DOLORES: Training-Free Meta-Scaffold Lifts 8B Models Past 32B Baselines
StellaLisy · x · 2026-11-04
A University of Washington team introduces Deep Reasoning and its DOLORES meta-harness at NeurIPS 2026: a training-free, inference-time approach that recursively decomposes tasks into formal and associative reasoning steps, building task-specific agent scaffolds just-in-time instead of hard-coding fixed structures.
Key results:
- 24.8% average improvement over SOTA scaffolds (including reasoning models and Deep Research-style methods)
- Fully training-free
- An 8B model beats same-family 32B baselines in more than half of settings
- Wins across 4 hard benchmarks: multi-hop reasoning, long-chain QA, long-context aggregation, deep-research-style retrieval
Mechanism: Baselines fail by overloading single LLM calls (premature termination, hallucination); DOLORES distributes cognition across structured, lower-load reasoning threads. Code open-sourced at github.com/DeanLight/dolores.
More from coding & agent
- YC-backed Simantic adds a simulation layer so AI agents can test firmware at software speed — ycombinator · 2026-11-04
- Bittensor Subnet 33 Launches Skill Store: Tested Skills Aim to Make Open Models Do Frontier Work — markjeffrey · 2026-11-04
- Building a 24-hour AI newsroom on Microsoft Copilot Harness: workflow beats prompts — WirelessLife · 2026-11-04
- Polyphonic to give AI agents their own Mac via open-source terminal-mac — RileyRalmuto · 2026-11-04
- Dev plans Rust open-source rewrite of ResMed CPAP firmware, slams Anthropic classifier blocks — ryunuck · 2026-11-04
- Salesforce open-sources Pronto sample app showcasing Headless Toolkit and hosted MCP servers — msrivastav13 · 2026-11-04