Confidence Calibration and Rationalization for LLMs via Multi-Agent Deliberation

Uncertainty estimation is a significant issue for current large language models (LLMs) that are generally poorly calibrated and over-confident, especially with reinforcement learning from human feedback (RLHF). Unlike humans, whose decisions and confidences not only stem from intrinsic beliefs but can also be adjusted through daily observations, existing calibration methods for LLMs focus on estimating or eliciting individual confidence without taking full advantage of the "Collective Wisdom": the interaction among multiple LLMs that can collectively improve both accuracy and calibration. In this work, we propose Collaborative Calibration, a post-hoc training-free calibration strategy that leverages the collaborative and expressive capabilities of multiple tool-augmented LLM agents in a simulated group deliberation process. We demonstrate the effectiveness of Collaborative Calibration on generative QA tasks across various domains, showing its potential in harnessing the rationalization of collectively calibrated confidence assessments and improving the reliability of model predictions.

Reinforcement Learning with Dynamic Multi-Reward Weighting for Multi-Style Controllable Generation

Style is an integral component of text that expresses a diverse set of information, including interpersonal dynamics (e.g. formality) and the author's emotions or attitudes (e.g. disgust). Humans often employ multiple styles simultaneously. An open question is how large language models can be explicitly controlled so that they weave together target styles when generating text: for example, to produce text that is both negative and non-toxic. Previous work investigates the controlled generation of a single style, or else controlled generation of a style and other attributes. In this paper, we expand this into controlling multiple styles simultaneously. Specifically, we investigate various formulations of multiple style rewards for a reinforcement learning (RL) approach to controlled multi-style generation. These reward formulations include calibrated outputs from discriminators and dynamic weighting by discriminator gradient magnitudes. We find that dynamic weighting generally outperforms static weighting approaches, and we explore its effectiveness in 2- and 3-style control, even compared to strong baselines like plug-and-play model. All code and data for RL pipelines with multiple style attributes will be publicly available.

Talk Through It: End User Directed Manipulation Learning

Training generalist robot agents is an immensely difficult feat due to the requirement to perform a huge range of tasks in many different environments. We propose selectively training robots based on end-user preferences instead. Given a factory model that lets an end user instruct a robot to perform lower-level actions (e.g. 'Move left'), we show that end users can collect demonstrations using language to train their home model for higher-level tasks specific to their needs (e.g. 'Open the top drawer and put the block inside'). We demonstrate this hierarchical robot learning framework on robot manipulation tasks using RLBench environments. Our method results in a 16% improvement in skill success rates compared to a baseline method. In further experiments, we explore the use of the large vision-language model (VLM), Bard, to automatically break down tasks into sequences of lower-level instructions, aiming to bypass end-user involvement. The VLM is unable to break tasks down to our lowest level, but does achieve good results breaking high-level tasks into mid-level skills. We have a supplemental video and additional results at talk-through-it.github.io.

Shallow Synthesis of Knowledge in GPT-Generated Texts: A Case Study in Automatic Related Work Composition

Numerous AI-assisted scholarly applications have been developed to aid different stages of the research process. We present an analysis of AI-assisted scholarly writing generated with ScholaCite, a tool we built that is designed for organizing literature and composing Related Work sections for academic papers. Our evaluation method focuses on the analysis of citation graphs to assess the structural complexity and inter-connectedness of citations in texts and involves a three-way comparison between (1) original human-written texts, (2) purely GPT-generated texts, and (3) human-AI collaborative texts. We find that GPT-4 can generate reasonable coarse-grained citation groupings to support human users in brainstorming, but fails to perform detailed synthesis of related works without human intervention. We suggest that future writing assistant tools should not be used to draft text independently of the human author.

Chain-of-Instructions: Compositional Instruction Tuning on Large Language Models

Fine-tuning large language models (LLMs) with a collection of large and diverse instructions has improved the model's generalization to different tasks, even for unseen tasks. However, most existing instruction datasets include only single instructions, and they struggle to follow complex instructions composed of multiple subtasks (Wang et al., 2023a). In this work, we propose a novel concept of compositional instructions called chain-of-instructions (CoI), where the output of one instruction becomes an input for the next like a chain. Unlike the conventional practice of solving single instruction tasks, our proposed method encourages a model to solve each subtask step by step until the final answer is reached. CoI-tuning (i.e., fine-tuning with CoI instructions) improves the model's ability to handle instructions composed of multiple subtasks. CoI-tuned models also outperformed baseline models on multilingual summarization, demonstrating the generalizability of CoI models on unseen composite downstream tasks.

Tuning Large Multimodal Models for Videos using Reinforcement Learning from AI Feedback

Recent advancements in large language models have influenced the development of video large multimodal models (VLMMs). The previous approaches for VLMMs involved Supervised Fine-Tuning (SFT) with instruction-tuned datasets, integrating LLM with visual encoders, and adding additional learnable modules. Video and text multimodal alignment remains challenging, primarily due to the deficient volume and quality of multimodal instruction-tune data compared to text-only data. We present a novel alignment strategy that employs multimodal AI system to oversee itself called Reinforcement Learning from AI Feedback (RLAIF), providing self-preference feedback to refine itself and facilitating the alignment of video and text modalities. In specific, we propose context-aware reward modeling by providing detailed video descriptions as context during the generation of preference feedback in order to enrich the understanding of video content. Demonstrating enhanced performance across diverse video benchmarks, our multimodal RLAIF approach, VLM-RLAIF, outperforms existing approaches, including the SFT model. We commit to open-sourcing our code, models, and datasets to foster further research in this area.

