Life Agents, 4D Video, and Sparse Domain Updates
- Life Agents Must Keep Working After Users Go Silent: VibeLifeBench tests proactive checks, timing, and hidden constraints across 200 multiweek tasks. All seven frontier models score poorly.
- Video Latents Can Connect Directly to 4D Decoders: About 1,000 reconstructed videos can train an interface that transfers across models sharing a VAE. It also improves geometric and temporal consistency.
- Domain Adaptation Barely Changes Singular Values: Most changes appear in singular vectors and selected attention heads. Up to 60% of head updates can be reverted without measurable quality loss.
Also Notable
- Mobile Assistants Must Stitch Evidence Across Apps Before Acting Correctly. SPIEval focuses on personal information scattered across multiple applications.
- Perceptual Consistency Can Be Built Directly Into Flow-Matching Paths. Flow Straight to Reality targets the tradeoff between restoration accuracy and visual realism with fewer sampling steps.
- Token-Level Hallucination Detection Can Replace Blanket Refusal or Regeneration. UniProbe traces relationships among image patches, query tokens, and generated tokens to find outputs lacking visual evidence.
- Packet-Loss-Tolerant Compression Must Spread Information More Evenly. Every Packet Counts prevents a few missing packets from destroying an entire image on satellite and emergency links.
- Apparently Textureless Surfaces May Still Support Passive Depth Estimation. Wave optics reveals second-order texture signals that conventional vision methods overlook.
- Separate Autoregressive Streams Can Improve Cloth Motion Outside the Training Distribution. DSAR propagates surface displacement and combines it with historical hidden states retrieved from memory.
- Grayscale Images May Not Have a Trustworthy Fixed Luminance Channel. Relaxing luminance reconstruction can outperform chroma-only prediction when colorizing historical imagery.
- Dehazing Diffusion Can Start From the Physics of Fog. HNDiff replaces pure Gaussian noise with a fog-to-noise process grounded in image degradation.
Don't miss what's next. Subscribe to AI Research Brief: