# Towards Understanding Momentum Acceleration in River-Valley Loss Landscape

arXiv: https://arxiv.org/abs/2609.30957 · HTML: https://mlsanity.com/paper/2609.30957/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 21 of its 89 references that are in the map.
- First posted on arXiv 25 Sep 2026.

## Builds on (most cited first, 21 of 21)

- [Adam: A Method for Stochastic Optimization](https://mlsanity.com/paper/1412.6980/) (2014)
- [The Pile: An 800GB Dataset of Diverse Text for Language Modeling](https://mlsanity.com/paper/2101.00027/) (2020)
- [Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour](https://mlsanity.com/paper/1706.02677/) (2017)
- [TinyStories: How Small Can Language Models Be and Still Speak Coherent English?](https://mlsanity.com/paper/2305.07759/) (2023)
- [Gradient Descent on Neural Networks Typically Occurs at the Edge of Stability](https://mlsanity.com/paper/2103.00065/) (2021)
- [A Walk with SGD](https://mlsanity.com/paper/1802.08770/) (2018)
- [Understanding Optimization in Deep Learning with Central Flows](https://mlsanity.com/paper/2410.24206/) (2024)
- [The AdEMAMix Optimizer: Better, Faster, Older](https://mlsanity.com/paper/2409.03137/) (2024)
- [Beyond the Quadratic Approximation: the Multiscale Structure of Neural Network Loss Landscapes](https://mlsanity.com/paper/2204.11326/) (2022)
- [When and Why Momentum Accelerates SGD:An Empirical Study](https://mlsanity.com/paper/2306.09000/) (2023)
- [Gradient Descent on Logistic Regression with Non-Separable Data and Large Step Sizes](https://mlsanity.com/paper/2406.05033/) (2024)
- [Training Dynamics of the Cooldown Stage in Warmup-Stable-Decay Learning Rate Scheduler](https://mlsanity.com/paper/2508.01483/) (2025)
- [Gradient descent with adaptive stepsize converges (nearly) linearly under fourth-order growth](https://mlsanity.com/paper/2409.19791/) (2024)
- [Momentum Further Constrains Sharpness at the Edge of Stochastic Stability](https://mlsanity.com/paper/2604.14108/) (2026)
- [Through the River: Understanding the Benefit of Schedule-Free Methods for Language Model Training](https://mlsanity.com/paper/2507.09846/) (2025)
- [Minimax Optimal Convergence of Gradient Descent in Logistic Regression via Large and Adaptive Stepsizes](https://mlsanity.com/paper/2504.04105/) (2025)
- [Accelerated Over-Relaxation Heavy-Ball Method: Achieving Global Accelerated Convergence with Broad Generalization](https://mlsanity.com/paper/2406.09772/) (2024)
- [Convergence Rates for Gradient Descent on the Edge of Stability in Overparametrised Least Squares](https://mlsanity.com/paper/2510.17506/) (2025)
- [Gradient Descent with Polyak's Momentum Finds Flatter Minima via Large Catapults](https://mlsanity.com/paper/2311.15051/) (2023)
- [Polyak's Heavy Ball Method Achieves Accelerated Local Rate of Convergence under Polyak-Lojasiewicz Inequality](https://mlsanity.com/paper/2410.16849/) (2024)
- [The Implicit Bias of Heterogeneity towards Invariance: A Study of Multi-Environment Matrix Sensing](https://mlsanity.com/paper/2403.01420/) (2024)

## Built on by (most cited first, 1 of 1)

- [The Best Optimizer Depends on Batch Size](https://mlsanity.com/paper/2610.08975/) (2026)

## Abstract (arXiv)

The empirical success of pretraining large language models has inspired a deeper investigation into the underlying loss landscapes and the optimization dynamics. Recent empirical and theoretical study suggest that the training loss landscape often exhibits a "river-valley" structure, which features a low-loss manifold (river) flanked by sharp orthogonal directions with higher loss (mountains). In the long term, the optimization progress is determined primarily by the progress along the river. Within such a landscape, gradient descent with large learning rates can move faster along the river despite high apparent loss due to vertical oscillations, while a subsequent sharp decay in the learning rate suppresses these oscillations, revealing genuine optimization progress. This explains the recent success of warmup-stable-decay (WSD) learning rate scheduler which, unlike cosine scheduling, keeps stable high learning rate and decays before producing intermediate checkpoints. Building on this foundation, in this work we take a step further and study the role of momentum within such a loss landscape. We establish theoretical analysis that characterizes how momentum accelerates optimization by stabilizing large learning rates that can not be tolerated by vanilla GD without deviating significantly from the river. The enabled large learning rate in-turn gives greater speed along the river and makes faster essential progress in the long run. Another intriguing observation from theory is that for a river-valley landscape with very flat and slow-spinning river, the momentum itself does not contribute directly to acceleration in terms of the speed of tracking the river, while the main acceleration comes from the admissible larger learning rate.