Extending Context by Dropping Positional Embeddings
DroPE introduces a method for extending a language model’s context window after pretraining without expensive long-context fine-tuning. The approach involves removing positional embeddings from a pretrained model and performing brief recalibration at the original context length.
Ask this paper
Core insight: Positional embeddings serve as a “training-time scaffold” - beneficial during pretraining but detrimental for extrapolation. RoPE enables faster attention non-uniformity development during training, but becomes problematic at test time when sequences exceed training length.
The length generalization problem: Popular RoPE scaling methods preserve perplexity but essentially “crop” effective context, failing at retrieval tasks requiring long-range attention. DroPE addresses this by completely removing the positional scaffold after training.
Simple methodology: The approach is straightforward: train or obtain a pretrained RoPE-based model, remove positional embeddings post-pretraining, then recalibrate briefly using as little as 0.5-2% of the original pretraining budget.
Strong recovery: Models regain 95%+ in-context performance after less than 5B recalibration tokens. On needle-in-haystack tasks, DroPE substantially outperforms RoPE-scaling methods that fail at long-range retrieval.
Scalability and benchmarks: Validated on models up to 7B parameters trained on trillions of tokens. Improves base SmolLM scores by 10x on LongBench and enables zero-shot context extension to 2x training length without task-specific fine-tuning.