AI 每日资讯 — 2026-09-23
🔥 HuggingFace 每日论文
1. RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
Peng Xia, Rujun Han, Zifeng Wang
An LLM agent's capability is largely magnified by its harness, namely the prompts, control flow, tooling, memory, a
nd context management surrounding the frozen backbone model. Recent methods increasinPDF · arXiv · 代码 · 项目 | ❤️ 142
2. WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory
Wangbo Yu, Kunhao Liu, Wenbo Hu
Video world models enable interactive exploration of dynamic environments, yet struggle to respect prior observatio
ns over long horizons and across viewpoints. We present WorldCrafter, a video world mPDF · arXiv · 代码 · 项目 | ❤️ 108
3. GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay
Yiran Wang, Xingyilang Yin, Junfu Pu
Modern video games provide a measurable testbed for AI models, combining abilities of visual understanding, instruc
tion decomposition, goal planning, and precise action control over multiple temporalPDF · arXiv · 代码 · 项目 | ❤️ 107
4. VideoGen-Agent: Reinforcing Video Generation Agents
Binxu Li, Haoyi Duan, Yuhui Zhang
Recent advances in video generative models have enabled high-fidelity, temporally coherent video generation. Howeve
r, these models often struggle to satisfy prompts requiring specialized knowledge, sp5. onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction
Lei Yang, Mengyin Liu, Jia Wang
We present onPanda, an interactive tool for efficiently annotating LLM alignment data and agent trajectories. onPan
da adopts token-level correction as its core interaction: while reading a model respo6. Harness-Zero: Harness Distillation via Agent-as-Harness
Haoran Ye, Yuxing Lu, Haonan Dong
Agent harnesses, the external systems that mediate model-environment interaction, can substantially improve agent p
erformance, but their gains remain tied to the harness at deployment. Because the bes7. GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation
Jiahao Lu, Minghao Yin, Wenbo Hu
We present a compact geometry-native latent space as a shared foundation for perception and generation. Visual gene
rators can produce photorealistic frames without preserving a consistent 3D scene. We8. Emergent Collusion in Long-Horizon LLM Agent Interaction
Xinrui Shi, Yanzhe Zhang, Diyi Yang
LLM agents are increasingly deployed in collaborative settings, yet long-term interaction may give rise to undesira
ble coordination. We study the emergence of collusion in a long-horizon multi-agent e🔥 arXiv 每日论文
📄 arXiv: cs.AI
1. Didactic knowledge or Clinical Cases? How Data Types Shape Medical Large Language Models
Yuzheng Fan, Haochun Wang, Sendong Zhao, Xiao Han, Ming Ma, Bing Qin
arXiv:2609.22161v1 Announce Type: new Abstract: Medical large language models are commonly trained on mixtures of
didactic data (e.g., textbooks) and clinical data (e.g., patient records), yet how th2. An Affordable AI-Integrated Smart Cane for Multimodal Mobility Assistance of Visually Impaired Users
Ali Akarma, Adeel Ahmad, Toqeer Ali Syed
arXiv:2609.22277v1 Announce Type: new Abstract: Visual impairment affects over 2.2 billion people worldwide, yet c
onventional white canes cannot detect elevated hazards or provide semantic environmen3. PAANI : On Device Visual Evidence Fusion and Explainable Guidance for River Robot Simulation
Savio Cardoz, Santhiya Rajan
arXiv:2609.22353v1 Announce Type: new Abstract: Mobile river monitoring robots must interpret obstacles and water
boundaries that geographic waypoints alone cannot describe. On resource constrained p📄 arXiv: cs.CL
1. Recognition, Simulation, and Refusal: A Contamination-Aware Study of Classic Psychological Effects in LLM Agents
Joy Bose
arXiv:2609.22090v1 Announce Type: new Abstract: An LLM producing the response pattern associated with a human psyc
hological effect is not the same claim as the LLM possessing that bias. We present Ps2. Memory That Looks Forward: A Zero-Inference Prospective Term for Personal Memory Retrieval
Jonathan Groff
arXiv:2609.22091v1 Announce Type: new Abstract: Retrieval over a personal memory store is retrospective: it surfac
es what resembles the query, and it is blind to what the user has committed to do. We3. Summarize, Judge, Refine: Decoupled Content Understanding and Policy Learning for Multimodal Content Moderation
Zeeshan Ahmed, Yang Qin, Hanqing Huang
arXiv:2609.22094v1 Announce Type: new Abstract: Content moderation systems traditionally entangle multimodal under
standing with policy-specific classification, requiring full pipeline retraining for📄 arXiv: cs.LG
1. PRQuant: Permutation Residual Quantization for Low-Overhead Inference
Peiran Wang, Anqi Wang, Jiaying Zhao, Huiwen Yang, Zhenyu Ming, Rongqian Wang, Yiwu Yao, Kun Tian, Xin Yao, Gong Zhang, Fan Yang, Zhongyi Huang
