Peer prediction mechanisms motivate high-quality feedback with provable guarantees. However, current methods only apply to rather simple reports, like multiple-choice or scalar numbers. We aim to broaden these techniques to the larger domain of text-based reports, drawing on the recent developments in large language models. This vastly increases the applicability of peer prediction mechanisms as textual feedback is the norm in a large variety of feedback channels: peer reviews, e-commerce customer reviews, and comments on social media. We introduce two mechanisms, the Generative Peer Prediction Mechanism (GPPM) and the Generative Synopsis Peer Prediction Mechanism (GSPPM). These mechanisms utilize LLMs as predictors, mapping from one agent's report to a prediction of her peer's report. Theoretically, we show that when the LLM prediction is sufficiently accurate, our mechanisms can incentivize high effort and truth-telling as an (approximate) Bayesian Nash equilibrium. Empirically, we confirm the efficacy of our mechanisms through experiments conducted on two real datasets: the Yelp review dataset and the ICLR OpenReview dataset. We highlight the results that on the ICLR dataset, our mechanisms can differentiate three quality levels -- human-written reviews, GPT-4-generated reviews, and GPT-3.5-generated reviews in terms of expected scores. Additionally, GSPPM penalizes LLM-generated reviews more effectively than GPPM.

Anomaly detection in continuous-time dynamic graphs is an emerging field yet under-explored in the context of learning-based approaches. In this paper, we pioneer structured analyses of link-level anomalies and graph representation learning for identifying anomalous links in these graphs. First, we introduce a fine-grain taxonomy for edge-level anomalies leveraging structural, temporal, and contextual graph properties. We present a method for generating and injecting such typed anomalies into graphs. Next, we introduce a novel method to generate continuous-time dynamic graphs with consistent patterns across time, structure, and context. To allow temporal graph methods to learn the link anomaly detection task, we extend the generic link prediction setting by: (1) conditioning link existence on contextual edge attributes; and (2) refining the training regime to accommodate diverse perturbations in the negative edge sampler. Building on this, we benchmark methods for anomaly detection. Comprehensive experiments on synthetic and real-world datasets -- featuring synthetic and labeled organic anomalies and employing six state-of-the-art learning methods -- validate our taxonomy and generation processes for anomalies and benign graphs, as well as our approach to adapting link prediction methods for anomaly detection. Our results further reveal that different learning methods excel in capturing different aspects of graph normality and detecting different types of anomalies. We conclude with a comprehensive list of findings highlighting opportunities for future research.

Off-policy reinforcement learning (RL) has achieved notable success in tackling many complex real-world tasks, by leveraging previously collected data for policy learning. However, most existing off-policy RL algorithms fail to maximally exploit the information in the replay buffer, limiting sample efficiency and policy performance. In this work, we discover that concurrently training an offline RL policy based on the shared online replay buffer can sometimes outperform the original online learning policy, though the occurrence of such performance gains remains uncertain. This motivates a new possibility of harnessing the emergent outperforming offline optimal policy to improve online policy learning. Based on this insight, we present Offline-Boosted Actor-Critic (OBAC), a model-free online RL framework that elegantly identifies the outperforming offline policy through value comparison, and uses it as an adaptive constraint to guarantee stronger policy learning performance. Our experiments demonstrate that OBAC outperforms other popular model-free RL baselines and rivals advanced model-based RL methods in terms of sample efficiency and asymptotic performance across 53 tasks spanning 6 task suites.

Lifelong learning aims to train a model with good performance for new tasks while retaining the capacity of previous tasks. However, some practical scenarios require the system to forget undesirable knowledge due to privacy issues, which is called selective forgetting. The joint task of the two is dubbed Learning with Selective Forgetting (LSF). In this paper, we propose a new framework based on contrastive strategy for LSF. Specifically, for the preserved classes (tasks), we make features extracted from different samples within a same class compacted. And for the deleted classes, we make the features from different samples of a same class dispersed and irregular, i.e., the network does not have any regular response to samples from a specific deleted class as if the network has no training at all. Through maintaining or disturbing the feature distribution, the forgetting and memory of different classes can be or independent of each other. Experiments are conducted on four benchmark datasets, and our method acieves new state-of-the-art.

Enhancing the attribution in large language models (LLMs) is a crucial task. One feasible approach is to enable LLMs to cite external sources that support their generations. However, existing datasets and evaluation methods in this domain still exhibit notable limitations. In this work, we formulate the task of attributed query-focused summarization (AQFS) and present WebCiteS, a Chinese dataset featuring 7k human-annotated summaries with citations. WebCiteS derives from real-world user queries and web search results, offering a valuable resource for model training and evaluation. Prior works in attribution evaluation do not differentiate between groundedness errors and citation errors. They also fall short in automatically verifying sentences that draw partial support from multiple sources. We tackle these issues by developing detailed metrics and enabling the automatic evaluator to decompose the sentences into sub-claims for fine-grained verification. Our comprehensive evaluation of both open-source and proprietary models on WebCiteS highlights the challenge LLMs face in correctly citing sources, underscoring the necessity for further improvement. The dataset and code will be open-sourced to facilitate further research in this crucial field.

