Phoebe Klett

Phoebe Klett is a machine-learning researcher and coauthor, with Thomas Ahle, of Extended Mind Transformers, which enable language models to retrieve external information through their own attention mechanisms without additional fine-tuning. Her research spans reliable text generation, transformer memory, Bayesian uncertainty, and physics-based computation.

Klett studied mathematics at Duke University, where her senior thesis implemented algorithms for monomial ideals in Macaulay2. She subsequently worked as a data scientist at Red Ventures and was a machine-learning engineer at Normal Computing in 2024. Her earlier work there included finite-state constrained generation, using regular expressions and finite-state machines to produce structurally valid model outputs.

• Transformer-native external memory. Klett and Ahle store external material as key-value representations, letting individual generated tokens retrieve relevant memories as an answer develops. Their open-source implementation supports relative positional methods used by Llama and MPT models without retraining.

• Counterfactual retrieval evaluation. Their benchmark replaces familiar facts with plausible alternatives to test whether models follow supplied evidence instead of defaulting to information memorized during pretraining.

• Causal citations and adaptive retrieval. Tracking retrieved memory tokens makes supporting evidence more directly inspectable; uncertainty-sensitive generation can retrieve additional information when a model becomes unsure. Klett detailed these mechanisms and their computational tradeoffs at AI Engineer World’s Fair 2024.

• Bayesian learning and thermodynamic inference. Klett coauthored ICLR 2025 research introducing the open-source posteriors library for scalable Bayesian learning. Her work on thermodynamic Bayesian inference investigates analog devices that sample probability distributions through physical Langevin dynamics, potentially reducing the energy and computational costs of uncertainty estimation.