Has there been any good analysis on attention in long chain of thought? If we can take a recent reasoning model, can we use attention to control the model performance?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{tan-attention-analysis-on-2026,
author = {Tan, Chenhao},
title = {Attention Analysis on Long Chain of Thought},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/wrQwn33aq498iqtKNELL}
}Please sign in to comment on this idea.
I have read related work in this direction, most notably using RL verification methods to evaluate the responses. These are mainly post-training methods, but there are papers such as https://arxiv.org/pdf/2510.03223 that use dynamic attention steering. I have emailed you about my interest in researching this idea.