@article{AgenticSuperoptimization,title={Agentic Superoptimization of Bioimaging Analysis Workflows},author={Wang, X. and Chen, J. and Farhang, A. R. and Stiles, S. and Horstmann, K. A. and Sehgal, A. and Light, J. and Van Valen, D. and Yue, Y. and Sun, J. J.},journal={Conference on Language Modeling (COLM) Workshop on LLM for Scientific Discovery},year={2025},}
preprint
Optimized Semantic Steering via Sparse Autoencoder Adapters
We study the problem of steering LLMs, where the goal is to intervene on hidden layer activations in order to improve specific behavioral properties. We are particularly interested in settings where one can optimize over a large candidate set of steering interventions to find valuable, task-related perturbations. We take the approach of using sparse autoencoder adapters coupled with natural language feature descriptions to identify disentangled latent dimensions, after which we select and optimize a subset of the the latent codes that are relevant for the target downstream behavior. A key benefit of our approach is the ability to leverage LLM priors to guide feature selection without manual inspection of relevant features. We empirically demonstrate that our approach can generate steered LLM variants that outperform unsteered LLMs on natural language tasks. Furthermore, the selected steering variants can exhibit cross-lingual transfer, providing task improvements on other languages, unseen during selection or tuning. Our method enables tractable, optimized LLM steering by decomposing the problem into discrete feature selection and continuous optimization. This work demonstrates how tools from mechanistic interpretability can be leveraged to improve model capabilities.
@article{OptimizedSemanticSteering,title={Optimized Semantic Steering via Sparse Autoencoder
Adapters},author={Farhang, A. R. and Erickson, A. L. and Yue, Y.},journal={preprint},year={2025},}
2024
IEEE CoG
Humanlike Behavior in a Third-Person Shooter with Imitation Learning
@article{farhang2024Humanlike,title={Humanlike Behavior in a Third-Person Shooter with Imitation Learning},author={Farhang, A. R. and Mulcahy, B. and Holden, D. and Matthews, I. and Yue, Y.},journal={IEEE Conference on Games (CoG)},year={2024},doi={10.1109/CoG60054.2024.10645651},}
2022
ICML
Investigating Generalization by Controlling Normalized Margin
@article{farhang2022margin,title={Investigating Generalization by Controlling Normalized Margin},author={Farhang, A. R. and Bernstein, J. and Tirumala, K. and Liu, Y. and Yue, Y.},journal={International Conference on Machine Learning (ICML)},year={2022},}
@article{bpm,title={Kernel Interpolation as a {B}ayes Point Machine},author={Bernstein, J. and Farhang, A. R. and Yue, Y.},journal={arXiv},year={2022},}
2019
Neuron
Songbird Ventral Pallidum Sends Diverse Performance Error Signals to Dopaminergic Midbrain
@article{CHEN2019266,title={Songbird Ventral Pallidum Sends Diverse Performance Error Signals to Dopaminergic Midbrain},journal={Neuron},volume={103},number={2},pages={266-276.e4},year={2019},issn={0896-6273},doi={https://doi.org/10.1016/j.neuron.2019.04.038},author={Chen, R. and Puzerey, P. A. and Roeser, A. C. and Riccelli, T. E. and Podury, A. and Maher, K. and Farhang, A. R. and Goldberg, J. H.},}
2016
Science
Dopamine Neurons Encode Performance Error in Singing Birds
@article{gadagkar_puzerey_chen_baird-daniel_farhang_goldberg_2016,title={Dopamine Neurons Encode Performance Error in Singing Birds},volume={354},doi={10.1126/science.aah6837},number={6317},journal={Science},author={Gadagkar, V. and Puzerey, P. A. and Chen, R. and Baird-Daniel, E. and Farhang, A. R. and Goldberg, J. H.},year={2016},pages={1278–1282},}