Breaking

NVIDIA Lands 74 Papers at ICML 2026 as Nemotron Is Cited by 145

July 7, 2026 at 10:22 EDT

Foundation Models
Open Source
Research & Papers

NVIDIA said 74 of its papers were accepted at ICML 2026, a leading machine learning conference held July 6–11, 2026 at COEX in Seoul, and that 145 papers built on its open Nemotron model family, according to its official blog. It added that roughly 2,000 of the more than 6,000 accepted papers—about a third—used NVIDIA GPUs in their research.

ICML 2026 · NVIDIA Research

Faster Video Diffusion, Flat Memory

A new method accelerates autoregressive video diffusion and world models by 5–10× while holding memory constant across long generations — pairing Temporal Cache Compression with Sparse Attention.

5–10×

Speedup over prior autoregressive generation

Flat

Constant memory even in long rollouts

NVIDIA papers accepted at the conference

Where the speedup comes from

Two techniques tackle the attention bottleneck of frame-by-frame generation.

Temporal Cache Compression

Compresses the cache along the time axis to keep memory flat.

Sparse Attention

Makes attention computation sparse to cut per-frame load.

→

5–10×

faster, flat memory

NVIDIA's footprint across the conference

Of ~6,352 accepted papers, how many leaned on NVIDIA hardware and open models.

~6,352

Total accepted

~2,000

Cited NVIDIA GPUs

145

Built on Nemotron

Why it excites creators

Cache compression may curb drift in character, clothing, and physics between clips — bringing studio-quality scenes within reach of individuals.

The caveat

Still at the research stage — large-scale real-world usage reports remain few for now.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.