• Encoding LLM Responses Instead of User Queries Lifts Embeddings by 9.3%. LLM2Vec-Gen uses purely self-supervised training to beat the best unsupervised methods on MTEB. Safety alignment transfers into embedding space as a bonus.

• The Real Bottleneck in STEM Visual Reasoning Is Perception, Not Reasoning. CodePercept's ablations show that scaling the perception component consistently outperforms scaling reasoning. Executable code as a perception scaffold works remarkably well.

• Differential Decomposition of Cross-Covariance Matrices for Attention Steering. Prism-Delta ties or beats SOTA on 19 of 20 evaluation configs, halves fluency degradation, and works with FlashAttention out of the box.

• Selecting Only "Just Hard Enough" Problems Caps Your RL Model's Ceiling. DPS uses training dynamics to balance efficiency and coverage. Validated across math, planning, and visual geometry tasks. Accepted at ICLR.

Also Notable

• V₀.₅ Uses a Pretrained Value Model as the Advantage Baseline for RLVR — removes the need for synchronous updates and reduces GRPO's variance problem.
• Video Reasoning Models Degrade Sharply Under Real-World Disturbances — weather occlusion and camera shake expose robustness gaps; ROVA provides a targeted training framework.
• Geometric Framework Unifies Three Latent Diffusion Optimization Objectives — semantic discriminability, reconstruction fidelity, and compression rate no longer need separate tuning.
• Differentiable Physics Framework Reconstructs 3D Material Properties From Surface Temperature — replaces traditional per-pixel 1D approximations for non-destructive testing.
• Multi-Agent RL Trains Humanoid Robots for Physical Assistance — requires continuous perception and adaptation to human partner pose changes.
• FP4 Quantization's Mean Bias Gets Amplified by LLM Anisotropic Distributions — blockwise schemes need to account for this systematic error.
• GLM-OCR: A 0.9B On-Device Document Understanding Model — CogViT encoder plus GLM decoder, targeting practical OCR scenarios.
• LLM User Simulators Show Systematic Bias Against Real Human Behavior — the Sim2Real gap in multi-turn agent evaluation deserves scrutiny.
• Semantically Degraded Conditions Replace Null Prompts for CFG Guidance — reduces geometric entanglement caused by null-prompt conditioning.
• Multi-Agent Collaboration Automates Comedy Short Video Production — LLM reviewers align output with real audience preferences.
• Real-Time Panoramic Scene Graph Generation — targeting edge deployment for embodied agent scenarios.

Read the full edition →