AI Paper Atlas

Research map / Vision-language and multimodal / Multimodal LLMs

OX-MABSR: A Benchmark for Open-domain Explainable Multimodal Aspect-Based Sentiment Reasoning

Xinjing Liu, ZiXin Xue, Pengyue Lin, Xinyu Tu, Siwei Xu, Ruifan Li

AAAI 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Multimodal LLM benchmarks · 1 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces OX-MABSR, a benchmark and task for open-domain explainable multimodal aspect-based sentiment reasoning covering expressive sentiment categories and implicit aspects.

Problem

Existing multimodal aspect-based sentiment analysis (MABSA) tasks usually use a fixed sentiment set and explicitly mentioned aspects, and rarely address expressive sentiment categories, implicit aspects, or explainability.

Approach

Defines a new open-domain explainable MABSA task that extracts aspects from text-image pairs (including implicit ones), assigns expressive sentiment categories, and requires explanations, released as a benchmark.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add images (multimodal input) to MABSR-LLM (ablation)MABSR-LLM · OX-MABSR-Bench test, aspect-sentimentF1_in: 31.15 -> 44.65Improved
add images (multimodal input) (ablation)MABSR-LLM · OX-MABSR-Bench test, reasoning pathsSC: 52.78 -> 72.17Improved
add GRPO RL stage after SFT (ablation)MABSR-LLM · OX-MABSR-Bench test, aspect-sentimentF1_in: 38.61 -> 44.65Improved
add GRPO RL stage after SFT (ablation)MABSR-LLM · OX-MABSR-Bench test, perceptual causesSC: 62.87 -> 71.21Improved
add reasoning-path generation task to multi-task training (ablation)MABSR-LLM · OX-MABSR-Bench test, reasoning pathsSC: 60.77 -> 72.17Improved
add reasoning-path generation task to multi-task training (ablation)MABSR-LLM · OX-MABSR-Bench test, aspect-sentimentF1_in: 41.72 -> 44.65Improved
use Qwen3 backbone instead of LLaMA3-8B or ChatGLM3-6B (ablation)MABSR-LLM · OX-MABSR-Bench testSCImproved
set LoRA rank to 8 instead of 4 (ablation)MABSR-LLM · OX-MABSR-Bench testreasoning qualityImproved
raise LoRA rank from 8 to 16 (ablation)MABSR-LLM · OX-MABSR-Bench testreasoning qualityHurt
train MABSR-LLM instead of prompting general MLLMsInternVL3-8B baseline · OX-MABSR-Bench test, aspect-sentimentF1_in: 15.08 -> 44.65Improved
implicit rather than explicit aspectsMABSR-LLM · OX-MABSR-Bench test, aspect-sentimentF1Hurt
longer, harder explanationsMABSR-LLM · OX-MABSR-Bench test, explanationsSCHurt

Experimental setup

Benchmarks: OX-MABSR

Builds on: MABSA

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy