The rapid advancement of large vision-language models (LVLMs) has driven significant progress in multimodal tasks, enabling models to interpret, reason, and generate outputs across both visual and textual domains. While excelling in generative tasks, existing LVLMs often face limitations in tasks requiring high-fidelity representation learning, such as generating image or text embeddings for retrieval. Recent work has proposed finetuning LVLMs for representational learning, but the fine-tuned model often loses its generative capabilities due to the representational learning training paradigm. To address this trade-off, we introduce CAFe, a contrastive-autoregressive fine-tuning framework that enhances LVLMs for both representation and generative tasks. By integrating a contrastive objective with autoregressive language modeling, our approach unifies these traditionally separate tasks, achieving state-of-the-art results in both multimodal retrieval and multimodal generative benchmarks, including object hallucination (OH) mitigation. CAFe establishes a novel framework that synergizes embedding and generative functionalities in a single model, setting a foundation for future multimodal models that excel in both retrieval precision and coherent output generation.

Recent advancements in large language models (LLMs) have demonstrated remarkable capabilities in generating coherent, contextually appropriate responses across a wide range of tasks (Brown et al., 2020). A key approach to further refine these models is Reinforcement Learning from Human Feedback (RLHF), which leverages human evaluations to guide the training process and align model outputs more closely with human preferences (Stiennon et al., 2020; Ouyang et al., 2022; Bai et al., 2022; Wang et al., 2024). RLHF typically involves training a reward model to capture human preferences, which is then used to fine-tune LLMs via reinforcement learning (RL) (Schulman et al., 2017; Chen et al., 2024b,f). Despite the success of RLHF, reward modeling is inherently prone to spurious correlations, which are associations in the training data that do not reflect true causal relationships (Veitch et al., 2021), and can lead to unintended biases and induce reward hacking (McMilin, 2022). Reward hacking occurs when RL agents exploit flaws or ambiguities in the reward function to maximize rewards without genuinely improving alignment with desired behaviors or completing designed tasks (Amodei et al., 2016; Weng, 2024). Consequently, this leads to misaligned models that exhibit biases such as favoring longer outputs (length bias) (Zheng et al., 2023), agreeing with user's incorrect assertions (sycophancy bias) (Perez et al., 2022), developing unintended shortcuts when making predictions (concept bias) (Zhou et al., 2023), and implicitly developing discrimination over certain demographic groups (discrimination bias) (Tamkin et al., 2023; Chen et al., 2024c). These biases, rooted in spurious correlations and reward hacking rather than true causal relationships, undermine the reliability and trustworthiness of LLMs, posing significant challenges for their safe and responsible deployment in real-world applications (Anwar et al., 2024; Qi et al., 2024). To understand and mitigate these issues, it is essential to consider the sources of error in reward modeling.