Valerio Pepe

Valerio Pepe

Computer Science · Cognitive Science · Robotics · AI Safety

I am an undergraduate student at Harvard University in the class of 2026, studying towards a concurrent BA/MS in Computer Science and Mind, Brain, and Behavior, and advised by Prof. Stuart Shieber.

I'm broadly passionate about the interface between language and thought, with a focus on the inductive biases that favor language learning and reasoning in both humans and machines.

I'm also interested in AI safety from both a technical (interpretability, control) and a political perspective, and work with the AI Safety Student Team at Harvard towards these goals.

I am currently affiliated with the Computational Cognitive Science Group at MIT Brain & Cognitive Sciences and the Language and Intelligence Group at the MIT Computer Science and AI Lab, where I am fortunate to be supervised by Gabriel Grand, Prof. Joshua Tenenbaum, and Prof. Jacob Andreas.

publications

Shoot First, Ask Questions Later? Building Rational Agents that Explore and Act Like People
Valerio Pepe*, Gabriel Grand*, Jacob Andreas, Joshua Tenenbaum
paper
Under Review
AnyTask: An Automated Task and Data Generation Framework for Advancing Sim-to-Real Policy Learning
Ran Gong*, Xiaohan Zhang*, Jigarkumar Patel*, ..., Valerio Pepe, ..., Karl Schmeckpeper
paper
Under Review
A Large-scale Investigation of Pronoun Interpretation Biases in LLMs
Valerio Pepe, Joshua Hartshorne
paper
AMLaP 2025
Loose LIPS Sink Ships: Asking Questions in Battleship with Language-Informed Program Sampling
Gabriel Grand, Valerio Pepe, Jacob Andreas, Joshua Tenenbaum
paper
System 2 Reasoning @ NeurIPS 2024 · CogSci 2024
SeqVerify: An accessible analysis tool for cell line genomic integrity, contamination, and gene editing outcomes
Merrick Pierson Smela*, Valerio Pepe*, Steven Lubbe, Evangelos Kiskinis, George Church
paper
Stem Cell Reports 2024

media & writing

As millions adopt Grok to fact-check, misinformation abounds
Valerio Pepe and Nilesh Christopher, Al Jazeera
July 2025
Read Article
Emergent Misalignment on a Budget
June 2025
Read Post

selected projects

🗺️
Linguapedia
Developed a website to automatically translate Wikipedia articles from multiple languages and collate them into a single article in any other language. Winner of the 2025 Anthropic × Harvard Hackathon (Grand Prize).
LLM Internationalization Knowledge Democratization Hackathon Winner
Try Linguapedia
🏎️
AdaSPEED
Adaptive self-speculative decoding for Llama 3 1B and 8B, resulting in up to 40% inference speedup. Developed a method to dynamically select the number of tokens to be generated by a speculative decoding system.
Speculative Decoding BranchyNet Efficient ML
View PDF
🔁
Controllable Benchmarks for LLM Unlearning
Developed methods and benchmarks for controlling the similarity of the retain and unlearn set in machine unlearning, suggesting that the focus of further unlearning research should be on better datasets, not only better algorithms.
Machine Unlearning NLP Benchmarking
View PDF
⚖️
Individual Fairness for Image Classification
Developed a method to use individual fairness to impose inference-time constraints on image-classifying neural networks, guaranteeing better adversarial robustness with no retraining, finetuning, or additional inference costs.
Algorithmic Fairness Adversarial Robustness FGSM
View PDF
🛡️
Adversarial Robustness in Self-Explaining Neural Networks
Explored which types of noise fool Self-Explaining Neural Networks, and what types of defenses could be implemented in order to mitigate their weaknesses.
Explainability Adversarial Robustness SENNs
View PDF
🗣️
Historical Linguistics-Informed Distinctive Feature Theory
Trained a sparse autoencoder on a graph representation of historical sound shifts in order to derive a historically-grounded distinctive feature set for phonology.
Computational Linguistics Interpretability Graph Representation Learning
View PDF