• Fixed block size leaves speedup on the table. BlockPilot predicts the best block per input instead of using one constant, reaching 4.20x lossless speedup on Qwen3-4B at temperature T=1.

• Treat "knowing what you don't know" as a training target, not a patch. Yale's RLMF uses metacognitive feedback as the RL signal so a model's stated "I'm unsure" matches its internal state, lifting faithful calibration by up to 63%.

• The tokenizer and generator stop training separately. GEAR's dual-readout design trains a VQ tokenizer and AR generator end-to-end, converging up to 10x faster on ImageNet gFID.

• Give "intermediate rewards" a checkup before you commit. QVal ranks dense-supervision signals by Q-value agreement without running downstream training, and across 1,200 experiments a simple prompting baseline beats most fancy methods.

Also Notable

• Same Target as BlockPilot: Against Fixed Policies — Stanford's LearnStop learns a hidden-state-agnostic checkpoint stopper and asks when a learned stopping rule actually beats a simple confidence or convergence threshold. Read it alongside the lead.
• GUI Agent Training and Eval Never Left Offline Trajectories — standard benchmarks are far from real apps' interfaces, interaction logic, and error-state distributions, so Xiaomi's GUI-0 technical report moves the whole pipeline onto real applications.
• Video World Models Lack Memory, So Long Scenes Drift — rule-based frame selection fails under occlusion and moving objects, so MemLearner switches to a learnable adaptive context-memory query.
• Text-Image Data Pipelines Crawl, Filter, Freeze, and Toss Rejects — DataEvolver uses a self-evolving multi-agent setup to recycle the signal buried in failed samples.
• Training-Free Photo Mosaics at Any Resolution — PhotoQuilt uses bootstrapped tiled denoising to satisfy both scales at once: each tile looks right alone, and the whole forms a scene.
• A Missing Domain Backbone for Sheet-Music Understanding — MuSViT is the first sheet-music vision foundation model, pretraining a ViT with MAE on 9.7 million IMSLP pages.
• Satellite-Image Synthesis Borrows the Natural-Image Recipe — dense rasters or sparse prompts both break the vector primitives geography relies on, so TerraDiT-Ω builds unified spatial control.
• End-to-End Self-Driving on Instant Sensor Data Only Reacts — PriorEye (ECCV) adds a geo-visual prior anchored to the streetscape, restoring the foresight humans get from experience.

Read the full edition →