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 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

Scientific discovery The Missing Layer in Autonomous Science Verified campaign environments convert search, trust, escalation, and revision into a multi-turn RL problem with physics-grounded oracle reward. Trajectory-level RL on 247 environments lifts held-out hypothesis accuracy from 55.2% to 79.3% with 3B active parameters, ahead of GPT-5.4 at 72.4%. blog paper code model

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 extracts band-gap signal from an unconditional crystal diffusion model and steers sampling without retraining. A 0.957-AUROC probe moves a model trained on 97.9% metals to 24-43% of samples in the target band-gap window, comparable to conditional generation at over 50x lower sampling cost. blog code model

Test-time scaling Simulator-verified skill acquisition for scientific instruments A persistent agent learns to operate an unfamiliar scientific instrument by drafting a procedure, running it against a simulator, and repairing it from what fails. It improves from every attempt but never grades its own work; a separate verifier the model cannot change controls what enters the skill library, and the whole loop self-verifies in simulation before any hardware is touched. Across three instrument families it acquired, repaired, and evolved skills without ever promoting an invalid one. technical report code skills

Post-training Self-improving pretraining as a substrate for agentic post-training Ordinary pretraining text becomes continuation-choice data, interleaved-thought supervision, and RL mid-training environments for shaping agentic behavior earlier in the model lifecycle. 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 2-machine async RL loop with NeMo RL, Polar, vLLM, trajectory contracts, and a 131x weight-sync speedup. recipe code

Rollout serving Optimizing throughput for async RL Serving DeepSeek-V4-Flash as a rollout engine on 2 DGX Spark machines, from 31 tok/s to 843.9 tok/s and then 1,106 tok/s after the decode-workspace patch. recipe code

ML systems recipes

Essays and working notes

A Theory on Becoming an Expert Learning systems / May 4, 2026 Building a World Model of Consequence Learning systems / Nov 19, 2025 My Agents Keep Failing. Yours Will Too. Learning systems / Jul 16, 2025 Everything is Changing...Again Learning systems / Apr 30, 2025

All writing