Explorative Models (XM)
Summary
Explorative Models (XMs) are generative models trained by exploring several latent-conditioned candidate outputs and updating only the candidate that best matches the observed target. The mechanism moves candidate search into training and is intended to reduce mode averaging without requiring extra candidate selection at inference.
Role In The Wiki
XM is the candidate-width branch of probabilistic generation in this corpus:
- VJEPA makes the predictive distribution explicit but currently uses a unimodal Gaussian head.
- FlowWM learns a conditional flow over future feature trajectories.
- XM instead trains a generator through best-of- candidate matching.
Forward XM emphasizes target coverage but grows in training cost with . Reverse XM can search many targets cheaply but is mode-seeking and needs entropy or coverage control. The implemented hard-min objective should be treated as a proposal-coverage mechanism, not as proof of calibrated mode probabilities.
Evidence Boundary
The source reports positive image, video, masked-language, behavior-cloning, and Maze2D results, including lower processed-sample/FLOP requirements in an RAE recipe and much lower inference NFE on the tested control tasks. The evidence is a fresh preprint, and the public code release does not yet include all headline RAE, XMDLM, and control experiments or any model weights.
The paper’s “third pretraining axis” framing is a promising finite-regime hypothesis. It is not yet a universal scaling law over parameters, unique data, training steps, exploration width, wall-clock cost, and model families.
Official Artifacts
- arXiv 2607.27372v1
- Official project page
- Official author blog
- Official Apache-2.0 code repository
- Official X thread
The verified 2026-07-31 repository snapshot had no public release, model weights, or checkpoints. Its README marks the RAE-specific implementation, masked-diffusion language-model experiments, and control-task code as forthcoming.
Evidence
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in Explorative Modeling rather than duplicating verdict rows here.
For time-series work, XM is a direct baseline for the averaged-future problem: draw raw or latent future trajectories, train on the one closest to the observed future, and evaluate whether this preserves separated regimes. It earns predictive-belief credit only if the resulting sample frequencies are calibrated and candidate actions move probability mass correctly; best-of- coverage alone is insufficient.