Reasoning Budgets
When should a system keep its answer, think again or seek new information?
Start with the first report → See all reports ↓Overview
More computation creates more opportunities to repair an answer, but it can also overwrite a good answer, add delay or repeat an unhelpful operation. Choosing effort therefore means choosing an action and evaluating its consequences, rather than treating a longer trace as an automatic improvement.
The earlier report examines a trained stop/refine procedure for a reasoner under outcome checks. The agent-focused report widens the choice: another candidate, a verification step, retrieval or human input may be more useful than further internal reasoning. Together, the two articles explain why uncertainty can help allocate effort while leaving the value and full cost of the particular intervention to be tested.
The next unit of effort can change the answer or change the action
Keep or refine an answer
Learn a stopping decision that preserves useful solutions.
Choose the next intervention
Compare extra reasoning with checking, retrieving or asking for help.
Two control problems at different scopes. Their interventions and guarantees are evaluated in different settings.
Reading order
A recommended path through the enquiry. Each report explains its part; continuation links preserve the original branches.
- Control the refinement loopHow a Reasoner Learns When to Stop
Introduces the repair-versus-damage problem and examines how training and inference use verdicts, confidence and stopping differently.
Start here - Price the next useful actionWhen is an agent's fast decision good enough?
Extends the decision to an interactive agent, where missing information or authority can require a different action rather than more thought.