Lauren Hyoseo Yoon

Agents stay generic and disposable; the only thing that persists is a shared, curated knowledge base. Transient agents attempt tasks and debate across them, and what survives is distilled into reusable bundles — higher solve rates at lower cost, transferring zero-shot across model families.

LLM-guided evolutionary search collapses onto a single program family. VendiEvolve puts the Vendi Score into the loop as a control signal, steering selection toward diversity so the search keeps its stepping-stones to better solutions.

What shapes a model's inquiry when no question is handed to it to begin with? This work introduces a bottom-up forum, modeled on Philosophy for Children, where LLM agents read a philosophical narrative and propose, select, and answer their own questions, reaching a shared conclusion with no facilitator or aggregator. The aim is to test whether the freedom to formulate questions genuinely produces self-directed inquiry, with the ECO framework as the lens for tracing where that inquiry draws from.

Aligns frozen, independently trained vision and language backbones via small autoencoders meeting at a shared bottleneck. A novel Spread Loss turns the Platonic Representation Hypothesis into a training objective — matching a fine-tuned CLIP at ~1/3 of the FLOPs. Google DeepMind; with Been Kim & Yisong Yue.

Writes spatiotemporally patterned percepts into cortex via an addressable micro-LED array and decodes them to behavioral choice — installing a representation and reading out what the animal made of it.

A perception-to-representation pipeline: lifts markerless video to 3D pose, then aligns it with unsupervised population-level neural trajectories — movement and neural dynamics in one shared low-dimensional space.

A real-time connector system for the brain — a GUI front-end over an NFC/RF link driving up to 256 battery-free implants, with behavior read back from video.