VecTree-RAG combines vector search and tree traversal for scientific QA
_reachsumit · x · 2026-07-28
VecTree-RAG proposes an agentic retrieval framework that separates two tasks in scientific QA: finding relevant papers and localizing evidence inside them.
- It combines vector search for candidate paper/section ranking with reasoning-guided tree traversal for evidence localization.
- Full text is kept in a page store and revealed progressively after structural localization.
- The paper evaluates on QASPER (300 questions), LitQA2 (54 questions), and MOSAIC (49 questions).
- It reports the best observed answer scores among compared systems, including 0.800 LLM-judge correctness on QASPER, 0.925 accuracy on LitQA2, and 0.547 composite score on MOSAIC.
The main contribution is a retrieval design that preserves document structure instead of flattening it into passages.
More from Research
- RL Does More Than Sharpen SFT: Longer Pretraining Yields Steeper RL Scaling Slope — gleech · 2026-08-24
- Reservoir & Photonic Computing: Escaping the O(n²) Trap of LLMs — navnt5 · 2026-08-24
- Open source Pelican SVG Env quantifies model drawing abilities — SergioPaniego · 2026-08-24
- Beyond Final Weights: The Value of Stage-Level Checkpoint Transparency — creditme7 · 2026-08-24
- 17-year-old who independently pretrained models gets ICML paper accepted — HanchungLee · 2026-08-24
- CAS-Spawned ScienceClaw Launches AutoProject: AI Research Agents Move From Tasks to Whole Projects — 量子位 · 2026-08-24