research

highlights from papers and research-adjacent systems I am working on.

TCG

Tool-call choice geometry: locating covert channels in the tiny surface-form decisions agents make when calling tools.

Covert Steganographic Communication Between Vision Models via Latent Space Perturbation

Training-free latent steganography probes for diffusion VAEs, with defenses that test whether hidden image channels survive decode-reencode.

TRAK-Traj

Scalable trajectory-level data attribution for robot imitation learning, using TRAK-style projected gradients to prune demonstrations.

SAE-Scope

Sparse autoencoder tooling for diffusion robot policies, aimed at finding editable features for objects, skills, and denoising-time corrections.

SafeSAE-VLA

Interpretable safety monitoring for vision-language-action models through sparse features tied to collisions, force, progress, and failures.

Vinsta

An identity, discovery, and trust layer for agents: handles, signed agent cards, DID docs, A2A messaging, MCP access, and searchable profiles.