Kevin Zhu

I founded Algoverse, where we mentor ambitious students into published ML researchers, and cofounded the Algoverse AI Safety Fellowship, funded by Open Philanthropy.

I think about AI safety, mechanistic interpretability, and how to build institutions that produce great research.

previously

Citadel — quantitative research
Palantir — software engineering
UC Berkeley — CS · 4x head TA, CS 170 · research with Ion Stoica

Always happy to chat — say hi!

writing

the bitter lesson of self-improvement
why scaling beats scaffolding in your own life too
A gentle introduction to algorithms
(former) official course notes for Berkeley CS 170
why I always have songs in my head
the most underdiagnosed sleep disorder: UARS
lucid dreaming is about to become insane alpha
train the world model in your brain before AI makes it the bottleneck
semantic self-consistency
LLM reasoning paper, selected by OpenAI for PaperBench

recent algoverse research

check out some of our recent work!

Figure 1 from Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation

NeurIPS 2026 Main

Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation

Figure 1 from Diagnostic Foundation for Evaluating LLMs' Research Integrity as Co-Scientists

NeurIPS 2026 Evals & Benchmarks

Diagnostic Foundation for Evaluating LLMs' Research Integrity as Co-Scientists

Figure 1 from Automated Stateful Specialization for Adaptive Agent Systems

Automated Stateful Specialization for Adaptive Agent Systems

Figure 1 from Emergent Misalignment via In-Context Learning

Emergent Misalignment via In-Context Learning

Figure 1 from Commitment to Cooperation with Self-Negotiated Contracts

Commitment to Cooperation with Self-Negotiated Contracts

Figure 1 from On Mitigation of Subliminal Learning in Large Language Models

EMNLP 2026 Findings

On Mitigation of Subliminal Learning in Large Language Models

Figure 1 from ProMoral-Bench: Evaluating Prompting Strategies for Moral Reasoning and Safety in LLMs

AACL 2026 Main · Oral

ProMoral-Bench: Evaluating Prompting Strategies for Moral Reasoning and Safety in LLMs

Figure 1 from Mitigating Over-Personalization in LLMs via Structured Memory

AACL 2026 Findings

Mitigating Over-Personalization in LLMs via Structured Memory

Figure 1 from MiSCHiEF: Minimal-Pairs of Safety and Culture for Fine-Grained Image-Caption Alignment

MiSCHiEF: Minimal-Pairs of Safety and Culture for Fine-Grained Image-Caption Alignment

Figure from MURMUR: Cross-Lingual and Multimodal Retrieval-Augmented Reasoning for Open Question Answering in Tamil and Yoruba

MURMUR: Cross-Lingual and Multimodal Retrieval-Augmented Reasoning for Open Question Answering in Tamil and Yoruba

Figure from Copying Before Suppression: What Drives a Below-Chance Dip During Language Model Training?

EMNLP 2026 Findings

Copying Before Suppression: What Drives a Below-Chance Dip During Language Model Training?

Figure 1 from NovelHopQA: Diagnosing Multi-Hop Reasoning Failures in Long Narrative Contexts

EMNLP 2025 Main

NovelHopQA: Diagnosing Multi-Hop Reasoning Failures in Long Narrative Contexts

more

EMNLP 2026 · IndustryCAST: Contrastive Adjudication via Selective Triggering for Efficient LLM Reasoning paper coming soon
EACL 2026 · IndustryAssertion-Conditioned Compliance: A Provenance-Aware Vulnerability in Multi-Turn Tool-Calling Agents
IASEAI 2026 · MainA Decision-Theoretic Approach for Managing Misalignment
EMNLP 2025 · FindingsEnDive: A Cross-Dialect Benchmark for Fairness and Performance in Large Language Models