Yizhong Wang

Joel Jang, Seungone Kim, Bill Yuchen Lin et al. · allen-ai, cmu

While Reinforcement Learning from Human Feedback (RLHF) aligns Large Language Models (LLMs) with general, aggregate human preferences, it is suboptimal for learning diverse, individual perspectives. In this work, we study Reinforcement Learning from Personalized Human Feedback (RLPHF) problem, wherein LLMs are aligned to multiple (sometimes conflicting) preferences by modeling alignment as a Multi-Objective Reinforcement Learning (MORL) problem. Compared to strong single-objective baselines, we show that we can achieve personalized alignment by decomposing preferences into multiple dimensions. These dimensions are defined based on personalizations that are declared as desirable by the user. In this work, we show that they can be efficiently trained independently in a distributed manner and combined effectively post-hoc through parameter merging. The code is available at https://github.com/joeljang/RLPHF.

Hamish Ivison, Yizhong Wang, Valentina Pyatkin et al. · allen-ai, uw

Since the release of TÜLU [Wang et al., 2023b], open resources for instruction tuning have developed quickly, from better base models to new finetuning techniques. We test and incorporate a number of these advances into TÜLU, resulting in TÜLU 2, a suite of improved TÜLU models for advancing the understanding and best practices of adapting pretrained language models to downstream tasks and user preferences. Concretely, we release: (1) TÜLU-V2-mix, an improved collection of high-quality instruction datasets; (2) TÜLU 2, LLAMA-2 models finetuned on the V2 mixture; (3) TÜLU 2+DPO, TÜLU 2 models trained with direct preference optimization (DPO), including the largest DPO-trained model to date (TÜLU 2+DPO 70B); (4) CODE TÜLU 2, CODE LLAMA models finetuned on our V2 mix that outperform CODE LLAMA and its instruction-tuned variant, CODE LLAMA-Instruct. Our evaluation from multiple perspectives shows that the TÜLU 2 suite achieves state-of-the-art performance among open models and matches or exceeds the performance of GPT-3.5-turbo-0301 on several benchmarks. We release all the checkpoints, data, training and evaluation code to facilitate future open efforts on adapting large language models.

Abhika Mishra, Akari Asai, Vidhisha Balachandran et al. · cmu, uw

Large language models (LMs) are prone to generate factual errors, which are often called hallucinations. In this paper, we introduce a comprehensive taxonomy of hallucinations and argue that hallucinations manifest in diverse forms, each requiring varying degrees of careful assessments to verify factuality. We propose a novel task of automatic fine-grained hallucination detection and construct a new evaluation benchmark, FavaBench, that includes about one thousand fine-grained human judgments on three LM outputs across various domains. Our analysis reveals that ChatGPT and Llama2-Chat (70B, 7B) exhibit diverse types of hallucinations in the majority of their outputs in information-seeking scenarios. We train FAVA, a retrieval-augmented LM by carefully creating synthetic data to detect and correct fine-grained hallucinations. On our benchmark, our automatic and human evaluations show that FAVA significantly outperforms ChatGPT and GPT-4 on fine-grained hallucination detection, and edits suggested by FAVA improve the factuality of LM-generated text.

Wenhao Wu, Yizhong Wang, Guangxuan Xiao et al. · uw

Despite the recent progress in long-context language models, it remains elusive how transformer-based models exhibit the capability to retrieve relevant information from arbitrary locations within the long context. This paper aims to address this question. Our systematic investigation across a wide spectrum of models reveals that a special type of attention heads are largely responsible for retrieving information, which we dub retrieval heads. We identify intriguing properties of retrieval heads:(1) universal: all the explored models with long-context capability have a set of retrieval heads; (2) sparse: only a small portion (less than 5\%) of the attention heads are retrieval. (3) intrinsic: retrieval heads already exist in models pretrained with short context. When extending the context length by continual pretraining, it is still the same set of heads that perform information retrieval. (4) dynamically activated: take Llama-2 7B for example, 12 retrieval heads always attend to the required information no matter how the context is changed. The rest of the retrieval heads are activated in different contexts. (5) causal: completely pruning retrieval heads leads to failure in retrieving relevant information and results in hallucination, while pruning random non-retrieval heads does not affect the model's retrieval ability. We further show that retrieval heads strongly influence chain-of-thought (CoT) reasoning, where the model needs to frequently refer back the question and previously-generated context. Conversely, tasks where the model directly generates the answer using its intrinsic knowledge are less impacted by masking out retrieval heads. These observations collectively explain which internal part of the model seeks information from the input tokens. We believe our insights will foster future research on reducing hallucination, improving reasoning, and compressing the KV cache.

Hamish Ivison, Yizhong Wang, Jiacheng Liu et al.

Learning from preference feedback has emerged as an essential step for improving the generation quality and performance of modern language models (LMs). Despite its widespread use, the way preference-based learning is applied varies wildly, with differing data, learning algorithms, and evaluations used, making disentangling the impact of each aspect difficult. In this work, we identify four core aspects of preference-based learning: preference data, learning algorithm, reward model, and policy training prompts, systematically investigate the impact of these components on downstream model performance, and suggest a recipe for strong learning for preference feedback. Our findings indicate that all aspects are important for performance, with better preference data leading to the largest improvements, followed by the choice of learning algorithm, the use of improved reward models, and finally the use of additional unlabeled prompts for policy training. Notably, PPO outperforms DPO by up to 2.5% in math and 1.2% in general domains. High-quality preference data leads to improvements of up to 8% in instruction following and truthfulness. Despite significant gains of up to 5% in mathematical evaluation when scaling up reward models, we surprisingly observe marginal improvements in other categories. We publicly release the code used for training (https://github.com/hamishivi/EasyLM) and evaluating (https://github.com/allenai/open-instruct) our models, along with the models and datasets themselves (https://huggingface.co/collections/allenai/tulu-v25-suite-66676520fd578080e126f618).

Seungone Kim, Juyoung Suk, Xiang Yue et al. · cmu

Given the increasing use of synthetic data in language model (LM) post-training, an LM's ability to generate high-quality data has become nearly as crucial as its ability to solve problems directly. While prior works have focused on developing effective data generation methods, they lack systematic comparison of different LMs as data generators in a unified setting. To address this gap, we propose AgoraBench, a benchmark that provides standardized settings and metrics to evaluate LMs' data generation abilities. Through synthesizing 1.26 million training instances using 6 LMs and training 99 student models, we uncover key insights about LMs' data generation capabilities. First, we observe that LMs exhibit distinct strengths. For instance, GPT-4o excels at generating new problems, while Claude-3.5-Sonnet performs better at enhancing existing ones. Furthermore, our analysis reveals that an LM's data generation ability doesn't necessarily correlate with its problem-solving ability. Instead, multiple intrinsic features of data quality-including response quality, perplexity, and instruction difficulty-collectively serve as better indicators. Finally, we demonstrate that strategic choices in output format and cost-conscious model selection significantly impact data generation effectiveness.

Lester James V. Miranda, Yizhong Wang, Yanai Elazar et al. · allen-ai, cambridge

Learning from human feedback has enabled the alignment of language models (LMs) with human preferences. However, collecting human preferences is expensive and time-consuming, with highly variable annotation quality. An appealing alternative is to distill preferences from LMs as a source of synthetic annotations, offering a cost-effective and scalable alternative, albeit susceptible to other biases and errors. In this work, we introduce HyPER, a Hybrid Preference routER that defers an annotation to either humans or LMs, achieving better annotation quality while reducing the cost of human-only annotation. We formulate this as an optimization problem: given a preference dataset and an evaluation metric, we (1) train a performance prediction model (PPM) to predict a reward model's (RM) performance on an arbitrary combination of human and LM annotations and (2) employ a routing strategy that selects a combination that maximizes the predicted performance. We train the PPM on MultiPref, a new preference dataset with 10k instances paired with humans and LM labels. We show that the selected hybrid mixture of synthetic and direct human preferences using HyPER achieves better RM performance compared to using either one exclusively by 7-13% on RewardBench and generalizes across unseen preference datasets and other base models. We also observe the same trend in other benchmarks using Best-of-N reranking, where the hybrid mix has 2-3% better performance. Finally, we analyze features from HyPER and find that prompts with moderate safety concerns or complexity benefit the most from human feedback.

Alisa Liu, Xiaochuang Han, Yizhong Wang et al.

Despite the general capabilities of large pretrained language models, they consistently benefit from further adaptation to better achieve desired behaviors. However, tuning these models has become increasingly resource-intensive, or impossible when model weights are private. We introduce proxy-tuning, a lightweight decoding-time algorithm that operates on top of black-box LMs to achieve the same end as direct tuning, but by accessing only its predictions over the output vocabulary, not its parameters. Our method tunes a smaller LM, then applies the difference between the predictions of the small tuned and untuned LMs to shift the original predictions of the larger untuned model in the direction of tuning, while retaining the benefits of larger-scale pretraining. In experiments, when we apply proxy-tuning to Llama2-70B using proxies of only 7B size, we can close 88% of the gap between Llama2-70B and its truly-tuned chat version, when evaluated across knowledge, reasoning, and safety benchmarks. We then demonstrate the generality of proxy-tuning by applying it to domain adaptation on code, and task-specific finetuning on question-answering and math problems. Finally, we show how to proxy-tune a truly black-box LM, GPT-3.5, for temporal adaptation, increasing its knowledge about recent events. Our work demonstrates the promise of using small tuned LMs to efficiently customize large, potentially proprietary LMs through decoding-time guidance.

Alon Talmor, Ori Yoran, Amnon Catav et al.

