Paper Reference
- Title: Interactive Reasoning: Visualizing and Controlling Chain-of-Thought Reasoning in LLMs
- Authors: RY Pang, KJK Feng, S Feng, C Li, W Shi, et al.
- Year: 2026
- Venue: IUI 2026 (ACM International Conference on Intelligent User Interfaces)
- URL: https://dl.acm.org/doi/10.1145/
Paper Summary
Introduces an interaction design that visualizes chain-of-thought outputs as a hierarchy of topics and enables user review and modification. User study shows users can quickly identify erroneous generations and efficiently steer models through direct manipulation of reasoning steps.
Proposed Feature
Implement interactive reasoning editing that allows users to modify agent reasoning in real-time:
Core Capabilities
- Hierarchical CoT Display: Show reasoning chains as expandable/collapsible topic hierarchies
- Inline Editing: Allow users to edit, remove, or add reasoning steps directly in the UI
- Live Replay: After editing, re-execute the agent from the modified point to see the effect
- What-If Scenarios: Save edited reasoning paths as named scenarios for comparison
Technical Approach
- Extend the session replay system to support mid-session modification
- Add editable reasoning step components in the frontend
- Implement "branch and replay" capability from any reasoning step
- Add scenario management for saving and comparing edited paths
Impact
Transforms debugging from passive observation to active experimentation. Users can test "what if I changed this decision?" directly in the debugger. This is a killer feature that no current agent debugger offers.
Labels
enhancement, paper-inspired, high-priority, frontend
Paper Reference
Paper Summary
Introduces an interaction design that visualizes chain-of-thought outputs as a hierarchy of topics and enables user review and modification. User study shows users can quickly identify erroneous generations and efficiently steer models through direct manipulation of reasoning steps.
Proposed Feature
Implement interactive reasoning editing that allows users to modify agent reasoning in real-time:
Core Capabilities
Technical Approach
Impact
Transforms debugging from passive observation to active experimentation. Users can test "what if I changed this decision?" directly in the debugger. This is a killer feature that no current agent debugger offers.
Labels
enhancement, paper-inspired, high-priority, frontend