Assertion Supported AI assessment confidence: 95% certainty 4/5 debate potential 2/5

Gaur: Scaling depth is more parameter- and sample-efficient than width in RL

Ishan Gaur · [NeurIPS Best Paper] 1000 Layer Networks for Self-Supervised RL — Kevin Wang et al, Princeton · Dec 31, 2025 · at 6:42

Ishan Gaur, co-author on the 1000-layer RL paper, explains the architectural trade-offs between depth and width scaling in reinforcement learning.

0:00 / 0:12exact quote · 12.9s

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

“But when you look at the number of parameters that your network has as you grow with, it's roughly a quadratic as opposed to something like growing depth, so it's more, in some sense, it's more parameter efficient, also more sample efficient from the experiments that we conducted.”

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

Growing network depth in self-supervised RL is more parameter efficient and sample efficient than growing network width, which scales parameters quadratically.

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

The speaker's claims regarding architectural scaling in self-supervised reinforcement learning (specifically goal-conditioned contrastive RL) are fully corroborated by the authors' paper '1000 Layer Networks for Self-Supervised RL' and associated OpenReview evaluations published between September 2025 and February 2026. First, standard fully connected layers inherently scale quadratically with width (W x W) versus linearly with depth. Second, empirical experiments showed that scaling depth outperformed scaling width while using roughly 50x fewer parameters. Third, depth-scaled models reached target performance thresholds using fewer environment transitions and samples than width-scaled alternatives.

In standard fully connected networks, parameter count scales quadratically with layer width (W^2) versus linearly with depth (D x W^2).
Experimental findings in goal-conditioned contrastive RL demonstrated that scaling network depth consistently outperformed widening up to width 2048 or 4096 while using ~50x fewer parameters.
OpenReview rebuttal evaluations (September 2025) confirmed that depth scaling achieved target performance with fewer environment samples and wall-clock time compared to width scaling.

Checked against

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

"1000 Layer Networks for Self-Supervised RL" "Table 1" OR "Figure" "depth" "width" "depth" "width" "quadratically" "self-supervised" "reinforcement learning" "1000 Layer Networks for Self-Supervised RL" "quadratically" "1000 Layer Networks for Self-Supervised RL" "width" "depth" "parameters"

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

Ishan Gaur network depth network width

⚑ Spot an error? Report it

Gaur: Scaling depth is more parameter- and sample-efficient than width in RL

Turn any episode into a week of clips.