ACTAVA ResearchFrontier research in vertical models and benchmarks.

Recursive self-improvement seeks to turn gains in an AI system into a stronger capacity to produce further gains. The survey organizes the field around automated design of AI infrastructure and the repair of existing systems, and finds that systems can keep useful changes but do not yet show reliably accelerating self-improvement.

A bidirectional gated delta network backbone for continuous language denoising, in place of full attention at every sampling step. On OpenWebText it lowers generated perplexity from 24.218 to 21.228, and it runs 2.72x faster at a sequence length of 16k.

Predicts whether a model's answer will be truthful before it decodes, by averaging layer-wise probe logits across the network. It achieves the highest AUROC in 23 of 24 model and benchmark pairs.

A one-trillion-parameter healthcare model post-trained through recursive self-improvement for patient care, clinical reasoning and agentic workflows. It leads on five of six healthcare panels.

A simulator of 21 healthcare applications and 75 long-horizon workflows across prior authorization, utilization management and care management. The best frontier agent solved 28% of tasks; end-to-end prior authorization, 0%.

A reproducible testbed for training and evaluating user-centric agents through gym environments paired with simulated users. SFT cold start, trajectory-level reward-to-go scoring and stronger simulated users all improve multi-turn reinforcement learning.

A gym environment for whether agents find out what users actually want when goals are underspecified, incremental or indirect. Models search and use tools well but elicit only a minority of hidden preferences.