| name | over-searching-control |
| title | Over-Searching in Search-Augmented Large Language Models |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2601.05503 |
| keywords | ["search-control","retrieval-efficiency","answer-confidence","hallucination-prevention","token-optimization"] |
| description | Diagnose and prevent over-searching—unnecessary search invocations that degrade LLM performance and waste tokens. Framework introduces Tokens Per Correctness (TPC) metric and OverSearchQA dataset to train models that know when NOT to search. Teaches agents to recognize answerable vs. unanswerable queries and selectively invoke search only when needed, improving both accuracy and efficiency. |
Problem
Search-augmented LLMs exhibit a critical but overlooked inefficiency:
- Over-Searching: Models invoke search tools unnecessarily for questions they could answer from training knowledge
- Context Pollution: Irrelevant search results introduce noise that degrades reasoning
- Contradictory Evidence: Retrieved documents may contradict model knowledge, confusing final answers
- Abstention Failure: Models particularly struggle to abstain (say "I don't know") when search returns irrelevant documents
- Token Waste: Unnecessary search rounds consume tokens that could improve other aspects of performance
Current systems treat search as always beneficial, but indiscriminate search actually harms performance on unanswerable questions.
Solution
Over-Searching Analysis introduces three components:
- Tokens Per Correctness (TPC): A cost-aware metric specifically designed for search-augmented systems
- Measures: How many tokens required per correctly answered question?
- Accounts for both search cost and reasoning quality
- Selective Search Triggering: Train models to recognize question types:
- Answerable from training knowledge → skip search
- Requires Current Information → search needed
- Factual Lookup → search optimal
- Unanswerable → don't search (prevents hallucination)
- Evidence Quality Management: Incorporate negative evidence (counter-signals) to teach models when to ignore irrelevant search results
When to Use
- Token-Constrained Systems: Production systems where search cost matters
- Hybrid QA Agents: Systems combining training knowledge with retrieval
- Retrieval-Augmented Generation: Any RAG pipeline seeking efficiency gains
- Low-Confidence Detection: Agents that must recognize and abstain on unanswerable queries
- Multi-Turn Dialogue: Long conversations where accumulated search costs compound
When NOT to Use
- For pure retrieval systems (search is the only knowledge source)
- When knowledge freshness is critical (always search for current info)
- In systems where computational cost is unconstrained
