R3Con: training-free harness lets 3B models beat Claude Sonnet agents at 3.7x lower cost
StellaLisy · x · 2026-10-15
University of Washington researchers introduce R3Con, a training-free harness grounded in the cognitive theory of Relevance Realization for building principled context representations over large corpora. Key points:
- Outperforms the strongest of nine SOTA baselines by 20 and 8.4 percentage points on two large-document reasoning benchmarks.
- R3Con with 4B/9B models beats all 35B-class baselines; R3Con + Qwen3.5-35B-A3B hits 71% on CorpusQA vs 70%/67% for Claude Code with Claude Sonnet, at $439 vs $1625/$1714 total cost (3.7-3.9x cheaper).
- Suggests principled context representation can reduce reliance on model scale.
Code is open-sourced on GitHub.
More from Research
- LLM4Code workshop at ICSE 2027 opens call for papers, due Nov 13, 2026 — LingmingZhang · 2026-10-15
- ELSA3D, a unified 3D generation and captioning model, hits SOTA at half the FLOPs, accepted at NeurIPS — CSProfKGD · 2026-10-15
- Moving ViT weights 1.3% without changing predictions scrambles representations, challenging Platonic hypothesis — neuroecology · 2026-10-15
- NeurIPS Spotlight: time-reversed LMs (TRLM) boost best-of-N reranking by up to 5% — LucaAmb · 2026-10-15
- Next-Concept Models Claim 31x Faster Text Generation Than Diffusion LMs — HazyResearch · 2026-10-15
- NCM iteratively refines its way to solving a sudoku puzzle — HazyResearch · 2026-10-15