AVERE: Improving Audiovisual Emotion Reasoning with Preference Optimization

The paper introduces EmoReAlM, a benchmark for evaluating spurious cue–emotion associations and hallucinations in multimodal large language models, alongside AVEm-DPO, a preference optimization technique that aligns model responses with audiovisual inputs and emotion-centric queries. Within the taxonomy, it resides in the Preference Optimization and Alignment leaf under Training Strategies and Optimization. This leaf contains only two papers total, indicating a relatively sparse research direction compared to more crowded areas like General Fusion Architectures or Comprehensive Emotion Understanding Benchmarks, which house four or more sibling works.

The taxonomy reveals that neighboring leaves—Instruction Tuning and Prompting, Knowledge Distillation and Transfer, and Contrastive and Representation Learning—collectively address training methodologies but do not explicitly target preference-based alignment for emotion reasoning. The Modality Conflict and Alignment Handling branch under Frameworks and Architectures focuses on cross-modal inconsistencies at the architectural level, whereas this work tackles alignment through optimization objectives. The scope_note for Preference Optimization emphasizes methods using DPO or alignment techniques, distinguishing it from instruction tuning without preference mechanisms, which clarifies that the paper's dual focus on benchmark and optimization sits at the intersection of evaluation and training innovation.

Among seventeen candidates examined, the EmoReAlM benchmark contribution shows one refutable candidate out of three examined, suggesting some overlap with existing evaluation frameworks. The AVEm-DPO technique examined ten candidates with one refutable match, indicating that while preference optimization for emotion reasoning is explored, the specific audiovisual alignment strategy may offer incremental novelty. The text-prior debiasing regularization examined four candidates with two refutable matches, pointing to more substantial prior work on mitigating language model biases. Given the limited search scope—seventeen candidates total, not hundreds—these statistics reflect a snapshot rather than exhaustive coverage, and the relatively low refutation counts suggest the contributions occupy a less saturated niche within the broader emotion MLLM landscape.

Overall, the work appears to address a genuine gap in preference-based alignment for audiovisual emotion reasoning, particularly given the sparse Preference Optimization leaf. However, the benchmark and debiasing components show moderate overlap with existing evaluation and bias-mitigation efforts based on the limited candidate pool. The analysis covers top-K semantic matches and does not claim comprehensive field coverage, so additional related work may exist beyond the seventeen candidates examined.