Training-free semantic trees beat full-context prompting: +1.72 QA-F1 with 58.55% less context
_reachsumit · x · 2026-10-28
An arXiv paper introduces a training-free framework that builds a reusable semantic hierarchy of a long document in a single generation, linking node descriptions to source spans. For a query, structural-semantic scoring and fine-grained lexical attribution select shared source nodes, unioned with ancestor context to preserve structure and complete local text. Across four frontier answer models on three LongBench tasks, it improves average QA-F1 by 1.72 points over full-context prompting while cutting answer-context length by 58.55%. Code is open-sourced.
More from Research
- Ex-OpenAI researcher's 26,000-word treatise claims AI's hardest interpretability problem just got easier — nordicinst · 2026-10-28
- Force field experts debate: zinc coordination fitting and non-scalable empirical parameterization — TimothyDuignan · 2026-10-28
- TTME beats test-time training on 5 math tasks while cutting FLOPs by 82.8-84.6% — UCSantaBarbara · 2026-10-28
- EnvLunch turns executable environments into free verifiable agent training data — KingsCollegeLondon · 2026-10-28
- Anthropic reveals it is developing its own experimental drugs using Claude — Polymarket · 2026-10-28
- Stanford launches TaskNet: 132 worker interviews build a large-scale occupational task inventory — soumitrashukla9 · 2026-10-28