Self-Consistency of the Internal Reward Models Improves Self-Rewarding Language Models

Aligning Large Language Models (LLMs) with human preferences is crucial for their deployment in real-world applications. Recent advancements in Self-Rewarding Language Models suggest that an LLM can use its internal reward models (such as LLM-as-a-Judge) \cite{yuanself} to generate preference data, improving alignment performance without costly human annotation. However, we find that different internal reward models within the same LLM often generate inconsistent preferences. This inconsistency raises concerns about the reliability of self-generated preference data, hinders overall alignment performance, and highlights the need for further research to ensure reliable and coherent alignment with human preferences. To address this limitation, we propose Self-Consistent Internal Rewards (SCIR), a novel framework designed to enhance consistency among internal reward models during training. In each training step, we collect preference predictions from multiple pre-defined internal reward models and enforce consistency and confidence through an inconsistency penalty mechanism, thereby improving the reliability of these internal reward models. We selectively use data with consistent predictions for preference optimization, ensuring the quality of the preference data. By employing self-consistent internal rewards, our method significantly improves the alignment performance and reward modeling capability of LLMs, outperforming baseline methods by a notable margin.

Enhancing Expressive Voice Conversion with Discrete Pitch-Conditioned Flow Matching Model

This paper introduces PFlow-VC, a conditional flow matching voice conversion model that leverages fine-grained discrete pitch tokens and target speaker prompt information for expressive voice conversion (VC). Previous VC works primarily focus on speaker conversion, with further exploration needed in enhancing expressiveness (such as prosody and emotion) for timbre conversion. Unlike previous methods, we adopt a simple and efficient approach to enhance the style expressiveness of voice conversion models. Specifically, we pretrain a self-supervised pitch VQVAE model to discretize speaker-irrelevant pitch information and leverage a masked pitch-conditioned flow matching model for Mel-spectrogram synthesis, which provides in-context pitch modeling capabilities for the speaker conversion model, effectively improving the voice style transfer capacity. Additionally, we improve timbre similarity by combining global timbre embeddings with time-varying timbre tokens. Experiments on unseen LibriTTS test-clean and emotional speech dataset ESD show the superiority of the PFlow-VC model in both timbre conversion and style transfer. Audio samples are available on the demo page https://speechai-demo.github.io/PFlow-VC/.

CueTip: An Interactive and Explainable Physics-aware Pool Assistant

We present an interactive and explainable automated coaching assistant called CueTip for a variant of pool/billiards. CueTip's novelty lies in its combination of three features: a natural-language interface, an ability to perform contextual, physics-aware reasoning, and that its explanations are rooted in a set of predetermined guidelines developed by domain experts. We instrument a physics simulator so that it generates event traces in natural language alongside traditional state traces. Event traces lend themselves to interpretation by language models, which serve as the interface to our assistant. We design and train a neural adaptor that decouples tactical choices made by CueTip from its interactivity and explainability allowing it to be reconfigured to mimic any pool playing agent. Our experiments show that CueTip enables contextual query-based assistance and explanations while maintaining the strength of the agent in terms of win rate (improving it in some situations). The explanations generated by CueTip are physically-aware and grounded in the expert rules and are therefore more reliable.

Multi-Task Model Merging via Adaptive Weight Disentanglement

Model merging has gained increasing attention as an efficient and effective technique for integrating task-specific weights from various tasks into a unified multi-task model without retraining or additional data. As a representative approach, Task Arithmetic (TA) has demonstrated that combining task vectors through arithmetic operations facilitates efficient capability transfer between different tasks. In this framework, task vectors are obtained by subtracting the parameter values of a pre-trained model from those of individually fine-tuned models initialized from it. Despite the notable effectiveness of TA, interference among task vectors can adversely affect the performance of the merged model. In this paper, we relax the constraints of Task Arithmetic Property and propose Task Consistency Property, which can be regarded as being free from task interference. Through theoretical derivation, we show that such a property can be approximately achieved by seeking orthogonal task vectors. Guiding by this insight, we propose Adaptive Weight Disentanglement (AWD), which decomposes traditional task vectors into a redundant vector and several disentangled task vectors. The primary optimization objective of AWD is to achieve orthogonality among the disentangled task vectors, thereby closely approximating the desired solution. Notably, these disentangled task vectors can be seamlessly integrated into existing merging methodologies. Experimental results demonstrate that our AWD consistently and significantly improves upon previous merging approaches, achieving state-of-the-art results. Our code is available at \href{https://github.com/FarisXiong/AWD.git}{https://github.com/FarisXiong/AWD.git}.

Unveiling and Consulting Core Experts in Retrieval-Augmented MoE-based LLMs

Retrieval-Augmented Generation (RAG) significantly improved the ability of Large Language Models (LLMs) to solve knowledge-intensive tasks. While existing research seeks to enhance RAG performance by retrieving higher-quality documents or designing RAG-specific LLMs, the internal mechanisms within LLMs that contribute to the effectiveness of RAG systems remain underexplored. In this paper, we aim to investigate these internal mechanisms within the popular Mixture-of-Expert (MoE)-based LLMs and demonstrate how to improve RAG by examining expert activations in these LLMs. Our controlled experiments reveal that several core groups of experts are primarily responsible for RAG-related behaviors. The activation of these core experts can signify the model's inclination towards external/internal knowledge and adjust its behavior. For instance, we identify core experts that can (1) indicate the sufficiency of the model's internal knowledge, (2) assess the quality of retrieved documents, and (3) enhance the model's ability to utilize context. Based on these findings, we propose several strategies to enhance RAG's efficiency and effectiveness through expert activation. Experimental results across various datasets and MoE-based LLMs show the effectiveness of our method.