When answering complex questions, people can seamlessly combine information from visual, textual and tabular sources. While interest in models that reason over multiple pieces of evidence has surged in recent years, there has been relatively little work on question answering models that reason across multiple modalities. In this paper, we present MultiModalQA(MMQA): a challenging question answering dataset that requires joint reasoning over text, tables and images. We create MMQA using a new framework for generating complex multi-modal questions at scale, harvesting tables from Wikipedia, and attaching images and text paragraphs using entities that appear in each table. We then define a formal language that allows us to take questions that can be answered from a single modality, and combine them to generate cross-modal questions. Last, crowdsourcing workers take these automatically-generated questions and rephrase them into more fluent language. We create 29,918 questions through this procedure, and empirically demonstrate the necessity of a multi-modal multi-hop approach to solve our task: our multi-hop model, ImplicitDecomp, achieves an average F1of 51.7 over cross-modal questions, substantially outperforming a strong baseline that achieves 38.2 F1, but still lags significantly behind human performance, which is at 90.1 F1

Yue Guo, Wei Qiu, Yizhong Wang et al.

Health literacy has emerged as a crucial factor in making appropriate health decisions and ensuring treatment outcomes. However, medical jargon and the complex structure of professional language in this domain make health information especially hard to interpret. Thus, there is an urgent unmet need for automated methods to enhance the accessibility of the biomedical literature to the general population. This problem can be framed as a type of translation problem between the language of healthcare professionals, and that of the general public. In this paper, we introduce the novel task of automated generation of lay language summaries of biomedical scientific reviews, and construct a dataset to support the development and evaluation of automated methods through which to enhance the accessibility of the biomedical literature. We conduct analyses of the various challenges in solving this task, including not only summarization of the key points but also explanation of background knowledge and simplification of professional language. We experiment with state-of-the-art summarization models as well as several data augmentation techniques, and evaluate their performance using both automated metrics and human assessment. Results indicate that automatically generated summaries produced using contemporary neural architectures can achieve promising quality and readability as compared with reference summaries developed for the lay public by experts (best ROUGE-L of 50.24 and Flesch-Kincaid readability score of 13.30). We also discuss the limitations of the current attempt, providing insights and directions for future work.

Qianying Liu, Sicong Jiang, Yizhong Wang et al.

In this paper, we introduce LiveQA, a new question answering dataset constructed from play-by-play live broadcast. It contains 117k multiple-choice questions written by human commentators for over 1,670 NBA games, which are collected from the Chinese Hupu (https://nba.hupu.com/games) website. Derived from the characteristics of sports games, LiveQA can potentially test the reasoning ability across timeline-based live broadcasts, which is challenging compared to the existing datasets. In LiveQA, the questions require understanding the timeline, tracking events or doing mathematical computations. Our preliminary experiments show that the dataset introduces a challenging problem for question answering models, and a strong baseline model only achieves the accuracy of 53.1\% and cannot beat the dominant option rule. We release the code and data of this paper for future research.

Swabha Swayamdipta, Roy Schwartz, Nicholas Lourie et al.

Large datasets have become commonplace in NLP research. However, the increased emphasis on data quantity has made it challenging to assess the quality of data. We introduce Data Maps---a model-based tool to characterize and diagnose datasets. We leverage a largely ignored source of information: the behavior of the model on individual instances during training (training dynamics) for building data maps. This yields two intuitive measures for each example---the model's confidence in the true class, and the variability of this confidence across epochs---obtained in a single run of training. Experiments across four datasets show that these model-dependent measures reveal three distinct regions in the data map, each with pronounced characteristics. First, our data maps show the presence of "ambiguous" regions with respect to the model, which contribute the most towards out-of-distribution generalization. Second, the most populous regions in the data are "easy to learn" for the model, and play an important role in model optimization. Finally, data maps uncover a region with instances that the model finds "hard to learn"; these often correspond to labeling errors. Our results indicate that a shift in focus from quantity to quality of data could lead to robust models and improved out-of-distribution generalization.

Eric Wallace, Yizhong Wang, Sujian Li et al.

The ability to understand and work with numbers (numeracy) is critical for many complex reasoning tasks. Currently, most NLP models treat numbers in text in the same way as other tokens---they embed them as distributed vectors. Is this enough to capture numeracy? We begin by investigating the numerical reasoning capabilities of a state-of-the-art question answering model on the DROP dataset. We find this model excels on questions that require numerical reasoning, i.e., it already captures numeracy. To understand how this capability emerges, we probe token embedding methods (e.g., BERT, GloVe) on synthetic list maximum, number decoding, and addition tasks. A surprising degree of numeracy is naturally present in standard embeddings. For example, GloVe and word2vec accurately encode magnitude for numbers up to 1,000. Furthermore, character-level embeddings are even more precise---ELMo captures numeracy the best for all pre-trained methods---but BERT, which uses sub-word units, is less exact.