AI reasoning has progressed rapidly, with large language models now supporting multi-step work across many domains. Yet long-horizon reasoning remains unreliable: early errors propagate silently through later steps, without principled local revision. This project replaces the text trace with an editable workspace: a graph based state of claims, evidence, and dependencies. When something changes, the system performs a minimal-change repair on the affected region instead of rerunning the chain.

Multi-step reasoning failures already cause real harm, from fabricated citations to unsafe medical advice. Current methods are limited in a crucial respect: reasoning is free form text with no explicit dependencies or constraints, so revision is not controlled but a byproduct of rerunning the chain. Recent work shows that making intermediate reasoning structured enables much more reliable correction, suggesting reliability comes from how reasoning is represented, not just searched.

We will implement this by first executing graph-based state construction based on structured belief maintenance. Next, we will develop the dynamics of this state through minimal repair dynamics, utilizing search and learning-based policy development to ensure cost-bounded updates. Finally, evaluation will feature mid-run interventions to test stability under sudden changes, alongside reusable motif distillation to capture and adapt optimized pathways.

Sayambhu Sen

‍