LongR: RL with dense utility rewards boosts long-context reasoning by 9% on LongBench v2
PKU · hf · 2026-10-14
Researchers from Peking University, SJTU and Tsinghua propose LongR, a framework addressing sparse outcome rewards in long-context reasoning. It combines a dynamic 'Think-and-Read' mechanism interleaving reasoning with document consultation, plus a contextual density reward based on relative information gain—no evidence annotations needed. LongR gains over 9% on LongBench v2 with consistent improvements on RULER and InfiniteBench, works across RL algorithms (DAPO, GSPO), and includes analyses of reasoning chain length and robustness to distractors.
More from Research
- RUC and Kling AI propose EventLens: event-structure RL to fix video LLMs' temporal understanding — Guowen Zhang · 2026-10-14
- JSIR dataset drops: 267 jazz recordings with beat, chord and structure annotations — umpedronosapato · 2026-10-14
- There is no such thing as an AI drug, argues researcher Patrick Kidger — PatrickKidger · 2026-10-14
- More thinking, worse scores: refusals tank closed-model evals at max reasoning effort — _lewtun · 2026-10-14
- Integer multiplication repo becomes an agent harness benchmark — and collaboration loses — iskander · 2026-10-14
- Recurrent self-conditioning: framing recent diffusion methods as depth decoupled from time — alec_helbling · 2026-10-14