ResearchMath-14K introduces largest research-level math dataset with multi-agent curation from Seoul National University; NEO-ov pioneers native one-vision VLMs for multi-image and video understanding; Understand-Anything leads GitHub with 4,465 stars/day as agent skills ecosystem explodes

ResearchMath-14K's most provocative finding is not the dataset itself but the fabrication behavior analysis. Across eight open-weight models, newer generations produce 5.6x more references and 5.0x more fake references per reasoning trace. This is a concrete, quantified demonstration of a widely suspected failure mode: as models become more capable at generating plausible-looking mathematical reasoning, they simultaneously become more prone to hallucinating citations and mathematical results. The multi-agent curation pipeline (sourcing from academic papers, filtering, and generating teacher trajectories) is a pragmatic approach to a real bottleneck — research-level math problems are scarce because they require domain expertise to formulate. The 9.2-point average improvement from filtered open-problem attempts on Qwen3 (4B-30B) shows that even imperfect supervision on hard problems provides useful signal, which challenges the conventional wisdom that training data must be fully correct.

NEO-ov's native pixel-to-language approach represents a genuine architectural bet. Current VLMs use a modular pipeline: image encoder produces features, adapter maps them to language space, decoder generates text. This works but fragments information — pixel-level signals are compressed through the encoder bottleneck before they can interact with language tokens. NEO-ov learns cross-frame representations directly from raw pixels, which in principle preserves more fine-grained visual information. The extension to multi-image, video understanding, and spatial intelligence is where this architecture could show its advantages, since temporal and spatial relationships across frames are precisely what gets lost in the encode-then-align paradigm. Whether the training cost of learning visual representations from scratch (rather than leveraging pre-trained encoders) is justified by the quality gains remains the key question.

Chartographer addresses a fundamental evaluation problem that extends beyond charts. The core insight — that models can answer chart questions correctly by exploiting shortcuts or prior knowledge rather than actually reading the chart — applies to virtually every VLM benchmark. By creating counterfactual charts where the visual content changes but the question structure remains fixed, Chartographer isolates visual reasoning from everything else. This is methodologically rigorous but also practically important: if we can't distinguish models that truly read charts from those that pattern-match, we can't trust them in production settings where the charts contain novel data.

The GitHub trending data reveals the agent skills ecosystem reaching a new phase. The combined scale is remarkable: Superpowers (210K stars), ECC (196K stars), FreeDomain (169K stars), Anthropic Skills (142K stars), and taste-skill (24K stars gaining 2,715/day). But the more interesting signal is the diversification from generic agent harnesses toward behavioral alignment tools. taste-skill (preventing generic output), stop-slop (removing AI tells from prose), and Anthropic-Cybersecurity-Skills (structured security capabilities) all focus on what agents should and shouldn't do, rather than how to make them run. This is a maturation signal — the community has moved past 'can we build agents' to 'how do we make agents that produce quality output.' Microsoft's agent-governance-toolkit (2,975 stars, 472 today) adds the enterprise governance layer.

The video generation space is converging on a common efficiency playbook. OSP-Next combines sparse attention, quantization (HiF8), and RL-based optimization — which is essentially the same toolkit that language model inference has used, now applied to Diffusion Transformers. Meituan's LongCat-Video-Avatar-1.5 (345 likes with zero downloads, suggesting it just launched) targets audio-driven avatar generation, while SulphurAI's Sulphur-2-base (1.4M downloads) continues to lead in open text-to-video. The pattern is clear: video generation is following the language model trajectory of capability → efficiency → specialization, compressed into a much shorter timeline.