LongR: Unleashing Long-Context Reasoning via Reinforcement Learning with Dense Utility Rewards
LongR adds reward signals for useful document consultation, not just final-answer success.
The paper proposes a long-context reasoning framework that interleaves “Think-and-Read” steps with reinforcement learning. Its dense utility reward estimates the value of relevant documents through relative information gain, without requiring evidence annotations. The authors report more than a 9% gain on LongBench v2, with improvements also on RULER and InfiniteBench. They also test the approach across RL algorithms including DAPO and GSPO, and analyze chain length efficiency and distractor robustness. HF Daily Papers' note
The paper proposes a long-context reasoning framework that interleaves “Think-and-Read” steps with reinforcement learning. Its dense utility reward estimates the value of relevant documents through relative information gain, without requiring evidence annotations. The authors report more than a 9% gain on LongBench v2, with improvements also on RULER and InfiniteBench. They also test the approach across RL algorithms including DAPO and GSPO, and analyze chain length efficiency and distractor robustness. HF Daily Papers' note
score 4