• Causal Pretraining Determines Whether Long Histories Improve Control: Extending video history from zero to 19.2 seconds raised RoboCasa GR-1 success from 63.3% to 78.7%. Bidirectional pretraining produced no net gain. Streaming execution kept action-chunk latency to 107.4 milliseconds on an RTX 5090.

• Visual Document Vectors Are Not Harmless Copies: Attackers recovered 47% of text and 45% of sensitive words. They could also use 98.4% of reconstructed pages to retrieve the source page. Shuffling can be reversed, while pooling has resisted attacks so far.

• Hybrid Attention Depends on How You Scale: Sliding-window hybrids suit training-free length extrapolation. Linear-attention hybrids benefit more from continued long-context pretraining. A 16× extrapolation result on 64k NIAH-SK1 does not prove broad capability.

Also Notable

• Keep the Existing Video DiT and Make Compressed Latents Compatible: GRACE adapts Wan2.1-I2V-14B at 480×832×81 resolution. It cuts tokens by 8× and latency by 11.1× while preserving the original pipeline's VBench quality.
• Separate Conflicting Gradients to Generate for 24 Hours: SGF+ assigns future-context writing and current-frame denoising to different parameters. Five-second training clips extrapolate to continuous videos up to 24 hours without extra data or long-video fine-tuning.
• Turn a Video Diffusion Model Into a Hand Tracker: RLHND uses Cosmos 3 as a deterministic clip-level feature extractor. It estimates pose, contact, and force from monocular egocentric video, with real-robot experiments validating its use for robot learning.
• Text Rendering Benchmarks Now Test Whole Layouts: UltraText Bench contains 432 bilingual Chinese-English prompts with four to 12 text regions. In tested settings, improved clarity sometimes reduced textual fidelity.
• Skill Libraries Need Retirement Policies: SkillForge moves skills through trial, stabilization, and retirement so outdated experience stops misleading agents. It delivered relative gains of up to 7.8% across interactive benchmarks while controlling library size.
• Tool Use Maps to an Internal Direction: An inhibition-shaped mechanism compresses an agent's tool-use decision into one internal direction. It appears across seven Qwen, Mistral, and Granite models, offering an intervention point when agents discuss actions instead of executing them.
• Visual Token Pruning Can Strengthen Malicious Anchors: Pruning may concentrate attention on hostile foreground content. The inference-time SAP defense reduced attack success by up to 62% in tested settings without sacrificing efficiency or utility.
• Reward Models Can Judge Without Writing Full Critiques: LatentGRM-8B compresses evaluation into continuous latent trajectories, shortening the judging process by about 9×. At vote@5, it cut total inference time by 6.1× to 7× while retaining competitive preference accuracy.
• Global Safety Scores Can Hide Local Regressions: PatchBench-Local tests both harmful-neighbor repair and benign-neighbor preservation. It exposes severe local failures even when overall capabilities barely change.
• Neural PDE Surrogates May Learn Solver Errors: Fourier-symbol diagnostics show that surrogates can imitate training-solver discretization errors beyond 99.8% of the theoretical maximum for perfect imitation.

Read the full edition →