Michelangelo: Long Context Evaluations Beyond Haystacks via Latent Structure Queries

TL;DR

Michelangelo introduces a novel framework for evaluating large language models' ability to understand and manipulate long contexts by revealing latent structures, providing more meaningful diagnostics than traditional retrieval tasks.

Contribution

The paper presents the Latent Structure Queries framework, a new method for creating long-context evaluation tasks that measure deep understanding beyond simple retrieval.

Findings

Evaluations are high-signal and effective for assessing long-context understanding.

State-of-the-art models show significant room for improvement in long-context reasoning.

The framework applies across code and natural language domains.

Abstract

We introduce Michelangelo: a minimal, synthetic, and unleaked long-context reasoning evaluation for large language models which is also easy to automatically score. This evaluation is derived via a novel, unifying framework for evaluations over arbitrarily long contexts which measure the model's ability to do more than retrieve a single piece of information from its context. The central idea of the Latent Structure Queries framework (LSQ) is to construct tasks which require a model to ``chisel away'' the irrelevant information in the context, revealing a latent structure in the context. To verify a model's understanding of this latent structure, we query the model for details of the structure. Using LSQ, we produce three diagnostic long-context evaluations across code and natural-language domains intended to provide a stronger signal of long-context language model capabilities. We perform evaluations on several state-of-the-art models and demonstrate both that a) the proposed evaluations are high-signal and b) that there is significant room for improvement in synthesizing long-context information.