Xiaohongshu and Fudan propose DR-Rubric for open-ended RL at NeurIPS 2026
Researchers from Xiaohongshu and Fudan University presented DR-Rubric (Deep Research as Rubric), a method accepted at NeurIPS 2026 that generates fine-grained scoring rubrics for open-ended reinforcement learning by first conducting deep research on the task. The approach, which has been used in Xiaohongshu's business, achieves competitive results on six benchmarks using only 1000 RL data per round. The paper and code are publicly available.
Coverage timeline
机器之心机器之心
让大模型回答一道数学题，通常还能拿标准答案判断对错。可如果问题换成「梳理一个领域的研究进展」「比较几种技术路线」或者「给出有依据的专业建议」，怎样才算答得好？答案没有唯一版本，遗漏关键事实、混淆概念、论证不充分，都可能藏在一段看起来很流畅的文字里。 这也是开放式任务做强化学习时最难的一步：模型可以生成很多回答，但训练系统得知道，哪些回答值得奖励。只给一个笼统的总分，模型很难知道该改哪里；把评判拆成细则，又会遇到另一道难题——这些细则由谁来写？ 专家手写可靠，却难以覆盖大量、不断变化的问题。直接让大模型写评分标准，速度快，但它可能只列出「清晰」「完整」「有逻辑」这样的通用要求，漏掉真正决定这道题质量的知识和边界条件。评分尺子本身不准，强化学习就可能朝着错误的方向优化。 针对这一问题，小红书与复旦大学等机构提出了 Deep Research as Rubric（DR-Rubric） ：让模型先围绕任务做一轮深度研究，收集外部事实、关键约束与常见错误，再把研究结果转成逐条可检查的评分细则，用作强化学习的奖励信号。这套方案也已在小红书业务中实际使用。 目前，该工作已被 NeurIPS 2026 接收，论文与代码已公开。 实验展示了它在六个基准上的实验结果：每轮仅用 1000 条 RL 数据，便在智能体研究和专家推理任务上取得有竞争力的表现。研究还进一步尝试让模型自己生成下一轮评分细则，并观察到一个有意思的现象：自举提升并不单调，评分标准会变化，模型能力也会随之重新平衡；继续迭代，还可能出现奖励越来越高、真实表现却崩溃的情况。 论文标题：Deep Research as Rubric for Reinforcement Learning 论文地址：https://arxiv.org/abs/2606.01091 代码地址：https://github.com/meiotoufa/DR-Rubric 研究背景：评分细则本身也需要研究 Rubric（评分细则）把一个笼统的质量判断拆解成细粒度、可解释的评价维度。回答每满足一条约束，就获得相应奖励，即便没有完全答对，也可以得到部分分数。这为缺乏唯一标准答案的开放式推理和长文本生成任务提供了更稠密的训练信号。 现有方法往往把 rubric 当成现成的、静态的产物：要么由专家手工撰写，要么直接通过 prompt 让模型生成。前者成本高，
Related stories
VisInteract: Interactive Text-to-Visualization for Imperfect Queries, NeurIPS 2026 Spotlight
Researchers from Hong Kong Polytechnic University and ByteDance propose VisInteract, a new paradigm for dynamic interactive text-to-visualization that handles imperfect user queries. They introduce VisInteract-Bench, a benchmark acknowledging ambiguous, incomplete, or erroneous inputs, and Vis-MCTS, a Monte Carlo tree search-based method that branches and backtracks between data retrieval, chart generation, and clarification. The work was accepted as a Spotlight paper at NeurIPS 2026 (292 out of 30,709 submissions).
1 reports · 1 sources
STILL: Content-Aware Intra-Layer Hybrid Attention Cuts Memory 80%, Speeds Prefill 83%
Researchers from Harbin Institute of Technology (Shenzhen) and collaborators present STILL, a NeurIPS 2026-accepted architecture that routes tokens to Softmax or linear attention within the same layer based on content importance via a self-saliency score. It achieves 80% memory reduction and 83% prefilling speedup at 16K context, with only 0.04B training tokens for linearization, and up to 86.2% improvement over prior linearization methods on long-context benchmarks.
ME-Decoding: Rethinking LLM Decoding as Ensemble Pruning
Researchers at Renmin University propose Mahalanobis-Ensemble Decoding (ME-Decoding), a method that selects candidate tokens by semantic redundancy in addition to probability, accepted at EMNLP 2026 Main Conference. The method formulates token selection as a dynamic subset optimization problem, using both model probabilities and token embeddings to find a sampling support with high confidence and low internal redundancy. In tests across three models and three temperature settings, ME-Decoding achieved the best average performance on reasoning and open-ended generation tasks, with only about 3% additional latency compared to the fastest probability-truncation baseline.
Recursive Self-Improvement Moves from Theory to Practice in 2026
In 2026, major AI labs disclosed concrete steps toward recursive self-improvement (RSI), a concept from I.J. Good's 1965 paper. OpenAI announced an automated research intern in September, Anthropic reported Claude leading 26% of its R&D, and Zhipu deployed an AI system that improves itself in production. These developments mark RSI's transition from philosophical speculation to measurable corporate KPIs.
News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe