Gemini 4 Argon reports a 15% hallucination rate, 3.4× lower than Astra (51%) and Sol (54%) on AA-Omniscience, and supports 1M context with 64K output. It ranks #1 on Vals Index (68.9%) and Finance Agent (65.4%) and can ingest massive CSVs and multi-file reports in 1 pass without dropping rows.
Key points
Hallucination Rate: 15% → 3.4× lower than Astra (51%) and Sol (54%) on AA-Omniscience
Context size: 1M context / 64K output, ingests massive CSVs and multi-file reports in 1 pass without dropping rows
Rankings: 1 on Vals Index (68.9%) & #1 on Finance Agent (65.4%)
Pricing: Native Workspace Memory connects across Google Sheets, Drive & Docs at $1.99/task
Sources
MAP2 pursuit controller with MPC cuts tracking error
MAP2 Model- and Acceleration-Based Pursuit with MPC and Gaussian Process Residual Learning reduces average lateral tracking error by up to 44.65%. Real-world tests show at least a 1.5% reduction in average lap time. The controller combines a curvature-based kinematic MPC with a Sparse Gaussian Process for residual correction.
Key points
Reduces average lateral tracking error by 37.99% versus MAP and 44.65% versus Pure Pursuit.
Reduces average lap time by at least 1.5%.
Sources
French Supermarket Refrigeration Controllers hit, 400
Hackers hijacked refrigeration controllers in French supermarkets, raising temperatures to 50 °C in about 400 stores. At roughly 150 stores the heat lasted six hours and caused several million euros of product loss.
Key points
Temperature rise: cabinets reached up to 50 °C.
Duration and impact: about 150 stores stayed hot for six hours, with losses estimated at several million euros.
Sources
Co-Adaptive Entropy Modeling reduces distortion 15-24% in perceptual audio coding
Co-Adaptive Entropy Modeling and Rate-Distortion-Optimal Bit Allocation for Perceptual Audio Coding presents a coder where entropy model and bit allocation co-adapt across blocks without side information. It lowers band noise-to-mask ratio by 15-24% versus a solver without entropy-coding awareness and an additional 3-4% from co-adaptation. A MUSHRA test at 80 kbps per channel confirms the ranking.
Key points
Reduction: entropy-coding-aware bit allocation lowers distortion by 15-24% relative to the same solver without it.
Additional gain: co-adaptation removes a further 3-4% (up to 15% on oboe) that a backward-adaptive coder alone cannot achieve.
Sources
Microsoft single-cell resource releases 572 checkpoints
Microsoft single-cell resource releases 572 pretrained checkpoints on Hugging Face. Checkpoints span Geneformer, scVI, SSL, and VAE to study dataset size and diversity effects.
Key points
Checkpoints: 572 pretrained checkpoints
Models: Geneformer, scVI, SSL, VAE
Sources
Content-Edited Layout Generation for Graphic Design announced
Content-Edited Layout Generation for Graphic Design introduces the CELG task and releases the InstructEdit dataset. The paper defines CELG as generating an adapted background and layout conditioned on text and optional foreground. It also presents a two-flow multimodal flow matching model that combines a pretrained image editor with a layout flow transformer.
Key points
Task: Content-edited layout generation (CELG) extends conventional content-aware layout generation.
Dataset: InstructEdit contains aligned source images, target canvases, instructions, and layouts created by the Search-Compare-Edit (SCE) framework.
