The exposure of a bot’s reasoning process as a legible trail of steps and decisions, transforming opaque computation into something the actor can inspect, trust, and challenge over time.
Steps are arranged sequentially — like a conversation or narrative — with separation between user inputs, model responses, and intermediate reasoning. Actors can scroll, expand, comment on, or revisit earlier steps, and if needed, branch the thread to explore alternatives.
Temporal modes
Breaking down reasoning transparency by when it occurs in the execution lifecycle makes the timing distinction explicit.
Before execution: planning
TODO: Presents complete, verified information before commitment, building trust through transparency of upcoming actions.
During execution: stream of thought
Reveals reasoning as work unfolds, enabling monitoring and early intervention.
After execution: Retrospective analysis
TODO: Condensed summary after completion, supporting learning and debugging.
Representation structures
Beyond temporal positioning, reasoning can be structured as linear sequences or branching explorations.
- Linear: presentation breaks down complex tasks into a series of simpler, sequential subtasks, making the model’s reasoning process more interpretable and transparent. Most implementations of Stream of Thought and Action Plan use linear structures.
- TODO: Branching: The nature of reasoning in LLMs, and human reasoning more broadly, is more nuanced than a simple linear chain.
Placement
In a chat
…
In a task
…
Actions
…
Related components
- Messaging - provides the conversational interface where transparent reasoning can be displayed.
Resources & references
Rock Yuren Pang, K. J. Kevin Feng, Shangbin Feng, Chu Li, Weijia Shi, Yulia Tsvetkov, Jeffrey Heer, Katharina Reinecke. 2025. Interactive Reasoning: Visualizing and Controlling Chain-of-Thought Reasoning in Large Language Models
To-do
- Workflow unpacking — in an LLM-driven system, reasoning processes can be viewed as dynamically generated workflows that emerge at runtime rather than being predefined.
Related patterns
Precedes
- Activity log — provides the detailed reasoning steps that get recorded in the activity log.
Enacts
- Learnability — by observing *why* the system made a decision, the actor implicitly learns its logic and capabilities
- Agency — visible reasoning is one of the levers for reliance calibration, helping the actor judge when to trust and when to override
Complements
- AI completion — explaining *why* a completion was suggested, especially for derived values
- Annotation — system annotations that reveal decision logic.
- Cognitive forcing functions — counterbalancing interventions when transparent reasoning slides into over-reliance
- Collaboration — Transparent reasoning makes AI decision-making processes inspectable for human collaborators
- Explanation — encourages understanding through interaction, while transparent reasoning shows the step-by-step process.
- Generated content — shows the step-by-step process behind the generated content.
- Live presentation — its "stream of thought" mode rides on the same rendering substrate
- Next-best action — explaining the rationale behind recommendations
- Progressive disclosure — a specialized high-detail form of disclosure for AI systems
- Assistance
- Delegation
- Suggestion — can emerge from the visible reasoning process.
Related
- Assisted task completion
- Data entry — explaining *why* a specific interpretation was made.
- Prose
Preceded by
- Bot — generates the reasoning that transparent reasoning makes visible and inspectable.