Presenting three papers on bias, fairness and safety of Large Language Models at NAACL 2024 in Mexico City, including detecting and mitigating bias in QA models with ground-truth bias labels, fingerprinting LLMs, and a pilot study on injecting backdoors by instruction tuning data poisoning. Click for schedule and location details.

Presenting a demo and a poster at AAAI 2024 in Vancouver, including a demo on information diffusion via community-level information pathways and a poster on improving low-resource information extraction by structure-to-text data generation with Large Language Models.

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.

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.

Oral presentation of the conf paper: Parameter-Efficient Low-Resource Dialogue State Tracking by Prompt Tuning. In the collaboration work with Amazon Alexa AI, we introduce a dialogue state tracking model tuning less than 1% of LM parameters and achieves better low-resource performance with prompt tuning techniques.

Check out course materials for the 2-week summer program “Data Science for ALL” that were just delivered by our joint team from UCI and UCLA!

We introduce a pioneering comprehensive benchmark to evaluate both intrinsic (within LLMs) and extrinsic (on downstream tasks) bias in LLMs for clinical decision tasks. Our experiments across popular and medically adapted LLMs, particularly from the Mistral and LLaMA families, unveil prevalent behaviors with both intrinsic and extrinsic bias. This work underscores the critical need to mitigate clinical bias and sets a new standard for future evaluations of LLMs' clinical bias.

we introduce MIRAI, a novel benchmark designed to systematically evaluate LLM agents as temporal forecasters in the context of international events. Our benchmark features an agentic environment with tools for accessing an extensive database of historical, structured events and textual news articles.

We introduce CliBench, a novel benchmark offering a comprehensive and realistic assessment of LLMs' capabilities in clinical diagnosis. This benchmark not only covers diagnosis from a diverse range of medical cases across various specialties but also incorporates tasks of clinical significance: treatment procedure identification, lab test ordering and medication prescriptions.

We introduce MERA, a clinical diagnosis prediction model that bridges pertaining natural language knowledge with medical practice. We apply hierarchical contrastive learning on a disease candidate ranking list to alleviate the large decision space issue. With concept memorization through fine-tuning, we bridge the natural language clinical knowledge with medical codes.

Existing works have been using decoding-free candidate selection methods to obtain candidate probability from initial output logits over vocabulary. Though these estimation methods are widely used, they are not systematically evaluated, especially on end tasks. We introduce an evaluation of a comprehensive collection of decoding-free candidate selection approaches.

We propose BMBI, an approach to mitigate the bias of multiple-choice QA models. Based on the intuition that a model would lean to be more biased if it learns from a biased example, we measure the bias level of a query instance by observing its influence on another instance. We then use the bias level detected as an optimization objective to form a multi-task learning setting in addition to the original QA task.

Our studies demonstrate that an attacker can inject backdoors by issuing very few malicious instructions among thousands of gathered data and control model behavior through data poisoning. Through such instruction attacks, the attacker can achieve over 90% attack success rate across four commonly used NLP datasets, and cause persistent backdoors that are easily transferred to 15 diverse datasets zero-shot.

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 prevents publisher overclaim, maintains robustness against fingerprint guessing and parameter-efficient training, and supports multi-stage fingerprinting akin to MIT License.

We propose STAR, a structure-to-text data generation method for complicated structure prediction tasks that first generates complicated event structures (Y) and then generates input passages (X), all with Large Language Models. We further reduce errors and improve data quality through self-reflection error identification and self-refinement with iterative revision. We show that the data generated by STAR significantly improves the performance of low-resource event extraction and relation extraction tasks, even surpassing the effectiveness of human-curated data.

We propose a computational model to infer users' susceptibility levels given their activities. Since user's susceptibility is a key indicator for their reposting behavior, we utilize the supervision from the observable sharing behavior to infer the underlying susceptibility tendency. Building upon such large-scale susceptibility labeling, we further conduct a comprehensive analysis of how different social factors relate to susceptibility.

We introduce DICE, a robust and data-efficient generative model for clinical event extraction, which specializes in clinical mention identification, and MACCROBAT-EE, the first clinical event extraction dataset with event argument annotation.

We present NBR, which converts biomedical relation extraction as natural language inference formulation through indirect supervision.

We use soft prompt tokens to learn task properties, incorporate segment information and reiterate the task before predicting value. Our method drastically reduces the number of parameters needed to less than 0.5% of prior works while achieving better low-resource dialogue state tracking performance.

A taxonomy expansion algorithm that seeks to preserve the structure of a taxonomy in a more expressive hyperbolic embedding space and learn to represent concepts and their relations with a Hyperbolic Graph Neural Network.