# 面壁智能与清华NLP提出长上下文推理数据配方，无需特殊奖励工程

- 来源：OpenBMB (@OpenBMB)
- 发布时间：2026-07-24 21:00
- AIHOT 分数：58
- AIHOT 链接：https://aihot.virxact.com/items/cmryz4nxe0080roeyt0r8v420
- 原文链接：https://x.com/OpenBMB/status/2080639139465015395

## AI 摘要

面壁智能（OpenBMB）联合清华NLP提出，仅靠精心设计的数据配方与极简GRPO即可显著提升长上下文推理，无需特殊奖励工程。该方法将推理分解为检索、多证据合成与推理三种能力，构建8个数据集（约1.4万条样本）训练。在Qwen3-4B/8B/30B-A3B上，七个长上下文基准平均提升+7.2/+3.2/+6.4，效果可泛化至512K+ token。

## 正文

Everyone improving long-context reasoning with RL keeps pouring effort into reward engineering， while the training data stays scarce and narrow. It is like perfecting the cooking while nobody preps good ingredients.🤔
Beyond Reward Engineering from @TsinghuaNLP （OpenBMB member） argues the opposite from a data-centric view： a well-designed data recipe alone， paired with a deliberately minimal outcome-based GRPO， is enough to substantially improve long-context reasoning， no special reward engineering needed.
🤗Paper： Beyond Reward Engineering： A Data Recipe for Long-Context Reinforcement Learning
📄arxiv： https://arxiv.org/abs/2606.18831

1⃣️ The core hypothesis is that long-context reasoning decomposes into three complementary abilities that must be exercised jointly： retrieval （locate evidence when it is semantically indirect or buried among distractors）， multi-evidence synthesis （integrate scattered evidence no single piece provides）， and reasoning （multi-step problem solving， e.g. math， over a long input）. One GAIA-like question needs all three at once.
2⃣️ They build 8 datasets， ~14K curated examples， targeting each ability's specific failure mode： FuzzyNeedle and MultiNeedle break lexical shortcuts and near-duplicate confusion； CrossEntity， WebSearch， MultiQuery， KeyChain and an extended LongDocQA attack surface-form shortcuts， incomplete coverage and distractor confusion； LongMath rewrites hard math into narrative fragments scattered across a long document.
3⃣️ The RL is intentionally simple： standard GRPO with no KL， task-balanced sampling and task-level advantage normalization to stabilize 8 heterogeneous reward distributions， and a plain recall reward （LLM judge only where recall can be gamed）. This minimalism is the point， if a bare RL setup gains this much just by swapping data， the data recipe is the real lever.
4⃣️ Average gains of +7.2 / +3.2 / +6.4 across seven long-context benchmarks on Qwen3-4B / 8B / 30B-A3B， beating two prior RL training sets. Despite 64K training the gains hold to 512K+ tokens， and they transfer to agents： continuing RL on an agent-tuned model lifts GAIA by +4.8 and BrowseComp by +7.0， largest on the hardest subsets. Datasets will be released.