This paper tackles the emerging challenge of training generative models within a self-consuming loop, wherein successive generations of models are recursively trained on mixtures of real and synthetic data from previous generations. We construct a theoretical framework to rigorously evaluate how this training regimen impacts the data distributions learned by future models. Specifically, we derive bounds on the total variation (TV) distance between the synthetic data distributions produced by future models and the original real data distribution under various mixed training scenarios. Our analysis demonstrates that this distance can be effectively controlled under the condition that mixed training dataset sizes or proportions of real data are large enough. Interestingly, we further unveil a phase transition induced by expanding synthetic data amounts, proving theoretically that while the TV distance exhibits an initial ascent, it declines beyond a threshold point. Finally, we specialize our general results to diffusion models, delivering nuanced insights such as the efficacy of optimal early stopping within the self-consuming loop.
Despite the success of reinforcement learning from human feedback (RLHF) in aligning language models with human values, reward hacking, also termed reward overoptimization, remains a critical challenge, which primarily stems from limitations in reward modeling, i.e., generalizability of the reward model and inconsistency in the preference dataset. In this work, we tackle this problem from an information theoretic-perspective, and propose a generalizable and robust framework for reward modeling, namely InfoRM, by introducing a variational information bottleneck objective to filter out irrelevant information and developing a mechanism for model complexity modulation. Notably, we further identify a correlation between overoptimization and outliers in the latent space, establishing InfoRM as a promising tool for detecting reward overoptimization. Inspired by this finding, we propose the Integrated Cluster Deviation Score (ICDS), which quantifies deviations in the latent space, as an indicator of reward overoptimization to facilitate the development of online mitigation strategies. Extensive experiments on a wide range of settings and model scales (70M, 440M, 1.4B, and 7B) support the effectiveness of InfoRM. Further analyses reveal that InfoRM's overoptimization detection mechanism is effective, potentially signifying a notable advancement in the field of RLHF. Code will be released upon acceptance.
Bridging the gap between diffusion models and human preferences is crucial for their integration into practical generative workflows. While optimizing downstream reward models has emerged as a promising alignment strategy, concerns arise regarding the risk of excessive optimization with learned reward models, which potentially compromises ground-truth performance. In this work, we confront the reward overoptimization problem in diffusion model alignment through the lenses of both inductive and primacy biases. We first identify the divergence of current methods from the temporal inductive bias inherent in the multi-step denoising process of diffusion models as a potential source of overoptimization. Then, we surprisingly discover that dormant neurons in our critic model act as a regularization against overoptimization, while active neurons reflect primacy bias in this setting. Motivated by these observations, we propose Temporal Diffusion Policy Optimization with critic active neuron Reset (TDPO-R), a policy gradient algorithm that exploits the temporal inductive bias of intermediate timesteps, along with a novel reset strategy that targets active neurons to counteract the primacy bias. Empirical results demonstrate the superior efficacy of our algorithms in mitigating reward overoptimization.
Time series modeling is uniquely challenged by the presence of autocorrelation in both historical and label sequences. Current research predominantly focuses on handling autocorrelation within the historical sequence but often neglects its presence in the label sequence. Specifically, emerging forecast models mainly conform to the direct forecast (DF) paradigm, generating multi-step forecasts under the assumption of conditional independence within the label sequence. This assumption disregards the inherent autocorrelation in the label sequence, thereby limiting the performance of DF-based models. In response to this gap, we introduce the Frequency-enhanced Direct Forecast (FreDF), which bypasses the complexity of label autocorrelation by learning to forecast in the frequency domain. Our experiments demonstrate that FreDF substantially outperforms existing state-of-the-art methods including iTransformer and is compatible with a variety of forecast models.
Supplying data augmentation to conversational question answering (CQA) can effectively improve model performance. However, there is less improvement from single-turn datasets in CQA due to the distribution gap between single-turn and multi-turn datasets. On the other hand, while numerous single-turn datasets are available, we have not utilized them effectively. To solve this problem, we propose a novel method to convert single-turn datasets to multi-turn datasets. The proposed method consists of three parts, namely, a QA pair Generator, a QA pair Reassembler, and a question Rewriter. Given a sample consisting of context and single-turn QA pairs, the Generator obtains candidate QA pairs and a knowledge graph based on the context. The Reassembler utilizes the knowledge graph to get sequential QA pairs, and the Rewriter rewrites questions from a conversational perspective to obtain a multi-turn dataset S2M. Our experiments show that our method can synthesize effective training resources for CQA. Notably, S2M ranks 1st place on the QuAC leaderboard at the time of submission (Aug 24th, 2022).
