Jarrod Barnes

I'm an ML researcher working on training and verification systems for open-ended scientific discovery.

I lead Dynamical Systems, a research lab building agentic discovery systems for self-driving labs.

Research threads

Scientific autoresearch Scientific Autoresearch: Virtual Laboratories for Physical R&D Dynamical CLI gives scientific agents one interface for virtual campaigns and independently authorized physical experiments. technical report code skills

Scientific post-training Training Scientific Judgment from Physical Experiments Replay environments compiled from physical experiments train a 35B open model for evidence-conditioned campaign control. technical report code model

Scientific evaluation Dynamical-SDL-1: Measuring How Scientific Agents Learn from Physical Experiments An environment compiler and closed-loop evaluation of how scientific agents convert physical evidence into better decisions. technical report

Benchmark construction Latent Mining for Verifiable Scientific Task Generation A benchmark-construction method that turns disagreement between public scientific models into verifier-backed agent tasks. The first release is 165 regulatory-genomics triage tasks where 6 frontier agents reach an 11.3% aggregate strict solve rate and none exceed 15%. paper blog dataset

Verification Scaling test-time verification for novel materials Probe-gradient guidance steers an unconditional crystal diffusion model toward a target band-gap window without retraining the generator. Reported gains are computational; physical qualification remains untested. blog code model

Post-training Self-improving pretraining as a substrate for agentic post-training A Qwen3-0.6B experiment in continuation selection, interleaved thoughts, and RL mid-training, with agentic transfer still untested. blog code model dataset

Interpretability Do language models know when to change their mind? Signal-detection experiments measure whether models revise for valid critique while resisting invalid critique when source pressure conflicts with evidence. blog code

Research index

Systems recipes

Async RL Running async RL on my DGX Spark cluster A topology-aware change cut weight sync from 226.5s to 1.73s and step time from 270s to 42s on two DGX Sparks. recipe code

Rollout serving Optimizing throughput for async RL DeepSeek-V4-Flash serving on two DGX Sparks: 731.5 tok/s sustained on distinct prompts and a separate 1,106 tok/s shared-prefix result after a workspace patch. recipe code

ML systems recipes

Essays and working notes

A Theory on Becoming an Expert Learning systems / Updated Sep 8, 2026 Building a World Model of Consequence Learning systems / Updated Sep 8, 2026 My Agents Keep Failing. Yours Will Too. Learning systems / Updated Sep 8, 2026 Everything is Changing...Again Learning systems / Updated Sep 8, 2026

All writing