Yanda Chen

I am a Member of Technical Staff (Research Scientist) on the Code RL team at Anthropic, where I train Claude models toward stronger coding and agentic capabilities. Previously, I was a Research Scientist on the Alignment team at Anthropic, working on safety-driven pre-training, chain-of-thought explainability, and AI control.

Before Anthropic, I completed my PhD in Computer Science at Columbia University in summer 2024, co-advised by Prof. Kathy McKeown, Prof. He He, and Prof. Zhou Yu. During my PhD, I worked on explainability and few-shot learning. I also received my bachelor's degree in Computer Science at Columbia University in April 2021.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

We remove harmful CBRN-related content from pretraining data using a classifier and pretrain models from scratch on the filtered dataset. This reduces performance on a harmful-capabilities evaluation by 33% relative to a random baseline, while causing no significant drop on standard benchmarks including MMLU, Code, and Prose.

We test whether reasoning models faithfully reveal their reasoning by inserting hints into evaluation questions and checking whether models acknowledge using them. Models mention the hints only 25% (Claude 3.7 Sonnet) and 39% (DeepSeek R1) of the time, and outcome-based RL improves faithfulness only up to a plateau, suggesting that chain-of-thought monitoring alone is insufficient to reliably catch misaligned behavior.

We find that the in-context learning ability of language models emerges from parallel structures in the pre-training data: pairs of phrases following similar templates within the same context window. Removing these structures reduces ICL accuracy by 51% (versus 2% from random ablation), and the drop persists even after excluding common patterns such as n-gram repetitions and long-range dependencies.

We propose explanation-consistency finetuning (EC-finetuning), which adapts LLMs to produce more consistent natural-language explanations across related examples by finetuning them on synthetic data constructed to contain consistent explanations. EC-finetuning improves explanation consistency by 10.0% on four finetuning datasets and by 4.5% on seven out-of-distribution datasets.

We propose evaluating the counterfactual simulatability of natural-language explanations: whether an explanation lets humans precisely infer a model's outputs on diverse counterfactuals of the explained input. Using two metrics, precision and generality, we find that (i) LLM explanations have low precision, and (ii) precision does not correlate with plausibility.

We find that label bias obscures true ICL sensitivity, and that ICL sensitivity is strongly and negatively correlated with accuracy. Motivated by this, we propose SenSel, a few-shot selective-prediction method based on ICL sensitivity.

We propose in-context learning distillation, which transfers in-context learning (ICL) ability from large language models to small ones by combining in-context tuning with teacher-student distillation. Experiments on LAMA and CrossFit show that it improves the ICL ability of small language models.

We propose in-context tuning, a novel few-shot meta-learning method in which training examples serve as prefix in-context demonstrations for task adaptation. In-context tuning outperforms MAML in accuracy and eliminates several well-known oversensitivity artifacts of few-shot language-model prompting.

We propose a data-augmentation strategy and a rationale-training strategy for cross-lingual sentence selection in low-resource settings, where no labeled relevance judgments are available for training. Our methods achieve state-of-the-art results on three language pairs.

We propose two synthetic training strategies: targeted synthetic pre-training (selecting useful synthetic examples to target the weaknesses of existing models) and synthetic knowledge distillation. Combined, the two techniques yield QA models that are simultaneously smaller, faster, and more accurate.

We propose a framework to systematically detect and reduce the language bias of deep-learning models in the high-stakes context of gang intervention.

Website design from Jon Barron