Preprint

In InstructionalFingerprint, we present a pilot study on LLM fingerprinting as a form of very lightweight instruction tuning. Model publisher specifies a confidential private key and implants it as an instruction backdoor that causes the LLM to generate specific text when the key is present. Results on 11 popularly-used LLMs showed that this approach is lightweight and does not affect the normal behavior of the model. It also prevents publisher overclaim, maintains robustness against fingerprint guessing and parameter-efficient training, and supports multi-stage fingerprinting akin to MIT License.

In BMBI, we propose to mitigate bias exhibited in QA models by observing the query instance’s influence on another instance, enabling bias mitigation with extremely low resources. With our method, bias levels in multiple bias categories can be reduced without using category-specific instance-level annotation.

In STAR, we propose to synthesize training data by structure-to-text generation using Large Language Models and we show that the generated data is even more effective than human-curated data instances to boost the low-resource event extraction performance.
In the new study about LLM’s backdoor attack, we demonstrate that an attacker can inject backdoors by issuing very few malicious instructions and control model behavior through data poisoning.

We propose to formulate biomedical relation extraction as the NLI task. We design a new evidence retrieval technique for cross-document relation extraction.

In DICE, we introduce a data-efficient generative model for clinical event extraction and propose the first benchmark in the domain.

In Summarization as Indirect Supervision for Relation Extraction, we convert the relation extraction task into a summarization formulation.