Articles

Some sparse autoencoder features fire on exactly one token, which makes them the one case where a causal claim can be checked against ground truth. Deleting one costs the model that token in 178 of 208 layer conditions, and a second dictionary trained on the same model mostly disagrees.

One automaton, built from agent traces, serves as a memory, prediction, and monitoring prior. A compact directly-follows automaton recovered from execution traces unifies workflow memory, next-step prediction, failure prediction, and runtime monitoring across 12 datasets and 8 domains.

Ask a model for the probability something succeeds, then for the probability it fails. The two numbers rarely sum to 100, and the gap leans one way. An interactive look at directional bias in language model probability judgment.

Language agents read every external tool result but are blind to their own internal state. PaaT exposes activation-derived safety probes as an agent-callable read_probe tool — and the same probe flips sign depending on whether the agent chooses to look or the system injects it.

When agents work the same task, they can't communicate in real time. An interactive look at coordination, not concurrency, as the active ingredient: a real-time coordination room over a CRDT-merged shared workspace.

Interactive exploration of how language models encode confidence — intrinsic dimension analysis, layer evolution, activation steering, and cross-architecture transfer.

Interactive exploration of Control Reinforcement Learning — learned per-token feature steering with branch analysis, critic trajectories, and layer-wise interpretability.

An interactive exploration of correlation-guided feature selection for controllable language model behavior using Sparse Autoencoders.

Investigating how dataset composition affects sparse autoencoder feature matching and density patterns.

Understanding in-context learning by reversing transformer representations, exploring phase changes and feature dimensionality.

Exploring how the superposition hypothesis in neural networks relates to steering language models using sparse autoencoders.