Plasticity, the ability of a neural network to evolve with new data, is crucial for high-performance and sample-efficient visual reinforcement learning (VRL). Although methods like resetting and regularization can potentially mitigate plasticity loss, the influences of various components within the VRL framework on the agent's plasticity are still poorly understood. In this work, we conduct a systematic empirical exploration focusing on three primary underexplored facets and derive the following insightful conclusions: (1) data augmentation is essential in maintaining plasticity; (2) the critic's plasticity loss serves as the principal bottleneck impeding efficient training; and (3) without timely intervention to recover critic's plasticity in the early stages, its loss becomes catastrophic. These insights suggest a novel strategy to address the high replay ratio (RR) dilemma, where exacerbated plasticity loss hinders the potential improvements of sample efficiency brought by increased reuse frequency. Rather than setting a static RR for the entire training process, we propose Adaptive RR, which dynamically adjusts the RR based on the critic's plasticity level. Extensive evaluations indicate that Adaptive RR not only avoids catastrophic plasticity loss in the early stages but also benefits from more frequent reuse in later phases, resulting in superior sample efficiency.
The swift advancement in the scales and capabilities of Large Language Models (LLMs) positions them as promising tools for a variety of downstream tasks. In addition to the pursuit of better performance and the avoidance of violent feedback on a certain prompt, to ensure the responsibility of the LLM, much attention is drawn to the robustness of LLMs. However, existing evaluation methods mostly rely on traditional question answering datasets with predefined supervised labels, which do not align with the superior generation capabilities of contemporary LLMs. To address this issue, we propose a novel rational evaluation approach that leverages pre-trained reward models as diagnostic tools to evaluate the longer conversation generated from more challenging open questions by LLMs, which we refer to as the Reward Model for Reasonable Robustness Evaluation (TREvaL). Longer conversations manifest the comprehensive grasp of language models in terms of their proficiency in understanding questions, a capability not entirely encompassed by individual words or letters, which may exhibit oversimplification and inherent biases. Our extensive empirical experiments demonstrate that TREvaL provides an innovative method for evaluating the robustness of an LLM. Furthermore, our results demonstrate that LLMs frequently exhibit vulnerability to word-level perturbations that are commonplace in daily language usage. Notably, we are surprised to discover that robustness tends to decrease as fine-tuning (SFT and RLHF) is conducted. The code of TREval is available in https://github.com/Harry-mic/TREvaL.
Previous studies regarding the perception of emotions for embodied virtual agents have shown the effectiveness of using virtual characters in conveying emotions through interactions with humans. However, creating an autonomous embodied conversational agent with expressive behaviors presents two major challenges. The first challenge is the difficulty of synthesizing the conversational behaviors for each modality that are as expressive as real human behaviors. The second challenge is that the affects are modeled independently, which makes it difficult to generate multimodal responses with consistent emotions across all modalities. In this work, we propose a conceptual framework, ACTOR (Affect-Consistent mulTimodal behaviOR generation), that aims to increase the perception of affects by generating multimodal behaviors conditioned on a consistent driving affect. We have conducted a user study with 199 participants to assess how the average person judges the affects perceived from multimodal behaviors that are consistent and inconsistent with respect to a driving affect. The result shows that among all model conditions, our affect-consistent framework receives the highest Likert scores for the perception of driving affects. Our statistical analysis suggests that making a modality affect-inconsistent significantly decreases the perception of driving affects. We also observe that multimodal behaviors conditioned on consistent affects are more expressive compared to behaviors with inconsistent affects. Therefore, we conclude that multimodal emotion conditioning and affect consistency are vital to enhancing the perception of affects for embodied conversational agents.
This paper surveys research works in the quickly advancing field of instruction tuning (IT), a crucial technique to enhance the capabilities and controllability of large language models (LLMs). Instruction tuning refers to the process of further training LLMs on a dataset consisting of \textsc{(instruction, output)} pairs in a supervised fashion, which bridges the gap between the next-word prediction objective of LLMs and the users' objective of having LLMs adhere to human instructions. In this work, we make a systematic review of the literature, including the general methodology of IT, the construction of IT datasets, the training of IT models, and applications to different modalities, domains and applications, along with an analysis on aspects that influence the outcome of IT (e.g., generation of instruction outputs, size of the instruction dataset, etc). We also review the potential pitfalls of IT along with criticism against it, along with efforts pointing out current deficiencies of existing strategies and suggest some avenues for fruitful research.