Semantic IDs: How to train an LLM-Recommender Hybrid with steerability and reasoning on recommendations.

This repository aims to develop CoT Steering based on CoT without Prompting. It focuses on enhancing the model’s latent reasoning capability without additional training by leveraging Test-Time Scaling techniques.

[COLM 2025] SEAL: Steerable Reasoning Calibration of Large Language Models for Free

KV Cache Steering for Inducing Reasoning in Small Language Models

Official implementation of "Steering LLM Reasoning Through Bias-Only Adaptation" and "Small Vectors, Big Effects: A Mechanistic Study of RL-Induced Reasoning via Steering Vectors"

Official github repo for "Fractional Reasoning via Latent Steering Vectors Improves Inference Time Compute"

The official implementation for "Mitigating Overthinking in Large Reasoning Models via Manifold Steering"

CREST is a training-free test-time steering framework that discovers cognitive heads via simple offline calibration and then rotates activations during decoding to guide the model’s reasoning—preserving norms to avoid per-model hyperparameter tuning. This improves accuracy and reduces tokens across models and datasets.

Detecting and steering temporal reasoning in LLMs

Code for Exploring Chain-of-Thought Reasoning for Steerable Pluralistic Alignment (EMNLP 2025)

CQoT is an approach that leverages argumentation theories to enhance LLMs' reasoning and mathematical capabilities. We share the code implementing the pipeline introduced in "Critical-Questions-of-Thought: Steering LLM reasoning with Argumentative Querying".

Source code for the paper "Understanding Reasoning in Thinking Language Models via Steering Vectors"

a plug-and-play intervention framework that adaptively steers LLM rea- soning in activation space

CRSM (Continuous Reasoning State Model): An asynchronous "System 2" architecture that implements Hierarchical State Sovereignty within a Mamba backbone. Unlike traditional search wrappers, CRSM uses Forward-Projected Planning and Sparse-Gated Injection to steer latent manifolds in real-time, decoupling strategic reasoning from token generation.

Minimal code to extract and apply steering vectors to a model with latent reasoning from the paper "Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach"

Amortized Latent Steering (ALS) for Efficient LLM Reasoning

No description available

Is activation steering more powerful than prompting at mitigating deception in some current reasoning LLMs?

Analysed determinism, faithfulness, reasoning patterns, & steering. Developed and tested methods to enhance control and fail-safes

Our research analyzes the step-wise causality to examine the faithfulness of reasoning in CoT and reveals a steering direction to mediate true thinking in LLMs.

The dataset and code for "Framing Bias in Arithmetic Reasoning: How Language and Identity Cues Steer LLM Outputs in Objective Tasks" paper

This project studies faithfulness in chain-of-thought reasoning by testing whether unfaithful reasoning forms a distinct representational direction in large language models. We fine-tune models on unfaithful CoT data, use model diffing to find correlated features, and develop methods to detect or steer faithfulness for more transparent AI reasoning

Repo for Julian Bernado and Michael Hwang's CSE 598 project.

Finding a direction in activation space that can improve reasoning in the ARC-AGI challenge

use steering vectors to try to understand reasoning models a bit more

No description available

No description available

Method to extract reasoning activation vector in R1-Distilled Llama and adaptively steer reasoning to mitigate unecessary reasoning on easy questions

No description available

Researching the role of steering vectors in the backtracking mechanism