DCI paper accepted at NeurIPS 2026: no retriever beats retrievers for agentic search
yuz9yuz · x · 2026-09-25
The DCI (Direct Corpus Interaction) paper, led by Zhuofeng Li with Wenhu Chen, Pan Lu, Jimmy Lin, Dongfu Jiang and others, is accepted at NeurIPS 2026.
Core claim: "the best retriever for agentic search is no retriever"
- Modern retrieval systems compress corpora into a fixed similarity interface with one top-k step, which bottlenecks agentic tasks: exact lexical constraints, sparse clue conjunctions, local context checks, and multi-step hypothesis refinement are hard via off-the-shelf retrievers, and evidence filtered out early can't be recovered by stronger downstream reasoning.
- DCI lets an agent search the raw corpus directly with terminal tools (grep, find, head, shell commands, light scripts) — no embedding model, no vector index, no retrieval API, no offline indexing, adapting naturally to evolving local corpora.
- In practice, just Claude Code or a simpler terminal-based coding harness plus grep/find beats competitive search agents, with no BM25/BERT/index involved; validated across IR benchmarks and end-to-end agentic search tasks.
Paper and code are available.
More from coding & agent
- Firecrawl's Alexandria lands in Claude: 113M+ sources, 21% better agent answers — devdigest · 2026-09-25
- Dev Builds Smart Home Controller With GitHub Copilot and Jev: 'Hype Is Well Deserved' — DanWahlin · 2026-09-25
- Hyperagent launches Rooms, a multiplayer space where agents and humans work together — aakashgupta · 2026-09-25
- Block joins x402 Foundation to build agentic payments rails with Bitcoin Lightning — kleffew94 · 2026-09-25
- SmolDataEnvs open-sources 5K+ verifiable RL tasks for training small code models — ben_burtenshaw · 2026-09-25
- Anthropic engineer: crank Claude's effort setting for risky code, dial it down for routine tasks — every · 2026-09-25