Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

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

Authors: Zhe Hu, Jing Li, Zhongzhu Pu, Hou Pong (Ken) Chan, Yu Yin

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments across diverse decision-making benchmarks demonstrate that Praxis-VLM substantially outperforms standard supervised fine-tuning, exhibiting superior performance and generalizability. Further analysis confirms that our models engage in explicit and effective reasoning, underpinning their enhanced performance and adaptability.
Researcher Affiliation Collaboration Department of Computing, The Hong Kong Polytechnic University, Tsinghua University Inspire Omni AI, Alibaba Group, Case Western Reserve University EMAIL
Pseudocode No The paper describes methods and processes in narrative text and diagrams (Figure 3), but does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code Yes https://github.com/Derekkk/Praxis-VLM
Open Datasets Yes To evaluate Praxis-VLM, we adopt challenging decision-making benchmarks spanning diverse tasks: VIVA [18] for human-centered situations, PCA-Bench for embodied agent tasks, and Ego Normia [22] for first-person video understanding. ... We employ the geometry3k dataset [33] for GRPO training...
Dataset Splits Yes We follow the original data splits and prompts provided by each benchmark. ... Key statistics for these benchmarks are presented in Table 3.
Hardware Specification Yes All models are trained on four NIVIDA A100 and H100 GPUs.
Software Dependencies No For GRPO implementation, we use the Easy RL Library 3. ... For SFT implementation, we employ the Hugging Face TRL 4. ... During inference, we leverage VLLM Library [37] with greedy decoding. ... use the Math-Verify Library 5 to calculate the binary reward
Experiment Setup Yes For GRPO implementation, we use the Easy RL Library 3. We adopt the default hyper-parameters, and set rollout N to 5 and KL divergence coefficient as 0.01. The learning rate is set as 1e-6. ... For SFT implementation, we employ the Hugging Face TRL 4. We set the number of training epochs as 3 and learning rate as 2e-5.