LPDP: Inference-Time Control for Variable-Length DNA Generation via Edit Flows

Forum topic · 二一 · 2026-05-13

Summary

Editing DNA with AI is typically limited to fixed-length outputs, which fails to capture the variable-length nature of real genomic elements such as enhancers and exon-intron boundaries. LPDP, a training-free method from KAIST researchers Jeongchan Kim, Yunkyung Ko, and Jong Chul Ye, addresses this by combining Edit Flows with inference-time reward control. At each step, the model proposes multiple insertion, deletion, and substitution operations on the DNA sequence, scores candidates with a reward model, performs local search, and retains the highest-scoring edit. This process allows length-flexible generation and real-time quality assessment without additional training. Two applications are demonstrated: enhancer design, where early edits matter most, and exon-intron boundary repair, where late edits dominate.

Key points

Problem with current DNA generation: Existing models are restricted to fixed-length outputs, while functional DNA elements (enhancers, exon–intron boundaries) are naturally variable in length.
Edit Flow framework: LPDP uses an Edit Flow formulation that lets the model perform three edit operations—insertion, deletion, and substitution—on a DNA sequence, producing arbitrary-length outputs.
Training-free inference-time control: The core contribution is a reward-guided search applied at each editing step. Multiple candidate edits are proposed, scored by a reward model, refined through local search, and the best is kept.
Two application scenarios:
*Enhancer optimization* — generating DNA sequences that strongly activate gene expression; prior (early-step) reward weighting works best because early edits have outsized influence.
*Exon–intron boundary repair* — generating correct splice sites given a boundary; posterior (late-step) reward weighting is most effective because late edits determine splice fidelity.
Practical impact: The method requires no additional training, enables real-time quality control during generation, and extends generative DNA modeling beyond fixed-length assumptions.

Paper information

Title: LPDP: Inference-Time Reward Control for Variable-Length DNA Generation with Edit Flows
Authors: Jeongchan Kim, Yunkyung Ko, Jong Chul Ye (KAIST)
Core contribution: Training-free inference-time control combined with edit-based, variable-length DNA sequence generation.

LPDP: Inference-Time Control for Variable-Length DNA Generation via Edit Flows

Summary

Key points

Paper information

Tags