Yadia / Research
Intelligence made inspectable.
Yadia is a research lab building multimodal AI that reasons causally. We post-train models with reinforcement learning to work out what causes what, and test by intervention that their reasoning drives their answers.
Multimodal reasoning post-training · Causal faithfulness · Factual representations
Present research
Three linked research programmes
We train multimodal models to reason about cause and effect, test by intervention whether their stated reasoning drives their answers, and measure what they represent inside.
Multimodal reasoning post-training
Can reinforcement learning teach multimodal models to reason about cause and effect, and how do we know the gain is real?
Current research
Causal faithfulness
Does the reasoning a model states actually cause its answer? We test it by intervention.
Current research
Factual representations
What do models represent internally, and which of it does their reasoning actually use?
Current research
Current work
Reinforcement learning for causal reasoning
Models that only learn what usually goes together break when conditions change. We post-train open multimodal models with reinforcement learning on tasks where the answer depends on what causes what, and test every gain causally, against a matched control that receives the same compute.
Each experiment is pre-registered: hypotheses, pass criteria and held-out test sets are fixed before training starts, and test sets are read once. Results are published when they are final.
Compute partnerships
Compute decides how much we can test
Our work is compute-bound: every claim needs a trained model, a matched control and a held-out evaluation. We welcome research credits and compute grants from cloud and GPU providers.
How compute is used- Post-trainingreinforcement learning runs
- Matched controlssame compute, no RL
- Pre-registrationcriteria fixed before training
- Evaluationheld-out tests, read once