Search agents improve reasoning by choosing evidence selectively
SEEDS: Preview-Guided Evidence Admission for Search Agents
Computation and Language
Summary
When search agents look for information to answer questions, they often take in all the top results they find, even if some are not useful for their current thinking. The authors noticed that this approach can slow down or mislead the agent. They created SEEDS, a method that helps these agents peek at summaries of the information first and then decide which details to fully read based on what they already know. This way, agents avoid wasting time on irrelevant documents and improve their answers. The authors tested SEEDS on many question-answering tasks and found it both speeds up processing and makes the answers better.
What this means in practice
- •For software engineers: Build AI assistants that selectively read documents to answer questions more efficiently and accurately by previewing content first.
- •For enterprise search teams: Improve internal search tools to deliver focused evidence for complex queries without processing all retrieved documents fully.
Authors
Jiajin Liu, Chenxi Xu, Dongzhe Fan, Yuanfu Sun, Qiaoyu Tan
Abstract
Search agents extend large reasoning models with multi-round access to external knowledge, yet most existing systems largely conflate retrieving evidence with consuming it. Retrieved documents are commonly incorporated according to fixed or rank-based policies, despite the fact that evidence utility is conditioned on the agent's current reasoning state. Consequently, a lower-ranked candidate may be critical for the current step, while a higher-ranked candidate may provide redundant, irrelevant, or misleading information. We formulate this overlooked problem as evidence admission, which asks, at each search step, which retrieved candidates should be allowed to enter the agent's evolving reasoning context. To enable such selective control, we introduce SEEDS (Selective Evidence Entry During Search), a preview-guided evidence admission framework for search agents. SEEDS first exposes lightweight previews of retrieved candidates and then lets the agent jointly determine which candidates and how many should be fully inspected, conditioned on its current reasoning state. Unlike fixed or adaptive prefix policies, SEEDS can select arbitrary subsets with flexible cardinality, and it can be instantiated directly with an existing reasoning agent or further optimized through learning. Across nine knowledge-intensive QA benchmarks under both local corpus and live web search, SEEDS consistently improves search effectiveness while substantially reducing full-evidence consumption. Experimental analyses show that SEEDS more effectively selects useful candidates beyond the corresponding retrieval prefix. These results establish evidence admission as a complementary dimension of agentic search, separating what information is retrieved from what information is ultimately allowed to shape reasoning.