Inference-Time Guidance

Inference-time guidance studies how to steer a model while it generates, often adapting a frozen diffusion or flow model, language model, or robot policy to a new reward, constraint, or user preference. Common approaches include classifier-free or gradient-based guidance, reward and verifier scoring with resampling, and direct interventions on logits, latent states, or action probabilities. For reasoning, methods use self-consistency, process verification, and adaptive search; for generation and control, they guide denoising trajectories or select among candidate plans, with evaluations spanning reasoning benchmarks, image and molecular generation, and simulated or real robots. A central challenge is making guidance reliable and calibrated without excessive compute or damage to sample quality: noisy rewards, conflicting objectives, and task-specific gains can make stronger intervention counterproductive.