Intelligence that accrues.

The year is 2026, and language models still don't learn from experience. This remains the most important unsolved problem in AI.

Accrue is a research project on continual learning. We aim to push the pareto frontier of learning plasticity and stability.

At FR8 Cohort 2.s, we developed CLUE, a method that is, to the best of our knowledge, the first to consistently achieve positive backward transfer on the Long Sequence Benchmark. This means that our method goes beyond catastrophic forgetting and lets different capabilities synergize instead of overwriting each other.