Meng Cao

Renjie Pi, Felix Bai, Qibin Chen et al.

The paradigm of using Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) as evaluative judges has emerged as an effective approach in RLHF and inference-time scaling. In this work, we propose Multimodal Reasoner as a Judge (MR. Judge), a paradigm for empowering general-purpose MLLMs judges with strong reasoning capabilities. Instead of directly assigning scores for each response, we formulate the judgement process as a reasoning-inspired multiple-choice problem. Specifically, the judge model first conducts deliberate reasoning covering different aspects of the responses and eventually selects the best response from them. This reasoning process not only improves the interpretibility of the judgement, but also greatly enhances the performance of MLLM judges. To cope with the lack of questions with scored responses, we propose the following strategy to achieve automatic annotation: 1) Reverse Response Candidates Synthesis: starting from a supervised fine-tuning (SFT) dataset, we treat the original response as the best candidate and prompt the MLLM to generate plausible but flawed negative candidates. 2) Text-based reasoning extraction: we carefully design a data synthesis pipeline for distilling the reasoning capability from a text-based reasoning model, which is adopted to enable the MLLM judges to regain complex reasoning ability via warm up supervised fine-tuning. Experiments demonstrate that our MR. Judge is effective across a wide range of tasks. Specifically, our MR. Judge-7B surpasses GPT-4o by 9.9% on VL-RewardBench, and improves performance on MM-Vet during inference-time scaling by up to 7.7%.

Meng Cao, Songcan Chen

Domain generalization addresses domain shift in real-world applications. Most approaches adopt a domain angle, seeking invariant representation across domains by aligning their marginal distributions, irrespective of individual classes, naturally leading to insufficient exploration of discriminative information. Switching to a class angle, we find that multiple domain-related peaks or clusters within the same individual classes must emerge due to distribution shift. In other words, marginal alignment does not guarantee conditional alignment, leading to suboptimal generalization. Therefore, we argue that acquiring discriminative generalization between classes within domains is crucial. In contrast to seeking distribution alignment, we endeavor to safeguard domain-related between-class discrimination. To this end, we devise a novel Conjugate Consistent Enhanced Module, namely Con2EM, based on a distribution over domains, i.e., a meta-distribution. Specifically, we employ a novel distribution-level Universum strategy to generate supplementary diverse domain-related class-conditional distributions, thereby enhancing generalization. This allows us to resample from these generated distributions to provide feedback to the primordial instance-level classifier, further improving its adaptability to the target-agnostic. To ensure generation accuracy, we establish an additional distribution-level classifier to regularize these conditional distributions. Extensive experiments have been conducted to demonstrate its effectiveness and low computational cost compared to SOTAs.