NVIDIA researchers present continuous diffusion LLM with 8B benchmark results
ArashVahdat · x · 2026-11-21
A talk for the paper "Large Language Continuous Diffusion Models" (Nov 23) by researchers from Cornell, Georgia Tech, and NVIDIA — Zhihan Yang, Wei Guo, Jean-Marie Lemercier, Simon Welker, and Mohammad Mahdi Kamani.
- Method: likelihood-based continuous diffusion over spherical token embeddings
- The slide shows 8B Base/Instruct radar charts comparing Qwen3-8B (autoregressive), Nemotron-Labs-Diffusion-8B (masked dLM), and their Nemotron-Labs-Sigma-8B across GSM8K, MMLU, GPQA, HumanEval, MBPP, AIME25, and more
- The authors say it unlocks distribution matching distillation and improves self-speculative decoding
More from Research
- Claim: Anthropic mathematician and Claude produce 250,000-page Lean 4 proof of Riemann Hypothesis, verification ongoing — Electronic-Weekend66 · 2026-11-21
- Levent Alpöge announces Claude-assisted claimed proof of Riemann Hypothesis — sudoraohacker · 2026-11-21
- Mathematicians spar over whether a 250,000-page Lean proof is just a useless oracle — jessi_cata · 2026-11-21
- Two new benchmarks push AI evals past static testing: recursive self-improvement and KradleBench — typewriters · 2026-11-21
- Claude's claimed Riemann proof mocked: 250,000 pages of unreviewed Lean, author calls it 'slop' — GaryMarcus · 2026-11-21
- GGUF 4-bit quantization is just GIF palette compression, 30 years later — carrigmat · 2026-11-21