arXiv Computer Vision

Multi-modal Knowledge Preserving Adapter for Embedding Backward Compatibility

The paper introduces the Multi-modal Knowledge Preserving Adapter (MKP-Adapter), an adapter-only approach that enables backward compatible training for multi-modal large language models without updating the backbone. It employs a multi-level preservation loss to maintain embedding geometry and a focal re-weighting strategy to focus on difficult samples. Experiments show strong backward compatibility across image, text, visual document, and video retrieval tasks with minimal latency overhead.

arXiv AI
Sep 10

MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models

MoEMB introduces a mixture‑of‑experts (MoE) approach to scale universal multimodal embeddings (UME) without increasing the size of the output vector or relying on autoregressive decoding. By expanding encoder capacity along the expert axis, MoEMB achieves state‑of‑the‑art performance on MMEB‑V2 and MRMR benchmarks with only 3 B active parameters, outperforming TTE‑based methods that use more than four times as many active parameters and require significantly more compute. The paper also presents the first comprehensive study of adaptive computation for MoE‑based embeddings, exploring training‑time and inference‑time strategies to further improve efficiency for large‑scale retrieval and recommendation systems.

By Xuanming Cui, Shlok Kumar Mishra, Wentao Bao, Aashu Singh, Zihao Wang, Xiangjun Fan, Jun Xiao, Ser-Nam Lim, Jianpeng Cheng
arXiv AI
Oct 26

Beyond Post-Hoc Rewards: Privileged Reasoning Exploration for Multimodal Embeddings

The paper introduces PILOT, a framework that treats reasoning as a latent variable to enhance multimodal embeddings. By modeling prior and posterior reasoning modes with a shared large language model, PILOT uses relational supervision during training to guide exploration of retrieval‑relevant reasoning traces. On the MMEB‑V2 benchmark, PILOT outperforms baselines across all task groups, notably improving visual‑document retrieval scores by over six points.

By Tianyu Yang, Yijun Tian, Lars Benedikt Kaesberg, Jan Philip Wahle, Bela Gipp, Terry Ruas, Daniel Kurzawe
arXiv Computer Vision
Aug 25

Learning Sample-wise Rank-aware Interpolation Weights for Composed Visual Data Retrieval

The paper introduces SRAIN, a framework that learns sample‑wise, rank‑aware interpolation weights for composed visual data retrieval. Instead of relying on complex multimodal large language models, SRAIN uses simple linear interpolation in embedding space, dynamically predicting query‑specific weights through batch‑wise rank‑aware estimation and a compact memory bank for hard negatives. This approach achieves state‑of‑the‑art performance on composed video retrieval and competitive results on composed image retrieval while significantly reducing query‑time latency.

By Boseung Jeong, Taegyu Park, Donghyeon Kwon, Hyunsouk Cho, Suha Kwak
arXiv Computation and Language
Oct 12

TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs

TokenSwap introduces a method to benchmark and reduce the modality gap in multimodal large language models (MLLMs) by interleaving semantically aligned images with text tokens. By transforming text-only benchmarks like MMLU into image‑interleaved versions (TokenSwap‑Bench), the authors reveal a pervasive performance drop—averaging 19.6%—across 42 MLLMs, with reasoning models showing smaller gaps. Training with TokenSwap mitigates this gap while maintaining strong text‑only and vision‑language performance.

By Andong Hua, Colton Bishop, Igor Mordatch, Arian Hosseini, Jindong Gu, Aleksandra Faust, Rebecca Roelofs, Yao Qin
arXiv AI
Oct 28

An Embarrassingly Simple Data Selection Strategy for Million-Scale Multi-Modal Instruction-Tuning

The paper introduces SSR, a simple yet effective data selection strategy for million‑scale multi‑modal instruction‑tuning. SSR uses interaction styles as a proxy for structural diversity and decomposes data quality into capability‑specific scores, forming a rich profile for each sample. By casting selection as submodular coverage maximization, SSR consistently outperforms random sampling across two architecture stacks, multiple data pools, and 14 benchmarks, achieving 99.1% of full‑data performance with only 30% of a 2.6M pool.

By Mengyao Lyu, Yan Li, Huasong Zhong, Hui Chen, Jungong Han, Guiguang Ding, Zhenheng Yang