Research approach

Start with the question

Start by defining the task, its context, and the limits of current methods. Then identify whether the next step calls for better data, a different training objective, a model change, or an inference improvement.

Character drift can involve context management and training data. Slow responses to motion prompts can involve causal modeling and scheduling. Each problem needs an experiment that tests the relevant mechanism.

Studying social intelligence

We study how models interpret context, relationships, intentions, and boundaries, and how they maintain roles and facts while expressing emotion, humor, and disagreement.

Natural interaction includes acknowledging uncertainty, accepting corrections, and giving users space. We study these behaviors through data, post-training, and evaluation.

Continuous generation and interaction

Real-time generation responds to input while earlier video is already playing. We study how identity, movement, timing, and spatial relationships carry across segments as instructions change.

We evaluate temporal consistency, causality, and control together. Speed measurements record hardware, resolution, duration, sampling, concurrency, and warm-up conditions, including both first-output latency and sustained throughput.

Comparable experiments

Meaningful comparisons hold inputs, versions, and procedures as consistent as possible. We record model identifiers, prompts, sampling settings, evaluation data, and judge configurations.

We assess subjective experience across raters, scenarios, and longer interactions. Automated evaluations account for judge preferences, prompt sensitivity, and ranking variation, alongside failure cases and individual capability scores.

• Record exploratory experiments and formal results separately.
• Report average performance together with the range of variation and the main failure modes.
• Run ablations or targeted retests on key changes to pin down where improvements come from.
• State the coverage of the samples clearly, and avoid generalizing limited experiments into universal conclusions.

Engineering as research

Deployed models face network changes, limited resources, and shifting user intent. Speed, memory use, deployment complexity, recovery, and cost determine whether a capability works in practice.

We study models and systems together, testing caching, scheduling, task partitioning, and compression against the final experience. Efficiency improvements must preserve the quality and safety the task requires.

Research others can review

Share research in a form that fits its stage of development. Papers explain methods, and evaluation reports document results. Releasing code or weights also requires reviewing licenses, data rights, and potential risks.

We distinguish research directions, experimental previews, released products, and formal commitments. Product documentation and release materials define available features, performance, and licensing.