ChenWei He, Xiangzhao Hao, Tianyu Yang, Yuxiang Ma, Yuheng Jia, Lingxiang Wu, Chaoyang Zhao, Haiyun Guo, Jinqiao Wang · ACM International Conference on Multimedia (ACM MM) 2026 · 2026
DOI: 10.1145/3767308.3835296
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Universal multimodal embedding (UME) maps heterogeneous inputs into a shared retrieval space with one model. Recent methods generate explicit chain-of-thought (CoT) rationales before extracting embeddings to resolve complex query intent, but this adds substantial inference cost and funnels rich multimodal evidence through a textual bottleneck. We propose PLUME, a latent reasoning framework that replaces verbalized CoT with a short autoregressive rollout of continuous latent states. A semantic-anchor-guided transition adapter steers the rollout along input-dependent reasoning trajectories under a fixed computation budget. A progressive explicit-to-latent curriculum uses verbalized reasoning only as a temporary training scaffold, then transfers it into hidden-state computation so no CoT is generated at inference. On the 78-task MMEB-v2 benchmark, PLUME outperforms strong explicit-CoT UME baselines while reducing hundreds of generated reasoning tokens to fewer than 10 latent steps and delivering a 30.3× inference speedup. It is especially effective when evidence is dense and structurally complex, as in video and visual document retrieval. These results show that structured latent computation preserves the benefits of intermediate reasoning without explicit rationale generation, offering a stronger accuracy--efficiency tradeoff for practical retrieval.
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
Last synced
This work has 15 recorded citations, but citing papers have not been linked locally yet.
No comments yet — start the discussion below.