@jiqizhixin - AI Dynamics - Page 10 of 114

Hallo-Live: streaming audio-video avatar generation from Fudan & Baidu

By

–

03 June 2026 9h54

What if your avatar could talk and move in real time?

Fudan & Baidu present Hallo-Live: streaming audio-video avatar generation.

Dual-stream diffusion + future-expanding attention syncs lips to upcoming speech.

Human-preference distillation preserves quality. 20 FPS, 0.94s… pic.twitter.com/0GHiOm5h2x
— 机器之心 JIQIZHIXIN (@jiqizhixin) 3 juin 2026

What if your avatar could talk and move in real time? Fudan & Baidu present Hallo-Live: streaming audio-video avatar generation. Dual-stream diffusion + future-expanding attention syncs lips to upcoming speech. Human-preference distillation preserves quality. 20 FPS, 0.94s

→ View original post on X — @jiqizhixin

3 June 2026

DeepSeek’s DeepGEMM Update: Developers Control Hardware Optimization for fp8_mqa_logits

By

@jiqizhixin

–

03 June 2026 5h52

Looks like DeepSeek is handing hardware optimization control directly to developers in the latest DeepGEMM update. For the fp8_mqa_logits function, the weights tensor dtype now explicitly dictates the accumulation precision.

→ View original post on X — @jiqizhixin

3 June 2026

FlashAR speeds up AI image generation 22x without retraining

By

@jiqizhixin

–

31 May 2026 20h45

What if you could speed up AI image generation by 22x without retraining from scratch? Researchers from Zhejiang University and the University of Adelaide introduce FlashAR. They add a lightweight vertical prediction head to existing autoregressive models, enabling parallel

→ View original post on X — @jiqizhixin

31 May 2026

Can AI See and Hear Simultaneously?

By

@jiqizhixin

–

31 May 2026 16h23

Can AI truly see and hear together? Researchers from NUS, Oxford, University of Toronto, and Microsoft Research present a comprehensive survey on Audio-Visual Intelligence. They unify the fragmented field of large foundation models that combine sound and vision—from speech.

→ View original post on X — @jiqizhixin

31 May 2026

World Action Models let robots imagine future before acting

By

@jiqizhixin

–

31 May 2026 5h44

What if robots could imagine the future before acting? Researchers from Fudan, NUS, and Shanghai Innovation Institute unveil World Action Models (WAMs). They merge predictive world models with action generation, so robots plan by simulating how the environment evolves, not

→ View original post on X — @jiqizhixin

31 May 2026

ATLAS: A Single Functional Token for Visual Reasoning Without Images or Code

By

@jiqizhixin

–

30 May 2026 20h19

What if visual reasoning could be done without generating endless images or code? Meta AI and CUHK researchers present ATLAS: a single "functional token" that works as both an agentic action and a latent reasoning unit. No visual supervision, no extra architecture—just one

→ View original post on X — @jiqizhixin

30 May 2026

Nemotron-Labs-Diffusion: tri-mode model boosts accuracy and throughput

By

@jiqizhixin

–

30 May 2026 16h29

Can one model switch between three decoding modes to crush both accuracy and throughput? NVIDIA researchers (with Georgia Tech, HKU, and MIT) introduce Nemotron-Labs-Diffusion — a tri-mode language model that unifies autoregressive (AR), diffusion, and self-speculation decoding

→ View original post on X — @jiqizhixin

30 May 2026

HiF-VLA: Robots remembering past and anticipating future via motion

By

@jiqizhixin

–

30 May 2026 10h15

What if your robot could remember the past and anticipate the future to master long tasks? Researchers from Westlake University and Zhejiang University introduce HiF-VLA. It uses motion as a compact representation to give robots hindsight, insight, and foresight—learning from

→ View original post on X — @jiqizhixin

30 May 2026

Theory of Agent: When AI Agents Should Use External Tools

By

@jiqizhixin

–

30 May 2026 6h23

When should AI agents actually reach for external tools? Researchers from Edinburgh, CUHK, UIUC, Northwestern & Princeton present Theory of Agent (ToA) at ICML 2026. Their answer: only when epistemically necessary — meaning the agent cannot complete the task reliably using its

→ View original post on X — @jiqizhixin

30 May 2026

WALL-WM: World Action Model using semantic events as atomic units

By

@jiqizhixin

–

29 May 2026 11h36

X Square Robot presents WALL-WM — a World Action Model that uses semantic events as atomic units, aligning language, video, and action granularities. https://t.co/JAKBdwnCKE
— 机器之心 JIQIZHIXIN (@jiqizhixin) 29 mai 2026

X Square Robot presents WALL-WM — a World Action Model that uses semantic events as atomic units, aligning language, video, and action granularities.

→ View original post on X — @jiqizhixin

29 May 2026