Skip to main content
Back

Discrete Guidance Matching Fixes Broken AI Steering With 1.6x Faster Sampling

A new posterior-based guidance framework for discrete flow matching delivers exact steering in one forward pass, outperforming approximate methods.

  • New guidance framework for discrete flow matching derives the exact target transition rate rather than a first-order approximation.
  • Sampling requires only one forward pass per step, roughly 1.6x faster than rate-based guidance.
  • Unified formulation covers classifier guidance, energy-weighted sampling, and RLHF as special cases via density ratios.
  • Guidance network is trained with a Bregman divergence loss using F(x)=<x, log x> plus optional target-sample regularization.
  • Applied to FUDOKI, it improved GenEval score from 0.77 to 0.78 and lifted all six multimodal understanding benchmarks tested.
  • Code released at github.com/WanZhengyan/Discrete-Guidance-Matching.

Guiding a generative model toward a specific target distribution, whether for class-conditional generation, energy-based sampling, or RLHF-style preference alignment, is a well-solved problem for continuous diffusion. In the discrete world, it has been considerably messier. A new framework called Discrete Guidance Matching shows why previous approaches were leaking accuracy at every step, and proposes a formulation that is both exact and cheaper to run.

The approximation tax on discrete guidance

Discrete diffusion models and discrete flow-based models operate on token-like state spaces (vocabulary indices, image tokens, protein residues) rather than continuous vectors. To steer them, you need to modify the transition rate that governs how one discrete state jumps to another. The main challenge is that discrete guidance is typically expressed through a transition probability or transition rate, which requires evaluating all possible target positions after the jump.

To dodge that combinatorial blowup, prior work from Vignac, Schiff, and Nisonoff et al. leaned on a first-order Taylor approximation, treating the model as if it were continuous. That trick keeps the number of forward passes small, but it introduces approximation errors that grow with guidance strength. In a state space where tokens do not sit in any meaningful Euclidean neighborhood, the approximation is also philosophically off, since there is no local linearity to exploit.