RISER: Orchestrating Latent Reasoning Skills for Adaptive Activation Steering

TL;DR

RISER tackles the challenge of adaptive, parameter-efficient reasoning in large language models by introducing a Router-based activation-intervention framework. It builds a compact library of latent reasoning primitives and learns a lightweight Router to dynamically compose and inject these primitives into the model’s activation space during inference, optimized via reinforcement learning. The approach achieves 3.4–6.5% absolute zero-shot accuracy gains and 2–3x improvements in token efficiency across seven benchmarks, while providing interpretable, compositional control over reasoning. This work advances controllability and efficiency in LLM reasoning, enabling dynamic, task-aware coordination of latent cognitive skills without weight updates.

Abstract

Recent work on domain-specific reasoning with large language models (LLMs) often relies on training-intensive approaches that require parameter updates. While activation steering has emerged as a parameter efficient alternative, existing methods apply static, manual interventions that fail to adapt to the dynamic nature of complex reasoning. To address this limitation, we propose RISER (Router-based Intervention for Steerable Enhancement of Reasoning), a plug-and-play intervention framework that adaptively steers LLM reasoning in activation space. RISER constructs a library of reusable reasoning vectors and employs a lightweight Router to dynamically compose them for each input. The Router is optimized via reinforcement learning under task-level rewards, activating latent cognitive primitives in an emergent and compositional manner. Across seven diverse benchmarks, RISER yields 3.4-6.5% average zero-shot accuracy improvements over the base model while surpassing CoT-style reasoning with 2-3x higher token efficiency and robust accuracy gains. Further analysis shows that RISER autonomously combines multiple vectors into interpretable, precise control strategies, pointing toward more controllable and efficient LLM reasoning.

Figures (7)

• Figure 1: Conceptual comparison between Standard Inference and the RISER framework. RISER (bottom) uses a learned Router to dynamically inject composed vectors, analogous to an explicit executive-control mechanism.
• Figure 2: An overview of the RISER framework, illustrating the process of offline extraction of reasoning vectors and offline training of the Router, followed by online inference where the pre-trained Router dynamically selects and combines vectors to intervene in the LLM's activation, guiding the final output.
• Figure 3: Latent space visualization of extracted vectors. We project the high-dimensional difference vectors onto a 2D plane using PCA. The visualization reveals naturally forming clusters, demonstrating that the extracted reasoning vectors possess strong semantic separability within the activation space.
• Figure 4: Reasoning vector library similarity heatmap. The low off-diagonal values confirm that the extracted vectors represent distinct and separable cognitive functions.
• Figure 5: Static steering validation. The performance sensitivity to steering strength ($\alpha$) confirms that the extracted vectors effectively modulate specific reasoning behaviors.