• Silicon Panels Match the Mean and Distort the Variance. Stanford used 277 professional philosophers as ground truth; seven open and closed models all replicate the aggregate distribution, but cross-question correlations come out systematically inflated and minority views collapse. Anything that depends on the shape of disagreement gets a smoothed signal.

• Emotion Probe Accuracy Drops From 82% to 5% Once Keywords Are Removed. MIT's AIPsy-Affect ships 480 paired stimuli with all emotion words surgically excised, and most published "emotion features" lose their signal under this baseline. New probing, SAE, and steering work without a keyword-free control deserves a discount.

• Metanetwork Symmetry Constraints Get Loosened a Notch. Quasi-equivariant networks abandon full theoretical symmetry and hold only on the equivalence classes that actually appear in real data, verified across feedforward, conv, and transformer architectures.

• FinGround Drops Hallucination Rate 68% by Grounding Atomic Facts to Regulatory Passages. An 8B distilled detector keeps 91.4% F1 at $0.003 per query. The EU AI Act's August 2026 deadline turns "fewer hallucinations" from product risk into compliance risk; the verify-then-ground decomposition ports cleanly to any vertical needing fact-level traceability.

Also Notable

• RouteNLP Closes the Routing-Distillation Loop. The large model only handles queries that genuinely need it, and the small model is continuously distilled from the real routing distribution rather than treating both steps as separate optimization targets.
• MTRouter Co-Embeds Conversation History and Candidate Models. Multi-turn routing stops looking only at the current turn; cumulative cost across turns gets explicit modeling for the first time.
• AgentEval Builds Agent Evaluation as a DAG With Explicit Error Propagation. Same family as recent MAS attribution work, but the focus is the step chain inside a single agent, not blame distribution across agents.
• ComplianceNLP Adds a Knowledge Graph to RAG for Compliance Gap Detection. Dependencies between regulatory clauses get explicit graph structure, which fits inter-policy conflict queries better than pure vector retrieval.
• S2G-RAG Makes "When to Stop Retrieving" a Learnable Decision. The model explicitly judges whether current evidence suffices and which gap remains; iterative-retrieval stopping conditions get structured for the first time in multi-hop QA.
• Deductive, Inductive, and Abductive Reasoning Have Distinct but Overlapping Representations Inside LLMs. Targeting a specific reasoning type opens an actionable training surface, instead of treating "logic" as one capability.

Read the full edition →