MIT Team Publishes LASER: Latent-Space Entropy-Regularized Offline RL Algorithm, Self-Reported SOTA on 40 Benchmarks

A research team from MIT submitted the preprint LASER (Latent Space Adjoint Matching for Support-Constrained Entropy-Regularized Offline RL) on October 6, 2026, introducing a new offline reinforcement learning algorithm. The work has been accepted at NeurIPS 2026.

Offline RL enables policy optimization from static datasets without costly online interaction but remains bottlenecked by the risk of executing out-of-distribution (OOD) actions. Recent approaches mitigate this by learning behavior-cloning policies through flow matching and performing RL within constrained latent spaces; however, naively optimizing latent policies can cause brittle mode collapse or exploit sharp artifacts of learned critics.

LASER identifies entropy regularization as essential for latent-space RL. The algorithm applies latent-space adjoint matching to achieve entropy-regularized RL with expressive flow policies while avoiding backpropagation through time. In experiments across 40 challenging OGBench tasks with varying dataset qualities, authors report that LASER achieves state-of-the-art performance using fixed method-specific hyperparameters, outperforming baselines that require task- and dataset-specific tuning.

This is an author-reported benchmark result (E2 evidence grade) without independent reproduction. Project website: mit-realm.github.io/laser.