A focused feed of models, agents, research and open-source releases for people building with AI.
Krea 3 Ultra debuts at #4 on Arena, Artificial Analysis, and Design Arena editing leaderboards . Finishes edits in ~8s, 3-5x faster than Nano Banana Pro and 11x faster than GPT Image 2.5. $80 per 1k edits vs $150 for Nano Banana Pro and $436 for GPT Image 2.5.
humans& released GPU-accelerated Weight Teleportation , lossless RL weight sync faster than intra-cluster RDMA. Sampler pause drops to 219.6 ms; end-to-end sync under 4 s on 445 GB NVFP4 GLM-5.3. Snappy plus Zstd double compression shrinks per-step delta payload to 2.65 GB, 0.60% of model weights.
Georgia Tech and ARL release MO-Playground , a JAX-native multi-objective RL toolkit for robotics. New MORLAX algorithm uses hypernetworks mapping trade-off vectors to full policy parameters. Reports 21x to 271x speedups over HYPER-MORL across five MuJoCo benchmarks.
Round-to-nearest fp32 to bf16 conversion silently discards most RL updates; only 0.4 to 0.5% of weights change per step. The served rollout policy lags the fp32 optimizer by 50 to 150 steps, invisible to existing mismatch fixes.
New paper proves first universal approximation theorems for data-driven smoothing and forecasting of dynamical systems. Authors: Calvello, Carlson, Kovachki, Manta, Stuart; arXiv 2603.20359 .
CMU-led team unifies pretraining, SFT, and RL into one maximum-likelihood objective called unified training. RL rollouts reframed as likelihood fitting to model-generated targets via MaxRL extension . Single trainer mixes static text and RL generations; staged training becomes a special case.
Prime Intellect released Prime Mega MoE , adding a backward kernel to DeepSeek's forward-only Mega MoE. Up to 3.11x faster MoE kernels versus a PyTorch baseline, 1.34x faster end-to-end training. Qwen3-30B-A3B steps 48% faster and Qwen3-235B-A22B 55% faster than TorchTitan.
The new release flips the rewritten NCCL backend on by default, adds one-sided transports, and accelerates NVFP4 decode on Blackwell.
Soft-MTP identifies and removes a noisy residual term buried in the standard multi-token prediction gradient. The fix uses a KL gradient-recovery trick, implemented as a one-line change to the loss. No new parameters, no measurable overhead, compatible with fused cross-entropy kernels.
A new research paper derives scaling laws showing that tool and database diversity, not task count, drive tool-use agent generalization.
Reflection AI reveals the raw infrastructure behind Beam's reinforcement learning run and commits to OpenEnv as its future environment standard.
Sakana AI unveils Sawara, a from-scratch Japanese foundation model, alongside a three-pillar strategy spanning pre-training, post-training, and orchestration.
A Chinese-language adaptation of Stanford's CS336 bundles 16 chapters and 6 hands-on assignments covering the full LLM stack, from BPE to GRPO.
OpenAI pulled frontier models offline, built a cached web for them to train in, then turned its own cyber agents loose to break the walls.
Two overlooked floating-point bugs in softmax and cross-entropy have been quietly corrupting gradients. A new paper fixes them in a few lines of JAX.
A new two-stage training recipe lets looped language models reuse a single KV-cache across every recurrence, cutting HBM use up to 66%.
IFM open-sources the agentic reinforcement learning stack behind K2 Horizon, running 8,192 concurrent sandboxes and 1.3M rollouts across seven days.
New research shows hidden model preferences slip through distillation even when teacher and student share no base model, data, or architecture.
Prime Intellect ships a new default RL loss built for async training, plus a 1.15x throughput bump over torchtitan on Qwen3-235B.
Meta's Prometheus supercluster hits 85% goodput and 47K GPU-hour mean time between interruptions across hundreds of thousands of GPUs training Muse Wonder.
A new analysis settles whether momentum-based optimizers can get stuck at saddle points, and designs a variant that escapes faster without hurting convergence.
A one-shot optimal transport method aggregates independently trained mixture-of-experts models into a single global estimator without pooling raw data.
A new unified framework reframes VAE design for diffusion as a spectrum-matching problem, boosting generation quality on CelebA and ImageNet.
Zephon plugs into TorchTitan to deliver identical global batches across GPU counts, so a job resumed on different hardware trains on the exact same data.
No stories match these filters.
