OTF-LAM
Summary
OTF-LAM is a latent action model family that first factorizes observation transitions into reusable local observed-transition primitives, then aggregates those primitives into state-aware action-like latents. OTF-LAM-Dino is the decoder-free variant that predicts future frozen DINOv2 representations.
Role In The Wiki
OTF-LAM sharpens the Latent Action Models page by naming the main ambiguity problem: observation-only transitions show effects, not causes. A monolithic latent action can therefore mix the controlled agent with distractors, camera motion, passive objects, and background dynamics.
The model is useful as a design pattern for missing-action settings, but it is also a warning for Alex’s time-series agenda. If a system can log actions, control inputs, interventions, outcomes, timing, and failures, the wiki should prefer that explicit data contract over recovering action-like codes from observations after the fact.
Official Artifacts
- Preprint: arXiv:2606.30544
- Official project page: hazel-heejeong-nam.github.io/LAM
- Official code: Hazel-Heejeong-Nam/lam_agent_ambiguity
- Public weights: not verified at ingest time.
- Dataset release: not verified at ingest time; the repository includes code for DCS environments and data collection.
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in otf-lam-2026 rather than duplicating verdict rows here.
At the entity level, OTF-LAM belongs to the action-discovery branch of world-model design. Its key contribution is the factorized transition-effect interface; its key boundary is that inferred latent actions are not typed actions, control inputs, or interventions until they are aligned and validated.