Goodfire

Goodfire experimentally applies model diff amplification to reveal rare post-training misalignment, reporting that in a Llama 3.2 1B experiment with 5% contaminated medical-advice training data, harmful responses increased from one in 10,000 ordinary samples to one in 30 amplified samples for the tested prompt.

Goodfire's RLFR research reports reducing hallucinations in Gemma-3-12B-IT by 58% when run with its probing harness, using internal-feature probes as reinforcement-learning rewards without degradation on its standard benchmarks.

Goodfire applies model interpretability tooling to frontier-model analysis and partner research, while broad production deployment remains early.

Goodfire’s 2026 research program advances parameter-level interpretability, predictive data debugging, scalable reward-hacking detection, uncertainty estimation and interpretability-guided training; its work includes reducing hallucinations by 58% using features as rewards and methods for targeted model editing.

Goodfire develops mechanistic interpretability and model-understanding techniques for analyzing and steering neural networks.

Goodfire develops mechanistic interpretability and model-understanding techniques for analyzing and steering neural networks.

Goodfire announced a US$150 million Series B at a US$1.25 billion valuation in February 2026, led by B Capital.

Goodfire identifies chief scientist Tom McGrath and documents that he previously led a research team at Google DeepMind, providing a named scientific leadership-capacity signal.

Goodfire has assembled a specialized interpretability research and engineering organization with frontier-model partnerships.

Goodfire's June 11, 2026 predictive-data-debugging research traces anticipated DPO behavior changes to preference-data clusters before training, and tests targeted dataset/training interventions using Dolci and Tulu examples across Llama models.

Interpretability, evaluation and deep-learning research skills transfer across frontier labs and safety teams, but the specialty remains narrower than general ML infrastructure.

Interpretability, evaluation and deep-learning research skills transfer across frontier labs and safety teams, but the specialty remains narrower than general ML infrastructure.