II-MMR: Identifying and Improving Multi-modal Multi-hop Reasoning in Visual Question Answering

Visual Question Answering (VQA) often involves diverse reasoning scenarios across Vision and Language (V&L). Most prior VQA studies, however, have merely focused on assessing the model's overall accuracy without evaluating it on different reasoning cases. Furthermore, some recent works observe that conventional Chain-of-Thought (CoT) prompting fails to generate effective reasoning for VQA, especially for complex scenarios requiring multi-hop reasoning. In this paper, we propose II-MMR, a novel idea to identify and improve multi-modal multi-hop reasoning in VQA. In specific, II-MMR takes a VQA question with an image and finds a reasoning path to reach its answer using two novel language promptings: (i) answer prediction-guided CoT prompt, or (ii) knowledge triplet-guided prompt. II-MMR then analyzes this path to identify different reasoning cases in current VQA benchmarks by estimating how many hops and what types (i.e., visual or beyond-visual) of reasoning are required to answer the question. On popular benchmarks including GQA and A-OKVQA, II-MMR observes that most of their VQA questions are easy to answer, simply demanding "single-hop" reasoning, whereas only a few questions require "multi-hop" reasoning. Moreover, while the recent V&L model struggles with such complex multi-hop reasoning questions even using the traditional CoT method, II-MMR shows its effectiveness across all reasoning cases in both zero-shot and fine-tuning settings.

Threads of Subtlety: Detecting Machine-Generated Texts Through Discourse Motifs

With the advent of large language models (LLM), the line between human-crafted and machine-generated texts has become increasingly blurred. This paper delves into the inquiry of identifying discernible and unique linguistic properties in texts that were written by humans, particularly uncovering the underlying discourse structures of texts beyond their surface structures. Introducing a novel methodology, we leverage hierarchical parse trees and recursive hypergraphs to unveil distinctive discourse patterns in texts produced by both LLMs and humans. Empirical findings demonstrate that, although both LLMs and humans generate distinct discourse patterns influenced by specific domains, human-written texts exhibit more structural variability, reflecting the nuanced nature of human writing in different domains. Notably, incorporating hierarchical discourse features enhances binary classifiers' overall performance in distinguishing between human-written and machine-generated texts, even on out-of-distribution and paraphrased samples. This underscores the significance of incorporating hierarchical discourse features in the analysis of text patterns. The code and dataset will be available at [TBA].

Under the Surface: Tracking the Artifactuality of LLM-Generated Data

This work delves into the expanding role of large language models (LLMs) in generating artificial data. LLMs are increasingly employed to create a variety of outputs, including annotations, preferences, instruction prompts, simulated dialogues, and free text. As these forms of LLM-generated data often intersect in their application, they exert mutual influence on each other and raise significant concerns about the quality and diversity of the artificial data incorporated into training cycles, leading to an artificial data ecosystem. To the best of our knowledge, this is the first study to aggregate various types of LLM-generated text data, from more tightly constrained data like "task labels" to more lightly constrained "free-form text". We then stress test the quality and implications of LLM-generated artificial data, comparing it with human data across various existing benchmarks. Despite artificial data's capability to match human performance, this paper reveals significant hidden disparities, especially in complex tasks where LLMs often miss the nuanced understanding of intrinsic human-generated content. This study critically examines diverse LLM-generated data and emphasizes the need for ethical practices in data creation and when using LLMs. It highlights the LLMs' shortcomings in replicating human traits and behaviors, underscoring the importance of addressing biases and artifacts produced in LLM-generated content for future research and development. All data and code are available on our project page.

SelectLLM: Can LLMs Select Important Instructions to Annotate?

Training large language models (LLMs) with a large and diverse instruction dataset aligns the models to comprehend and follow human instructions. Recent works have shown that using a small set of high-quality instructions can outperform using large yet more noisy ones. Because instructions are unlabeled and their responses are natural text, traditional active learning schemes with the model's confidence cannot be directly applied to the selection of unlabeled instructions. In this work, we propose a novel method for instruction selection, called SelectLLM, that leverages LLMs for the selection of high-quality instructions. Our high-level idea is to use LLMs to estimate the usefulness and impactfulness of each instruction without the corresponding labels (i.e., responses), via prompting. SelectLLM involves two steps: dividing the unlabelled instructions using a clustering algorithm (e.g., CoreSet) to multiple clusters, and then prompting LLMs to choose high-quality instructions within each cluster. SelectLLM showed comparable or slightly better performance on the popular instruction benchmarks, compared to the recent state-of-the-art selection methods. All code and data are publicly available (https://github.com/minnesotanlp/select-llm).