2026-09-03 今日值得读

1智能体arXiv 2608.296220/4

AGENTIC R AG-R1: Agentic Reinforcement Learning with Stack Memory for Multi-Step Reasoning, Retrieval and Memorizing

Xinke Jiang, Yue Fang, Zhibang Yang, Jiaran Gao, Zhixin Zhang, Tao Feng, Rihong Qiu, Wentao Zhang, Hongxin Ding, Ruizhe Zhang, Yongxin Xu, Yuheng Huang, Xu Chu, Junfeng Zhao, Yasha Wang

为什么值得读

The design may help agents revise misleading intermediate retrievals instead of accumulating them during long reasoning chains. This could benefit question-answering and research agents that need adaptive retrieval and memory control. Its value remains conditional on substantial training, judging, retrieval, checkpoint, and evaluator infrastructure that is not fixed in this compilation.

核心研究结论

AGENTIC R AG-R1 reports higher or near-highest F1 than listed baselines across seven multi-hop, open-domain, and agentic QA benchmarks.
The paper reports increasing AGENTIC R AG-R1 F1 as the maximum reasoning-step budget grows on selected 2Wiki, Bamboogle, and TriviaQA evaluations.
The paper reports lower average inference time than TC-RAG and downstream results on MedQA, DeepResearch Bench, ALFWorld, and WebShop.

人工智能机器学习 (cs.LG)多智能体系统机器学习 (stat.ML)

2大语言模型arXiv 2608.295390/6

LoGo: Token-Level Dynamic Local-Global Attention

Yuqi Pan, Zheng Li, Bohao Tang, Zhen Qin, Guoqi Li

为什么值得读

If its reported conditions can be reproduced, LoGo could lower the cost of long-context language modeling by reserving full-context attention for tokens that appear to need distant information rather than applying it uniformly. This may help model builders serving long documents or retrieval-heavy prompts, while preserving ordinary local processing for predictable tokens. The value is conditional on the learned router, sparse kernels, training recipe, checkpoints, and evaluation data matching the paper; the fixed snapshot supplies none of the author implementation or in-house pretraining assets, so the claims remain unverified here.

核心研究结论

The architecture ablation table reports that removing ContextNorm fails to converge, while additive fusion, auxiliary-loss budget control, and removing P-mask reduce the reported recall average relative to LoGo.
On the paper's 32k-checkpoint short-context recall aggregate, LoGo is reported as competitive with the other methods and scores 57.65 percentage points across six recall tasks.
In the controlled 1.5B comparison, LoGo reports the best language-modeling loss and Lambada perplexity among the listed full-attention and matched-budget static hybrid variants, with a higher commonsense-reasoning average than the full Transformer.

计算与语言机器学习 (cs.LG)分布式、并行与集群计算