Search for: All records

Language models are aligned to emulate the collective voice of many, resulting in outputs that align with no one in particular. Steering LLMs away from generic output is possible through supervised finetuning or RLHF, but requires prohibitively large datasets for new ad-hoc tasks. We argue that it is instead possible to align an LLM to a specific setting by leveraging a very small number (< 10) of demonstrations as feedback. Our method, Demonstration ITerated Task Optimization (DITTO), directly aligns language model outputs to a user's demonstrated behaviors. Derived using ideas from online imitation learning, DITTO cheaply generates online comparison data by treating users' demonstrations as preferred over output from the LLM and its intermediate checkpoints. Concretely, DITTO operates by having an LLM generate examples that are presumed to be inferior to expert demonstrations. The method iteratively constructs pairwise preference relationships between these LLM-generated samples and expert demonstrations, potentially including comparisons between different training checkpoints. These constructed preference pairs are then used to train the model using a preference optimization algorithm (e.g. DPO). We evaluate DITTO's ability to learn fine-grained style and task alignment across domains such as news articles, emails, and blog posts. Additionally, we conduct a user study soliciting a range of demonstrations from participants (N = 16). Across our benchmarks and user study, we find that win-rates for DITTO outperform few-shot prompting, supervised fine-tuning, and other self-play methods by an avg. of 19% points. By using demonstrations as feedback directly, DITTO offers a novel method for effective customization of LLMs. 

Online mental health support communities, in which volunteer counselors provide accessible mental and emotional health support, have grown in recent years. Despite millions of people using these platforms, the clinical effectiveness of these communities on mental health symptoms remains unknown. Although volunteers receive some training on the therapeutic skills proven effective in face-to-face environments, such as active listening and motivational interviewing, it is unclear how the usage of these skills in an online context affects people's mental health. In our work, we collaborate with one of the largest online peer support platforms and use both natural language processing and machine learning techniques to examine how one-on-one support chats on the platform affect clients' depression and anxiety symptoms. We measure how characteristics of support-providers, such as their experience on the platform and use of therapeutic skills (e.g. affirmation, showing empathy), affect support-seekers' mental health changes. Based on a propensity-score matching analysis to approximate a random-assignment experiment, results shows that online peer support chats improve both depression and anxiety symptoms with a statistically significant but relatively small effect size. Additionally, support providers' techniques such as emphasizing the autonomy of the client lead to better mental health outcomes. However, we also found that the use of some behaviors, such as persuading and providing information, are associated with worsening of mental health symptoms. Our work provides key understanding for mental health care in the online setting and designing training systems for online support providers. 

Abstract Large language models (LLMs) are capable of successfully performing many language processing tasks zero-shot (without training data). If zero-shot LLMs can also reliably classify and explain social phenomena like persuasiveness and political ideology, then LLMs could augment the computational social science (CSS) pipeline in important ways. This work provides a road map for using LLMs as CSS tools. Towards this end, we contribute a set of prompting best practices and an extensive evaluation pipeline to measure the zero-shot performance of 13 language models on 25 representative English CSS benchmarks. On taxonomic labeling tasks (classification), LLMs fail to outperform the best fine-tuned models but still achieve fair levels of agreement with humans. On free-form coding tasks (generation), LLMs produce explanations that often exceed the quality of crowdworkers’ gold references. We conclude that the performance of today’s LLMs can augment the CSS research pipeline in two ways: (1) serving as zero-shot data annotators on human annotation teams, and (2) bootstrapping challenging creative generation tasks (e.g., explaining the underlying attributes of a text). In summary, LLMs are posed to meaningfully participate in social science analysis in partnership with humans. 

Recent abstractive conversation summarization systems generally rely on large-scale datasets with annotated summaries. However, collecting and annotating these conversations can be a time-consuming and labor-intensive task. To address this issue, in this work, we present a sub-structure level compositional data augmentation method, COMPO, for generating diverse and high-quality pairs of conversations and summaries. Specifically, COMPO first extracts conversation structures like topic splits and action triples as basic units. Then we organize these semantically meaningful conversation snippets compositionally to create new training instances. Additionally, we explore noise-tolerant settings in both self-training and joint-training paradigms to make the most of these augmented samples. Our experiments on benchmark datasets, SAMSum and DialogSum, show that COMPO substantially outperforms prior baseline methods by achieving a nearly 10% increase of ROUGE scores with limited data.