LPDP: Training-Free Inference-Time Reward Control for Variable-Length DNA Sequence Generation with Edit Flows

Forum topic · 二一 · 2026-05-13

Summary

LPDP is a KAIST research paper by Jeongchan Kim, Yunkyung Ko, and Jong Chul Ye that introduces a training-free, inference-time reward control method for generating variable-length DNA sequences. Unlike existing DNA generative models restricted to fixed lengths, LPDP is built on the Edit Flow framework, which generates DNA through iterative insertions, deletions, and base substitutions—analogous to editing text—enabling arbitrary-length outputs. During generation, the method samples multiple candidate edits at each step, scores them with a reward model, retains high-scoring candidates, and applies local search for refinement before proceeding. The authors demonstrate two applications: enhancer optimization, where generated sequences potently activate gene expression (front-loaded reward, with early edits mattering most), and exon-intron boundary repair, generating correct splice sites at given boundaries (back-loaded reward, where later edits are most important). The approach requires no additional model training, making reward-guided DNA design practical and flexible.

> Feynman once quipped about genetics: "DNA is a manual written by nature—as long as you understand its language, you can read it, modify it, and even write it." The LPDP paper tackles the "writing" part: using AI to generate DNA sequences with specific functions, while controlling their quality in real time during generation.

---

The Problem: How to Generate "Useful" DNA?

DNA sequence generation is not a new topic. But existing methods have a key limitation: they can only generate at fixed lengths, whereas real functional DNA elements (such as enhancers and exon-intron boundaries) are naturally variable in length.

LPDP solves this with a framework called Edit Flow. Edit Flow lets the AI perform three operations during generation—inserting, deleting, and substituting bases—editing DNA just like editing text. This allows the generated DNA to have arbitrary lengths.

---

Core Innovation: Score-As-You-Generate Inference-Time Control

The key challenge: how do you judge in real time whether a DNA sequence is "good" while it is being generated?

LPDP's core is a training-free inference-time reward control method. At every edit step (insertion/deletion/substitution), it: 1. Tries multiple candidate edit operations 2. Scores each candidate with a reward model 3. Keeps high-scoring candidates and refines them with local search 4. Picks the best one and continues

Two application scenarios:

Enhancer optimization: generating DNA sequences that strongly activate gene expression ("front-loaded reward"—early edits matter most)
Exon-intron repair: generating correct splice sites at a given boundary ("back-loaded reward"—later edits matter most)

---

*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: Training-free inference-time control; variable-length, edit-style DNA generation

LPDP: Training-Free Inference-Time Reward Control for Variable-Length DNA Sequence Generation with Edit Flows

Summary

The Problem: How to Generate "Useful" DNA?

Core Innovation: Score-As-You-Generate Inference-Time Control

Tags