New 90-page review unifies statistics and foundation models via context-adaptive inference
anshulkundaje · x · 2026-11-03
- A new 90-page review offers a unified view of statistics and foundation models through the lens of context-adaptive inference.
- Core claim: statistical estimation is being absorbed into context. Classical tools (varying-coefficient models, local/kernelized smoothing, hierarchical multitask sharing) are framed alongside meta-learning (bilevel training, MAML/adapters, hypernetworks) and foundation-model mechanisms (prompted inference/ICL, linear readout on prompt-induced features, retrieval and tools as contextual signals).
- Unifying bridge: context acts as parameters, transmitted via support, similarity and structure; shared tuning knobs are context information, inductive bias and compute.
More from Research
- New study probes 10 vision foundation models across 4 levels: how much spatial information survives? — liuziwei7 · 2026-11-03
- Schmidhuber revives his 1991 fast weight programmers: the first linear Transformer variant — SchmidhuberAI · 2026-11-03
- Google & HKU's VeriCUA cuts false task-completion rate to 5.4% by checking saved files, not agent claims — rohanpaul_ai · 2026-11-03
- Direct Weight Editing via Codex+Astra Lifts Audio Classifier Accuracy From 84.3% to 88.75% Without Retraining — dansitu · 2026-11-03
- TSFLab: A Self-Improving Agent-Agnostic Workbench for Time Series Forecasting — D2I-Group · 2026-11-03
- YC's Momentic launches QA-Bench: its AI QA agent finds 2.7x more reproducible bugs than Codex — ycombinator · 2026-11-03