Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data

2609.11917 ↗ · 2026-09-10 · cs.LG, cs.CL

Atindra Jha, Margaret Li, Jure Leskovec, Percy Liang, Luke Zettlemoyer

The finding

Headline claim

Mixture-of-Experts language models overfit faster than dense models when training data is repeated, with overfitting severity governed by total rather than active parameters.

The test

Compute-matched controlled sweeps training dense and MoE transformers (80M, 200M, 1B active parameters; sparsity 2x to 32x) on OLMoE data domains (DCLM web crawl, StarCoder code, peS2o, Wikipedia), fixing total token budget T=20*N_active and varying repetition rates R from 1 to 1024, with held-out cross-entropy, downstream tasks, and mechanistic router/expert analyses.

Effect size & scope

80M dense models tolerate up to 8x repetition with minimal degradation; MoEs begin degrading at 4x and lose their unique-data advantage entirely at 32x, underperforming dense. With masking regularization, MoEs outperform dense even past 64x repetition, but no regularizer matches all-unique-data performance. Effects are consistent across data domains and data-quality filtering, and depend on total parameters: 200M dense sits between 80M-active MoEs of 158M and 244M total parameters.

Limitations

The authors concede no regularization method fully matches all-unique-data performance and that regularization only mitigates, not eliminates, overfitting. Unstated limits: models are tiny by frontier standards (max 1B active, 8.5B total), results are pre-training cross-entropy oriented with downstream tasks relegated to appendix, compute-matching uses Chinchilla-style ratios that may not transfer to frontier-scale training, and the reversal of validation loss at extreme repetition suggests the degradation story is non-monotonic and not fully understood.

Counterpoint

Small-scale lab models trained on academic corpora may not predict behavior of frontier MoEs trained once on hundreds of billions of unique tokens, where repetition rates are far below 4x. The interaction could be an artifact of the fixed-token compute-matching regime rather than an intrinsic property of sparsity. Validation loss recovery at extreme repetition complicates the clean overfitting narrative.

Economy implication

Modest near-term relevance: if unique-text supply is a binding constraint, MoE compute-efficiency gains may be partially clawed back by higher data requirements, slightly raising the effective data cost of sparse frontier training; this is an incremental cost-curve adjustment, not a regime change in the next 12 to 24 months.

Who this matters for

Technical Capability: Practitioners choosing between dense and MoE architectures under data constraints learn that sparsity's efficiency benefit is conditional on unique-data budget and can be partly restored with dropout or output masking.

Market & Valuation: To the extent MoE compute-efficiency underpins frontier cost claims, a data-side penalty on sparsity slightly complicates the compute-efficiency advantage cited in lab cost narratives.

Bears on (advisory · papers never write evidence)

[6.2] Algorithmic / training-efficiency gains reduce the compute (and chip-access) needed for frontier capability.

Abstract

As the supply of human-written text is exhausted, it has become standard practice to repeat language model training data. Prior work has studied data repetition for densely activated Transformers, but the effects of data repetition remains largely unexplored for recently dominant sparse architectures such as Mixture-of-Experts (MoE), despite their increased compute efficiency. We vary data repetition rates across single- and multi-domain data mixes, and across MoE settings, including expert count and granularity. We consistently find, for models ranging from 80M to 1B active (8.5B total) parameters, that MoEs degrade more rapidly under data repetition. This effect increases with sparsity, dictated by total rather than active parameters. While 80M dense models can repeat data over 8x with minimal degradation, MoEs instead begin to suffer at 4x, and deteriorate rapidly, ceding their performance benefits in all-unique data settings to underperform dense models after 32x. We experiment with existing regularization methods as a potential remedy. We find that some methods, such as dropout, can mitigate overfitting. In particular, with strong masking-based regularization, MoEs are able to outperform dense models even when data is repeated more than 64 times. However, no method fully matches the performance of all-unique training data. Finally, we analyze internal mechanisms correlated with MoE overfitting in high repetition regimes, and find that MoE routing universally stabilizes early in training, and that expert specialization correlates with overfitting to repeated data. In sum, our work addresses the adverse interactions between sparsity and data repetition: we present evidence for the core mechanisms of overfitting and its potential remediation, and suggest promising avenues for future methods to reduce over-specialization in model parameters by disrupting memorization patterns.