The paper introduces the World-Model Policy Arbiter (WMPA), a test‑time framework that selects among a portfolio of frozen goal‑conditioned policies by simulating each policy’s future trajectories in a learned world model and evaluating them with a shared value function. WMPA avoids frequent policy switching by committing to the chosen policy for a short interval and does not require retraining or task‑specific knowledge. On the OGBench benchmark, WMPA raises the macro‑average success rate from 44% to 58%, achieving significant improvements on 12 of 18 datasets, including large gains on cube‑double‑play and scene‑play.
arXiv:2602. 05459v2 Announce Type: replace Abstract: Offline goal-conditioned reinforcement learning (GCRL) is typically benchmarked by the best tuned success rate of each method.
arXiv:2607. 25415v1 Announce Type: new Abstract: Production LLM agents are increasingly assembled from a frozen model wrapped in a harness: a prompt template, a tool set, a memory/retrieval layer, a planning strategy, and a verification policy.
The paper introduces Generalized Implicit Temporal Abstraction (GITA), a method for goal-conditioned reinforcement learning that conditions a single value function on multiple temporal abstraction levels (k). By aggregating advantage-weighted supervision across various k values, GITA preserves both long-range signal and local resolution without committing to a single k. Experiments on OGBench show that GITA outperforms existing offline GCRL baselines, improving average success rates by 25 percentage points over HIQL and 7 percentage points over OTA.
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%.
FIRM-WM is a compact pixel world model that separates a goal‑comparable configuration from a 128‑dimensional dynamic fiber, enabling reward‑free visual planning from offline videos. It addresses two key mismatches: aligning planning states with goal images and reconciling factual trajectories with interventional sampling. In experiments, FIRM‑WM achieves high success rates on TwoRoom, Reacher, and OGBench‑Cube while using fewer parameters and faster planning times than prior models.
The paper introduces Imagine-then-Plan (ITP), a framework that lets agents learn by interacting with a learned world model to generate multi-step imagined trajectories. ITP features an adaptive lookahead mechanism that balances ultimate goals with task progress, producing richer signals about future outcomes. Experiments on various benchmarks show that ITP outperforms existing baselines, and analyses suggest the adaptive lookahead improves reasoning for complex tasks.
The paper introduces QWM, a framework that integrates world models with standard Q‑learning to perform test‑time search over imagined trajectories. By training the policy and value function solely on real transitions, QWM avoids compounding model bias while still benefiting from predictive search. Experiments on the Robomimic and LIBERO manipulation benchmarks show that QWM outperforms strong prior state‑of‑the‑art methods in both sample efficiency and performance.
DreamAvoid introduces a test‑time dreaming framework for Vision‑Language‑Action models to anticipate and avoid failures during critical manipulation phases. It uses a Dream Trigger to detect critical phases, samples candidate action chunks via an Action Proposer, and evaluates short‑horizon futures with a Dream Evaluator trained on success, failure, and boundary data. Experiments on real‑world and simulated tasks show DreamAvoid improves task success rates, achieving 72.5% success versus 48.8% for the base policy and 54.4% for GPC‑RANK.
WorldGuide is a closed‑loop video world model that, given an initial image and a task goal, predicts an atomic action, generates the corresponding video clip, and uses that clip to decide the next action or terminate the task. It trains a Planner to predict the next action or completion from visual progress and an Executor to realize those actions, while a hierarchical visual memory tracks state over long horizons. The authors introduce WorldGuide‑Bench, a dataset of ~59K step‑annotated videos across 245 tasks, and show that WorldGuide outperforms the recent MiniMax‑H3 model on both WorldGuide‑Bench and VideoCraft‑Bench, highlighting the benefit of coupling planning with learned execution for goal‑directed procedural video generation.
arXiv:2608.29061v1 Announce Type: new Abstract: Offline goal-conditioned reinforcement learning (GCRL) aims to learn policies for reaching diverse goals entirely from fixed trajectory data. Long-hori...
The paper identifies an informational curse of horizon in goal-conditioned policy learning, showing that longer goal relabeling horizons during training can hurt policy generalization and performance. Controlled experiments with oracle planners separate training horizons from test horizons, revealing that behavioral cloning policies degrade with longer horizons while reinforcement learning mitigates this effect. The authors attribute this to reduced conditional mutual information between actions and hindsight-relabeled goals, and demonstrate that distilling short-horizon policy Jacobians into long-horizon policies improves performance, especially in combinatorial manipulation tasks.
arXiv:2609.13845v1 Announce Type: cross Abstract: World models trained with joint-embedding predictive architectures learn compact, structured latent representations from physical interaction, yet pl...