• Watermarks Change Generation Paths and Can Cause Factual Errors: All six tested methods showed this effect. At a 0.90 detection true positive rate and 1% false positive rate, two combined plug-ins reduced relative factual errors from watermarked decoding by about 90%.

• Verified Deployment Trajectories Can Become Lasting Skills: ASCENT lets a frozen model reinterpret an attempt using its final outcome, then distills that hindsight distribution into LoRA fast weights. Its gains depend on related tasks and trustworthy verification signals.

• Prism Adapts Sparsity to Shifting Audio-Visual Information: It adjusts block shapes using visual change directions and audio-visual coupling. A mixed-block policy sets sparsity per query, delivering 2.5× faster training than full attention while improving generation quality.

Also Notable

• Future-Action Disagreement Decides When to Wake the Slow Generalist — TUD used 75% fewer calls than the strongest fixed-interval baseline in real-robot experiments while achieving a higher success rate.
• Prune Visual Tokens Only After Attention Becomes Reliable — SAPrune requires no training. It removes 87.5% of visual tokens and delivers up to 1.718× faster inference while maintaining competitive task success rates.
• Where Extra Reasoning Resumes Changes Accuracy — A learned router beat uniform placement but showed no gain over always restarting from the last qualified step.
• Multiscale Log-Signatures Compress Irregular High-Frequency Sequences First — LogSig-SSM then models long-range dependencies with a selective SSM. On the longest sequences, it reports up to 30× faster training and 37× lower memory use than Mamba.
• Euclidean Symmetry Goes Directly Into the Diffusion Solver — EDISCO handles TSP and CVRP with 33%–50% of the training instances. It also improves resilience to changes in position, orientation, and spatial distribution.
• Graph Evidence Does Not Mean the Model Uses It — ColdDDI tests masking, drug replacement, and channel sensitivity to diagnose gaps in mechanistic evidence use for cold-start drug-interaction prediction.
• LLMs Sometimes Win GNN Contests but Lack Consistency — GNN-CB covers 18 real competition-style tasks. Under its evaluation protocol, humans still achieve the highest score on most tasks.
• Each Emotion Maps to Only 23–27 Components — These attention heads and MLPs make up about 5% of the examined components. Activation patching finds sparse causal paths and links internal interventions to pitch, energy, and spectral brightness in speech.
• Messy Interleaved Logs Can Still Produce Skills — TeleTune iteratively updates a text skill library using action-prediction errors. Workflow retrieval then finds examples that cover subtasks in new tasks.
• Decodable Operations Knowledge May Not Drive Predictions — Residual-swapping experiments on Toto show that recoverability and causal influence need separate tests.

Read the full edition →