• Tencent's Hy-Embodied-VLM-1.0 uses MoE to activate just 3B parameters and match a previous model that activated 32B, taking best-in-class on 19 of 38 embodied benchmarks. The abstract only reports leaderboard scores, so the deployment questions — latency, throughput — still need the full paper.

• AVQ makes attention codebooks adaptive by importance: VQ attention drops O(N²) to O(MN) but spends the same codebook budget everywhere. AVQ refines only the most important codes, pushing attention compression into a spatially adaptive regime.

• MobileSAM2 distills SAM2 down to phones and laptops, building the teacher's temporal and multi-granularity knowledge into a hypergraph before aligning it. The idea is clean, but the abstract skips the accuracy-drop-for-speed ratio that decides everything.

• Bayesian accuracy gives multiple-choice evals a drop-in fix for length bias: token-level log-prob penalizes long answers, length normalization often overcorrects, and an explicit length prior cancels the linear effect.

Also Notable

• Stop Letting MLLMs Score Image Quality From Semantics Alone — IQA-T1 calls tools to find visual evidence before scoring, easing the old problem of unexplainable, non-generalizing scores in open scenes.
• 4D Gaussian Splatting Drops Fast-Moving Objects — This Puts Them Back — large inter-frame displacements leave Gaussian attributes poorly learned and fast objects lost; an implicit spatiotemporal position network fills the gap.
• 3D Detection for Self-Driving Can Drop the Heavy 2D-Pretrained Backbone — LiDAR-camera fusion has relied on massive vision backbones; DeGuNet uses a depth-guided, ultra-compact backbone to cut the parameter redundancy.
• Text-Edited Images Keep Losing the Subject — A Flow-Matching Model Holds Key Attributes — DiTailed uses three improvements to keep the subject's key attributes intact through the whole edit.

Read the full edition →