ACL 2026: Alibaba DAMO I²B-LPO Improves RLVR Exploration

ACL 2026｜打破推理同质化！阿里达摩院新作让RLVR从重复采样走向有效探索

Alibaba DAMO Academy introduced I²B-LPO, an RLVR post-training framework that branches rollouts at high-entropy nodes and filters them with an information-bottleneck self-reward, reporting up to 5.3% accuracy gains and 7.4% semantic-diversity gains on math benchmarks using Qwen2.5-7B and Qwen3-14B.

Why it matters: HKR-H/K/R all pass: the ACL 2026 DAMO paper has a clear RLVR exploration hook, concrete I²B-LPO mechanics, and benchmark gains. It is still a training-method paper, not a major model or product release, so 78 fits the lower good-quality band.