• Community citation signals can train "taste." RLCF uses 700K paper pairs for preference modeling, producing a judge that outperforms GPT-5.2. The paradigm transfers to any domain requiring taste-based decisions.

• Classifier blind spots hide in the null space. SING translates geometric invariants of linear mappings into natural language. Auditing what a model ignores before deployment beats chasing accuracy.

• Model behavior is far more sensitive to context wording than expected. Changing task descriptions systematically shifts performance. Whether or not this constitutes "motivation," the manipulability itself is an alignment problem.

• VLA multi-task bottleneck is the inference system, not model architecture. OxyGen manages cross-task KV cache as a shared resource, computes shared visual observations once, and achieves up to 3.7x speedup.

Also Notable

• Multi-Agent Research Framework Reaches 300+ Skill Modules with Artifact Lineage Tracking. Zero-central-coordination architecture with impressive engineering completeness. Source
• RL Plus Visual Perception Prompts Improve Spatiotemporal Grounding in Video Reasoning. No extra annotation data or inference-time external tools required. Source
• FOMC Statement Hawk-Dove Analysis Modeled as Relative Change, Not Absolute Classification. Better matches how markets actually react to incremental wording shifts. Source
• EPFL Adds Spectral Clipping to AdamW. Targets gradient spectral concentration and norm explosion, two persistent problems in large-scale training. Source
• Black-Box Trust-Region Search Aligns Diffusion Models at Inference Time. Doesn't require a differentiable reward model, broadening applicability (ICLR). Source
• Bayesian Network Classifiers Compiled into Logic Formulas. xAI team turns classification decisions from statistical black boxes into verifiable logical derivations. Source
• Computational Argumentation Frameworks Add a Reasoning Exoskeleton to High-Stakes LLM Decisions. Every conclusion is challengeable and traceable (Imperial College). Source
• Multi-Vendor Mammography Dataset LUMINA Released. Energy calibration protocol tackles cross-device generalization, a persistent barrier to medical AI deployment (CVPR). Source

Read the full edition →