🧑‍🔬 Joey Tianyi Zhou

Joey Tianyi Zhou studies how neural language models behave under pressure. The two papers here span 2019 to 2025. One line of work asks how easily text classifiers can be misled by small changes to their input that leave its meaning intact. The other asks whether reinforcement learning with automatically checkable rewards can bring out step-by-step reasoning in large models. Both rely on tightly controlled test material: crafted adversarial sentences in one case, synthetic puzzles with verifiable answers in the other. The papers fall under Language models & pretraining and Reasoning.

The most cited paper here is Is BERT Really Robust? (2019). It presented TextFooler, a black-box attack that needs nothing beyond a model's output scores. The method finds the words whose removal most weakens the correct prediction and replaces them with synonyms taken from counter-fitted embeddings. A substitution is kept only when the part of speech matches and the altered sentence stays close in meaning to the original. The attack was applied to both text classification and entailment, including pre-trained BERT.

The newer work moves from attacking models to training them to reason. Logic-RL (2025) asks whether the rule-based reinforcement learning approach behind DeepSeek-R1, whose training code and data were never released, can be reproduced with a 7B model. Instead of math problems, whose difficulty varies without control, the study uses procedurally generated Knights and Knaves puzzles. These can be tuned to a chosen difficulty, and each has one answer a script can verify. Training combines a modified REINFORCE++ algorithm, a system prompt that keeps thinking and answering apart, and a strict reward for format.

Zhou's bylines list the Agency for Science, Technology and Research and its Institute of High Performance Computing in Singapore from 2015 onward, with earlier papers at Nanyang Technological University and 2026 papers at Singapore Management University. Co-authors on the TextFooler study include Di Jin and Peter Szolovits.

Joey Tianyi Zhou has 2 papers in our library, in 2019–2025, most of them on Language models & pretraining, Reasoning; often with Di Jin, Peter Szolovits.