How the advisor learns, and how it is used
Training happens once, offline. At test time the agent is unchanged; it only receives a short plan.
01Hindsight
Study failures
Run the agent on past questions, grade every answer, and note why weak answers fell short. The notes become a checklist of recurring gaps.
→
02Hindsight
Try strategies
Answer every question several ways and record which way actually scored best.
→
03Structured advice
Write the ideal plan
A teacher that can see the grading and every answer writes the plan that would have helped, as a structured form.
→
04Distillation
Train the advisor
A small model learns to write that plan from the question alone.
At test time
Question
→
Advisor writes a plansmall model, one short call
→
Samaya agent answersquestion + plan, same tools
Figure 2: Hindsight structured advice distillation. The advisor learns from the agent's graded history which plan would have helped, and at test time it writes that plan from the question alone.