Skip to main content
Back

GILC Steers Biological AI Without Retraining, Beating Fine-Tuned Rivals

GILC steers pretrained discrete diffusion models toward target biological properties at inference time, no retraining required, beating fine-tuned baselines on DNA, protein, and molecule design

Paper, Full Access
  • GILC is a plug-and-play guidance framework for discrete diffusion models that requires no retraining and works at inference time.
  • It corrects the model's output logits using a Jacobian-free gradient mechanism, avoiding the instability of backpropagating through high-dimensional discrete spaces.
  • Supports both differentiable and black-box reward functions, making it compatible with physics simulators and wet-lab oracles.
  • Benchmarked on DNA sequence design, protein engineering, and molecular generation -- outperforms training-free baselines and matches fine-tuned models.
  • Compared against strong baselines including DRAKES, SMC, SVDD, and CFG; achieves state-of-the-art without parameter updates.
  • No public code repository released yet; method is fully described in the arxiv paper.

Diffusion models have conquered images, audio, and video. But biology runs on sequences -- DNA, proteins, small molecules -- and those are discrete: a finite alphabet of letters, not a smooth continuous space. Getting a pretrained discrete diffusion model to generate sequences that hit a specific target property (high binding affinity, drug-likeness, promoter activity) has been surprisingly hard. The standard playbook either demands retraining the model or bolts on guidance methods so expensive they barely scale. A new paper from Hongkun Dou and colleagues proposes a cleaner path.

The work introduces GILC (Gradient-Informed Logit Correction), a plug-and-play guidance framework for discrete diffusion that requires zero additional training and works with virtually any reward function -- differentiable or not. On benchmarks spanning regulatory DNA design, protein engineering, and molecular generation, GILC matches or beats methods that required expensive fine-tuning.

Why steering discrete diffusion is so painful

Recent studies show that diffusion models hold significant potential in discrete spaces, spanning domains such as language generation, biological sequence synthesis, and molecular design. But while discrete diffusion effectively captures complex categorical data distributions, many scientific and industrial applications require generated samples that are not only realistic but also optimized for specific task objectives, moving beyond unconditional generation.

The core difficulty is mathematical. In continuous diffusion (think image generation), you can compute a gradient -- a direction to nudge the generation -- with respect to any differentiable objective. In discrete spaces, tokens are integers. There is no gradient to take. Applying guidance to discrete diffusion is challenging because guidance terms are not differentiable with respect to the discrete representations.

The workarounds each carry a steep cost. Existing efforts based on Sequential Monte Carlo (SMC) or importance sampling often suffer from prohibitively high computational costs and demonstrably limited steering effectiveness. The other route -- fine-tuning the model with a reward signal -- works but classical approaches typically employ classifier/classifier-free guidance or extensive model fine-tuning; however, these established methods demand additional training stages, which fundamentally limit their generalizability and introduce significant computational overhead.