JEPA · 2026-10-26

DT-JEPA: Learning to Predict Future Representations for Decision-Making

Changxu Ni, Yiyang Ni, Jun Li, Feiran You, Hongyang Du, Haitao Zhao, Shi Jin

arXiv:2610.32657PDF

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

TL;DR

DT-JEPA is an offline pretraining and online fine-tuning framework for decision-making that predicts future latent representations from reward-free offline trajectories using a probabilistic JEPA objective. The predicted future latents condition a Decision Transformer via AdaLN for action generation, and critic-guided fine-tuning adapts to unseen reward objectives with limited online interactions.

Why it matters

The authors address limitations of return-conditioned Decision Transformers: expected future returns can be ambiguous across different future behaviors, and reward-conditioned pretraining depends on reward-labeled data. DT-JEPA instead learns predictive, uncertainty-aware future latent representations in latent space from reward-free trajectories, then uses these latents to guide DT’s action generation and supports adaptation to unseen rewards via critic-guided fine-tuning.

Method

  • Offline pretraining: learn a variational JEPA-style module that models the conditional distribution of future latent representations qη(F̂t|Ht) as a probabilistic predictor, trained to maximize a variational lower bound on mutual information between historical representations and future trajectories.
  • Latent-conditioned action generation: use AdaLN to inject the sampled predicted future latent F̂t into DT’s hidden representations, via feature-wise scaling and shifting of normalized features.
  • Online fine-tuning: adapt the pretrained policy to a new unseen reward objective using task-specific critics trained with temporal-difference targets and a value-guided policy loss using critic Q-value guidance.
Abstract (from arXiv)

Decision Transformer (DT) has demonstrated strong performance in various offline reinforcement learning (RL) tasks by conditioning action generation on expected future returns and historical trajectories. However, treating expected future returns as the condition provides an incomplete description of the future, as trajectories with different state transitions and behaviors may share similar returns. Moreover, expected future returns inherently depend on task-specific reward functions, which makes policy learning reliant on reward-labeled datasets and restricts its scalability to tasks with similar but unseen rewards. Recently, Joint Embedding Predictive Architectures (JEPA) have emerged as a promising paradigm of self-supervised learning-based world models, which learn predictive representations of future dynamics in latent space rather than reconstructing future states in the original space. Inspired by JEPA, we propose DT-JEPA, an offline pretraining and online fine-tuning framework that learns general knowledge from reward-free offline trajectories and achieves excellent adaptation to unseen tasks. Instead of predicting a single deterministic future embedding, we design a novel probabilistic predictor to model the conditional distribution of future representations in the latent space, thereby capturing both predictive information and uncertainty about future dynamics. Theoretically, we show that 1) the JEPA objective seeks to maximize a variational lower bound on the mutual information between historical representations and future trajectories; and 2) predictive future-conditioned policy learning admits a finite-sample performance bound with an O(N-1/4) estimation term. Experiments on D4RL benchmarks demonstrate that DT-JEPA achieves efficient adaptation across varying offline RL tasks, outperforming traditional DT-based and other offline RL methods.

Related papers

Publication record

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

First public release: 2026-10-23 · Indexed by Pub-AI: 2026-10-26 UTC

preprint

DT-JEPA: Learning to Predict Future Representations for Decision-Making

Changxu Ni, Yiyang Ni, Jun Li, Feiran You, Hongyang Du, Haitao Zhao, Shi Jin

Preprint posted: 2026-10-23 · Original record ↗

arxiv
2610.32657
Explicit source versions (1)
  • v1 · 2026-10-23
  1. Preprint available

    Source date: 2026-10-23 · First detected: 2026-10-27 UTC · Historical baseline

    Source evidence (2)
Abstract · arxiv

Decision Transformer (DT) has demonstrated strong performance in various offline reinforcement learning (RL) tasks by conditioning action generation on expected future returns and historical trajectories. However, treating expected future returns as the condition provides an incomplete description of the future, as trajectories with different state transitions and behaviors may share similar returns. Moreover, expected future returns inherently depend on task-specific reward functions, which makes policy learning reliant on reward-labeled datasets and restricts its scalability to tasks with similar but unseen rewards. Recently, Joint Embedding Predictive Architectures (JEPA) have emerged as a promising paradigm of self-supervised learning-based world models, which learn predictive representations of future dynamics in latent space rather than reconstructing future states in the original space. Inspired by JEPA, we propose DT-JEPA, an offline pretraining and online fine-tuning framework that learns general knowledge from reward-free offline trajectories and achieves excellent adaptation to unseen tasks. Instead of predicting a single deterministic future embedding, we design a novel probabilistic predictor to model the conditional distribution of future representations in the latent space, thereby capturing both predictive information and uncertainty about future dynamics. Theoretically, we show that 1) the JEPA objective seeks to maximize a variational lower bound on the mutual information between historical representations and future trajectories; and 2) predictive future-conditioned policy learning admits a finite-sample performance bound with an O(N-1/4) estimation term. Experiments on D4RL benchmarks demonstrate that DT-JEPA achieves efficient adaptation across varying offline RL tasks, outperforming traditional DT-based and other offline RL methods.