Online Feedback-Efficient Sampling for Search and Alignment

• Generative Modelling• Sampling• Applied Stochastic Process

Objective: We study online, feedback-efficient sampling from generative (diffusion/flow) models when the object of interest — a search target or a user’s preference — is unknown in advance and is only revealed through sequential feedback under a limited query budget. Existing samplers either explore locally around a small region of the distribution or rely on rigid, non-adaptive proposal dynamics, causing them to waste budget or collapse to a single mode once feedback arrives. This line of work develops feedback-driven samplers that use each new piece of feedback to steer generation toward high-utility regions while preserving sample diversity and computational efficiency, enabling effective online search and personalized preference alignment within a fixed sampling budget.

Related Publications:

  1. Sampling Meets Interaction: Sequentially-Controlled Multi-Particle Flow-Maps for Efficient Inference-time Search (arXiv 2026)

  2. BFMT: Enhancing Search Capabilities of Tree Sampler via Bootstrap Flow-Map Tree (arXiv 2026)

  3. PAPA: Online Personalized Active Preference Alignment (ECML PKDD 2026)

Figure 1 from each paper: IMPFM, BFMT, and PAPA.