Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion

简体中文

L1 alignment/diversity set objective; no training of the 53.9M-parameter diffusion retriever.

Paper information

Field	Details
Paper	Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion
Institution	University of Illinois Urbana-Champaign / Google Research (first author).
First public release	2026-03-06 (arXiv v1)
Author code	No public author repository found as of 2026-09-16.
Local adapter or method key	r4t
Local reproduction code	src/auto_research/foundation_latest_20260916.py

Paper method and evidence

Define alignment with a composite set-level reward, then compile that target into a single diffusion fan-out retrieval pass.

flowchart LR
  R[Set-level alignment and diversity rewards] --> A[Aligned target]
  A --> D[Compile into single diffusion fan-out]

Original paper figure

The original figure illustrates compiling the objective into retrieval. From the original paper; copyright remains with the authors.

Core equation

The reference preserves the defining set decision and reports auditable intermediate states.

Reported results

Paper capability, speed and cost do not transfer to these fixtures or production.

Local reproduction

Three-seed metrics use the same public fixtures, budget, and seeds. The deterministic reference exposes decisions and state transitions; tests verify invariants and boundaries. These are not capability or production results. Results are diagnostic_only=true and not eligible as formal dashboard improvements.