The 1,000 Layer Secret: Unlocking RL Scaling with Self-Supervised Objectives

The decade-long anomaly of shallow reinforcement learning networks has finally been broken. The conventional wisdom, that deep networks inherently fail in RL, was challenged and overturned by a team from Princeton that successfully scaled models to 1,000 layers, a feat previously thought impossible. This breakthrough, which earned the team the Best Paper award at NeurIPS 2025, represents a fundamental paradigm shift that promises to accelerate the capabilities of autonomous systems.

Kevin Wang, Ishaan Javali, Michał Bortkiewicz, and their advisor Benjamin Eysenbach spoke with Swyx of Latent Space at NeurIPS 2025 about their award-winning work, "1000 Layer Networks for Self-Supervised RL," detailing how they achieved massive scaling by fundamentally rethinking the core objective function. For years, deep RL models stagnated at two or three layers, an anomaly compared to the revolution seen in vision and language models. Eysenbach noted that he was "kind of skeptical it was going to work" when the students proposed pushing depth.

The key barrier, the team discovered, was not depth itself, but the traditional value-based objective function. Calculating temporal difference (TD) errors is prone to noisy, spurious, and biased signals that degrade performance as networks deepen. The Princeton team shifted the learning burden away from regression toward classification. As Kevin Wang explained, they moved from "regressing to like TD errors... to fundamentally like a classification problem." This self-supervised approach trains the network to learn representations where states along the same trajectory are pushed together, while states along different trajectories are pushed apart, proving vastly more stable under scale.

The scalability was unlocked by incorporating architectural choices proven in other domains, specifically residual connections and layer normalization, which mitigate vanishing gradients. This architectural robustness allowed them to prove that scaling depth is far more parameter-efficient than scaling width, growing parameters linearly versus quadratically for the same performance gains.

Achieving massive depth also revealed a non-linear scaling phenomenon the team dubbed "critical depth." Performance gains were not incremental; they multiplied significantly once the models crossed a threshold of around 15 million transitions and possessed the correct architectural components. This research was only feasible due to modern infrastructure, specifically, the use of JAX and GPU-accelerated environments that allowed them to collect "hundreds of millions of times of the data within just a few hours," enabling the data abundance necessary for deep networks to truly pay off.

This new paradigm challenges the strict definition of reinforcement learning itself. While the resulting system is an actor-critic, goal-conditioned RL algorithm, the learning signal is derived purely from self-supervised representation learning, rather than direct reward maximization. Eysenbach pointed out that their core code "doesn't have a line of code saying maximize rewards here," suggesting a profound blurring of the lines between classic RL and self-supervised methods. The result is a system capable of solving complex robotic tasks without the need for manually crafted reward functions or expert demonstrations.

The shift in objective also unlocked another scaling dimension: batch size. Traditionally, scaling batch size in RL yielded diminishing returns because the shallow networks lacked the capacity to fully utilize the parallel signal. The depth afforded by RL1000, however, provided the necessary capacity, demonstrating that large batch sizes become significantly more effective once network complexity is sufficient. The team emphasized that this approach opens the door for training robust robotic agents, moving away from the expensive and non-scalable practice of relying on human supervision and demonstrations toward architectural scaling.

The practical deployment of such large, deeply trained models remains a subject of ongoing research. The team suggests that knowledge distillation offers a promising pathway: training frontier capabilities using the resource-intensive 1,000-layer "teacher" model, then distilling those insights into smaller, more efficient "student" models suitable for real-time inference and embedded systems. This maintains deployment efficiency by leveraging the representational power of scale. The convergence of self-supervised learning methods with reinforcement learning architectures suggests that the field of deep RL is finally poised to follow the exponential scaling trajectories observed in large language and vision models.