The Challenge

Traditional approaches to understanding LLM reasoning often struggle to separate the semantic content of what models are processing from the behavioral patterns of how they process it. When we intervene in a model’s hidden states, are we changing what it “thinks about” or how it “thinks”? This fundamental question has important implications for both interpretability and control.

Our Approach: Meaning-Removed Control Vectors

Our key innovation is constructing control vectors that isolate behavioral patterns from semantic content. Here’s how it works:

1. Position-Matched Rephrasing

We create pairs of sentences that have the same syntactic structure and reasoning requirements but different semantic content. For example:

• Original: “If all roses are flowers, and this is a rose, then…”
• Rephrased: “If all cars are vehicles, and this is a car, then…”

2. Vector Construction

By comparing the hidden state representations of these matched pairs, we extract what we call “meaning-removed control vectors” (v_r^-). These vectors capture the reasoning process while factoring out the specific content.

3. Causal Interventions

We then inject or ablate these vectors at different layers of the model to observe their effects on reasoning behavior.

Key Findings

Our preliminary results show that these interventions can:

• Substantially improve reflective reasoning (ΔR) - the model becomes better at double-checking its own logic
• Minimally affect transition and execution behaviors (ΔT, ΔE) - the model’s basic language generation remains intact
• Provide interpretable insights into where and how reasoning happens in transformer architectures

Validation and Reproducibility

We’ve implemented rigorous validation through:

• Cosine similarity analysis to verify that our vectors capture meaningful patterns
• Nearest-neighbor analyses to understand the semantic space around our interventions
• Re-sampling stability filters to ensure robustness across different examples
• Layer-wise ablation studies to map reasoning processes across the model

Broader Impact

This work has implications beyond just understanding LLMs:

• AI Safety: Better control over reasoning processes could help prevent hallucinations
• Education: Understanding how models reason could inform how we teach reasoning to humans
• Tool Development: More controllable reasoning could enable better AI assistants

Open Science

As part of our commitment to reproducible research, we’re releasing:

• Complete tooling for vector extraction and analysis
• Comprehensive tutorials and documentation
• Example datasets and validation scripts

This work is also contributing to the NSF C2D project (Award #2321054), where I’ve been developing educational materials and tutorials to help other researchers understand these techniques.

What’s Next?

We’re currently preparing our manuscript for submission to EACL 2026. The next steps include:

• Expanding to more diverse reasoning tasks
• Testing on larger model architectures
• Developing real-time intervention techniques for practical applications

Conclusion

Understanding how LLMs reason is one of the most important challenges in modern AI research. By separating the “what” from the “how” of reasoning, meaning-removed steering vectors offer a new lens for both interpreting and controlling these powerful systems.

If you’re interested in learning more about this work or discussing potential collaborations, feel free to reach out at hyoo@nd.edu.

This blog post describes ongoing research at the MINE Lab, University of Notre Dame, under the supervision of Prof. Xiangliang Zhang. The work is currently under submission for peer review.