DualGraph 架构在 Sol 与 Opus 5.5 上跑出高分
微软研究者提出双图记忆架构，用 GPT-6 Sol、Luna 与 Claude Opus 5.5 组合在 DeepResearch Bench 拿到 53.08 RACE 分数。
this is pure f*cking insane Claude Opus 5.5 and GPT-6 Sol broke the autonomous research ceiling Standard deep research agents suffer from outline-centric tunnel vision. When an agent locks into predetermined section headings before gathering data, it issues narrow queries, ignores high-signal peripheral entities, and hallucinates text to fill structural gaps Researchers from Microsoft introduced DualGraph, an architecture that splits agent memory into an Outline Graph for document presentation and a dynamic Knowledge Graph for semantic exploration. Running this system across Claude Opus 5.5 and OpenAI GPT-6 Sol sets an unprecedented 53.08 RACE score on DeepResearch Bench Outline Graph directives → Topological curiosity search → GPT-6 Luna extraction → Cross-graph attention sync → Claude Opus 5.5 synthesis Four structural mechanisms driving the dual-memory research engine: 1. Decoupled Presentation and Discovery The Outline Graph governs section boundaries while the Knowledge Graph maps entities independently. This structural separation stops premature outline locks and eliminates search blind spots 2. Topological Curiosity Engine The agent calculates graph sparsity and node degree variance to target unmapped concept clusters instead of issuing generic conversational search queries 3. High-Volume Extraction via GPT-6 Luna High-frequency entity extraction runs on GPT-6 Luna at $0.10 per million tokens. This specialized worker layer slashes exploratory compute expense by 68.2% while maintaining complete factual density 4. Harness Provenance and Factual Verification Claude Opus 5.5 synthesizes final report sections using verified citations from the Knowledge Graph. Factual grounding reaches 98.6% while $0.20 cache reads reduce inference expense by 40% The master agent architecture proves that unconstrained prompt loops produce expensive hallucinations. Unifying deterministic harness verification, curiosity-driven loops, and dual-graph memory turns stochastic LLM calls into rigorous research systems Read the complete breakdown in the article below ↓
这条是引用着另一条写的：x.com/marfinxx/status/2081687570488954915