InjectRBP: Steering Large Language Model Reasoning Behavior via Pattern Injection

2026-02-12

AI Summary

Current approaches to reasoning optimization in large language models often rely on heuristic prompting, lacking systematic analysis of the models’ reasoning behavior patterns. This work addresses this gap by modeling adaptive reasoning from a behavioral perspective and proposes two parameter-free, injection-based guidance mechanisms: InjectCorrect, which mimics historically correct responses, and InjectRLOpt, which injects actions guided by a value function. To dynamically generate injection signals, the authors introduce a reliability-aware Softmax strategy. Experimental results demonstrate that the proposed method significantly enhances performance across diverse reasoning tasks, achieving accuracy gains of up to 5.34% and 8.67% respectively, without any model training or parameter updates.

Technology Category

• Search and Optimization: Metareasoning and Metaheuristics
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning
• Reasoning under Uncertainty: Stochastic Optimization

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Reasoning can significantly enhance the performance of Large Language Models. While recent studies have exploited behavior-related prompts adjustment to enhance reasoning, these designs remain largely intuitive and lack a systematic analysis of the underlying behavioral patterns. Motivated by this, we investigate how models'reasoning behaviors shape reasoning from the perspective of behavioral patterns. We observe that models exhibit adaptive distributions of reasoning behaviors when responding to specific types of questions, and that structurally injecting these patterns can substantially influence the quality of the models'reasoning processes and outcomes. Building on these findings, we propose two optimization methods that require no parameter updates: InjectCorrect and InjectRLOpt. InjectCorrect guides the model by imitating behavioral patterns derived from its own past correct answers. InjectRLOpt learns a value function from historical behavior-pattern data and, via our proposed Reliability-Aware Softmax Policy, generates behavioral injectant during inference to steer the reasoning process. Our experiments demonstrate that both methods can improve model performance across various reasoning tasks without requiring any modifications to model parameters, achieving gains of up to 5.34% and 8.67%, respectively.

Problem

• reasoning behavior
• behavioral patterns
• large language models
• prompt adjustment
• reasoning enhancement

Innovation

• behavioral pattern injection
• parameter-free reasoning optimization
• reliability-aware softmax
• reasoning steering
• large language model