CLMay 19

GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment

arXiv:2605.1957783.01 citationsHas Code
Predicted impact top 59% in CL · last 90 daysOriginality Incremental advance
AI Analysis

This work provides an open-source, systematic approach to long-context RL training for practitioners, addressing the need for diverse task coverage and reward alignment in post-training.

GoLongRL introduces a capability-oriented long-context RL training recipe with a 23K-sample dataset spanning 9 task types, and TMN-Reweight for heterogeneous multitask optimization. Their Qwen3-30B-A3B model achieves long-context performance comparable to DeepSeek-R1-0528 and Qwen3-235B-A22B-Thinking-2507, with the dataset alone outperforming the closed-source QwenLong-L1.5 dataset.

We present GoLongRL, a fully open-source, capability-oriented post-training recipe for long-context reinforcement learning with verifiable rewards (RLVR). Existing long-context RL methods often treat data construction as a matter of designing increasingly complex retrieval paths, leading to homogeneous task coverage and reward formulations that inadequately reflect practical long-context requirements. Our work offers two contributions. (1) Capability-oriented data construction with full open release. We openly release a dataset of 23K RLVR samples, the complete construction pipeline, and all training code. Guided by a taxonomy of long-context capabilities, the dataset spans 9 task types, each paired with its natural evaluation metric. It comprises curated open-source samples from established corpora and synthetic samples whose QA pairs are generated from real source documents such as books, academic papers, and multi-turn dialogues. Under the same vanilla GRPO setup, our dataset alone outperforms the closed-source QwenLong-L1.5 dataset. Moreover, our Qwen3-30B-A3B model trained on this data delivers long-context performance comparable to DeepSeek-R1-0528 and Qwen3-235B-A22B-Thinking-2507, suggesting that broader coverage and greater reward diversity substantially benefit long-context capability improvement. (2) TMN-Reweight for heterogeneous multitask optimization. To address optimization challenges from heterogeneous rewards, we propose TMN-Reweight, which combines task-level mean normalization for cross-task reward scale alignment with difficulty-adaptive weighting for more reliable advantage estimation. TMN-Reweight further improves average performance over vanilla GRPO, with general capabilities preserved or improved across reported evaluations.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes