8 Must-Read ACL 2025 Papers from Meituan: Generative Retrieval, Multimodal LLMs & More

Meituan’s technology team presented eight papers at the international ACL 2025 conference, covering generative retrieval algorithms, multi‑objective preference‑aligned training, rich‑text image understanding, search‑term recommendation, cross‑language transfer, multimodal mathematical reasoning, third‑person tasks, and related topics.

01. Multi-level Relevance Document Identifier Learning for Generative Retrieval

Paper type: Long Paper (ACL 2025 Main)

Paper link: https://aclanthology.org/2025.acl-long.497.pdf

Abstract: This work proposes MERGE, a multi‑level relevance‑based generative retrieval document identifier learning method. It addresses the semantic insufficiency of discrete DocIDs by using queries as bridges and building a hierarchical relevance learning framework with three core modules: a multi‑relevance query‑document alignment module, an outer contrastive learning module for binary relevance features, and an inner multi‑level relevance learning module for fine‑grained document distinction. Experiments on a multilingual e‑commerce search dataset show MERGE significantly outperforms existing baselines.

02. HierGR: Hierarchical Semantic Representation Enhancement for Generative Retrieval in Food Delivery Search

Paper type: Industry Oral

Paper link: https://aclanthology.org/2025.acl-industry.31.pdf

Abstract: To tackle challenges of generative retrieval in food‑delivery search, HierGR introduces a hierarchical semantic identifier system and a joint offline‑online optimization pipeline. It improves large‑model inference latency and respects strict LBS constraints by using offline training, query caching, and integration with dense retrieval models. Offline and online A/B tests on Meituan Waimai show a 0.68 % increase in order conversion.

03. AMoPO: Adaptive Multi-objective Preference Optimization without Reward Models and Reference Models

Paper type: Long Paper

Paper link: https://arxiv.org/pdf/2506.07165

Abstract: AMoPO eliminates the need for reference and reward models in multi‑objective preference optimization. It uses an adaptive sampling strategy to dynamically adjust dimension weights, achieving efficient single‑pass optimization. Experiments on public datasets and real‑world tasks demonstrate superior effectiveness, efficiency, and multi‑dimensional adaptability.

04. Multimodal Large Language Models for Text‑rich Image Understanding: A Comprehensive Review

Paper type: Findings

Paper link: https://arxiv.org/pdf/2502.16586

Abstract: This survey systematically reviews multimodal large language models (MLLMs) for text‑rich image understanding (TIU). It outlines timelines, architectures, and pipelines of existing TIU‑MLLMs, compares benchmark performances, and discusses future directions, challenges, and limitations.

05. Consistency‑Aware Online Multi‑Objective Alignment for Related Search Query Generation

Paper type: Industry Track

Paper link: https://aclanthology.org/2025.acl-industry.96.pdf

Abstract: CMAQ proposes an online multi‑objective alignment framework for related‑search query generation. It builds precise reward models for CTR and topic extension, aligns query style via supervised fine‑tuning, and employs a consistency‑aware multi‑objective DPO strategy with adaptive β weighting. Offline and online experiments on a large‑scale platform show significant CTR gains (+2.3 %) and higher query quality.

06. Investigating and Scaling up Code‑Switching for Multilingual Language Model Pre‑Training

Paper type: Findings

Paper link: https://aclanthology.org/2025.findings-acl.575.pdf

Abstract: The paper studies spontaneous cross‑language transfer in large language models and identifies low‑frequency code‑switching data as a key factor. It introduces SynCS, a low‑cost synthetic code‑switching data generator, and demonstrates that scaling such data markedly improves multilingual capabilities.

07. The Role of Visual Modality in Multimodal Mathematical Reasoning: Challenges and Insights

Paper type: ACL 2025 Main (Oral)

Paper link: https://aclanthology.org/2025.acl-long.1102.pdf

Abstract: By perturbing visual inputs, the study reveals that current multimodal math‑reasoning models rely heavily on textual cues, with visual information contributing little. The newly created HC‑M3D dataset exposes this weakness, showing that state‑of‑the‑art models often fail to adjust predictions when visual content changes.

08. From Observation to Understanding: Front‑Door Adjustments with Uncertainty Calibration for Enhancing Egocentric Reasoning in LVLMs

Paper type: Findings

Paper link: https://aclanthology.org/2025.findings-acl.979.pdf

Abstract: FRUIT introduces a causal front‑door adjustment with uncertainty calibration to improve egocentric reasoning in large visual‑language models. It first locates interaction regions, builds hierarchical visual‑text cues, filters noise via uncertainty estimation, and integrates the refined observation into prompts, boosting first‑person understanding.

The article also mentions Meituan’s “Beidou” recruitment program for top campus talent and the Meituan research collaboration initiative, inviting academic partners to explore AI, big data, IoT, autonomous driving, and optimization.