NeurIPS Spotlight Paper DCI: Ditching Top-k Retrieval for grep-Style Agentic Search
stanfordnlp · x · 2026-09-25
Stanford NLP researcher Dongfu Jiang announced that "Beyond Semantic Similarity" (DCI) was accepted to NeurIPS as a Spotlight. The paper argues that top-k retrieval is a compressed interface that bottlenecks agentic search: exact lexical constraints, sparse clue conjunctions, and multi-step hypothesis refinement are hard to express through a fixed semantic-similarity retriever, and evidence filtered out early can't be recovered.
The team's DCI-Agent drops the embedding model, vector index, and retrieval API entirely — instead letting the agent operate directly on large-scale local corpora via terminal-style tools like grep, read, bash, regex, and lightweight scripts. The result: higher accuracy at lower token cost.
Key points:
- NeurIPS Spotlight, co-authored by Wenhu Chen, Jimmy Lin, Yejin Choi, James Zou, Jiawei Han and others
- Code and paper are openly available
- Premise: as agents get stronger at planning, coding, and tool use, retrieval should expose the raw corpus rather than a retriever's compressed view
Related event: DCI Paper Named NeurIPS 2026 Spotlight: grep Beats top-k Retrieval(3 posts)→
More from coding & agent
- Opus 5.5 draws a collage-style animation frame by frame with p5.js code — oran_ge · 2026-09-25
- NeurIPS 26 paper accepted; workshop organizers ask what agent behaviors lie beyond planning and reasoning — DanielKhashabi · 2026-09-25
- Building an Obsidian carousel editor with just Claude Code, skills and markdown files — evielync · 2026-09-25
- Baseten Loops fine-tunes open models on LangSmith traces for self-improving agents — BraceSproul · 2026-09-25
- Pseudo Self-Distillation Trains Small Models to Match GPT-4.1-mini on Agent Memory Retrieval — burny_tech · 2026-09-25
- Replit Animation + Opus: a working JS animation video generated in under 30 minutes — amasad · 2026-09-25