A latent-GRPO training step for the Qwen 3.8 27B model dropped from a median of 36 minutes on a single H100 to 2.6 minutes on one 8× NVIDIA B300 node, a 14× speedup, according to the training team's a…
A custom inference stack running Qwen 3.8 27B on NVIDIA B300 SXM6 GPUs reached 9,825 tokens per second on 8× B300 with DP4×TP2 plus Suffix4, and a maximum cluster ceiling of 16,931 tokens per second o…
VLLM's PagedAttention and continuous iteration-level batching address the KV cache memory bottleneck that limits LLM inference throughput, according to a technical deep dive on the inference engine's …
Engineering teams deploying reinforcement learning with verifiable rewards (RLVR) post-training have found that building the training harness accounts for only 20% of the work, with the remaining 80% …
Enterprise teams including Shopify, Meta, Intercom, Ramp, Snowflake, and ONESTRUCTION have replaced frontier LLM API prompting with compact open models of 8B to 35B parameters trained via reinforcemen…
Long-horizon AI agents fail at an exponential rate because a 95% per-step success probability yields only a 35.8% chance of completing 20 turns and 7.7% at 50 steps, and the chance of a second error j…
A controlled pilot on a dedicated NVIDIA H200 GPU found ISO-AdamW scored 758 correct answers versus 754 for baseline AdamW on a 1,000-question held-out test, a +0.4% delta with a 95% confidence interv…
Physical AI is emerging as the next major frontier because frontier language models have already solved long-horizon tool use, allowing reasoning agents to treat hardware such as 6-axis robotic arms, …
A 2026 position paper from Google DeepMind titled "LLMs Can't Jump," authored by research scientist Tom Zahavy, argues that today's foundation models have mastered induction and deduction but are stru…
A technical explainer on distributed training and inference details how work is divided across CPUs, GPUs, and multi-node clusters, distinguishing data, pipeline, and tensor parallelism and sharding, …
Reinforcement learning for language models has moved from Reinforcement Learning from Human Feedback (RLHF) through LLM-as-a-Judge grading to Reinforcement Learning with Verifiable Rewards (RLVR), whe…
New research by Zhu et al. (July 2026, arXiv:2607.19331, "ISO: An RLVR-Native Optimization Stack") provides mathematical proof that Reinforcement Learning with Verifiable Rewards (RLVR) post-training …
Enterprise software leaders are shifting from renting frontier AI APIs to building sovereign, self-hosted models after Mythos shut down its model API in June 2026, according to an analysis citing Fire…
A technical analysis of multi-reward reinforcement learning for LLM agents finds that summing multiple reward channels into standard algorithms like PPO or vanilla GRPO causes scale dominance, where t…
Latent-GRPO, a reinforcement-learning method that replaces discrete text tokens with continuous recurrent thought vectors in embedding space, eliminates the token tax that causes 19% to 34% of generat…