Large language models achieve state-of-the-art performance but are costly to fine-tune due to their size. Parameter-efficient fine-tuning methods, such as prompt tuning, address this by reducing trainable parameters while maintaining strong performance. However, prior methods tie prompt embeddings to the model's dimensionality, which may not scale well with larger LLMs and more customized LLMs. In this paper, we propose Ultra-Low-dimensional Prompt Tuning (ULPT), which optimizes prompts in a low-dimensional space (e.g., 2D) and use a random but frozen matrix for the up-projection. To enhance alignment, we introduce learnable shift and scale embeddings. ULPT drastically reduces the trainable parameters, e.g., 2D only using 2% parameters compared with vanilla prompt tuning while retaining most of the performance across 21 NLP tasks. Our theoretical analysis shows that random projections can capture high-rank structures effectively, and experimental results demonstrate ULPT's competitive performance over existing parameter-efficient methods.

Alzheimer's Disease (AD) early detection is critical for enabling timely intervention and improving patient outcomes. This paper presents a DMV framework using Llama3-70B and GPT-4o as embedding models to analyze clinical notes and predict a continuous risk score associated with early AD onset. Framing the task as a regression problem, we model the relationship between linguistic features in clinical notes (inputs) and a target variable (data value) that answers specific questions related to AD risk within certain topic categories. By leveraging a multi-faceted feature set that includes geolocation data, we capture additional environmental context potentially linked to AD. Our results demonstrate that the integration of the geolocation information significantly decreases the error of predicting early AD risk scores over prior models by 28.57% (Llama3-70B) and 33.47% (GPT4-o). Our findings suggest that this combined approach can enhance the predictive accuracy of AD risk assessment, supporting early diagnosis and intervention in clinical settings. Additionally, the framework's ability to incorporate geolocation data provides a more comprehensive risk assessment model that could help healthcare providers better understand and address environmental factors contributing to AD development.

Despite extensive safety alignment efforts, large language models (LLMs) remain vulnerable to jailbreak attacks that elicit harmful behavior. While existing studies predominantly focus on attack methods that require technical expertise, two critical questions remain underexplored: (1) Are jailbroken responses truly useful in enabling average users to carry out harmful actions? (2) Do safety vulnerabilities exist in more common, simple human-LLM interactions? In this paper, we demonstrate that LLM responses most effectively facilitate harmful actions when they are both actionable and informative--two attributes easily elicited in multi-step, multilingual interactions. Using this insight, we propose HarmScore, a jailbreak metric that measures how effectively an LLM response enables harmful actions, and Speak Easy, a simple multi-step, multilingual attack framework. Notably, by incorporating Speak Easy into direct request and jailbreak baselines, we see an average absolute increase of 0.319 in Attack Success Rate and 0.426 in HarmScore in both open-source and proprietary LLMs across four safety benchmarks. Our work reveals a critical yet often overlooked vulnerability: Malicious users can easily exploit common interaction patterns for harmful intentions.

We present an approach to identifying which ransomware adversaries are most likely to target specific entities, thereby assisting these entities in formulating better protection strategies. Ransomware poses a formidable cybersecurity threat characterized by profit-driven motives, a complex underlying economy supporting criminal syndicates, and the overt nature of its attacks. This type of malware has consistently ranked among the most prevalent, with a rapid escalation in activity observed. Recent estimates indicate that approximately two-thirds of organizations experienced ransomware attacks in 2023 \cite{Sophos2023Ransomware}. A central tactic in ransomware campaigns is publicizing attacks to coerce victims into paying ransoms. Our study utilizes public disclosures from ransomware victims to predict the likelihood of an entity being targeted by a specific ransomware variant. We employ a Large Language Model (LLM) architecture that uses a unique chain-of-thought, multi-shot prompt methodology to define adversary SKRAM (Skills, Knowledge, Resources, Authorities, and Motivation) profiles from ransomware bulletins, threat reports, and news items. This analysis is enriched with publicly available victim data and is further enhanced by a heuristic for generating synthetic data that reflects victim profiles. Our work culminates in the development of a machine learning model that assists organizations in prioritizing ransomware threats and formulating defenses based on the tactics, techniques, and procedures (TTP) of the most likely attackers.

Transformers have demonstrated remarkable in-context learning (ICL) capabilities, adapting to new tasks by simply conditioning on demonstrations without parameter updates. Compelling empirical and theoretical evidence suggests that ICL, as a general-purpose learner, could outperform task-specific models. However, it remains unclear to what extent the transformers optimally learn in-context compared to principled learning algorithms. To bridge this gap, we introduce a new framework for quantifying optimality of ICL as a learning algorithm in stylized settings. Our findings reveal a striking dichotomy: while ICL initially matches the efficiency of a Bayes optimal estimator, its efficiency significantly deteriorates in long context. Through an information-theoretic analysis, we show that the diminishing efficiency is inherent to ICL. These results clarify the trade-offs in adopting ICL as a universal problem solver, motivating a new generation of on-the-fly adaptive methods without the diminishing efficiency.

