Apple proposes self-reflective program search to improve long-context language model performance

Apple’s Machine Learning Research team proposes Self-Reflective Program Search for Long Context (SRLM), a framework designed to address the persistent challenge of long-context understanding in language models. Even with extended context windows, models often struggle to reliably extract, reason over, and use information across long contexts. While recent approaches like Recursive Language Models (RLMs) attempt to tackle this by decomposing long contexts into recursive sub-queries through programmatic interaction at inference, their success hinges on how these trajectories of context-interaction programs are selected—a problem that has remained underexplored until now.

SRLM augments programming-based context interaction with uncertainty-aware self-reflection, leveraging three intrinsic signals: self-consistency, reasoning trace length, and verbalized confidence. These signals serve as complementary indicators of a model’s internal uncertainty, enabling it to evaluate and compare candidate context-interaction programs. By using these signals, SRLM aims to steer reasoning more effectively in challenging long-context scenarios, particularly those with semantically intensive tasks where heuristic program search falls short.

In extensive experiments across diverse benchmark datasets, context lengths, and backbone models, SRLM consistently outperforms state-of-the-art baselines. The framework yields up to a 22% improvement over RLMs under the same time budget. Notably, the authors find that recursion itself is not the primary driver of performance in RLMs, and a simple self-reflective program search can match or surpass RLM without requiring self-query or explicit recursion mechanisms.

The findings also reveal that RLMs with recursion can degrade performance relative to the base model for context lengths within the model’s context window, whereas SRLM delivers consistent and robust gains across both short and long contexts. This suggests that uncertainty-aware self-reflection provides a more reliable signal for program selection than heuristic or recursive approaches, particularly in tasks demanding deeper contextual understanding.