Assertion Supported AI assessment confidence: 90% certainty 4/5 debate potential 2/5

Wang: Deep Self-Supervised Method Beats SOTA on Goal-Conditioned RL Significantly

Kevin Wang · [NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton · Dec 31, 2025 · at 21:24

Kevin Wang (Princeton researcher and NeurIPS Best Paper co-author) highlights the benchmark performance gains of their 1,000-layer reinforcement learning architecture.

0:00 / 0:05exact quote · 5.2s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“We do achieve state-of-the-art performance on goal-conditioned RL and Jack's CCRL by a significant amount.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

Kevin Wang et al.'s deep self-supervised reinforcement learning method achieves state-of-the-art performance on goal-conditioned RL and Jax-based RL benchmarks by a significant amount.

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

In the NeurIPS 2025 Best Paper award-winning publication '1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities' by Kevin Wang et al., the authors evaluated their 1,024-layer contrastive reinforcement learning (CRL) method against standard goal-conditioned RL baselines (including SAC, SAC+HER, TD3+HER, GCBC, and GCSL) on the JaxGCRL benchmark suite across 10 goal-conditioned tasks. The method outperformed all baseline algorithms on 8 of the 10 benchmark environments by substantial margins (achieving performance gains ranging from 2x to over 50x, and up to 1051x on extreme tasks like Humanoid Big Maze), while reaching comparable asymptotic performance on the remaining two tasks.

Kevin Wang et al. published '1000 Layer Networks for Self-Supervised RL: Scaling Depth Can Enable New Goal-Reaching Capabilities' (arXiv:2503.14858 / OpenReview, September 2025), which received a NeurIPS 2025 Best Paper Award.
On the JaxGCRL benchmark suite across 10 locomotion, navigation, and manipulation tasks, scaled CRL outperformed all baseline algorithms in 8 of 10 environments by 2x to >50x (and up to 1051x on Humanoid Big Maze).
On the remaining 2 tasks, scaled CRL matched the asymptotic performance of existing baselines like SAC.

Checked against

[1] neurips.cc [2] arxiv.org [3] nips.cc [4] openreview.net [5] github.io [6] arxiv.org [7] youtube.com [8] virtuslab.com

"1000 Layer Networks for Self-Supervised RL" "outperforming" "baselines" Kevin Wang NeurIPS 2025 self supervised RL JAX Kevin Wang NeurIPS 2025 goal conditioned reinforcement learning "1000 Layer Networks for Self-Supervised RL" "Table 1" OR "Table 2" OR "Table" "1000 Layer Networks for Self-Supervised RL" baselines results OpenReview "1000 Layer Networks for Self-Supervised RL" "JaxGCRL"

assessed as of 2026-09-02

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

Mentioned

goal-conditioned RL Jack's CCRL Kevin Wang

⚑ Spot an error? Report it

More from Kevin Wang

Insight

Kevin Wang: Cross-entropy trajectory classification enables scalable deep reinforcement learning

“I think it's because we're fundamentally shifting the burden of learning from something like, Q-learning or, like, regressing to, like, TD errors, which we know is quite spurious and noisy and biased, to fundamentally, like, a classification problem. We're try…”

Kevin Wang Dec 31, 2025 ▶ 9:56 [NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton

Assertion Supported

Kevin Wang: Scaling RL network depth unlocks effective batch size scaling

“We notice that we see that scaling width actually also improves performance, and we also find that actually by scaling depth, we actually unlock the ability to scale along batch size as well.”

Kevin Wang Dec 31, 2025 ▶ 22:38 [NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton

Assertion Supported

Kevin Wang: 1000-layer RL networks can train on single 80GB H100

“The nice thing is that all of our experiments, even the thousand layer networks, can be run on one single, 80 gigabyte, each 100 GPU.”

Kevin Wang Dec 31, 2025 ▶ 24:19 [NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton

Assertion Not checkable as stated

Wang: Traditional value-based reinforcement learning fails to scale

“And so what we did is that we know that traditional RL, like let's say like value value-based RL doesn't really scale, right? This is pretty clear from the literature.”

Kevin Wang Dec 31, 2025 ▶ 4:29 [NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton

Assertion Supported

Wang: 64 layers saturate performance in most reinforcement learning tasks

“Within our paper, like, for most environments we are able to, like, saturate, like, get to, like, almost perfect performance within just, you know, we don't even need to get to, like, a thousand layers. Like, maybe just 64 layers, for example, is sufficient.”

Kevin Wang Dec 31, 2025 ▶ 15:25 [NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton

Assertion Supported

Wang: GPU environments collect hundreds of millions of RL timesteps hourly

“With these, like, GPU accelerated environments, we can collect hundreds of millions of time steps of data within just a few hours”

Kevin Wang Dec 31, 2025 ▶ 17:43 [NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton

Wang: Deep Self-Supervised Method Beats SOTA on Goal-Conditioned RL Significantly

More from Kevin Wang

Turn any episode into a week of clips.