The paper investigates the closed‑loop behavior of a recurrent state‑space model (RSSM) trained on human demonstrations in LunarLander‑v3. Using the model for zero‑shot CEM MPC and actor‑critic training in imagination, the authors find that a model‑based A2C policy trained on a single checkpoint achieves returns comparable to the best model‑free A2C policy while requiring far fewer real transitions. They introduce the Reward Observability Fraction (ROF) to quantify how much of the reward gradient lies in the observable subspace of the latent dynamics, but show that ROF does not predict the collapse of MPC in offline settings, leaving the problem of predicting closed‑loop failure unresolved.
arXiv:2607. 01736v1 Announce Type: cross Abstract: We study how to predict the downstream closed-loop performance of a learned latent world model from validation-time diagnostics alone.
We study how to predict the downstream closed-loop performance of a learned latent world model from validation-time diagnostics alone. Choosing the right checkpoint from a world-model training run is difficult: validation loss and multi-step prediction RMSE keep improving long after closed-loop performance has collapsed.
The paper introduces the concept of intervention fidelity in latent world models, measuring whether a model’s open‑loop transitions align with actual environment interventions. Experiments on TD‑MPC2, Cheetah, and DreamerV3 show that high reward fit does not guarantee fidelity, and that self‑supervised models can outperform task‑anchored ones in preserving intervention effects. The authors propose a capture‑gated audit to localize failures and argue that fidelity must be directly audited on the model’s native interface.
AD-WM is a new action‑discriminative joint‑embedding world model designed for counterfactual model predictive control. It augments residual latent dynamics with action‑recovery regularization based on inverse dynamics and conditional mutual information, while discarding auxiliary heads at test time so that MPC remains unchanged. Experiments on OGBench‑Cube and other simulation environments show substantial gains in hard‑start success and mean success, and zero‑shot transfer to a Franka robot improves pick‑and‑place success from 42.2% to 71.1%.
arXiv:2606. 03238v1 Announce Type: cross Abstract: Reinforcement learning from human feedback (RLHF) makes large-scale post-training possible by replacing an underspecified human objective with learned and scalable proxies.
CausalDreamer enhances generative world models by freezing the original video tokenizer and re-encoding its latent space into a factored representation that separates controllable versus uncontrollable dynamics and reward-relevant versus reward-irrelevant information. The pretrained dynamics model is fine‑tuned to predict this factored representation, enabling more accurate model‑predictive planning. On the MMBench2 benchmark, CausalDreamer improves normalized scores by 14% on clean tasks and 25% on manipulated variants compared to the baseline, while both models perform poorly on entirely new environments.
arXiv:2607. 04978v1 Announce Type: new Abstract: Joint-Embedding Predictive Architectures (JEPAs) underpin a growing family of latent world models for control from raw pixels, but every existing JEPA world model commits at training time to a single inference paradigm: either trajectory optimisation in a learned dynamics model, or direct behaviour cloning.
The paper challenges the common practice of selecting latent world models based on one‑step prediction loss, arguing that this metric poorly predicts planning performance. Using a frozen‑decoder protocol, the authors decompose one‑step error into encoder‑inherent and predictor‑added components and evaluate goal‑conditioned planning across three continuous‑control environments. Their experiments show that improvements in one‑step fidelity do not correlate with planning success, with correlation coefficients near zero and inconsistent exchange rates across tasks.
arXiv:2608.29937v1 Announce Type: new Abstract: Latent world-action models avoid rendering future pixels by predicting an action-relevant visual subgoal in feature space. LaWAM established this formu...
Joint-Embedding Predictive Architectures (JEPAs) underpin a growing family of latent world models for control from raw pixels, but every existing JEPA world model commits at training time to a single inference paradigm: either trajectory optimisation in a learned dynamics model, or direct behaviour cloning. A single checkpoint that serves both would defer this choice to inference, when deployment constraints (rollout cost, observation accessibility) determine which path wins.
arXiv:2607. 19749v1 Announce Type: cross Abstract: Model-based reinforcement-learning agents of the DreamerV3 family forget catastrophically when trained on task sequences, even when an unbounded replay buffer preserves every earlier experience.
Agentic Critical Training (ACT) is a reinforcement learning approach that trains language‑model agents to judge actions directly by pairing expert actions with alternatives and rewarding correct selection. Unlike imitation learning or self‑reflection, ACT does not require reference rationales and can be reused across model sizes. Experiments on Qwen3‑8B and Olmo‑3‑7B‑Instruct show that ACT improves performance on ALFWorld, WebShop, ScienceWorld, and outperforms supervised reflection baselines, with significant gains over IL and IL→RL baselines.