aiqbench analysis
R2IF addresses the critical misalignment between reasoning traces and tool-execution decisions in LLMs by introducing a composite reward framework optimized via GRPO. This approach significantly boosts function-calling accuracy and interpretability on standard benchmarks, offering a more reliable foundation for deploying tool-augmented models. The method's focus on aligning Chain-of-Thought effectiveness with actual outcomes provides a concrete path forward for improving the trustworthiness of agentic systems.
R2IF: Aligning Reasoning with Decisions via Composite Rewards for Interpretable LLM Function Calling
arXiv:2604.20316v2 Announce Type: replace Abstract: Function calling empowers large language models (LLMs) to interface with external tools, yet existing RL-based approaches suffer from misalignment between reasoning processes and tool-call decisions. We propose R2IF, a reasoning-aware RL framework for interpretable function calling, adopting a composite reward integrating format/correctness constraints, Chain-of-Thought Effectiveness Reward (CER), and Specification-Modification-Value (SMV) reward, optimized via GRPO. Experiments on BFCL/ACEBench show R2IF outperforms baselines by up to 34.62% (Llama3.2-3B on BFCL) with positive Average CoT Effectiveness (0.05 for Llama3.2-3B), enhancing both function-calling accuracy and interpretability for reliable tool-augmented LLM deployment.
Abstract:Function calling empowers large language models (LLMs) to interface with external tools, yet existing RL-based approaches suffer from misalignment between reasoning processes and tool-call decisions. We propose R2IF, a reasoning-aware RL framework for interpretable function calling, adopting a composite reward integrating format/correctness constraints, Chain-of-Thought Effectiveness Reward (CER), and Specification-Modification-Value (SMV) reward, optimized via GRPO. Experiments on BFCL/ACEBench show R2IF outperforms baselines by up to 34.62% (Llama3.2-3B on BFCL) with positive Average CoT Effectiveness (0.05 for Llama3.2-3B), enhancing both function-calling accuracy and interpretability for reliable tool-augmented LLM deployment.
Machine Learning (cs.LG)
arXiv:2604.20316 [cs.LG]
(or arXiv:2604.20316v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2604.20316 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 179 words
Related coverage
Theorem-Grounded Execution Ontologies for Interpretable Machine Reasoning
arXiv · Jun 16, 2026
Outcome-Guided Distillation: A Teacher-Student Framework to Advance VLM Reasoning in Autonomous Driving
arXiv · Aug 3, 2026
Latent reasoning architectures would likely undermine CoT, our strongest oversight tool
AI Alignment Forum · Sep 23, 2026
Retro-Expert: Collaborative Reasoning for Interpretable Retrosynthesis
arXiv · Jun 16, 2026