# Overtrained Language Models Are Harder to Fine-Tune

arXiv: https://arxiv.org/abs/2503.19206 · HTML: https://mlsanity.com/paper/2503.19206/

## Key facts

- Cited by 58 papers in the map (data through 14 Oct 2026).
- First posted on arXiv 24 Mar 2025.

## Built on by (most cited first, 30 of 58)

- [Tracing the Representation Geometry of Language Models from Pretraining to Post-training](https://mlsanity.com/paper/2509.23024/) (2025)
- [RL Fine-Tuning Heals OOD Forgetting in SFT](https://mlsanity.com/paper/2509.12235/) (2025)
- [Pre-training under infinite compute](https://mlsanity.com/paper/2509.14786/) (2025)
- [Mid-Training of Large Language Models: A Survey](https://mlsanity.com/paper/2510.06826/) (2025)
- [Unveiling the Basin-Like Loss Landscape in Large Language Models](https://mlsanity.com/paper/2505.17646/) (2025)
- [Activation Control for Efficiently Eliciting Long Chain-of-thought Ability of Language Models](https://mlsanity.com/paper/2505.17697/) (2025)
- [Replaying pre-training data improves fine-tuning](https://mlsanity.com/paper/2603.04964/) (2026)
- [Pre-training LLM without Learning Rate Decay Enhances Supervised Fine-Tuning](https://mlsanity.com/paper/2603.16127/) (2026)
- [Midtraining Bridges Pretraining and Posttraining Distributions](https://mlsanity.com/paper/2510.14865/) (2025)
- [Sharpness-Aware Pretraining Mitigates Catastrophic Forgetting](https://mlsanity.com/paper/2605.02105/) (2026)
- [LLM Fingerprinting via Semantically Conditioned Watermarks](https://mlsanity.com/paper/2505.16723/) (2025)
- [The Butterfly Effect: Neural Network Training Trajectories Are Highly Sensitive to Initial Conditions](https://mlsanity.com/paper/2506.13234/) (2025)
- [Scaling with Collapse: Efficient and Predictable Training of LLM Families](https://mlsanity.com/paper/2509.25087/) (2025)
- [The wall confronting large language models](https://mlsanity.com/paper/2507.19703/) (2025)
- [Can You Learn to See Without Images? Procedural Warm-Up for Vision Transformers](https://mlsanity.com/paper/2511.13945/) (2025)
- [Posterior Behavioral Cloning: Pretraining BC Policies for Efficient RL Finetuning](https://mlsanity.com/paper/2512.16911/) (2025)
- [Universal Properties of Activation Sparsity in Modern Large Language Models](https://mlsanity.com/paper/2509.00454/) (2025)
- [Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay](https://mlsanity.com/paper/2605.26097/) (2026)
- [Reinitializing weights vs units for maintaining plasticity in neural networks](https://mlsanity.com/paper/2508.00212/) (2025)
- [Towards Active Synthetic Data Generation for Finetuning Language Models](https://mlsanity.com/paper/2512.00884/) (2025)
- [What Do Agents Learn from Trajectory-SFT: Semantics or Interfaces?](https://mlsanity.com/paper/2602.01611/) (2026)
- [(How) Learning Rates Regulate Catastrophic Overtraining](https://mlsanity.com/paper/2604.13627/) (2026)
- [Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner](https://mlsanity.com/paper/2506.01301/) (2025)
- [A Theory of Inference Compute Scaling: Reasoning through Directed Stochastic Skill Search](https://mlsanity.com/paper/2507.00004/) (2025)
- [Conda: Column-Normalized Adam for Training Large Language Models Faster](https://mlsanity.com/paper/2509.24218/) (2025)
- [Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less](https://mlsanity.com/paper/2605.06654/) (2026)
- [Massively Multilingual Adaptation of Large Language Models Using Bilingual Translation Data](https://mlsanity.com/paper/2506.00469/) (2025)
- [From Memorization to Parameter Interference: How Overtraining Experts Harms Model Merging](https://mlsanity.com/paper/2506.14126/) (2025)
- [Can Generalist Vision Language Models (VLMs) Rival Specialist Medical VLMs? Benchmarking and Strategic Insights](https://mlsanity.com/paper/2506.17337/) (2025)
- [The Effect of Mini-Batch Noise on the Implicit Bias of Adam](https://mlsanity.com/paper/2602.01642/) (2026)

## Abstract (arXiv)

Large language models are pre-trained on ever-growing token budgets under the assumption that better pre-training performance translates to improved downstream models. In this work, we challenge this assumption and show that extended pre-training can make models harder to fine-tune, leading to degraded final performance. We term this phenomenon catastrophic overtraining. For example, the instruction-tuned OLMo-1B model pre-trained on 3T tokens leads to over 2% worse performance on multiple standard LLM benchmarks than its 2.3T token counterpart. Through controlled experiments and theoretical analysis, we show that catastrophic overtraining arises from a systematic increase in the broad sensitivity of pre-trained parameters to modifications, including but not limited to fine-tuning. Our findings call for a critical reassessment of pre-training design that considers the downstream adaptability of the model.