Alignment of large language models (LLMs) with human values and preferences, often achieved through fine-tuning based on human feedback, is essential for ensuring safe and responsible AI behaviors. However, the process typically requires substantial data and computation resources. Recent studies have revealed that alignment might be attainable at lower costs through simpler methods, such as in-context learning. This leads to the question: Is alignment predominantly superficial? In this paper, we delve into this question and provide a quantitative analysis. We formalize the concept of superficial knowledge, defining it as knowledge that can be acquired through easily token restyling, without affecting the model's ability to capture underlying causal relationships between tokens. We propose a method to extract and isolate superficial knowledge from aligned models, focusing on the shallow modifications to the final token selection process. By comparing models augmented only with superficial knowledge to fully aligned models, we quantify the superficial portion of alignment. Our findings reveal that while superficial knowledge constitutes a significant portion of alignment, particularly in safety and detoxification tasks, it is not the whole story. Tasks requiring reasoning and contextual understanding still rely on deeper knowledge. Additionally, we demonstrate two practical advantages of isolated superficial knowledge: (1) it can be transferred between models, enabling efficient offsite alignment of larger models using extracted superficial knowledge from smaller models, and (2) it is recoverable, allowing for the restoration of alignment in compromised models without sacrificing performance.

Recent advances in text-based large language models (LLMs), particularly in the GPT series and the o1 model, have demonstrated the effectiveness of scaling both training-time and inference-time compute. However, current state-of-the-art TTS systems leveraging LLMs are often multi-stage, requiring separate models (e.g., diffusion models after LLM), complicating the decision of whether to scale a particular model during training or testing. This work makes the following contributions: First, we explore the scaling of train-time and inference-time compute for speech synthesis. Second, we propose a simple framework Llasa for speech synthesis that employs a single-layer vector quantizer (VQ) codec and a single Transformer architecture to fully align with standard LLMs such as Llama. Our experiments reveal that scaling train-time compute for Llasa consistently improves the naturalness of synthesized speech and enables the generation of more complex and accurate prosody patterns. Furthermore, from the perspective of scaling inference-time compute, we employ speech understanding models as verifiers during the search, finding that scaling inference-time compute shifts the sampling modes toward the preferences of specific verifiers, thereby improving emotional expressiveness, timbre consistency, and content accuracy. In addition, we released the checkpoint and training code for our TTS model (1B, 3B, 8B) and codec model publicly available.

The GPT-4 technical report from OpenAI suggests that model performance on specific tasks can be predicted prior to training, though methodologies remain unspecified. This approach is crucial for optimizing resource allocation and ensuring data alignment with target tasks. To achieve this vision, we focus on predicting performance on Closed-book Question Answering (CBQA) tasks, which are closely tied to pre-training data and knowledge retention. We address three major challenges: 1) mastering the entire pre-training process, especially data construction; 2) evaluating a model's knowledge retention; and 3) predicting task-specific knowledge retention using only information available prior to training. To tackle these challenges, we pre-train three large language models (i.e., 1.6B, 7B, and 13B) using 560k dollars and 520k GPU hours. We analyze the pre-training data with knowledge triples and assess knowledge retention using established methods. Additionally, we introduce the SMI metric, an information-theoretic measure that quantifies the relationship between pre-training data, model size, and task-specific knowledge retention. Our experiments reveal a strong linear correlation ($\text{R}^2 > 0.84$) between the SMI metric and the model's accuracy on CBQA tasks across models of varying sizes (i.e., 1.1B, 1.6B, 7B, and 13B). The dataset, model, and code are available at https://github.com/yuhui1038/SMI.

As Language Model (LM) capabilities advance, evaluating and supervising them at scale is getting harder for humans. There is hope that other language models can automate both these tasks, which we refer to as "AI Oversight". We study how model similarity affects both aspects of AI oversight by proposing a probabilistic metric for LM similarity based on overlap in model mistakes. Using this metric, we first show that LLM-as-a-judge scores favor models similar to the judge, generalizing recent self-preference results. Then, we study training on LM annotations, and find complementary knowledge between the weak supervisor and strong student model plays a crucial role in gains from "weak-to-strong generalization". As model capabilities increase, it becomes harder to find their mistakes, and we might defer more to AI oversight. However, we observe a concerning trend -- model mistakes are becoming more similar with increasing capabilities, pointing to risks from correlated failures. Our work underscores the importance of reporting and correcting for model similarity, especially in the emerging paradigm of AI oversight.

The current research on Role-Playing Conversational Agents (RPCAs) with Large Language Models (LLMs) primarily focuses on imitating specific speaking styles and utilizing character backgrounds, neglecting the depiction of deeper personality traits.~In this study, we introduce personality-infused role-playing for LLM agents, which encourages agents to accurately portray their designated personality traits during dialogues. We then propose PsyPlay, a dialogue generation framework that facilitates the expression of rich personalities among multiple LLM agents. Specifically, PsyPlay enables agents to assume roles with distinct personality traits and engage in discussions centered around specific topics, consistently exhibiting their designated personality traits throughout the interactions. Validation on generated dialogue data demonstrates that PsyPlay can accurately portray the intended personality traits, achieving an overall success rate of 80.31% on GPT-3.5. Notably, we observe that LLMs aligned with positive values are more successful in portraying positive personality roles compared to negative ones. Moreover, we construct a dialogue corpus for personality-infused role-playing, called PsyPlay-Bench. The corpus, which consists of 4745 instances of correctly portrayed dialogues using PsyPlay, aims to further facilitate research in personalized role-playing and dialogue personality detection.