Latent Dynamics · 2026-10-08

LeCuration: A Tiny World Model as a Data Curation Multi-Tool

Mayank Sengupta, Nirmit Desai, Eric Song, Kunal Sawarkar

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors train LeCuration, a small predictive world model for data curation on CS:GO gameplay. Using LeWorldModel-style latent encoder/predictor with a DiT decoder and TAESD latents, they show that CLS embeddings support content-based clustering and anomaly detection via embedding divergence during autoregressive rollouts. The paper is a qualitative proof-of-concept and does not report quantitative curation metrics or downstream results.

Why it matters

The work targets an unmet need in physical AI data curation: learning what is typical, rare, or surprising in a dataset without physics-engine state access or labeled annotations. If embeddings and prediction divergence reliably highlight problematic or novel data, a small scout model could enable scalable curation before training larger downstream world models.

Method

Train LeCuration on CS:GO (de_dust2) gameplay by compressing frames into TAESD latents, encoding them to a 196-dim CLS world state, and using a causal transformer predictor over a 6-step history plus action embeddings to predict the next CLS embedding.
Use CLS embeddings for content-based clustering (UMAP + HDBSCAN) to organize gameplay into semantically coherent clusters and support degenerate-content filtering and rebalancing heuristics.
Run the predictor autoregressively and score unsupervised anomaly candidates using Euclidean embedding divergence between predicted and ground-truth encoded world states; decode embeddings with a DiT (for visualization).

Limitation

The current limitation is that the anomaly signal is not physics-specific: spikes can arise from accumulated rollout error or out-of-distribution actions, and isolating physics-caused divergence would require a model pretrained on a wide corpus of real-world data rather than just gameplay.

Abstract (from arXiv)

Many applications of physical AI run within finite or closed physical worlds with a limited set of physical laws governing object behavior. Examples include robots working in a warehouse and agents moving around in a video game. In order to better organize, filter, and curate data for physical AI applications, we propose a new approach centered on the unique settings and physical laws of individual datasets. We train LeCuration, a small world model intended to serve as a data curation tool for a separate, larger downstream model. To build this model, we choose LeWorldModel (LeWM)as our latent encoder and predictor, adding a diffusion transformer (DiT) decoder to add visuals to autoregressive gameplay rollout. We find that the embeddings of this model can be used as an anomaly detection signal and as a content-based clustering heuristic, and that auto-regressively predicting the game state with this model allows us to qualitatively check for action-state consistency. This paper presents a qualitative, proof-of-concept case study on CS:GO gameplay data; we do not yet report quantitative curation metrics or downstream training results, which we identify as the key next step.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-07 · Indexed by Pub-AI: 2026-10-08 UTC

preprint