Mind · Reasoning
ExtendedThinking
Scale inference compute via extended reasoning
Definition
Mechanism
Inference-time compute scaling where the model generates extended reasoning traces (potentially thousands of tokens) before producing a final answer. Unlike standard Chain-of-Thought, this pattern explicitly trades latency and compute cost for accuracy by allowing the model to 'think longer'. The reasoning budget can be user-controlled or adaptive based on Task complexity.
Preconditions
- Complex prompt
- High token budget
Postconditions
- High-accuracy answer
Invariants
- The reasoning budget is fixed before the trace begins.
- A trace is produced and logged for every answer.
- Emitted, not certified: the trace is what the model produced, not proof of how it concluded.
Risks
Failure modes
Confabulation Chains
Long reasoning traces that sound plausible but contain errors.
Overthinking
Spending compute on simple problems that don't benefit.
Hidden Reasoning
Model reaches conclusions via paths not shown in the trace.
Relationships
Builds on
Accepts
Infrastructure
The atomic unit of intent
References
Infrastructure
Sequential data container (linked list)
Used by
References
High-confidence classification from minimal data
Reference
Parameters
budget_modeEnum, {UserControlled, Adaptive}- Whether the caller sets the reasoning budget or the {{task}}'s complexity does.