Supervised Fine-Tuning or Contrastive Learning? Towards Better Multimodal LLM Reranking

*arXiv, October 2025 — https://arxiv.org/abs/2510.14824*

Authors / Affiliations: Ziqi Dai, Xin Zhang, Mingxin Li, Yanzhao Zhang, Dingkun Long, Pengjun Xie, et al. (9 authors in total)

Category: Ranking for Search

TL;DR

This work investigates the question posed in its title: for multimodal LLM-based reranking, should one prefer supervised fine-tuning (SFT) or contrastive learning? It systematically examines the trade-offs between these two training paradigms in the context of large-scale search, recommendation, and personalization systems.

Background and Motivation

In large-scale search, recommendation, and personalization systems, ranking has long faced challenges in efficiency, scalability, and user-intent understanding. Traditional pipeline approaches tend to treat retrieval, ranking, and generation as separate stages, making it hard to satisfy LLM-era requirements for natural language interaction, multi-hop reasoning, and real-time knowledge. This paper is proposed against this backdrop, aiming to systematically advance the theoretical and practical boundaries of this intersection.

From a problem-definition perspective, the core scenarios include open-domain information access, enterprise knowledge retrieval, conversational search, semantic understanding in recommendation, and end-to-end architectures that coordinate external knowledge sources with generative models.

Core Contributions

• Proposes or consolidates a unified perspective for the problem domain, bringing scattered related work into a comparable framework.
• Provides a clear decomposition of method components (representation learning, retrievers, rerankers, planners, generators, feedback mechanisms) to facilitate engineering adoption.
• Offers reproducible benchmarks, datasets, or taxonomy tables in its experimental protocol, lowering the entry cost for follow-up researchers.
• Discusses interfaces with emerging paradigms such as LLM tool calling, reinforcement learning, and multi-agent collaboration, pointing to paths from research prototypes to industrial systems.
• Explicitly lists open problems: evaluation trustworthiness, latency and cost, hallucination and safety, and cross-lingual and multimodal extension.

Method / System Architecture

Methodologically, the work follows a four-step pattern: problem formalization → model/system design → training or construction pipeline → inference pipeline.

1. Input and representation: encode queries, documents, and user context as dense or sparse representations, or build structured prompts. 2. Core modules: may include retrievers, rerankers, planners, memory modules, and tool interfaces, chained or combined in parallel by task. 3. Learning strategies: supervised fine-tuning, contrastive learning, distillation, reinforcement learning (including process rewards), and bootstrapped data synthesis. 4. Inference strategies: single-round retrieval, iterative retrieval, parallel sub-queries, early stopping, and budget control.

Experiments and Evaluation

Typical experimental setups in this research area include:

• Datasets: MS MARCO, BEIR, Natural Questions, domain-specific corpora, and public recommendation datasets.
• Metrics: nDCG@10, MRR, Recall@k, Hit@k, human preference, task success rate, latency, and token cost.
• Baselines: BM25, dense retrieval, cross-encoder rerankers, retrieval-free LLMs, and commercial search APIs.
• Ablations: verifying the contribution of modules such as number of retrieval steps, reranking depth, and training-data scale.

Specific numerical results should be checked against the original paper's tables; this report is compiled from the abstract and public metadata, so please verify quantitative claims in the PDF before citing.

Key Takeaways for Search / Rec / Personalization

1. Architecture: cascaded retrieval + reranking + generation remains mainstream, but the agentic paradigm is making "retrieval frequency and policy" itself a learnable object. 2. Data: high-quality instruction data and click/session logs are equally critical; synthetic data must guard against knowledge leakage and distribution shift. 3. Evaluation: the gap between offline metrics and online satisfaction is widening; LLM-as-judge needs cross-validation with human assessment. 4. Product: latency, cost, explainability, and safety policies are hard constraints for industrial deployment — academic benchmarks alone are not enough.

Limitations and Future Work

Possible limitations include experiment scale constrained by GPU budgets, mismatch between benchmarks and real user distributions, unknown cross-lingual generalization from English-centric data, and safety risks of agentic systems on the open web. Future directions include more efficient test-time compute allocation, deeper integration with knowledge graphs and structured databases, and causal/fairness constraints for recommendation systems.

Related Entries

• Deep Learning to Rank in Industrial Search Engines, Recommender Systems (doi:10.1145/3797895)
• Multi-Objective Recommendation in the Era of Generative AI: A Survey (arXiv:2506.16893)
• A Generative Re-ranking Model for List-level Multi-objective Optimization (arXiv:2505.07197)
• A Thorough Comparison of Cross-Encoders and LLMs for Reranking SPLADE (arXiv:2403.10407)
• Accelerating Listwise Reranking: Reproducing and Enhancing FIRST, SIGIR (doi:10.1145/3726302.3730287)
• Adaptive Neural Ranking Framework: Toward Maximized Business Goal (doi:10.1145/3589334.3645605)

Glossary

| Term | Meaning | |------|---------| | IR | Information Retrieval | | RAG | Retrieval-Augmented Generation | | LTR | Learning to Rank | | nDCG | Normalized Discounted Cumulative Gain, a ranking quality metric | | Agentic Search | Modeling search as sequential decision-making and tool calling | | Gen-IR | Generative Information Retrieval |

Actionable Advice for Readers

1. Researchers: reproduce the core comparison experiments and check whether statistical significance and compute cost are reported. 2. Engineers: extract pluggable modules (encoders, rerankers, planners) and evaluate integration cost with your existing stack. 3. Product managers: focus on user-perceivable benefits (latency, answer trustworthiness, multi-turn consistency) rather than offline nDCG alone.