Bridging Auxiliary Constraints to Resolve Instruction Following in Large Reasoning Models

A TACL 2026 paper by Zhengyi Zhao, Shubo Zhang, Huimin Wang, Zezhong Wang, Yutian Zhao, Yefeng Zheng, Binyang Li, Yulan He, Kam-Fai Wong and Xian Wu introduces Constraint Relationship Graph Completion (CRGC), a framework that models instructions as a structured knowledge graph of constraints and adds "bridge constraints" as auxiliary instructions to make primary constraints more salient and compatible. Across three popular instruction following datasets, CRGC reduced constraint violations by 39% compared to standard prompting while maintaining the reasoning abilities of large reasoning models, addressing what the authors formalize as the Constraint Adherence Problem (CAP). The work appears in Transactions of the Association for Computational Linguistics, Volume 14, pages 1329–1347, published by MIT Press.

Abstract Large Reasoning Models LRMs have demonstrated impressive capabilities in many tasks, yet they struggle with reliably following multiple instructions, either by failing to satisfy individual constraints or by struggling to balance competing constraints simultaneously. We formalize this challenge as the Constraint Adherence Problem CAP . This paper introduces a novel framework that addresses CAP by representing instructions as a structured knowledge graph of constraints. Our approach, Constraint Relationship Graph Completion CRGC , explicitly models relationships between constraints, identifies adherence challenges, and discovers “bridge constraints” that help the model better focus on and reconcile requirements. Bridge constraints act as auxiliary instructions that make primary constraints more salient and compatible. Unlike existing approaches that enhance instruction following through general training methods, CRGC specifically improves constraint satisfaction by leveraging the model’s own knowledge to create better pathways for generation. Experiments across three popular instruction following datasets demonstrate that our approach reduces constraint violations by 39% compared to standard prompting while maintaining reasoning abilities of large reasoning models. - Anthology ID: - 2026.tacl-1.60 - Volume: - Transactions of the Association for Computational Linguistics, Volume 14 https://aclanthology.org/volumes/2026.tacl-1/ - Month: - Year: - 2026 - Address: - Cambridge, MA - Venue: - TACL https://aclanthology.org/venues/tacl/ - SIG: - Publisher: - MIT Press - Note: - Pages: - 1329–1347 - Language: - URL: - https://aclanthology.org/2026.tacl-1.60/ https://aclanthology.org/2026.tacl-1.60/ - DOI: - 10.1162/tacl.a.730 https://doi.org/10.1162/tacl.a.730 - Cite ACL : - Zhengyi Zhao, Shubo Zhang, Huimin Wang, Zezhong Wang, Yutian Zhao, Yefeng Zheng, Binyang Li, Yulan He, Kam-Fai Wong, and Xian Wu. 2026. Bridging Auxiliary Constraints to Resolve Instruction Following in Large Reasoning Models https://aclanthology.org/2026.tacl-1.60/ . Transactions of the Association for Computational Linguistics , 14:1329–1347. - Cite Informal : - Bridging Auxiliary Constraints to Resolve Instruction Following in Large Reasoning Models https://aclanthology.org/2026.tacl-1.60/ Zhao et al., TACL 2026 - PDF: - https://aclanthology.org/2026.tacl-1.60.pdf https://aclanthology.org/2026.tacl-1.60.pdf