Discrete Diffusion Inference-Time Control with Nested Sequential Monte Carlo
The paper says nested SMC can steer discrete diffusion text generation better than best-of-n or bootstrap SMC without retraining.
The authors target sequence-level rewards at inference time, with experiments on toxicity and fluency steering. They argue earlier particle methods can run into overoptimism or weight degeneracy. Their nested SMC and fully adapted nested SMC formulations are presented as fixes, including corrections to prior biased final estimates. ArXiv · AI/CL/LG's note
The authors target sequence-level rewards at inference time, with experiments on toxicity and fluency steering. They argue earlier particle methods can run into overoptimism or weight degeneracy. Their nested SMC and fully adapted nested SMC formulations are presented as fixes, including corrections to prior biased final estimates. ArXiv · AI/CL/LG's note
score 5