# LogicVista ![Category](https://img.shields.io/badge/Category-Reasoning-1f6feb) ![Metric](https://img.shields.io/badge/Metric-%25%20Accuracy-2da44e) ![Tasks](https://img.shields.io/badge/Tasks-448-8957e5) ![Saturation](https://img.shields.io/badge/Saturation-Low-d29922) ![Created](https://img.shields.io/badge/Created-Jul%202024-555) [![Paper](https://img.shields.io/badge/Paper-arXiv%202407.04973-b31b1b)](https://arxiv.org/abs/2407.04973) [![GitHub](https://img.shields.io/badge/GitHub-Yijia--Xiao%2FLogicVista-181717?logo=github&logoColor=white)](https://github.com/Yijia-Xiao/LogicVista) > **Quick answer:** LogicVista (Xiao, Sun, Liu & Wang, 2024) is a multimodal LLM benchmark of 448 visual multiple-choice questions evaluating integrated logical reasoning — spatial, deductive, inductive, numeric, and mechanical — across diagrams, OCR, patterns, graphs, tables, 3D shapes, puzzles, sequences, and physics scenarios. Liquid AI's LFM3-VL-2B scores 60.4% as of October 2026. ## At a Glance **What it tests:** Five logical reasoning skills — inductive, deductive, numerical, spatial, and mechanical reasoning — applied to nine visual task formats (diagrams, OCR, patterns, graphs, tables, 3D shapes, puzzles, sequences, physics). **Why it matters:** Many multimodal benchmarks focus on perception or knowledge recall; LogicVista specifically isolates logical reasoning that requires visual input, closer to the kind of structured thinking needed for navigation, puzzle-solving, and real-world planning tasks. **Known limitations:** At 448 questions it is a relatively small benchmark, so per-skill breakdowns can be noisy, and each question is annotated with both a correct answer and the human-written reasoning behind it, enabling open-ended as well as multiple-choice evaluation — but most reported scores to date use the multiple-choice setting only. ## What LogicVista Measures LogicVista was built to close a gap in multimodal LLM evaluation: while MLLMs had shown fascinating emergent abilities (from writing poetry about an image to solving math problems), there was no systematic benchmark isolating their proficiency at integrated logical reasoning in visual contexts — reasoning essential for tasks like navigation and puzzle-solving. The benchmark organizes its 448 multiple-choice questions across five reasoning-skill categories (inductive: 107, deductive: 93, numerical: 95, spatial: 79, mechanical: 74) and nine visual task-format capabilities (diagrams: 330, OCR: 234, puzzles: 256, patterns: 105, sequences: 76, tables: 70, physics: 69, graphs: 67, 3D shapes: 45 — formats overlap within questions). Each question is manually annotated with the correct answer and the reasoning behind it, so models can be evaluated in either an open-ended or strict multiple-choice mode. Because the original paper evaluated only 8 MLLMs at release, LogicVista remains a comparatively low-saturation benchmark; scores reported since by individual labs (not yet aggregated into a public third-party leaderboard) should be read as self-reported unless independently re-run. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Reasoning (multimodal logical reasoning) | | Metric | % accuracy | | Number of tasks | 448 | | Saturation | Low | | Created by | Xiao et al. | | Source paper | [Xiao et al. 2024](https://arxiv.org/abs/2407.04973) | | GitHub | [Yijia-Xiao/LogicVista](https://github.com/Yijia-Xiao/LogicVista) | | Dataset | [Yijia-Xiao/LogicVista on GitHub](https://github.com/Yijia-Xiao/LogicVista) | ## How LogicVista Is Scored Each of the 448 multiple-choice questions is scored for correctness against the annotated answer key; the overall score is the percentage of questions answered correctly, with breakdowns available by reasoning skill (inductive/deductive/numerical/spatial/mechanical) and by visual task format. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [LFM3-VL-2B](/models/liquid-ai/lfm3-vl-2b) | 60.4% | [LFM3-VL-2B model card](https://huggingface.co/LiquidAI/LFM3-VL-2B) | 2026-10 | | 2 | [LFM2.5-VL-3B](/models/liquid-ai/lfm2-5-vl-3b) | 37.4% | [LFM2.5-VL-3B launch blog](https://huggingface.co/LiquidAI/LFM2.5-VL-3B) | 2026-08 | *Scores above are self-reported by Liquid AI on their respective model cards/launch posts and shown for context. They are not Benchgen measurements, and are not drawn from a shared third-party leaderboard — only the 2 scores above are currently tracked locally.* ## LogicVista on Benchgen > No Benchgen results yet — [be the first to run LogicVista](/benchmarks/xiao-et-al/logicvista). ## LogicVista vs Other Benchmarks | Benchmark | What it tests | Saturation | |---|---|---| | **LogicVista** | Integrated logical reasoning (spatial/deductive/inductive/numeric/mechanical) in visual contexts | Low | | [MMStar](/benchmarks/chen-et-al/mmstar) | Curated, vision-indispensable general VQA | Low | | [BLINK](/benchmarks/fu-et-al/blink) | Core visual perception abilities humans do effortlessly | Low | LogicVista is narrower and more reasoning-focused than general VQA benchmarks like MMStar — use it when you specifically need to assess a model's structured logical reasoning over visual input, not general perception or knowledge recall. ## Run LogicVista on Your Model Benchgen lets you evaluate your own multimodal model against LogicVista's 448-question logical-reasoning set and track results across model versions. ## Related Models & Benchmarks - Top performers: [LFM3-VL-2B](/models/liquid-ai/lfm3-vl-2b) · [LFM2.5-VL-3B](/models/liquid-ai/lfm2-5-vl-3b) - Related benchmarks: [MMStar](/benchmarks/chen-et-al/mmstar) · [BLINK](/benchmarks/fu-et-al/blink) ## Frequently Asked Questions

What is LogicVista?

LogicVista is a multimodal LLM benchmark of 448 visual multiple-choice questions evaluating integrated logical reasoning — spatial, deductive, inductive, numeric, and mechanical — introduced by Xiao, Sun, Liu & Wang in 2024.

What does a good LogicVista score look like?

Liquid AI's LFM3-VL-2B reports 60.4% as of October 2026, the highest currently tracked on Benchgen; LFM2.5-VL-3B, an earlier and smaller model, scores 37.4%.

Who created LogicVista?

LogicVista was created by Yijia Xiao, Edward Sun, Tianyu Liu, and Wei Wang; see the [original paper](https://arxiv.org/abs/2407.04973).

Is LogicVista saturated?

No — LogicVista remains a low-saturation benchmark; the original paper evaluated only 8 MLLMs, and no public third-party leaderboard aggregates results broadly yet.

How does LogicVista differ from MMStar?

MMStar tests general vision-indispensable VQA across six broad capability dimensions, while LogicVista narrows specifically to integrated logical reasoning (spatial, deductive, inductive, numeric, mechanical) using puzzle-like visual formats.

--- *Benchmark definition paraphrased from [Xiao et al. 2024](https://arxiv.org/abs/2407.04973). State-of-the-art scores sourced from published model cards and attributed inline. Last updated 2026-10-31.*

Rank	Model	Score
1	lfm3-vl-2b	60.4
2	lfm2-5-vl-3b	37.4

LogicVista — Results

LogicVista — Evaluations