AI Shorts

What would you construct in the event you might run Reinforcement Learning (RL) post-training on a 32B LLM in 4-bit NVFP4—on a single H100—with BF16-level accuracy and 1.2–1.5× step speedups? NVIDIA researchers (with collaborators from MIT, HKU, and Tsinghua) have open-sourced QeRL (Quantization-enhanced Reinforcement Learning), a coaching framework that pushes Reinforcement Learning (RL) post-training into…

NVIDIA AI has launched Reinforcement Learning Pretraining (RLP), a coaching goal that injects reinforcement studying into the pretraining stage reasonably than deferring it to post-training. The core concept is easy and testable: deal with a quick chain-of-thought (CoT) as an motion sampled earlier than next-token prediction and reward it by the data achieve it supplies…

SwiReasoning is a decoding-time framework that lets a reasoning LLM resolve when to suppose in latent house and when to write specific chain-of-thought, utilizing block-wise confidence estimated from entropy developments in next-token distributions. The methodology is training-free, model-agnostic, and targets Pareto-superior accuracy/effectivity trade-offs on arithmetic and STEM benchmarks. Reported outcomes present +1.5%–2.8% common accuracy enhancements…

A important improvement is about to remodel AI in healthcare. Researchers at Stanford University, in collaboration with ETH Zurich and tech leaders together with Google Research and Amazon, have launched OpenTSLM, a novel household of Time-Series Language Models (TSLMs). This breakthrough addresses a crucial limitation in present LLMs by enabling them to interpret and cause…

TL;DR: Skala is a deep-learning change–correlation practical for Kohn–Sham Density Functional Theory (DFT) that targets hybrid-level accuracy at semi-local value, reporting MAE ≈ 1.06 kcal/mol on W4-17 (0.85 on the single-reference subset) and WTMAD-2 ≈ 3.89 kcal/mol on GMTKN55; evaluations use a mounted D3(BJ) dispersion correction. It is positioned for main-group molecular chemistry at the…

Can an iterative draft–revise solver that repeatedly updates a latent scratchpad outperform far bigger autoregressive LLMs on ARC-AGI? Samsung SAIT (Montreal) has launched Tiny Recursive Model (TRM)—a two-layer, ~7M-parameter recursive reasoner that studies 44.6–45% check accuracy on ARC-AGI-1 and 7.8–8% on ARC-AGI-2, surpassing outcomes reported for considerably bigger language fashions reminiscent of DeepSeek-R1, o3-mini-high, and…