SEEDS: search agents preview and pick evidence by reasoning state, beating fixed top-k
_reachsumit · x · 2026-10-12
The SEEDS paper tackles overlooked 'evidence admission' in search agents: existing systems conflate retrieving with consuming evidence, taking fixed or rank-based top-k prefixes even though evidence utility depends on the agent's current reasoning state — a low-ranked candidate may be critical while a high-ranked one is redundant or misleading.
SEEDS exposes lightweight previews of retrieved candidates and lets the agent jointly decide which candidates, and how many, to inspect in full. Unlike prefix policies it selects arbitrary subsets with flexible cardinality, and can be instantiated with an existing reasoning agent or further optimized via learning.
Across nine knowledge-intensive QA benchmarks under both local corpus and live web search, SEEDS consistently improves search effectiveness.
More from coding & agent
- Meta's ReHarness adds backward revision to agent harnesses, beating memory and reflection baselines — Yuxin Xiao · 2026-10-12
- Dev open-sources tool using local VLMs to catalog and auto-suggest his 5,600-LoRA collection — EricRollei · 2026-10-12
- Tencent Yuanqi agent platform shuts down Nov 9, stranding developers' bots and users — sujingshen · 2026-10-12
- Blender MCP didn't work for zeeg—he's getting better results with Tripo — zeeg · 2026-10-12
- "I don't want to look at code, just state machines and live docs": one giant diagram — _AustinCalvert_ · 2026-10-12
- VectorDash: open-source Windows 11 search replacement powered by EmbeddingGemma 2 — andy2na · 2026-10-12