✨ GoLongRL-4B

We present GoLongRL, a fully open-source, capability-oriented post-training recipe for long-context reinforcement learning with verifiable rewards (RLVR).

Resource	Link
📝 Preprints	Paper
🤗 Daily Paper	Paper
🤗 Model Hub	GoLongRL-4B
🤗 Model Hub	GoLongRL-30B-A3B
🤗 Dataset Hub	GolongRL Dataset
📧 Contact	xiao_xuan_zi_666@163.com & suzhenpeng13@163.com

📌 Overview

_{Overall performance comparison on long-context benchmarks (DocMath, LongBench-V2, Frames, MRCR, CorpusQA, LBV1-QA).}

GoLongRL-4B achieves strong long-context performance at the 4B scale.

Model	Avg.	DocMath	LBV2	Frames	MRCR	CorpusQA	LBV1-QA
Qwen3-4B-Thinking-2507	53.0	61.0	40.2	64.4	38.4	49.9	64.0
QwenLong-L1.5-4B (w. GRPO)	56.1	61.3	44.3	67.1	40.9	58.8	64.1
QwenLong-L1.5-4B (w. AEPO)	59.4	62.5	47.9	67.4	47.9	64.7	65.8
GoLongRL-4B (w. GRPO)	62.2	62.5	45.5	66.6	67.5	65.1	65.9
GoLongRL-4B (w. TMN-Reweight, Ours)	63.0	62.3	47.1	67.4	65.5	69.6	65.9

Our framework combines the following.

1. Capability-Oriented Dataset (23K samples, 9 task types). Guided by a taxonomy of long-context capabilities, the dataset covers precise retrieval, comprehension, exhaustive retrieval, numerical reasoning, structured extraction, structured matching, graded ranking, sequence ordering, and summarization. Each task is paired with its natural evaluation metric (EM, Accuracy, F1, math_verify, IoU, SubEM, NDCG, Pairwise, ROUGE-L) as the reward function, rather than being collapsed into a single indicator.

2. TMN-Reweight. When training on heterogeneous reward types, per-prompt normalization in standard GRPO can mix up cross-task scale differences with prompt difficulty. TMN-Reweight is a simple modification that normalizes advantages at the task level instead of the prompt level, and adds a difficulty-adaptive weight to reduce noise from very easy or very hard prompts. It provides a modest but consistent improvement over vanilla GRPO in our ablations (+0.8 avg. at 4B scale), with gains mainly on aggregation-intensive benchmarks like CorpusQA.

3. Full Open Release. We publicly release the complete dataset, the four-phase construction pipeline, and all training code.

Key Results

• Under the same vanilla GRPO setup, our dataset alone outperforms the closed-source QwenLong-L1.5 dataset at both 4B and 30B scales (+6.1 avg. at 4B, +2.6 avg. at 30B).
• TMN-Reweight further improves average performance to 63.0 at 4B scale, surpassing QwenLong-L1.5 with its specialized AEPO algorithm (59.4).
• General capabilities (MMLU-Pro, AIME24/25, GPQA) are preserved or improved, with substantial gains on dialogue memory (LongMemEval +13.6) and agentic memory benchmarks.

🔍 Evaluation

Evaluation uses QwenLong-Benchmarks, covering three capability dimensions:

Dimension	Benchmarks
Long-Context	LongBench-V2, MRCR (≤128K / 128K–512K / 512K–1M), Frames, LongBench QA, DocMath, CorpusQA (≤128K / ≤1M)
General	MMLU-Pro, AIME 2024/2025, GPQA-Diamond
Memory	BFCL-V4 (memory subset), LongMemEval

🤝 Citation

@misc{lv2026golongrlcapabilityorientedlongcontext,
      title={GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment}, 
      author={Minxuan Lv and Tiehua Mei and Tanlong Du and Junmin Chen and Zhenpeng Su and Ziyang Chen and Ziqi Wang and Zhennan Wu and Ruotong Pan and jian Liang and Ruiming Tang and Han Li},
      year={2026},
      eprint={2605.19577},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.19577}, 
}