Area 01
Efficient reasoning
Models that spend compute where the difficulty is, and stop when more thinking won't help. This work produces the Kaer Reasoners.
02Research
What we study
We work on two questions: how to make models reason well on less compute, and how to tell whether a model is doing its job once people rely on it. We'll publish methods and results here as the work matures.
Area 02
Evaluation
How models behave on everyday questions, judged by the people who do the work. It's also the method behind the studies we offer public services.
Area 03
Alignment and interpretability
Why models behave as they do, and whether that behaviour holds up as they are updated over months.
- Can a model learn when more thinking will help?Training a halting signal that decides, question by question, whether to keep reasoning, and when to hand over instead.
- Which training data improves reasoning?Measuring each part of a training mix by what changes when it's taken out.
- How much does wording change an answer?Testing the same question in many phrasings, and reporting the spread rather than one score.
- How should an assistant be judged in real use?Realistic questions, scored by practitioners as helped, neutral or harmed.
- Does the model answer the question that was asked?Catching the moment a model quietly answers an easier question instead.
- How does behaviour change after launch?Monthly checks that pick up changes in manner before they show up in accuracy.
What every study includes
Five things will ship with everything we publish, so anyone can check the result or rerun it.
- 01The protocol
Written before the results, so the method can't bend to fit them.
- 02The cost
What each score cost to produce, at batch pricing.
- 03The variance
Spread across phrasings and seeds, never a mean on its own.
- 04The limits
Where the result stops holding, stated up front.
- 05The failures
Variants that didn't work, and bugs we found in our own tools.