MEMORY INDUSTRY INTELLIGENCE

Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve-for-Train

초안 발행 2026.09.16 05:00 사건 미확인수집 2026.10.09 22:22원문 언어 en

한국어 번역·요약·분석

분석을 준비하고 있습니다. 완료된 결과는 이 페이지에 표시됩니다.

원문 텍스트

원문 열기 ↗

v1 · 10.09 20:26 v2 · 10.09 22:22

Algorithms & Theory

이 자료의 근거

초안원문 v2

In our [ICML 2026](https://icml.cc/virtual/2026/poster/66354) paper, “[Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion](https://arxiv.org/abs/2603.06397)”, we address this decomposition bottleneck via a reward-to-data compilation framework.

초안원문 v2

The pipeline operates in three distinct steps:

초안원문 v2

A compact, 53.9M-parameter diffusion model learns to map a query embedding directly to a complete set of target embeddings in one non-autoregressive pass, officially bypassing the need for text-based CoT reasoning tokens.

초안원문 v2

For our open-ended abstract retrieval tasks, this composite reward is a weighted balance of three competing pillars:

초안원문 v2

For the language models, the query fan-out process was driven by 4B open-source models, specifically [Gemma3-4B](https://huggingface.co/google/gemma-3-4b-it) and [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B), which were tasked with generating exactly 10 sub-queries for every single main search prompt they processed.

초안원문 v2

For the multimodal embedding backbones, we conducted experiments across two domains: A large-scale fashion dataset of user-curated outfits used for text-to-image experiments (evaluated using a [CLIP](https://openai.com/index/clip/)-based retriever), and a proprietary industrial dataset of expert-generated music playlists used for text-to-music evaluations (evaluated using [MuLan](https://research.google/pubs/mulan-a-joint-embedding-of-music-audio-and-natural-language/)).

초안원문 v2

Across both retrieval tasks, Retrieve-for-Train outperformed traditional single-query search, zero-shot expansion, and even the heavily optimized [Best-of-N baseline](https://openai.com/index/measuring-goodharts-law/).

초안원문 v2

Because the diffusion model generates all target directions simultaneously in a single, non-autoregressive parallel pass in continuous embedding space, it delivers a massive 12 to 20 speedup over autoregressive approaches.

초안원문 v2

At scale, while autoregressive fan-out latency expands linearly to nearly 50 seconds under large context batches, Retrieve-for-Train-Diffusion stays between sub-second to a few seconds, delivering production-ready, expert-level search at a fraction of the computational cost.

초안원문 v2

Without a diversity term, the model quickly collapses into generating degenerate, nonsensical strings (like _"line ending line ending"_) to mathematically exploit the vector coordinates of the database.

초안원문 v2

We demonstrated that RL can be highly effective when used as a one-time "objective transducer" rather than an online inference engine.