Research map / Vision-language and multimodal / Multimodal LLMs

Praxis-VLM: Vision-Grounded Decision Making via Text-Driven Reinforcement Learning

NeurIPS 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Reinforcement learning for visual 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.

Shows VLMs do comparably when images are replaced by text, then trains Qwen2.5-VL with text-only GRPO (multi-stage, adaptive rewards) and transfers the learned decision-making to multimodal benchmarks, outperforming SFT baselines.

Problem

Vision-language models lack explicit situational reasoning for vision-grounded decision-making, and training them requires scarce, costly paired image-text data with reasoning annotations.

Approach

Train the VLM's language component with GRPO on GPT-4o-synthesized text-only decision scenarios (after a math geometry cold start with adaptive format/accuracy/length rewards), then apply the model to image inputs at inference, transferring text-learned reasoning to visual decision tasks.

Key results

• Praxis-VLM-7B reaches 84.03 on VIVA, 60.25 on PCA-Bench, and 54.33 on EgoNormia, versus 80.97, 46.37, and 46.19 for Qwen2.5-VL-7B.
• SFT baselines drop sharply on out-of-domain EgoNormia (e.g., Qwen2.5-VL-7B w/ SFT 34.83, Reason SFT 34.08), while Praxis-VLM stays above the base model.
• Praxis-VLM-3B reaches 79.03 on VIVA, 50.79 on PCA-Bench, and 54.27 on EgoNormia, versus 76.61, 48.58, and 51.92 for the 3B base.

Measured effects reported in the paper

Surprising findings

• Replacing images with textual captions gives comparable or slightly better VLM accuracy on VIVA and PCA-Bench in the preliminary analysis.
• Directly running GRPO on an instruction-tuned VLM with adaptive rewards works without an SFT warm-up stage.

Limitations

• Only 3B and 7B Qwen2.5-VL backbones tested; scaling behavior unexplored.
• Visual perception is not improved, so transfer depends on the base model's visual grounding; error analysis shows misinterpretation of context and safety prioritization failures.
• Long reasoning chains can truncate at 1024 tokens and show overthinking on the hardest samples.

Experimental setup

Datasets: Geometry3K GPT-4o-synthesized text decision-making dataset (10K train, 1K val) VIVA PCA-Bench EgoNormia

Benchmarks: VIVA PCA-Bench EgoNormia

Metrics: accuracy majority vote accuracy pass@1

Baselines: Qwen2.5-VL-3B Qwen2.5-VL-7B SFT Reason SFT one-stage GRPO

Builds on: GRPO DeepSeek-R1 Qwen2.5-VL

Closest papers by meaning

• PeRL: Permutation-Enhanced Reinforcement Learning for Interleaved Vision-Language Reasoning (NeurIPS 2025 · similarity 0.90)
• More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
• PRPO: Perception-Reinforced Policy Optimization via Token-Level Dynamic Advantage Reshaping (NeurIPS 2026 · similarity 0.90)
• Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
• Visual Planning: Let's Think Only with Images (ICLR 2026 · similarity 0.89)
• Grounded Reinforcement Learning for Visual Reasoning (NeurIPS 2025 · similarity 0.89)