Title: Connecting Auxiliary Constraints to Enhance Instruction Adherence in Large Reasoning Models

Abstract:

While Large Reasoning Models (LRMs) have shown remarkable proficiency across a variety of tasks, they frequently encounter difficulties in consistently executing multiple instructions. These challenges often manifest as failures to meet specific requirements or an inability to harmonize conflicting demands simultaneously. We define this issue as the Constraint Adherence Problem (CAP). To tackle CAP, this study proposes a new framework that encodes instructions into a structured knowledge graph composed of constraints. Our method, termed Constraint Relationship Graph Completion (CRGC), explicitly maps the relationships among constraints, pinpoints areas of adherence difficulty, and uncovers "bridge constraints." These bridge constraints serve as auxiliary directives that enhance the prominence and compatibility of primary instructions, thereby aiding the model in focusing on and reconciling its tasks. In contrast to conventional methods that rely on broad training techniques to boost instruction following, CRGC improves constraint satisfaction by utilizing the model’s internal knowledge to construct more effective generation pathways. Evaluations across three widely used instruction-following datasets reveal that our approach cuts constraint violations by 39% relative to standard prompting, all while preserving the robust reasoning capabilities inherent in large reasoning models.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC