Research
Papers and studies from the lab.
Claims research study: 9.4× faster and 90% lower cost, with 10.9% higher verdict accuracy
We trained a Qwen3.5-4B model for claims research. On 500 held-out development claims from the public AVeriTeC benchmark, it delivered a 10.9% relative verdict-accuracy gain over Claude Haiku 4.5, with 9.4× faster processing and 90% lower run cost.
4 minComputer use study: 97% cheaper than Opus 4.8, and 50% better than baseline
We trained a macOS computer use agent to help a customer get off a frontier API and onto a 9B model. When evaluated in closed-loop simulation on tasks resembling production, we measured a +50% gain from training.
3 minSize Doesn't Matter: Cosine-Scored Sparse Autoencoders
Inner product scoring conflates directional alignment with input magnitude. We propose cosine scoring for dictionary learning on normalized representations. ICML Spotlight.
2 minCLaaS: Continual Learning as a Service for Sample Efficient Online Learning
Online continual learning for deployed LLM agents behind a chat API: experience replay for gradient reuse, superior forward transfer, and less forgetting than in-context learning.
2 min