Dynamic Causal Interventions: On-Policy Editing and Real-Time Resampling for Interactive LLM Reasoning

TL;DR: Imagine letting a human interactively tweak reasoning steps and instantly see how new resampled CoTs adapt—like real-time model debugging! We’ll build an interface for dynamic, on-policy intervention and measure how LLMs adapt their downstream reasoning.

Research Question: How do LLM reasoning distributions dynamically adapt to real-time, user-driven edits of intermediate CoT steps, and can this process be used for targeted model steering or debugging?

Hypothesis: Interactive, on-policy resampling after user interventions yields more stable and interpretable downstream adaptations than traditional off-policy edits, especially for critical reasoning junctures.

Experiment Plan: Develop an interface where users can edit, insert, or delete reasoning steps in a sampled CoT, then trigger resampling of downstream steps. Compare the stability, faithfulness, and causal influence of edits using on-policy (resample-from-edit) vs. off-policy (edit-then-complete) approaches. Test on tasks with known critical steps (e.g., blackmail scenarios, complex planning) and measure resilience and model steerability. Expected outcome: On-policy resampling produces more predictable and causally coherent adaptations, enabling new forms of LLM explanation and intervention.

References:

• 1. Debjit Paul, West, R., Bosselut, A., & Faltings, B. (2024). Making Reasoning Matter: Measuring and Improving Faithfulness of Chain-of-Thought Reasoning. Conference on Empirical Methods in Natural Language Processing. 2. Tamera Lanham, Chen, A., Radhakrishnan, A., et al. (2024). DeCoT: Debiasing Chain-of-Thought for Knowledge-Intensive Tasks in Large Language Models via Causal Intervention. Annual Meeting of the Association for Computational Linguistics. 3. Tinghui Zhu, Zhang, K., Xie, J., & Su, Y. (2024). Deductive Beam Search: Decoding Deducible Rationale for Chain-of-Thought Reasoning. arXiv.org.