Deciphering the Chaos: Enhancing Jailbreak Attacks via Adversarial Prompt Translation

Automatic adversarial prompt generation provides remarkable success in jailbreaking safely-aligned large language models (LLMs). Existing gradient-based attacks, while demonstrating outstanding performance in jailbreaking white-box LLMs, often generate garbled adversarial prompts with chaotic appearance. These adversarial prompts are difficult to transfer to other LLMs, hindering their performance in attacking unknown victim models. In this paper, for the first time, we delve into the semantic meaning embedded in garbled adversarial prompts and propose a novel method that "translates" them into coherent and human-readable natural language adversarial prompts. In this way, we can effectively uncover the semantic information that triggers vulnerabilities of the model and unambiguously transfer it to the victim model, without overlooking the adversarial information hidden in the garbled text, to enhance jailbreak attacks. It also offers a new approach to discovering effective designs for jailbreak prompts, advancing the understanding of jailbreak attacks. Experimental results demonstrate that our method significantly improves the success rate of jailbreak attacks against various safety-aligned LLMs and outperforms state-of-the-arts by large margins. With at most 10 queries, our method achieves an average attack success rate of 81.8% in attacking 7 commercial closed-source LLMs, including GPT and Claude-3 series, on HarmBench. Our method also achieves over 90% attack success rates against Llama-2-Chat models on AdvBench, despite their outstanding resistance to jailbreak attacks. Code at: https://github.com/qizhangli/Adversarial-Prompt-Translator.

Recent Advances in Speech Language Models: A Survey

Large Language Models (LLMs) have recently garnered significant attention, primarily for their capabilities in text-based interactions. However, natural human interaction often relies on speech, necessitating a shift towards voice-based models. A straightforward approach to achieve this involves a pipeline of ``Automatic Speech Recognition (ASR) + LLM + Text-to-Speech (TTS)", where input speech is transcribed to text, processed by an LLM, and then converted back to speech. Despite being straightforward, this method suffers from inherent limitations, such as information loss during modality conversion and error accumulation across the three stages. To address these issues, Speech Language Models (SpeechLMs) -- end-to-end models that generate speech without converting from text -- have emerged as a promising alternative. This survey paper provides the first comprehensive overview of recent methodologies for constructing SpeechLMs, detailing the key components of their architecture and the various training recipes integral to their development. Additionally, we systematically survey the various capabilities of SpeechLMs, categorize the evaluation metrics for SpeechLMs, and discuss the challenges and future research directions in this rapidly evolving field.

Enabling Beam Search for Language Model-Based Text-to-Speech Synthesis

Tokenising continuous speech into sequences of discrete tokens and modelling them with language models (LMs) has led to significant success in text-to-speech (TTS) synthesis. Although these models can generate speech with high quality and naturalness, their synthesised samples can still suffer from artefacts, mispronunciation, word repeating, etc. In this paper, we argue these undesirable properties could partly be caused by the randomness of sampling-based strategies during the autoregressive decoding of LMs. Therefore, we look at maximisation-based decoding approaches and propose Temporal Repetition Aware Diverse Beam Search (TRAD-BS) to find the most probable sequences of the generated speech tokens. Experiments with two state-of-the-art LM-based TTS models demonstrate that our proposed maximisation-based decoding strategy generates speech with fewer mispronunciations and improved speaker consistency.

Two Optimizers Are Better Than One: LLM Catalyst for Enhancing Gradient-Based Optimization

Learning a skill generally relies on both practical experience by doer and insightful high-level guidance by instructor. Will this strategy also work well for solving complex non-convex optimization problems? Here, a common gradient-based optimizer acts like a disciplined doer, making locally optimal update at each step. Recent methods utilize large language models (LLMs) to optimize solutions for concrete problems by inferring from natural language instructions, akin to a high-level instructor. In this paper, we show that these two optimizers are complementary to each other, suggesting a collaborative optimization approach. The gradient-based optimizer and LLM-based optimizer are combined in an interleaved manner. We instruct LLMs using task descriptions and timely optimization trajectories recorded during gradient-based optimization. Inferred results from LLMs are used as restarting points for the next stage of gradient optimization. By leveraging both the locally rigorous gradient-based optimizer and the high-level deductive LLM-based optimizer, our combined optimization method consistently yields improvements over competitive baseline prompt tuning methods. Our results demonstrate the synergistic effect of conventional gradient-based optimization and the inference ability of LLMs. The code is released at https://github.com/guozix/LLM-catalyst.

Improved Generation of Adversarial Examples Against Safety-aligned LLMs

Despite numerous efforts to ensure large language models (LLMs) adhere to safety standards and produce harmless content, some successes have been achieved in bypassing these restrictions, known as jailbreak attacks against LLMs. Adversarial prompts generated using gradient-based methods exhibit outstanding performance in performing jailbreak attacks automatically. Nevertheless, due to the discrete nature of texts, the input gradient of LLMs struggles to precisely reflect the magnitude of loss change that results from token replacements in the prompt, leading to limited attack success rates against safety-aligned LLMs, even in the white-box setting. In this paper, we explore a new perspective on this problem, suggesting that it can be alleviated by leveraging innovations inspired in transfer-based attacks that were originally proposed for attacking black-box image classification models. For the first time, we appropriate the ideologies of effective methods among these transfer-based attacks, i.e., Skip Gradient Method and Intermediate Level Attack, for improving the effectiveness of automatically generated adversarial examples against white-box LLMs. With appropriate adaptations, we inject these ideologies into gradient-based adversarial prompt generation processes and achieve significant performance gains without introducing obvious computational cost. Meanwhile, by discussing mechanisms behind the gains, new insights are drawn, and proper combinations of these methods are also developed. Our empirical results show that the developed combination achieves >30% absolute increase in attack success rates compared with GCG for attacking the Llama-2-7B-Chat model on AdvBench.