Research Briefing — July 23, 2026

Masked Visual Actions (MVA) introduces a pixel-space control interface for video world models, expressing action as a partially revealed trajectory of an arbitrary entity. Revealing robot motion makes the model predict scene response (forward dynamics); revealing desired object motion makes it recover robot behavior (inverse dynamics). Fine-tuned with only 15 hours of manipulation data, it unifies forward/inverse modeling.

ABot-World-0 presents an action-conditioned video world model enabling real-time, long-horizon closed-loop interaction on a single desktop GPU. It uses multi-source data from AAA games, simulations, and internet videos, with a unified pipeline applying 14 quality checks and VLM-based assessment. The model progressively distills a bidirectional teacher into a causal student using teacher forcing and ODE distillation, with LongForcing to align long self-rollouts and mitigate distribution shift.

SkewAdam reduces optimizer state memory for MoE training by 97.4% (50.6 GB → 1.29 GB) and peak training memory from 81.4 GB to 31.3 GB. It assigns different state configurations to MoE's three parameter populations: float32 momentum + factored second moment for dense backbone (5% params), factored second moment alone for experts (95%), exact second moment for router (<0.01%).

This paper proves asymptotically optimal horizon-free regret for finite-horizon tabular MDPs: Õ(√(SAK) + S⁸A³), completely removing log H dependence from prior Õ(√(SAK log H) + S²A log H) and drastically improving prior horizon-free Õ(√(S⁹A³K)). Matches contextual bandit lower bound Ω(√(SAK)) up to log factors.

ISO (Isospectral Optimization) is an RLVR-native optimization framework building on the discovery of spectral inheritance: RLVR reuses base model weight spectra while acquiring new behaviors through changes in input/output singular frames. ISO-Merger combines frame changes of specialists offline; ISO-Online applies fixed-spectrum updates online. This rethinks the optimization layer converting reward feedback to weight updates.

In DreamerV3 continual RL, the world model retains everything measurable about old tasks (reward discrimination ~1.0, value estimates, termination structure) while the actor's behavior collapses. Forgetting is a channel problem, not memory. Freezing world model with identical imagined rollouts and retraining actor recovers performance.

Mage-Flow introduces a compact 4B-scale generative stack for efficient text-to-image generation and instruction-based editing. It co-designs Mage-VAE (lightweight latent tokenizer with one-step diffusion encoding/decoding and anchor-latent regularization) with a Native-Resolution Multimodal Diffusion Transformer trained via rectified flow matching. Achieves >10x tokenization cost reduction with native-resolution packing and CUDA kernel fusion.

HOST (Human-to-robot One-Shot Skill Acquisition) enables robots to acquire novel manipulation skills from a single human video in ~29 seconds, achieving 62% average success on 50 novel tasks while preserving previously mastered skills without policy parameter updates.

This paper introduces a causal interpretability framework for diffusion transformers (DiTs) combining attention decomposition with targeted interventions. It discovers that structural template tokens, despite carrying little prompt-specific information at encoder output, emerge as dominant image-to-text attention sinks and act as implicit semantic registers maintaining object identity. These tokens acquire identity indirectly through prompt semantics injected into image latents.

SLAI T-Rex presents full-parameter post-training of trillion-parameter MoE models (DeepSeek-V4 family) on Ascend NPU SuperPOD. Develops hierarchical optimization across model parallelism, computation-communication orchestration, and kernel execution, achieving 34.22% MFU with 2.93x improvement over open-source baseline while maintaining training stability.

HijackKV reveals a new threat in position-independent KV cache reuse: KV caches retrieved by token match encode attacker-controlled prefix context. When reused in victim queries, contaminated KV silently hijacks model behavior without attacker-controlled tokens in the victim query. Affects system optimizations for cache efficiency.

AlayaWorld presents a 15B video diffusion transformer for interactive long-horizon world modeling, generating 24-fps video at 540p/720p. It generates short latent chunks autoregressively under camera trajectories and switchable text prompts, with bounded memory via a sliding window. The full technical report covers architecture, data, training, and evaluation for persistent spatiotemporal consistency and stable long-horizon generation.