• Post-Training Can Improve One-Shot Accuracy While Shrinking Solution Coverage. Across 42 comparisons spanning four model families and three agent benchmarks, pass@1 and pass@K sometimes moved in opposite directions. With sufficient test-time budget, base models often surpassed their post-trained counterparts in pass@K solution coverage.

• Image Generators Can Simulate GUI Environments. AutoGUIWorld creates interaction traces without installing software. Fine-tuning raised the average OSWorld task score from 33.0% to 40.8%, but reliable state transitions, action targets, and quality filters remain essential.

• 4Director Turns Directorial Intent Into Executable 3D Constraints. It reuses a canonical mesh and applies rigid transformations frame by frame, improving cross-view consistency. The generative model still handles non-rigid changes.

Also Notable

• Continuous Latents Become the Only Persistent State, While Discrete Tokens Add Scaffolding. HC-DLM targets both broken dependencies in parallel decoding and validity gaps in continuous representations.
• More Distinct Text Embeddings Are Not Always Better for Diffusion. Soft-label distillation brings plausible substitute words closer, creating a more connected latent space that is easier to sample.
• Long-Horizon Agents Need More Than Compressed History. PoS explicitly tracks the current world, unresolved requirements, and stalled states. It checks belief consistency and selects recovery actions for different failure modes.
• Training Scenarios Should Change With Agent Failure Modes. ActiveSaddler's non-stationary curriculum improved test Pass@1 by 4.4 percentage points on GAIA2 and 7.5 points on Terminal-Bench 2.0.
• Fitting a Catalog Into Long Context Does Not Make It Usable. RPTune first learns product ranking and pruning, then applies post-training with rewards relative to the current context.
• Execution Traces Can Test Whether Videos Follow Rules. PROWBench uses program traces and timestamped events to move world-model evaluation from visual plausibility toward faithful execution.
• Long Audio-Video Reasoning Does Not Require Watching Everything. OmniSeek decides when to look and when to listen. Its rewards discourage shortcuts that solve tasks through only one modality.
• Compiling Lean Code Does Not Prove the Right Claim. Isolated execution, statement comparison, assumption audits, and independent review turn successful outputs into traceable evidence.
• Hybrid Transformer-SSM Architectures Can Reuse μP Learning Rates. Experiments span widths from 256 to 2,048, depths from 4 to 32, and billion-parameter models. Empirical results currently run ahead of theory.
• VideoLLMs Learn Temporal Order, Then Lose It Across Layers. Relevant representations peak in intermediate layers before fading. Reinjecting temporal-difference vectors at inference partially restores temporal reasoning without training.
• A Shared SE(3) Trajectory Representation Can Cover Objects, People, Cameras, and Robots. World Motion Models use different masks to support prediction, completion, control, and cross-embodiment transfer.
• Open-Weight Models Still Struggle With Exact Kali Commands. Across 1,642 Kali tools, none exceeded 42% exact-command accuracy without explicit tool hints. KaliBench also provides verifiable training rewards without running commands.

Read the full edition →