Back to AI Research

AI Research

Generative Latent Priors expose a protein model steering gap

Key Takeaways

  • A preprint tests diffusion priors on ESM-2 and finds that plausible internal activations can still produce less plausible protein sequences after decoding.
  • The authors used diffusion priors to probe and edit activations in two ESM-2 models.
  • Sequence plausibility remained a problem after repeated decoding.
  • A protein language model preprint tests whether a diffusion model trained on internal activations can make those activations easier to interpret and edit.
  • Its most useful result is a mismatch: an edit can make an activation look more natural to the model while failing to produce a similarly plausible protein sequence after decoding.

A protein language model preprint tests whether a diffusion model trained on internal activations can make those activations easier to interpret and edit. Its most useful result is a mismatch: an edit can make an activation look more natural to the model while failing to produce a similarly plausible protein sequence after decoding. The preprint was submitted to arXiv on November 6, 2026.

What the diffusion prior learns

The researchers trained Generative Latent Priors on frozen ESM-2 activations from UniRef50. They worked with ESM-2 hosts at 8 million and 650 million parameters. The prior learns where real activations tend to occur rather than requiring the representation to be described in terms of a set of sparse features.
The team used the learned representation in two ways. As a probe, it lets researchers test how much individual features reveal about annotated protein concepts. As an editing prior, it can move an edited activation back toward the range of activations the original model normally encounters.

How well the features explain proteins

The authors report that single prior features were stronger probes of discrete Swiss-Prot concepts than the released InterPLM sparse autoencoder features at both host sizes. For continuous properties, they report an advantage at 8 million parameters that did not hold at 650 million. Their falsification tests suggest the features mostly concentrate information already present, rather than uncovering new information.

What changes after a sequence is decoded

The authors tested steering for hydrophobicity, instability and charge. Correcting the activation helped in isolation, but repeated decoding and re-embedding reduced structural plausibility. In additional structural steering experiments with sparse autoencoder clamping, only the helix motif succeeded. The authors trace much of the difficulty to the scope of sequence rewriting and to the decoder. For anyone using representation editing in protein research, a useful check is to evaluate the resulting sequence separately instead of treating a plausible activation as the final result.

Comments