Entity · paper

Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering

paperactiveprovisionalcan-we-break-llms-out-of-self-loops-fine-grained-reasoning-control-with-activation-steering-09d65c96·1 events·first seen 9h ago

Aliases: Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering

Co-occurring entities

SOPHIA

More like this (12)

Agentic Chain-of-Thought Steering for Efficient and Controllable LLM Reasoning Forecasting With LLMs: Improved Generalization Through Feature Steering Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs Predicting Future Behaviors in Reasoning Models Enables Better Steering Beyond the Commitment Boundary: Probing Epiphenomenal Chain-of-Thought in Large Reasoning Models Can LLMs Judge Better Than They Generate? Evaluating Task Asymmetry, Mechanistic Interpretability and Transferability for In-Context QA When the Chain of Thought Knows Better: Failure Modes in Multi-Turn Reasoning Models Extending LLM Context via Associative Recurrent Memory Scaling LLM Reasoning from Minimal Labels: A Semi-Supervised Framework with a Lightweight Verifier Operads for compositional reasoning in LLMs ExpRL: Exploratory RL for LLM Mid-Training Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning

Recent events (1)

6arXiv · cs.AI·9h ago·source ↗

SOPHIA: Activation steering to break LLMs out of reasoning self-loops at inference time

Researchers introduce SOPHIA (Steering Of reasoning Processes via Hidden-state Intervention and Activations), a method for fine-grained control over LLM reasoning traces via inference-time activation steering. The approach models each reasoning trace as a sequence of latent states, builds a bank of steering vectors indexed by state-pair transitions, and uses a controller to detect and intervene on self-loops — failure modes where models exhaust their token budget without progress. Experiments show SOPHIA reliably breaks self-loops and improves both end-task accuracy and token efficiency, with steering vectors that generalize across state pairs.

Evaluation and Benchmarking Inference Economics Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering SOPHIA +1 more