Research map / Vision-language and multimodal / Multimodal LLMs
MuSLR: Multimodal Symbolic Logical Reasoning
Read the paperCheck your idea against 108,967 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 MuSLR, a benchmark of 1,093 image+text instances requiring formal logical reasoning (PL, FOL, NM) across 7 domains, and LogiCAM, a modular GPT-4.1-based CoT framework that lifts performance by 14.13 points.
Problem
Current VLMs are not evaluated on deducing new facts from joint visual and textual input via formal symbolic logic, a capability needed in high-stakes settings such as autonomous driving and medical diagnosis.
Approach
Grounds formal logical rules (modus ponens, hypothetical syllogism, FOL and default-reasoning patterns) in web-sourced image-text scenarios, and uses a modular VLM pipeline (premise selection, reasoning-type identification, symbolic/heuristic reasoner) that approximates symbolic deduction directly on multimodal inputs instead of translating images to text for a theorem prover.
Key results
- GPT-4.1 is the best of seven evaluated VLMs at 46.8% on MuSLR-Bench with three-shot CoT, while all models struggle.
- LogiCAM on GPT-4.1 raises average CoT accuracy by 14.13 points, with FOL gains of 48.93 points.
- Around 70% of failures are attributed to logical misalignment between visual and textual premises.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply LogiCAM to GPT-4.1 | GPT-4.1 · MuSLR-Bench | accuracy: 14.13% | Helped |
| Apply LogiCAM to GPT-4.1 | GPT-4.1 · MuSLR-Bench PL | accuracy: 31.93% | Helped |
| Apply LogiCAM to GPT-4.1 | GPT-4.1 · MuSLR-Bench FOL | accuracy: 48.93% | Helped |
| Apply LogiCAM to GPT-4.1 | GPT-4.1 · MuSLR-Bench NM | accuracy: 26.17% | Helped |
| Use LogiCAM instead of Logic-LM + VLM | MuSLR-Bench PL · MuSLR-Bench PL | accuracy: 35.14 -> 60.44 | Helped |
| Use LogiCAM instead of Logic-LM + VLM | MuSLR-Bench FOL · MuSLR-Bench FOL | accuracy: 32.65 -> 42.55 | Helped |
| Add a symbolic reasoning module (ablation) | LogiCAM · MuSLR-Bench | accuracy: 5.14% | Helped |
| Add heuristic reasoning (ablation) | LogiCAM · MuSLR-Bench | accuracy: 3.45% | Helped |
| Add premise selection (ablation) | LogiCAM · MuSLR-Bench | accuracy: 3.27% | Helped |
Surprising findings
- Open-weight InternVL (45.20%) nearly matches closed-weight GPT-4.1 (46.84%).
- Logic-LM with a VLM translating images to text performs far worse than LogiCAM (35.14% vs 60.44% on PL).
Limitations
- Heuristic shortcuts are frequent in LogiCAM (13% of errors), showing difficulty deciding when to use symbolic versus commonsense reasoning.
- Performance drops sharply with reasoning depth, with LogiCAM falling from 71.91% to 54.61%.
- Visual-textual logical misalignment remains the dominant failure mode across models.
Experimental setup
Datasets: COCO Flickr30k nocaps MIMIC RVL-CDIP ScienceQA
Benchmarks: MuSLR-Bench
Metrics: accuracy ROUGE-L BERTScore ROSCOE
Baselines: GPT-4.1 GPT-4o Claude-3.7-Sonnet Qwen2.5-VL-7B-Instruct InternVL3-8B Llava-1.5-7B InstructBLIP-Vicuna-13B Logic-LM + VLM
Builds on: chain-of-thought prompting Logic-LM
Closest papers by meaning
- AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.86)
- Logic Unseen: Revealing the Logical Blindspots of Vision-Language Models (AAAI 2026 · similarity 0.86)
- Can MLLMs Reason in Multimodality? EMMA: An Enhanced MultiModal ReAsoning Benchmark (ICML 2025 · similarity 0.86)
- Large Language Models Meet Symbolic Provers for Logical Reasoning Evaluation (ICLR 2025 · similarity 0.85)
- TVI-CoT: Text-Visual Interleaved Chain-of-Thought Reasoning for Multimodal Understanding (ICML 2026 · similarity 0.85)
- MMR-Life: Piecing Together Real-life Scenes for Multimodal Multi-image Reasoning (ICLR 2026 · similarity 0.85)
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.