Today’s capable LLMs are powered by automated data curation pipelines, where large-scale datasets synthesized by LLMs themselves play a crucial role. But the prevalent noise and redundancy contained in synthetic datasets haven’t received enough attention. There’s also a lack of understanding on how various properties of synthetic data, including correctness, diversity and generation sources (e.g., real-world, on-policy or off-policy generators) contribute to the ultimate learning outcome.

• Can we build principled data selection techniques that not only identify individual data points with favorable properties, but also model the interactions inside a subset of data and their joint influence?
• Is it possible that certain human-perceived noise and redundancy in synthetic data are actually beneficial to language models?
• In light of this, how can we further design efficient curation pipelines with better quality and diversity control?

The success of current LLMs is a combined accomplishment of multiple learning paradigms, including unsupervised pre-training, supervised post-training (e.g., instruction tuning and preference learning) and inference-time context optimization (e.g., in-context learning and RAG). But it still remains unclear what the language model essentially learns in these learning stages, and how various forms of data contribute differently to each stage.

• How can we scientifically study their respective working mechanism?
• What are the shared and exclusive characteristics of “good” data for different stages?
• Can we build novel algorithms and objectives functions that can utilize data more efficiently given the specific mechanism of each learning stage?

The following two lines of work are good examples of designing data-efficient learning algorithms and objectives:

• Advance the understanding of in-context learning: In-context Vectors.
• Improve the efficiency of supervision for preference learning: from DPO to KTO, NCA, SimPO, etc.

Attributing model behavior to specific training data is of vital importance to the trustworthy deployment of LLMs. But the efficiency of traditional data attribution techniques have become a major bottleneck when it comes to LLM-scale applications. It’s also questionable whether they can be well transferred across models of different sizes and architectures, and generalizable across different tasks and application domains. I’m thus interested in designing a more accessible and generalizable data attribution framework tailored to LLMs, and applying such a framework to promote trustworthy applications of LLMs in various social and scientific domains.

What cannot be measured, cannot be improved. I aim to improve the evaluation of LLM capabilities with respect to both trustworthiness and efficiency.

• Evaluation of LLMs is prone to error and noise from many sources, including flawed experimental design, dataset bias, and even improper interpretation of results. How well do current evaluation metrics and datasets truly reflect the LLM’s performance on the capabilities they benchmark? And how should we run evaluation experiments and report their results in a way that minimizes statistical noise and maximizes informativeness?
• There’s also a trade-off between trustworthy and efficient evaluation. Can we maintain the robustness of evaluation results while reducing the cost of exhaustive runs, possibly by removing noisy and over-represented samples in evaluation datasets?