The Evidence Path

Agentic search policies

7 of 12 · Agents · 12 named systems

Agentic search systems learn or specify how to query, inspect evidence, continue reasoning, and stop under external budgets.

Agentic search policies: architecture comparisonEach lane reads from input to output.

Reason and act

  1. reasonUnresolved question
  2. actSearch query
  3. observeRetrieved result
  4. repeatUpdated state
  5. stopAnswer or abstain
  • ReAct
  • Self-Ask

Outcome RL

  1. sampleSearch trajectory
  2. scoreFinal answer reward
  3. optimizePolicy update
  4. auditSearch cost and faithfulness
  • Search-R1
  • ReSearch
  • R1-Searcher

Process and value RL

  1. stateEvidence state
  2. valueRetrieve or stop
  3. rewardProcess quality
  4. selectNext action
  5. terminateBounded result
  • StepSearch
  • Q-RAG
  • DeepRAG
  • HiPRAG

Integrated control

  1. emitControl token
  2. retrieveExternal evidence
  3. resolveConflict and utility
  4. citeEvidence selection
  5. answerFinal output
  • GRIP
  • Knowledgeable-R1
  • RAG-RL

Read the full explanation