# Replay on Demand: An Emergent Curriculum for Balancing Adaptation and Forgetting in Continued Pretraining

arXiv: https://arxiv.org/abs/2609.40089 · HTML: https://mlsanity.com/paper/2609.40089/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 26 of its 34 references that are in the map.
- First posted on arXiv 30 Sep 2026.

## Builds on (most cited first, 26 of 26)

- [Overcoming catastrophic forgetting in neural networks](https://mlsanity.com/paper/1612.00796/) (2016)
- [Don't Stop Pretraining: Adapt Language Models to Domains and Tasks](https://mlsanity.com/paper/2004.10964/) (2020)
- [Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time](https://mlsanity.com/paper/2203.05482/) (2022)
- [Experience Replay for Continual Learning](https://mlsanity.com/paper/1811.11682/) (2018)
- [Dark Experience for General Continual Learning: a Strong, Simple Baseline](https://mlsanity.com/paper/2004.07211/) (2020)
- [Progress & Compress: A scalable framework for continual learning](https://mlsanity.com/paper/1805.06370/) (2018)
- [Prioritized Training on Points that are Learnable, Worth Learning, and Not Yet Learnt](https://mlsanity.com/paper/2206.07137/) (2022)
- [Rho-1: Not All Tokens Are What You Need](https://mlsanity.com/paper/2404.07965/) (2024)
- [Simple and Scalable Strategies to Continually Pre-train Large Language Models](https://mlsanity.com/paper/2403.08763/) (2024)
- [Functional Regularisation for Continual Learning with Gaussian Processes](https://mlsanity.com/paper/1901.11356/) (2019)
- [Lifelong Pretraining: Continually Adapting Language Models to Emerging Corpora](https://mlsanity.com/paper/2110.08534/) (2021)
- [Reuse, Don't Retrain: A Recipe for Continued Pretraining of Language Models](https://mlsanity.com/paper/2407.07263/) (2024)
- [Investigating Continual Pretraining in Large Language Models: Insights and Implications](https://mlsanity.com/paper/2402.17400/) (2024)
- [Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning](https://mlsanity.com/paper/2606.15007/) (2026)
- [D-CPT Law: Domain-specific Continual Pre-Training Scaling Law for Large Language Models](https://mlsanity.com/paper/2406.01375/) (2024)
- [A Unifying Bayesian View of Continual Learning](https://mlsanity.com/paper/1902.06494/) (2019)
- [CMR Scaling Law: Predicting Critical Mixture Ratios for Continual Pre-training of Language Models](https://mlsanity.com/paper/2407.17467/) (2024)
- [Adaptive Memory Replay for Continual Learning](https://mlsanity.com/paper/2404.12526/) (2024)
- [CoLoR-Filter: Conditional Loss Reduction Filtering for Targeted Language Model Pre-training](https://mlsanity.com/paper/2406.10670/) (2024)
- [Towards Effective and Efficient Continual Pre-training of Large Language Models](https://mlsanity.com/paper/2407.18743/) (2024)
- [TiC-LM: A Web-Scale Benchmark for Time-Continual LLM Pretraining](https://mlsanity.com/paper/2504.02107/) (2025)
- [Revisiting Replay and Gradient Alignment for Continual Pre-Training of Large Language Models](https://mlsanity.com/paper/2508.01908/) (2025)
- [Velocitune: A Velocity-based Dynamic Domain Reweighting Method for Continual Pre-training](https://mlsanity.com/paper/2411.14318/) (2024)
- [FOREVER: Forgetting Curve-Inspired Memory Replay for Language Model Continual Learning](https://mlsanity.com/paper/2601.03938/) (2026)
- [Data Mixing Agent: Learning to Re-weight Domains for Continual Pre-training](https://mlsanity.com/paper/2507.15640/) (2025)
- [When to Review: Spaced Repetition for Continual Pre-Training of Language Models](https://mlsanity.com/paper/2608.17530/) (2026)

## Abstract (arXiv)

Continued pretraining enables language models to adapt to new domains and knowledge, but often at the cost of forgetting previously acquired capabilities. Replay can mitigate this trade-off, but fixed replay mixtures allocate training independently of the model's actual retention needs. We introduce Replay on Demand (RoD), which instead derives the replay allocation from the model's learning dynamics. RoD jointly prioritizes adaptation samples by their remaining learning potential and replay samples by their observed forgetting. Their competition for a shared training budget yields an online curriculum that determines what to train on at each step. Across models, scales, and adaptation domains, RoD reaches or improves upon the adaptation-forgetting frontier of tuned fixed-replay baselines and model merging without prescribing a replay allocation in advance. Replay concentrates on sources that are more vulnerable to forgetting and dynamically increases and redistributes as forgetting emerges during training. Together, our results show that replay can be allocated online from the model's evolving state, targeting what is needed, when it is needed.