Reasoning

Raises the score

• Deliberate-inference systems now sustain far longer chains before compounding error becomes fatal, and can allocate more computation to harder instances rather than answering at fixed cost.
• Competition-grade mathematics and formal proof assistance have moved from novelty to routine, including problems whose statements post-date the training corpus.
• Self-verification — a model checking its own intermediate steps and revising them — improves accuracy measurably rather than merely restating the original answer.

Holds it down

• Accuracy still degrades sharply when a familiar problem is restated with unfamiliar surface structure, suggesting pattern retrieval doing work we attribute to inference.
• Chains that look valid frequently reach correct answers through invalid intermediate steps, so scoring the output overstates the reasoning.
• Reliability, not peak capability, remains the binding constraint: the same system solves a problem and fails its near-neighbour.