arXiv:2609.22106v1 Announce Type: new Abstract: Accuracy of Low-bit quantization of linear layers is often dominat
ed by a small number of outliers. Although existing methods, such as smoothing, rotat2. Generalized Multimodal Foundation Model
Huizi Cui, Zongbo Han, Chenggong Ding, Naichuan Xiao, Jialong Yang, Jingdong Chen, Guangyu Wang, Qinghua Hu, Changqing Zhang
arXiv:2609.22107v1 Announce Type: new Abstract: Making prediction with multimodal data is widely used in diverse s
cenarios. Existing multimodal fusion models, once deployed, can only handle predefine3. Correcting Learning-based Perception for Safety
Yan Miao, Hussein Darir, Sayan Mitra
arXiv:2609.22108v1 Announce Type: new Abstract: Learning-enabled perception is important in many autonomous system
s. Unlike traditional sensors, the boundary where ML perception does or does not work📄 arXiv: cs.CV
1. Did You Steal My Shot? Pioneering Camera Motion Plagiarism Detection in Generative Videos
Chengguo Zhang, Ping Ping
arXiv:2609.22267v1 Announce Type: new Abstract: Camera motion often reflects directorial intent and requires profe
ssional equipment, making it a high value form of intellectual property. However, gen2. Enabling Vision and Cross-Modal Learning for Multimodal Stroke Recurrence Prediction: An Interpretable Two-Step Framework
Christian Gapp, Elias Tappeiner, Martin Welk, Karl Fritscher, Stephanie Mangesius, Constantin Eisenschink, Philipp Deisl, Michael Knoflach, Astrid E. Grams, Elke R. Gizewski, Rainer Schubert
arXiv:2609.22271v1 Announce Type: new Abstract: Multimodal stroke recurrence prediction requires effective integra
tion of heterogeneous clinical and imaging data, yet modality imbalance often causes3. Moonworks Lunara: Modeling Artistic Intelligence
Yan Wang, Yanzu Wang, Maitreyee Joshi, Samiha Sadeka, Partho Hassan, Reza Jarral, Sayeef Abdullah, Sabit Hassan
arXiv:2609.22272v1 Announce Type: new Abstract: We formulate \emph{Artistic Intelligence} as exploration driven wo
rld realization, leaving space for creative possibility while preserving the semantic📝 AI 官方博客
1. New experts join Google’s AI & Economy team
📝 Google AI Blog
Text “AI & Economy Research Program” all over a green grid background, with the Google G logo in the bottom right corner
2. Co-creating the future of fashion with Google
📝 Google AI Blog
Jane Wade and Sergio Hudson
3. Making global data easier to explore
📝 Google AI Blog
UN System Data Commons Data webpage
4. What We Learned Trying to Catch AI Liars: An Aletheia’s Quest Retrospective
📝 EleutherAI Blog
What we learned while building black-box and white-box detectors for AI deception during Aletheia’s Quest.
5. A Dynamical Model of AI Governability
📝 EleutherAI Blog
A toy dynamical model of whether the AI workforce that builds future AI ends up cooperative or uncooperative: where the …
basin boundary lies, what current evidence says about which side we are on, and6. Early Indicators of Reward Hacking via Reasoning Interpolation
📝 EleutherAI Blog
Using importance sampling with fine-tuned donor prefills to predict reward hacking emergence during training
7. Introducing Claude Opus 5.5AnnouncementsSep 22, 2026Opus 5.5 performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.
📝 Anthropic
AnnouncementsSep 1, 2026Introducing Claude Fable 5.1 and Claude Mythos 5.1Our most advanced models for coding and knowle…
dge work. Their research capabilities also offer an early glimpse of how AI mode8. AnnouncementsSep 1, 2026Introducing Claude Fable 5.1 and Claude Mythos 5.1Our most advanced models for coding and knowledge work. Their research capabilities also offer an early glimpse of how AI models will contribute to scientific progress.
📝 Anthropic
暂无摘要
9. Sep 17, 2026Measurements for understanding the pace of AI development inside frontier labsToday, the world can’t see what’s going on inside AI labs. Anthropic is proposing new metrics that would give the public visibility into frontier AI development.
📝 Anthropic
暂无摘要
📬 TLDR AI 精选
1. one daily email
one daily email
💬 Hacker News AI 热门
1. OpenAI GPT–6 Astra breaks Enigma message that has resisted solution since 2005
🔥 305 分 · 💬 269 评论
OpenAI GPT–6 Astra breaks Enigma message that has resisted solution since 2005
2. OpenAI is well positioned to fast-follow Jev
🔥 128 分 · 💬 96 评论
OpenAI is well positioned to fast-follow Jev
📰 TechCrunch AI 新闻
1. Anthropic releases Opus 5.5 with lower prices and Fable-level performance
Anthropic called it “the strongest-performing model we’ve tested to date.”
2. AstroForge is putting AI in command of its next spacecraft
Autonomy-1 will have a small, transformer-based AI model taking charge of a space probe.