Poisson Labs

Poisson Labs is an independent research group in Brooklyn working on multi-agent RL, sim-to-real, and adversarial simulation for LLM agents, robotics, and non-deterministic systems.

We take on research and engineering work for labs, startups, and teams shipping agentic systems.

Recent research

  1. 2026.08 Map the Failure Boundary We froze a Go1 locomotion policy and measured where it fails across 6,400 combinations of friction and lateral push, then retrained on the gaps and measured again. With a live map you can probe.
  2. 2026.05 Monte found a decorative channel and a reward exploit in one benchmark An adversarial run exposed two failure modes at once: a communication channel that looked load-bearing but wasn't, and a reward function that rewarded trivial policies.
  3. 2026.04 Where V-JEPA 2.1's dense features hold up — and where they don't A pre-registered robustness study across all four V-JEPA 2.1 sizes, with practical implications for robotics deployment.
  4. 2026.01 10,924x: the instability bomb at 1.7B scale Scaling the mHC reproduction to 1.7B parameters on 8x H100s. Hyper-Connections explode; Sinkhorn-projected mHC stays flat.
  5. 2026.01 DeepSeek's mHC: when residual connections explode Reproducing manifold-constrained Hyper-Connections at 10M parameters. Why unconstrained mixing matrices break at scale.

Currently building

Monte — adversarial simulation for MARL research and LLM-based agents. Co-evolving adversaries that find failure modes before deployment does.

Contact

— for new engagements
— everything else