Large language models are known to hallucinate when faced with unfamiliar queries, but the underlying mechanism that govern how models hallucinate are not yet fully understood. In this work, we find that unfamiliar examples in the models' finetuning data -- those that introduce concepts beyond the base model's scope of knowledge -- are crucial in shaping these errors. In particular, we find that an LLM's hallucinated predictions tend to mirror the responses associated with its unfamiliar finetuning examples. This suggests that by modifying how unfamiliar finetuning examples are supervised, we can influence a model's responses to unfamiliar queries (e.g., say ``I don't know''). We empirically validate this observation in a series of controlled experiments involving SFT, RL, and reward model finetuning on TriviaQA and MMLU. Our work further investigates RL finetuning strategies for improving the factuality of long-form model generations. We find that, while hallucinations from the reward model can significantly undermine the effectiveness of RL factuality finetuning, strategically controlling how reward models hallucinate can minimize these negative effects. Leveraging our previous observations on controlling hallucinations, we propose an approach for learning more reliable reward models, and show that they improve the efficacy of RL factuality finetuning in long-form biography and book/movie plot generation tasks.

Large language models (LLMs) have shown remarkable performance on many tasks in different domains. However, their performance in closed-book biomedical machine reading comprehension (MRC) has not been evaluated in depth. In this work, we evaluate GPT on four closed-book biomedical MRC benchmarks. We experiment with different conventional prompting techniques as well as introduce our own novel prompting method. To solve some of the retrieval problems inherent to LLMs, we propose a prompting strategy named Implicit Retrieval Augmented Generation (RAG) that alleviates the need for using vector databases to retrieve important chunks in traditional RAG setups. Moreover, we report qualitative assessments on the natural language generation outputs from our approach. The results show that our new prompting technique is able to get the best performance in two out of four datasets and ranks second in rest of them. Experiments show that modern-day LLMs like GPT even in a zero-shot setting can outperform supervised models, leading to new state-of-the-art (SoTA) results on two of the benchmarks.

We consider the problem of online multiclass U-calibration, where a forecaster aims to make sequential distributional predictions over $K$ classes with low U-calibration error, that is, low regret with respect to all bounded proper losses simultaneously. Kleinberg et al. (2023) developed an algorithm with U-calibration error $O(K\sqrt{T})$ after $T$ rounds and raised the open question of what the optimal bound is. We resolve this question by showing that the optimal U-calibration error is $\Theta(\sqrt{KT})$ -- we start with a simple observation that the Follow-the-Perturbed-Leader algorithm of Daskalakis and Syrgkanis (2016) achieves this upper bound, followed by a matching lower bound constructed with a specific proper loss (which, as a side result, also proves the optimality of the algorithm of Daskalakis and Syrgkanis (2016) in the context of online learning against an adversary with finite choices). We also strengthen our results under natural assumptions on the loss functions, including $\Theta(\log T)$ U-calibration error for Lipschitz proper losses, $O(\log T)$ U-calibration error for a certain class of decomposable proper losses, U-calibration error bounds for proper losses with a low covering number, and others.

When neural networks are confronted with unfamiliar data that deviate from their training set, this signifies a domain shift. While these networks output predictions on their inputs, they typically fail to account for their level of familiarity with these novel observations. This challenge becomes even more pronounced in resource-constrained settings, such as embedded systems or edge devices. To address such challenges, we aim to recalibrate a neural network's decision boundaries in relation to its cognizance of the data it observes, introducing an approach we coin as certainty distillation. While prevailing works navigate unsupervised domain adaptation (UDA) with the goal of curtailing model entropy, they unintentionally birth models that grapple with calibration inaccuracies - a dilemma we term the over-certainty phenomenon. In this paper, we probe the drawbacks of this traditional learning model. As a solution to the issue, we propose a UDA algorithm that not only augments accuracy but also assures model calibration, all while maintaining suitability for environments with limited computational resources.

Learning from feedback reduces the hallucination of multimodal large language models (MLLMs) by aligning them with human preferences. While traditional methods rely on labor-intensive and time-consuming manual labeling, recent approaches employing models as automatic labelers have shown promising results without human intervention. However, these methods heavily rely on costly proprietary models like GPT-4V, resulting in scalability issues. Moreover, this paradigm essentially distills the proprietary models to provide a temporary solution to quickly bridge the performance gap. As this gap continues to shrink, the community is soon facing the essential challenge of aligning MLLMs using labeler models of comparable capability. In this work, we introduce RLAIF-V, a novel framework that aligns MLLMs in a fully open-source paradigm for super GPT-4V trustworthiness. RLAIF-V maximally exploits the open-source feedback from two perspectives, including high-quality feedback data and online feedback learning algorithm. Extensive experiments on seven benchmarks in both automatic and human evaluation show that RLAIF-V substantially enhances the trustworthiness of models without sacrificing performance on other tasks. Using a 34B model as labeler, RLAIF-V 7B model reduces object hallucination by 82.9\% and overall hallucination by 42.1\%, outperforming the labeler model. Remarkably, RLAIF-V also reveals the self-alignment potential of open-source MLLMs, where a 12B model can learn from the feedback of itself to achieve less than 29.5\% overall hallucination rate, surpassing GPT-4V (45.9\%) by a large margin. The results shed light on a promising route to enhance the efficacy of leading-edge MLLMs.