Large Language Models

2024-07-04

AI Summary

Existing LLM benchmarks (e.g., Open LLM Leaderboard) suffer from high item intercorrelation and information redundancy, hindering precise characterization of latent capability dimensions. Method: We propose the first capability-aware sparse benchmark construction framework, integrating Item Response Theory (IRT), factor analysis, and sparse optimization to select highly informative items from 28,632 questions across six mainstream benchmarks. Contribution/Results: Our method yields Metabench—a hyper-sparse evaluation benchmark containing only 2.8% of the original items—while enabling joint compression across multiple benchmarks and disentangled modeling of latent capabilities (e.g., reasoning, knowledge). Despite extreme sparsification, it achieves an average RMSE of 1.24% in reconstructing individual benchmark scores and only 0.58% for overall scores; a single dominant common factor correlates strongly with total scores (Spearman’s *r* = 0.94). This significantly enhances evaluation efficiency and interpretability.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Metareasoning and Metaheuristics
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the Open LLM Leaderboard aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly). However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed. We use data from n>5000 LLMs to identify the most informative items of six benchmarks, ARC, GSM8K, HellaSwag, MMLU, TruthfulQA and WinoGrande (with d = 28,632 items in total). From them we distill a sparse benchmark, metabench, that has less than 3% of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities. We show that these estimators (1) can be used to reconstruct each original individual benchmark score with, on average, 1.24% root mean square error (RMSE), (2) reconstruct the original total score with 0.58% RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is r = 0.94.

Problem

• Identify most informative benchmarks
• Create sparse benchmark metabench
• Estimate underlying benchmark-specific abilities

Innovation

• Distills sparse benchmark
• Identifies informative items
• Yields ability estimators