Abraham Yeung

I'm a junior at Stanford doing a B.S. in Mathematics and a coterm M.S. in Computer Science. I work on post-training and AI safety. I'm a research assistant in the Scaling Intelligence Lab at SAIL, and next year I'm joining Redwood Research part-time as an ML research intern, after a summer on the Data Platform team at Databricks.

I grew up in Hong Kong and went to Eton College. I sing baritone with the Mendicants, Stanford's oldest a cappella group. I speak English, Cantonese, and Mandarin natively, plus some German and beginner Japanese.

Email GitHub LinkedIn CV
Abraham Yeung

Research

What Happens to a Monitor's Accuracy When You Train Against It

Abraham Yeung, Anagha Ramaswamy
NeurIPS 2026, InterpScience Workshop · code

Using a correctness probe as an RL reward cuts true accuracy by 31 points, and the probe takes about 40 steps to notice.

How Small Can We Go? Calibrating Activation-Cache Compression for SAE Training

Abraham Yeung
NeurIPS 2026, InterpScience Workshop

Naive int4 activation caches lose more SAE features than reseeding does; standardizing channels first fixes it at the same size.

Coding Agents for Coding Theory

Abraham Yeung
NeurIPS 2026, MATH-AI Workshop · bounds tables

An LLM coding agent improved 13 published lower bounds for quaternary edit-metric codes, one of which had stood since 2012.

Projects

Edit-metric code tablescoding theory · maintained

Current bounds on the size of binary, ternary, and quaternary codes under the edit (Levenshtein) distance, continuing Houghten's tables from Brock University. Every improved entry comes with a code file and a short script to check it. Tables

Research Frontier Mineragents · in progress

A resumable agent that mines recent ML papers, walks the citation graph to find proposed extensions nobody has attempted, and ranks the open problems. SQLite plus vector search, content-addressed caching, best-first frontier.

Prediction Markets Agentmarkets · top 5 of 100+

Agent that evaluates live Polymarket prices and flags mispriced contracts from incoming information and microstructure signals. Finalist at the NVIDIA, Vercel, and Brex hackathon.

Multi-Agent Causal Modelingagents · Bridgewater AI Hackathon

Selected as 1 of 24. A multi-agent system that maps causal graphs for policy-impact scenarios by decomposing macro questions into testable sub-claims.

Maestroai / music · TreeHacks runner-up

Multimodal music coach that analyzes instrument audio and video with NVIDIA vision and audio models. Devpost

gigabpesystems · Rust · PyPI

BPE tokenizer trainer in Rust with Python bindings. A 32k vocabulary on 12.9 GB of FineWeb in 38 s against 257 s for HuggingFace tokenizers, with every merge byte-identical and a fraction of the memory. GitHub

Transformer from scratchsystems · CS 336

A language model built end to end in PyTorch: tokenizer, training loop, FlashAttention as Triton kernels, distributed data parallel with optimizer sharding, profiling, scaling laws, and SFT plus GRPO post-training.

AQI Forecastingtime series · CS 229

LSTM, GNN, and CNN models over 545k+ observations for spatiotemporal air quality forecasting, with robustness testing under distribution shift.

Experience

Honors

Ask a model about my work

A small open-weights model with a short bio in its context. Ask it about my work. It can be wrong, so trust the sections above over it.