Report prepared for the Whitebox Research AI Interpretability Fellowship 2025 - Project Phase
Chain-of-Thought (CoT) reasoning by Large Language Models is not always faithful, i.e., the reasoning steps don't necessarily reflect actual internal processes. Simultaneously, there has been a rapid uptake of multimodal models that integrate both visual and textual information. These systems pose new challenges and opportunities for studying CoT reasoning across modalities.
This report examines unfaithful shortcuts in multimodal models' CoT reasoning when presented with semantically equivalent visual and textual tasks. We evaluate Gemini 2.0 Flash Experimental and Claude 3.7 Sonnet using paired, semantically equivalent visual and textual versions of PutnamBench math problems to test the faithfulness of reasoning and performance across modalities.
Please see the report for more details.
Prepared by Ishita Pal, Aditya Thomas & Angel Martinez. Mentored by Angel Martinez.
We find that for the models tested, Google Gemini 2.0 Flash Experimental (with thinking enabled) and Anthropic Claude 3.7 Sonnet (with extended thinking), there was no difference in the accuracy of the results for the text and semantically similar image inputs (see section 3 Results and Discussion of our report).
| Claude 3.7 Sonnet (Thinking Mode) | Gemini 2.0 Flash Experimental (Thinking Mode) | |||
|---|---|---|---|---|
| Metric | Texts | Images | Texts | Images |
| Problems analysed | 27 | 27 | 181 | 181 |
| Correct responses | 24 | 26 | 114 | 110 |
The distribution of our “Faithful Metric” across the two modes for the two models is similar. The scores concentrating around 2 or 3 for both models suggest some degree of reliance on partial shortcuts. However, complete failures or perfect reasoning were also shown to be uncommon. Severely unfaithful steps are rare and so are highly faithful ones. More large-scale analysis is needed to validate these trends in various tasks and domains.
For how to install and run the scripts, please see Appendix D of the report.
We are grateful to the AI Interpretability Fellowship from WhiteBox Research for their support, which included training in interpretability and evaluation methodologies as well as financial assistance for this project. We are especially thankful to our project mentor, Angel Martinez, who conceived the project and provided guidance throughout its development. We also thank PutnamBench, whose transcription of the problems formed the basis of our dataset. Finally, we thank Arcuschin et al. (2025) for their codebase, which served as the foundation for our implementation.
Arcuschin, I., Janiak, J., Krzyzanowski, R., Rajamanoharan, S., Nanda, N., & Conmy, A. (2025). Chain-of-thought reasoning in the wild is not always faithful. arXiv preprint arXiv:2503.0867
Tsoukalas, G., Lee, J., Jennings, J., Xin, J., Ding, M., Jennings, M., ... & Chaudhuri, S. (2024). Putnambench: Evaluating neural theorem-provers on the putnam mathematical competition. arXiv preprint arXiv:2407.11214.