Introspection Adapters: Training LLMs to Report Their Learned Behaviors

Keshav Shenoy, Li Yang, Abhay Sheshadri, Jack Lindsey, Samuel Marks, Rowan Wang
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:111624-111655, 2026.

Abstract

Can we train LLMs to introspect, i.e. to faithfully describe their own behaviors in natural language? Prior work has shown some, limited, success. However, it is difficult to scale introspection training due to a lack of ground-truth labels. In this work, we study an approach to introspection training which side-steps this data bottleneck. Given a target model $M$, our method works by fine-tuning models $M_i$ from $M$ with implanted behaviors $b_i$ (such as downplaying medical problems); the $(M_i, b_i)$ pairs serve as labeled introspection training data. We then train an introspection adapter (IA): a LoRA adapter jointly optimized across the fine-tunes $M_i$ which causes them to verbalize their implanted behaviors. This IA induces faithful introspection in fine-tunes of $M$ that were trained in very different ways from the $M_i$, as well as in $M$ itself. This is surprising because the IA was never trained on $M$. To demonstrate the utility of IAs, we use them to successfully audit misaligned models introduced in prior work. IAs can also be used to detect fine-tuning API attacks which train models to comply with encrypted harmful requests. Notably, IAs are more effective when applied to larger models. Overall, our results suggest that IAs are a scalable, effective, and practically useful approach to LLM introspection training.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-shenoy26a, title = {Introspection Adapters: Training {LLM}s to Report Their Learned Behaviors}, author = {Shenoy, Keshav and Yang, Li and Sheshadri, Abhay and Lindsey, Jack and Marks, Samuel and Wang, Rowan}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {111624--111655}, 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/shenoy26a/shenoy26a.pdf}, url = {https://proceedings.mlr.press/v306/shenoy26a.html}, abstract = {Can we train LLMs to introspect, i.e. to faithfully describe their own behaviors in natural language? Prior work has shown some, limited, success. However, it is difficult to scale introspection training due to a lack of ground-truth labels. In this work, we study an approach to introspection training which side-steps this data bottleneck. Given a target model $M$, our method works by fine-tuning models $M_i$ from $M$ with implanted behaviors $b_i$ (such as downplaying medical problems); the $(M_i, b_i)$ pairs serve as labeled introspection training data. We then train an introspection adapter (IA): a LoRA adapter jointly optimized across the fine-tunes $M_i$ which causes them to verbalize their implanted behaviors. This IA induces faithful introspection in fine-tunes of $M$ that were trained in very different ways from the $M_i$, as well as in $M$ itself. This is surprising because the IA was never trained on $M$. To demonstrate the utility of IAs, we use them to successfully audit misaligned models introduced in prior work. IAs can also be used to detect fine-tuning API attacks which train models to comply with encrypted harmful requests. Notably, IAs are more effective when applied to larger models. Overall, our results suggest that IAs are a scalable, effective, and practically useful approach to LLM introspection training.} }
Endnote
%0 Conference Paper %T Introspection Adapters: Training LLMs to Report Their Learned Behaviors %A Keshav Shenoy %A Li Yang %A Abhay Sheshadri %A Jack Lindsey %A Samuel Marks %A Rowan Wang %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-shenoy26a %I PMLR %P 111624--111655 %U https://proceedings.mlr.press/v306/shenoy26a.html %V 306 %X Can we train LLMs to introspect, i.e. to faithfully describe their own behaviors in natural language? Prior work has shown some, limited, success. However, it is difficult to scale introspection training due to a lack of ground-truth labels. In this work, we study an approach to introspection training which side-steps this data bottleneck. Given a target model $M$, our method works by fine-tuning models $M_i$ from $M$ with implanted behaviors $b_i$ (such as downplaying medical problems); the $(M_i, b_i)$ pairs serve as labeled introspection training data. We then train an introspection adapter (IA): a LoRA adapter jointly optimized across the fine-tunes $M_i$ which causes them to verbalize their implanted behaviors. This IA induces faithful introspection in fine-tunes of $M$ that were trained in very different ways from the $M_i$, as well as in $M$ itself. This is surprising because the IA was never trained on $M$. To demonstrate the utility of IAs, we use them to successfully audit misaligned models introduced in prior work. IAs can also be used to detect fine-tuning API attacks which train models to comply with encrypted harmful requests. Notably, IAs are more effective when applied to larger models. Overall, our results suggest that IAs are a scalable, effective, and practically useful approach to LLM introspection training.
APA
Shenoy, K., Yang, L., Sheshadri, A., Lindsey, J., Marks, S. & Wang, R.. (2026). Introspection Adapters: Training LLMs to Report Their Learned Behaviors. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:111624-111655 Available from https://proceedings.mlr.press/v306/shenoy26a.html.

Related Material