Mohammad Taufeeque

Timeline

1. 2021

   A Reconnaissance Blind Chess win

   At IIT Bombay, he earns a B.Tech in computer science, and his bachelor's thesis with Prof. Shivaram Kalyanakrishnan wins the NeurIPS 2021 Reconnaissance Blind Chess competition. His internships include Microsoft Research, where he works on adapting deployed neural text classifiers to out-of-distribution data with Prof. Sunita Sarawagi and Dr. Sriram Rajamani, and TU Braunschweig, with Prof. Thomas Deserno.

2. November 2022

   The imitation library

   In November 2022, he coauthors imitation: Clean Imitation Learning Implementations with Adam Gleave, Juan Rocamonde, Stuart Russell and others. The open-source PyTorch library implements three inverse reinforcement learning algorithms, three imitation learning algorithms and a preference comparison algorithm.

3. October 2023

   Codebook features for interpretability

   In October 2023, he coauthors Codebook Features: Sparse and Discrete Interpretability for Neural Networks with Alex Tamkin and Noah D. Goodman. The method finetunes neural networks with vector quantization bottlenecks at each layer, so that hidden features become sums of a few discrete codes drawn from a larger codebook.

4. December 2023

   Red-teaming GPT-4 APIs

   In December 2023, he coauthors Exploiting Novel GPT-4 APIs with Kellin Pelrine, Michał Zając, Euan McLean and Adam Gleave. The team red-teams GPT-4's fine-tuning, function calling and knowledge retrieval, and finds that as few as 15 harmful examples or 100 benign examples can remove core safeguards from GPT-4.

5. 2024–2025

   Planning in a Sokoban RNN

   He is first author of Planning in a recurrent neural network that plays Sokoban, submitted in July 2024 and revised in May 2025, with Philip Quirke, Maximilian Li, Chris Cundy, Aaron David Tucker, Adam Gleave and Adrià Garriga-Alonso. The paper finds that the network has a causal plan representation that predicts its future actions about 50 steps in advance and that it paces in cycles to give itself extra computation at the start of a level.

6. 2025–2026

   Path channels in Sokoban

   As first author, he publishes Path Channels and Plan Extension Kernels at ICLR 2026, after a Spotlight at the Mechanistic Interpretability Workshop at NeurIPS 2025. With Aaron David Tucker, Adam Gleave and Adrià Garriga-Alonso, he partially reverse-engineers a Sokoban-playing RNN and finds that it stores future moves as activations in path channels.

7. 2026

   An ICML honorable mention

   In 2026, he is first author of The Obfuscation Atlas: Mapping Where Honesty Emerges in RLVR with Deception Probes, written with Stefan Heimersheim, Adam Gleave and Chris Cundy. The paper introduces a taxonomy of outcomes when training against a deception detector, and shows that sufficiently high KL regularization together with a detector penalty can yield honest policies. ICML 2026 names it an Outstanding Paper Honorable Mention and presents it as an oral.

8. As of October 2026

   Research engineer at FAR.AI

   He is a research engineer at FAR.AI, listed as Member of Technical Staff. His current research interests are scalable interpretability, post-training interventions that robustly preserve values like honesty across contexts and personas, and the introspective awareness of language models.