AI Research Papers & Highlights

The non-obvious finding is that data composition, not raw task count, drives tool-use agent performance, giving teams a concrete recipe for allocating synthetic data budgets. Builders training or fine-tuning agents can apply SynthAPI's control axes to prioritize tool and database diversity over volume. Compared with prior synthetic environment generation work, this approach offers a quantitative scaling law rather than ad-hoc data curation.

#SynthAPI #agents #open-source models

arXiv cs.LG@Andrew Szot et al. (8)Source

Papers Builders10:55

Chinese contentEnvSched：环境感知调度提升 Agent 强化学习 rollout 效率

训练 agent RL 太烧 rollout 预算的可以看这篇，EnvSched 在 Qwen3-8B 上 40 步就跑赢随机选任务 120 步的效果，省了三倍训练量。

#EnvSched #Qwen3-8B #强化学习 #智能体

arXiv cs.LG@Sihan Yang et al. (12)Source

Papers Builders10:51

Chinese contentZero-shot框架用多模态模型识别足球转播事件，F1达72.0%

不训练就能看球赛自动挑事件，靠解说文本加分镜省72%推理成本，还超过专门训练的SoccerNet模型，做视频集锦的可以看看。

#多模态 #zero-shot #SoccerNet #视频理解

arXiv cs.LG@Sadia Kamal et al. (3)Source

Papers Builders10:51

Chinese contentSpectral Braids：用拓扑辫子刻画动态有向网络的新方法

作者把复数特征值随时间的缠绕路径变成一条拓扑辫子，再用 Burau 表示转成特征向量，四分类做到 97.1%，比不追踪特征值的基线 44.5% 高出一大截，图机器学习方向值得看看。

#图神经网络 #谱方法 #辫子理论 #动态网络

arXiv cs.LG@Mehmet Emin Aktas, Esra AkbasSource

Papers Builders10:51

Chinese contentLCM-Promo：用微调 LLM 做促销分配的增益模型

把 LLM 用在电商促销分配上，单 token 输出当转化概率用，A/B 测试里转化涨 1% 还省了 6% 促销费，思路挺有意思。

#LLM #LCM-Promo #微调 #增益模型

arXiv cs.LG@Nesta Midavaine et al. (7)Source

Papers Builders10:50

Chinese contentPneuma：面向商用飞机预测性维护的大规模真实时序基准发布

航空业终于有了公开的预测性维护数据集，9.2 亿时间步、9 个真实故障场景，想搞工业时序的同学可以拿来跑 baseline 了。

#Pneuma #时序预测 #异常检测 #预测性维护

arXiv cs.LG@Jiafeng Lin et al. (8)Source

Papers PolicySelected10:50

Theory Explains Why Unsupervised RLVR Collapses on Large Datasets

The paper turns a known failure mode of self-rewarding training into a predictive formula, giving practitioners a way to estimate safe dataset sizes before a run collapses. It also identifies two concrete mitigations, removing the shared coupling direction or adding a marginal-entropy bonus, that were validated on an actual Transformer. Researchers working on verifiable-reward training and anyone training reasoning models without labels will find the scaling laws directly applicable.

#RLVR #reinforcement learning #reasoning models #open-source models

arXiv cs.LG@Peize DingSource

Papers Builders10:50

Chinese contentGroundControl：用卫星图像辅助多视角扩散模型生成新视角

Niantic Spatial 做了个多视角扩散模型，把卫星图当额外条件喂进去，生成看不到的街景视角时布局不再乱编，连 3D 重建都不需要。

#GroundControl #扩散模型 #多模态 #Niantic Spatial

arXiv cs.LG@Akshay Krishnan et al. (10)Source

Papers Builders10:50

Chinese contentGTFormer 用图网络替代自注意力，实现 8192 倍长度外推

用图网络替掉自注意力，训练只到 512 长度就能零样本跑 400 万 token，线性内存还不爆显存，思路挺新鲜。

#GTFormer #GNN #Transformer #长上下文

arXiv cs.LG@Elisabeth FinkSource

PapersMulti-sourceSelected10:47

Study Explains Why DPO Updates Stay Sparse via NTK Theory

The practical payoff is a theoretical licence to merge DPO-aligned checkpoints: sparser, disentangled updates mean less interference when combining model variants. Researchers working on preference alignment or model merging can use the support-pair bounds to reason about how many samples actually shape a fine-tune. Teams choosing between DPO and SFT for multi-task pipelines get a concrete 35× interference reduction figure to weigh.

#DPO #NTK #Llama-3-8B #Mistral-7B

Event

arXiv: Anthropic@Kensuke Sasaki, Issei Sato6 sourcesSource

PapersMulti-source10:38

Chinese contentarXiv 论文用 OpenAI 证明 Haldane 能隙的方法证明 S=1 链基态唯一性

用 OpenAI 证明 Haldane 能隙那套方法，把 S=1 海森堡链基态唯一性证出来了，还坦承 ChatGPT 参与写证明，数学和 AI 协作两个角度都挺有意思。

#Haldane gap #Heisenberg chain #ChatGPT #arXiv

Event

arXiv: OpenAI@Hal Tasaki11 sourcesSource

Papers10:33

Chinese contentVideoLVR：用潜空间视觉证据改进视频推理

一篇讲视频推理的新论文，思路是让模型在思考时能回看关键画面片段，不用塞进一堆帧，做视频理解方向的可以看看。