• Pivot-SD Beats Full-Sequence SFT With Sparse Supervision. It used 200 problems and four samples each to outperform full-sequence SFT and same-budget diffusion RL. Results cover only math and coding benchmarks on LLaDA-8B-Instruct.

• Better Text Alignment Can Weaken Physical Regression. Contrastive training improved retrieval across 8,000 controlled simulations from 80 families, but hurt physical attribute regression. Better reference-video retrieval also improved MiniMax-H3’s physical fidelity.

• Queen Climbs From 1782 to 2697 Elo in Seven Rounds. The 4B chess language model connects a chess encoder to an instruction-tuned language model through cross-attention. Language-model-based fluency and coherence scores cannot establish explanatory faithfulness.

Also Notable

• Static Reconstruction Does Not Transfer to Time-Varying Scenes — 4DCodeBench asks agents to turn video deformation, fluids, and fractures into executable graphics programs.
• Target Frames Help Autoregressive Video Models Plan Ahead — ProAR constrains long-range outcomes and short-range transitions separately. It surpassed a fully trained standard autoregressive baseline using only 25% of the training steps.
• Fine-Grained Multi-Model Selection Beats Single-Model Bias Baselines — CBM selects and organizes multiple models by fine-grained behavior. Its gains must be weighed against the inference cost of multi-model coordination.
• Action Policies Can Pretrain Without Action Labels — NAVA-WAM transfers supervision from future video changes into policy learning, reducing dependence on labeled robot trajectories.
• Program Evolution Should Optimize Improvement per Cost — FrugalEvo assigns strategy exploration to stronger models and implementation to cheaper ones. BA-AUC explicitly measures progress across the cost curve.
• Browser Agents Now Need Cross-Modal Evidence Chains — HyperBrowseComp contains 423 human-written and verified questions across 13 languages. Sources include videos, scanned documents, images, maps, and obscure material.
• Ordinal Filtering Can Replace Repeated Group Sampling — FTW reuses trajectories through a replay buffer. It trades CPU memory for fewer samples in critic-free reinforcement fine-tuning for stateful environments.
• Models Can Learn Which Modality to Acquire First — ECHO-k uses pretrained internal representations as a proxy objective, producing acquisition orders independent of labels and specific downstream tasks.
• Reasoning Data Synthesis Changes Tasks and Tooling Together — Online improvement converts intermediate failures into reusable skills. Post-batch updates revise skills, prompts, and workflows while model weights and verification criteria stay fixed.
• Continually Adapting World Models Should Forget Stale Facts — The paper separates permanent physical laws from revisable instance facts by invariance timescale. It proposes reporting both invariant regression and revision delay.
• Unlabeled Target Recordings Can Generate Multimodal Adapters — ZeroMAG adds accompanying physiological signals while keeping the EEG encoder frozen. It trails supervised multimodal adaptation by only 0.50 percentage points on average.
• A Restricted Decoder Preserves More Transferable 3D Geometry — SNAP combines a local decoder with latent reconstruction. Under camera shifts, it degrades more slowly than standard 2D representations.

Read the full edition →