closed-form

Strong Helps Weak: Directional Cross-Modal Alignment Transfer in Multi-modal LLMs (NeurIPS 2026 Spotlight) — training-free, closed-form merging that transfers alignment from a strong donor modality to a weak recipient modality.

NormaStream: streaming time-series forecasting without backpropagation or gradient training - closed-form ridge init + online RLS/Woodbury adaptation, with an optional fixed complex-valued LRU reservoir (NormaRC) for longer temporal memory.

CERATA (formerly PAL-MoE): learning after deployment as an API call on a frozen model. Closed-form edits, exact forgetting, and measured locality/reversibility guards on every write and forget. Pre-registered continual-learning research code in PyTorch.

oeis closed form miner

Six Independent Appearances of One Integer Across Scale

Closed-form nth-prime estimator built on invariant-normalization logic, with deterministic refinement and exact benchmarks across a contract grid spanning n = 10^2 to 10^24.

First closed-form solution to the Rate-Distortion-Perception tradeoff for Gaussian sources, extending Blau & Michaeli (2019).

High-Performance O(1) / O(log N) Algebraic Loop Acceleration & Exact Rational Recurrence Engine for Python and C