ATLAS: All-round Testing of Long-context Abilities across Scales

2026-05-27

AI Summary

Existing evaluations of long-context language models are often limited to a single context length or narrow task sets, failing to reveal performance degradation as context scales and the inability of retrieval capabilities to generalize to downstream tasks. To address this, this work proposes ATLAS, a comprehensive evaluation framework that enables fine-grained, multidimensional assessment across context lengths from 8K to 1M tokens. ATLAS introduces a hierarchical capability taxonomy, length-grid-based AUC scoring, and ATLAScore—a nonlinear aggregation metric incorporating uncertainty propagation. Experiments on 26 models demonstrate substantial shifts in model rankings across different context-length regimes and reveal only 61% shared variance between the two capability dimensions, underscoring the necessity of jointly reporting long-context performance along both capability and length axes.

Abstract

Long-context language models now advertise context windows up to millions of tokens, yet evaluations typically report a single length or a narrow task family, masking two failure modes: performance can collapse as length grows, and strong retrieval need not transfer to downstream use. We present ATLAS, a benchmarking framework that redefines long-context evaluation as length-dependent capability profiling. ATLAS contributes three methodological principles:(i) a layered taxonomy separating foundational operations from application workloads so failures can be attributed, (ii) length-aware AUC scoring that integrates score-length curves over a fixed 8K-1M grid, replacing single-point metrics with full degradation profiles, and (iii) ATLAScore, a harmonic-mean aggregate over taxonomy categories that penalizes imbalanced profiles, with end-to-end uncertainty propagation from subset scores through the nonlinear final aggregate. We instantiate the framework across eight capability dimensions with nine auditable components and 6,438 instances, and evaluate 26 models. Gemini-3.1-Pro-Preview leads at 128K, Claude-Opus-4.6 leads at 1M. Rankings reshuffle substantially between ATLASscore@8K-128K and ATLASscore@8K-1M: 7 models move by at least two ranks, and the two taxonomy layers share only 61% of cross-model variance, with individual rank gaps up to 12 positions. These results support reporting long-context quality by capability and length, not by a single headline score.

Problem

• long-context evaluation
• performance degradation
• retrieval-transfer gap
• length-dependent capability
• benchmarking

Innovation

• long-context evaluation
• length-aware AUC
• layered taxonomy
• ATLAScore
• capability profiling