Reverse Distillation: Consistently Scaling Protein Language Model Representations

Darius Catrina, Christian Bepler, Samuel Sledzieski, Rohit Singh

TL;DR

Researchers have developed Reverse Distillation (RD), a framework that fixes the frustrating tendency of large protein language models (PLMs) to underperform their smaller siblings. By nesting representations like Russian Matryoshka dolls—where a large model’s embedding literally contains the smaller model’s embedding as its prefix—they ensure that adding parameters consistently leads to better performance.

What Problem Does This Paper Address?

In the world of Natural Language Processing (NLP) and Computer Vision, "bigger is better" is a reliable law. If you scale up a model, its performance generally follows a predictable, improving curve. However, Protein Language Models (PLMs) are notoriously rebellious.

On benchmarks like ProteinGym (which measures how well models predict the effects of mutations), researchers often find that mid-sized models (e.g., 650M parameters) actually outperform the massive flagship models (e.g., 15B parameters). This "non-monotonic scaling" makes model selection a guessing game.

Comparison of standard entangled features in a large model versus the proposed orthogonalized, Matryoshka-style decomposition.

The authors argue this happens because large models "entangle" broad, fundamental biological features with rare, niche patterns. Small models are constrained by capacity and preferentially encode broadly-shared protein features. When a downstream task only needs the basics, the extra complexity in a large model's representation acts like noise, drowning out the useful signal and causing performance to drop.

Methods: The Reverse Distillation Pipeline

To solve this, the authors introduce Reverse Distillation. Unlike traditional distillation, which tries to squeeze a big model's knowledge into a small one, RD uses the small model as a "foundation" to organize the big model's knowledge.

The four-step Reverse Distillation pipeline, from computing representations to constructing the final Matryoshka-style embedding.

The process works in four main steps:

  1. Compute Model Representations: The system extracts representations from both a small model and a larger model from the same family (e.g., ESM-2).
  2. Learn Linear Mapping: A linear mapping is trained to approximate the large model's representations using the small model's representations.
  3. Summarize Residual Subspace: Using Singular Value Decomposition (SVD), the framework isolates the "residual"—the unique, higher-order information the large model has that the small model lacks.
  4. Construct Matryoshka-style Embedding: The final embedding is built iteratively. If a small model has a 320-dimensional embedding, the "reverse-distilled" version of a larger model will have those exact 320 dimensions as its first coordinates, followed by new, orthogonal dimensions that capture the residual biological features.

Key Insights and Results

The study primarily utilized the ESM-2 model family (ranging from 8M to 15B parameters), training the linear mappings on UniRef50 sequences and evaluating on ProteinGym and various structural/functional tasks.

  • Restored Scaling: In ProteinGym benchmarks, the reverse-distilled models showed nearly perfect monotonic scaling. The rd.15B model consistently achieved the strongest performance, finally justifying its massive parameter count.
  • Superior Baselines: Even at the same dimensionality, reverse-distilled models often beat their original counterparts. For example, rd.650M generally outperformed the standard ESM-2 650M.
  • The Bias-Variance Tradeoff: Small models are "biased" toward frequent, broadly shared biological regularities (like hydrophobicity). Large models have the "variance" to capture rare epistatic interactions but often fail to prioritize the basics. RD gives researchers the best of both worlds.
  • Orthogonality Prevents Interference: By forcing the new information from large models to be orthogonal (perpendicular) to the small model’s basis, RD prevents "destructive interference." The model stores broad features and niche features in separate, non-competing subspaces.
  • Better Feature Disentanglement: Using Sparse Autoencoders (SAEs), the authors demonstrated that RD embeddings capture more unique Gene Ontology (GO) terms and specific biological functions than standard embeddings.

Limitations and Future Work

The current framework relies on linear decomposition. While this makes the models highly interpretable and mathematically optimal in a mean-squared-error sense, non-linear mappings (like neural networks) might extract even more information, albeit at the cost of simplicity.

Future directions include applying Reverse Distillation to other biological domains, such as genomics or drug discovery, and exploring parameter-efficient versions (like LoRA) to generate these embeddings without needing to run a full chain of multiple models during inference.

Summary updated

Abstract

Unlike the predictable scaling laws in natural language processing and computer vision, protein language models (PLMs) scale poorly: for many tasks, models within the same family plateau or even decrease in performance, with mid-sized models often outperforming the largest in the family. We introduce Reverse Distillation a principled framework that decomposes large PLM representations into orthogonal subspaces guided by smaller models of the same family. The resulting embeddings have a nested, Matryoshka-style structure: the first $k$ dimensions of a larger model's embedding are exactly the representation from the smaller model. This ensures that larger reverse-distilled models consistently outperform smaller ones. A motivating intuition is that smaller models, constrained by capacity, preferentially encode broadly-shared protein features. Reverse distillation isolates these shared features and orthogonally extracts additional contributions from larger models, preventing interference between the two. On ProteinGym benchmarks, reverse-distilled ESM-2 variants outperform their respective baselines at the same embedding dimensionality, with the reverse-distilled 15 billion parameter model achieving the strongest performance. Our framework is generalizable to any model family where scaling challenges persist. Code and trained models are available at https://github.com/rohitsinghlab/plm_reverse_distillation.
BibTeX
@inproceedings{catrina2026reverse,
  author = {Darius Catrina and Christian Bepler and Samuel Sledzieski and Rohit Singh},
  title = {{Reverse Distillation: Consistently Scaling Protein Language Model Representations}},
  booktitle = {International Conference on Learning Representations},
  year = {2026},
  url = {https://openreview.net/forum?id=f12Lo7ZUX5},
}
Figure