Research
Research that makes AI more reliable.
Our team publishes research on making AI more accurate and trustworthy. Here is what each paper means for the systems we build for you.
Why it matters
AI that shows its reasoning and corrects itself before giving you an answer.
Process Supervision via Verbal Critique Improves Reasoning in LLMs
- What we studied
- A training-free generate–critique–refine framework for structured reasoning at inference time.
Why it matters
More dependable AI assistants for software teams.
Verbal Process Supervision Elicits Better Coding Agents
- What we studied
- CURA, a code understanding and reasoning agent enhanced with verbal process supervision.
Why it matters
Fewer made-up answers from AI, without retraining the model.
Delta — Contrastive Decoding Mitigates Text Hallucinations in LLMs
- What we studied
- An inference-time decoding method designed to suppress hallucinations without retraining or additional data.
Next step
Start with a 4–8 week sprint.
No hardware to buy. Tell us one task that takes too much time, and we will show you what AI can realistically do for it — and how to keep your data safe.
You will leave with
- A clear view of what AI can do for you
- Your privacy and security questions answered
- A practical first step, with no hardware to buy