Research map / Vision-language and multimodal / Multimodal LLMs

What MLLMs Learn about When they Learn about Multimodal Reasoning

NeurIPS 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Chain-of-thought multimodal reasoning · 0 citations (OpenAlex)

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 MathLens, a symbolically grounded geometry benchmark that decomposes multimodal reasoning into perception, text-only reasoning, and multimodal-specific residual errors, and uses it to compare training strategies (RL, textual SFT, multimodal SFT) on 7-9B open MLLMs and several larger closed models.

Problem

Aggregate accuracy on multimodal reasoning benchmarks hides whether gains come from perception, reasoning, or their integration, and existing benchmarks rarely allow controlled component-wise diagnosis of training strategies.

Approach

Generate every geometry problem from a formal symbolic state (FormalGeo-7K), so that diagrams, text-only descriptions, atomic perception probes, and semantically perturbed diagrams are aligned; perception and reasoning are measured separately, and multimodal-specific errors are defined as the residual failures when both succeed in isolation.

Key results

• MathLens correlates with MathVista (Spearman rho=0.83) and MathVerse (rho=0.86) across eight models, and weaker with MathVision (rho=0.67).
• Revisual-R1-7B text SFT raises text-description accuracy from 38.9 to 62.3 while diagram accuracy rises from 34.7 to 43.8 before multimodal RL; MM-RL further lifts diagram accuracy to 51.7.
• Consistency under diagram modifications rises after multimodal RL (e.g., VL-Rethinker-7B 31.0 to 37.3; ShareVL-R1-7B 31.0 to 41.3), while multimodal SFT lowers it (R1-OneVision-7B 31.0 to 24.0; Vision-R1 SFT 21.1).

Surprising findings

• Multimodal SFT lowers diagram consistency and accuracy relative to the backbone, with RL only partially recovering.
• Textual SFT, which never sees images, improves perception probes, and the gains are associated with more reflective reasoning.

Limitations

• Focuses on geometry; generalization to charts, science figures, and natural images is tested only on the small MathLens-NS set (107 problems).
• Multimodal-specific category is a conditional residual that may absorb unmeasured perceptual factors.
• Analyzes released open-weight checkpoints rather than controlled retraining, so comparisons are cross-model rather than ablations.

Experimental setup

Datasets: FormalGeo-7K Geometry3K BLINK V* SpatialEval-Real MMMU-Pro EMMA MathVista

Benchmarks: MathLens MathLens-NS MathVista MathVerse MathVision

Metrics: accuracy perception probe accuracy cross-variant consistency error-type categorization Spearman correlation

Baselines: MathVista MathVerse MathVision VL-Rethinker Vision-R1 R1-Onevision ShareVL-R1 Revisual-R1 Open-Vision-Reasoner GLM-4.1V-Thinking Gemini-2.5-Flash Claude 4 Sonnet

Builds on: FormalGeo-7K MathVerse Vision-R1 VL-Rethinker Revisual-R1

Closest papers by meaning

• AStar: Boosting Multimodal Reasoning with Automated Structured Thinking (AAAI 2026 · similarity 0.90)
• MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · similarity 0.90)
• Can MLLMs Reason in Multimodality? EMMA: An Enhanced MultiModal ReAsoning Benchmark (ICML 2025 · similarity 0.90)
• From Where Things Are to What They Are For: Benchmarking Spatial-Functional Intelligence in Multimodal LLMs (CVPR 2026 · similarity 0.90)
• TableMix: Enhancing Multimodal Table Reasoning in MLLMs from a Data-Centric Perspective (CVPR 2026 · similarity 0.89)
• Perception-R1: Advancing Multimodal Reasoning Capabilities of MLLMs via Visual Perception Reward (ICLR 2026 · similarity 0.89)