• A single spectral condition unifies μP scaling across width and depth. No more per-architecture, per-optimizer derivations for hyperparameter transfer. Code included.

• Data curation itself leaks membership information. Anthropic shows that even models trained only on public data expose the composition of the original dataset through the selection process.

• VLMs give dexterous hands natural language instructions. UniHM uses a unified tokenizer to generalize across hand morphologies, trained only on human-object interaction video. No teleoperation data needed.

Also Notable

• GRPO Moves From LLM Alignment to 3D Mesh Generation — asynchronous advantage-guided preference optimization replaces offline DPO for artistic quad-mesh generation.
• Idempotent Experience Replay Mitigates Catastrophic Forgetting — more stable under high-reliability requirements in continual learning.
• Mamba/SSM Handles Industrial-Scale CAD Sequences — efficiency advantage over Transformers pays off in fine-grained part modeling.
• Wavelet Transform Detects Semantic Boundaries for Video Frame Selection — preserves narrative structure better than query-relevance-based selection.
• Cross-Modal Counting Benchmark for MLLMs — unified image-text-audio counting evaluation reveals basic numeracy gaps.
• Gaussian Splatting Reconstructs Radar-Quality Precipitation Fields From Sparse Weather Stations — a new path for low-cost weather monitoring.
• Molecular Representation Shifts From Atom-Centric to Bond-Centric — resonance and stereoselectivity are no longer ignored at the bond level.
• Visual Autoregressive Next-Scale Prediction for Super-Resolution — addresses global consistency in upscaling.

Read the full edition →