Nemotron-3.5-Lightning-30B-A3B

Headlines · 2

1. Oct 15
   DeepMind and partners unveil IdeaLens, a detector for AI-generated research ideas

   Researchers from the University of Maryland, Google DeepMind and Simon Fraser University have posted an arXiv preprint, IdeaLens: Detecting AI Ideas in Long-form Writing, describing a detector that converts an article into an outline and judges idea provenance rather than prose style. The team trained on about one million English long-form texts using Pangram 3.3.2 labels, fine-tuned Nvidia's Nemotron-3.5-Lightning-30B-A3B with LoRA, and built a prose-reading control model called ProseLens. In the authors' reported tests, IdeaLens flagged 6.8% of articles as AI when AI wrote from a complete human outline, down from 94.9% with only a topic prompt; when humans rewrote stories from AI-generated outlines, it judged 68% as AI versus 8% for Pangram 4. The authors caution that training labels came from an existing detector, the TwiceTold set had only 50 human rewrites, and the system cannot yet apportion credit in multi-turn human-AI co-creation.

   MIT科技评论中文 · 🔥 0
2. Oct 14
   vLLM open-sources bitwise-consistent RL training kernels and an agent skill

   A vLLM blog post argues that inference-side optimizations in high-performance RL — low-precision compute, low-precision KV cache, kernel and communication tuning — shift the numerics, so the training engine assigns different probabilities to the same token than the engine that sampled it; on DeepSeek-V4-Flash the high-performance baseline's train-inference KL rose from 2e-3 to 29.9 as reward fell back. The proposed fix has the training forward call vLLM's inference kernels directly, modified so rollout and training produce bitwise-identical logprobs for every sampled token (diff = 0). The authors report diff of 0 at every one of 100 DeepSeek-V4-Flash steps at 91.3% of the baseline's rollout throughput, and say that on Nemotron-3.5 (a Mamba hybrid with NVFP4) the consistent run's SWE-bench Pro score held near 0.6 while the baseline's fell from 0.56 to 0.22. The workflow is packaged as an agent skill and released with recipes and code, and the team says DeepSeek-V4.1 has reached 0 diff with performance work still in progress.

   vLLM Blog · 🔥 0

Experience and discussion from the community

Share my Nemotron-3.5-Lightning-30B-A3B experienceAsk about Nemotron-3.5-Lightning-30B-A3B