What Does Vision Tool-Use Reinforcement Learning Really Learn? Disentangling Tool-Induced and Intrinsic Effects for Crop-and-Zoom

Yan Ma, Weiyu Zhang, Tianle Li, Du Linge, Xuyang Shen, Pengfei Liu
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:85238-85261, 2026.

Abstract

Vision tool-use reinforcement learning (RL) can equip vision–language models with visual operators such as crop-and-zoom and achieves strong performance gains, yet it remains unclear whether these gains are driven by improvements in tool use or evolving intrinsic capabilities. We introduce MED (Measure–Explain–Diagnose), a coarse-to-fine framework that disentangles intrinsic capability changes from tool-induced effects, decomposes the tool-induced performance difference into gain and harm terms, and probes the mechanisms driving their evolution. Across checkpoint-level analyses on two VLMs with different tool priors and six benchmarks, we find that improvements are dominated by intrinsic learning, while tool-use RL mainly reduces tool-induced harm (e.g., fewer call-induced errors and weaker tool schema interference) and yields limited progress in tool-based correction of intrinsic failures.Overall, current vision tool-use RL learns to coexist safely with tools rather than master them.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-ma26as, title = {What Does Vision Tool-Use Reinforcement Learning Really Learn? {D}isentangling Tool-Induced and Intrinsic Effects for Crop-and-Zoom}, author = {Ma, Yan and Zhang, Weiyu and Li, Tianle and Linge, Du and Shen, Xuyang and Liu, Pengfei}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {85238--85261}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/ma26as/ma26as.pdf}, url = {https://proceedings.mlr.press/v306/ma26as.html}, abstract = {Vision tool-use reinforcement learning (RL) can equip vision–language models with visual operators such as crop-and-zoom and achieves strong performance gains, yet it remains unclear whether these gains are driven by improvements in tool use or evolving intrinsic capabilities. We introduce MED (Measure–Explain–Diagnose), a coarse-to-fine framework that disentangles intrinsic capability changes from tool-induced effects, decomposes the tool-induced performance difference into gain and harm terms, and probes the mechanisms driving their evolution. Across checkpoint-level analyses on two VLMs with different tool priors and six benchmarks, we find that improvements are dominated by intrinsic learning, while tool-use RL mainly reduces tool-induced harm (e.g., fewer call-induced errors and weaker tool schema interference) and yields limited progress in tool-based correction of intrinsic failures.Overall, current vision tool-use RL learns to coexist safely with tools rather than master them.} }
Endnote
%0 Conference Paper %T What Does Vision Tool-Use Reinforcement Learning Really Learn? Disentangling Tool-Induced and Intrinsic Effects for Crop-and-Zoom %A Yan Ma %A Weiyu Zhang %A Tianle Li %A Du Linge %A Xuyang Shen %A Pengfei Liu %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-ma26as %I PMLR %P 85238--85261 %U https://proceedings.mlr.press/v306/ma26as.html %V 306 %X Vision tool-use reinforcement learning (RL) can equip vision–language models with visual operators such as crop-and-zoom and achieves strong performance gains, yet it remains unclear whether these gains are driven by improvements in tool use or evolving intrinsic capabilities. We introduce MED (Measure–Explain–Diagnose), a coarse-to-fine framework that disentangles intrinsic capability changes from tool-induced effects, decomposes the tool-induced performance difference into gain and harm terms, and probes the mechanisms driving their evolution. Across checkpoint-level analyses on two VLMs with different tool priors and six benchmarks, we find that improvements are dominated by intrinsic learning, while tool-use RL mainly reduces tool-induced harm (e.g., fewer call-induced errors and weaker tool schema interference) and yields limited progress in tool-based correction of intrinsic failures.Overall, current vision tool-use RL learns to coexist safely with tools rather than master them.
APA
Ma, Y., Zhang, W., Li, T., Linge, D., Shen, X. & Liu, P.. (2026). What Does Vision Tool-Use Reinforcement Learning Really Learn? Disentangling Tool-Induced and Intrinsic Effects for Crop-and-Zoom. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:85238-85261 Available from https://proceedings.mlr.press/v306/ma26as.html.

Related Material