August 2026 Summaries

FigmaTrace is an open-source dataset of 3,469 Figma design trajectories representing more than 200 hours of human work across 10 creative skills, including accessibility fixes, template corrections, prototyping, sketch conversion, and platform adaptation. It addresses the shortage of workflow data for training vision-language models on subjective design tasks by capturing the sequence of expert decisions rather than only final design artifacts. Researchers found that fine-tuning the open Qwen3.8-27B model on these trajectories improved both design and broader agentic navigation performance, surpassing Claude Opus 5 on GUI-Odyssey and AndroidControl while achieving gains on an in-domain creative benchmark. Segmenting long recordings according to design phases outperformed maximum-context-length sharding by 7.3 percentage points on average, suggesting that intent-based curation better preserves task context and teaches transferable skills. Manual analysis associated performance improvements with more accurate element selection, stronger coordinate understanding, and greater decisiveness in producing actions. The release includes 2,883 training trajectories, 586 evaluation trajectories, and the fine-tuned Qwen3.8-27B model.

Engineers adapted the 744B-parameter GLM-5.2 mixture-of-experts model, whose routed experts use 4-bit NVFP4 quantization, to play Super Mario Bros. through bf16 LoRA reinforcement learning on two 8×B200 GPU nodes. Making the system trainable required resolving several hardware and software issues: Transformer Engine’s dual quantized weight layouts doubled expected memory use, so the team retained only the forward layout and used bf16 dequantization during backward passes; chunked kernels eliminated large quadratic-memory buffers in a reference sparse-attention implementation; and specialized SGLang settings enabled serving an NVFP4 MoE model with continuously synchronized LoRA adapters. They also fixed distributed adapter loading, optimizer initialization that overwrote warm-started LoRA weights with zeros, and sequence-length-driven memory failures. Although standard GRPO training initially improved reward before collapsing into deterministic, identical rollouts with zero learning signal, switching to DAPO-style dynamic sampling and removing an entropy regularization term sustained training performance. Tests on smaller models indicated little meaningful difference between bf16 and NVFP4 bases for supervised fine-tuning or reinforcement learning, while the resulting full-scale policy completed roughly 77% of Mario level 1-1, demonstrating an end-to-end approach for parameter-efficient RL on a 744B model.