arXiv subjects

Yiming Zhang

Publications and source records attributed to Yiming Zhang.

At least 145 records · Page 8Linked to original sources

arXiv2024

Unconstrained Model Merging for Enhanced LLM Reasoning

Recent advancements in building domain-specific large language models (LLMs) have shown remarkable success, especially in tasks requiring reasoning abilities like logical inference over complex relationships and multi-step problem solving. However, creating a powerful all-in-one LLM remains challenging due to the need for proprietary data and vast computational resources. As a resource-friendly alternative, we explore the potential of merging multiple expert models into a single LLM. Existing studies on model merging mainly focus on generalist LLMs instead of domain experts, or the LLMs under the same architecture and size. In this work, we propose an unconstrained model merging framework that accommodates both homogeneous and heterogeneous model architectures with a focus on reasoning tasks. A fine-grained layer-wise weight merging strategy is designed for homogeneous models merging, while heterogeneous model merging is built upon the probabilistic distribution knowledge derived from instruction-response fine-tuning data. Across 7 benchmarks and 9 reasoning-optimized LLMs, we reveal key findings that combinatorial reasoning emerges from merging which surpasses simple additive effects. We propose that unconstrained model merging could serve as a foundation for decentralized LLMs, marking a notable progression from the existing centralized LLM framework. This evolution could enhance wider participation and stimulate additional advancement in the field of artificial intelligence, effectively addressing the constraints posed by centralized models.

cs.CL↗

arXiv2024

Persistent Pre-Training Poisoning of LLMs

Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web. Prior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets. Our work evaluates for the first time whether language models can also be compromised during pre-training, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO). We pre-train a series of LLMs from scratch to measure the impact of a potential poisoning adversary under four different attack objectives (denial-of-service, belief manipulation, jailbreaking, and prompt stealing), and across a wide range of model sizes (from 600M to 7B). Our main result is that poisoning only 0.1% of a model's pre-training dataset is sufficient for three out of four attacks to measurably persist through post-training. Moreover, simple attacks like denial-of-service persist through post-training with a poisoning rate of only 0.001%.

cs.CR↗

arXiv2024

Human-aligned Chess with a Bit of Search

Chess has long been a testbed for AI's quest to match human intelligence, and in recent years, chess AI systems have surpassed the strongest humans at the game. However, these systems are not human-aligned; they are unable to match the skill levels of all human partners or model human-like behaviors beyond piece movement. In this paper, we introduce Allie, a chess-playing AI designed to bridge the gap between artificial and human intelligence in this classic game. Allie is trained on log sequences of real chess games to model the behaviors of human chess players across the skill spectrum, including non-move behaviors such as pondering times and resignations In offline evaluations, we find that Allie exhibits humanlike behavior: it outperforms the existing state-of-the-art in human chess move prediction and "ponders" at critical positions. The model learns to reliably assign reward at each game state, which can be used at inference as a reward function in a novel time-adaptive Monte-Carlo tree search (MCTS) procedure, where the amount of search depends on how long humans would think in the same positions. Adaptive search enables remarkable skill calibration; in a large-scale online evaluation against players with ratings from 1000 to 2600 Elo, our adaptive search method leads to a skill gap of only 49 Elo on average, substantially outperforming search-free and standard MCTS baselines. Against grandmaster-level (2500 Elo) opponents, Allie with adaptive search exhibits the strength of a fellow grandmaster, all while learning exclusively from humans.

cs.LG↗

arXiv2024

Backtracking Improves Generation Safety

Text generation has a fundamental limitation almost by definition: there is no taking back tokens that have been generated, even when they are clearly problematic. In the context of language model safety, when a partial unsafe generation is produced, language models by their nature tend to happily keep on generating similarly unsafe additional text. This is in fact how safety alignment of frontier models gets circumvented in the wild, despite great efforts in improving their safety. Deviating from the paradigm of approaching safety alignment as prevention (decreasing the probability of harmful responses), we propose backtracking, a technique that allows language models to "undo" and recover from their own unsafe generation through the introduction of a special [RESET] token. Our method can be incorporated into either SFT or DPO training to optimize helpfulness and harmlessness. We show that models trained to backtrack are consistently safer than baseline models: backtracking Llama-3-8B is four times more safe than the baseline model (6.1\% $\to$ 1.5\%) in our evaluations without regression in helpfulness. Our method additionally provides protection against four adversarial attacks including an adaptive attack, despite not being trained to do so.

cs.LG↗

arXiv2024

Multi-Agent Reinforcement Learning for Multi-Cell Spectrum and Power Allocation

This paper introduces a novel approach to radio resource allocation in multi-cell wireless networks using a fully scalable multi-agent reinforcement learning (MARL) framework. A distributed method is developed where agents control individual cells and determine spectrum and power allocation based on limited local information, yet achieve quality of service (QoS) performance comparable to centralized methods using global information. The objective is to minimize packet delays across devices under stochastic arrivals and applies to both conflict graph abstractions and cellular network configurations. This is formulated as a distributed learning problem, implementing a multi-agent proximal policy optimization (MAPPO) algorithm with recurrent neural networks and queueing dynamics. This traffic-driven MARL-based solution enables decentralized training and execution, ensuring scalability to large networks. Extensive simulations demonstrate that the proposed methods achieve comparable QoS performance to genie-aided centralized algorithms with significantly less execution time. The trained policies also exhibit scalability and robustness across various network sizes and traffic conditions.

cs.MA↗

arXiv2024

The HitchHiker's Guide to High-Assurance System Observability Protection with Efficient Permission Switches

Protecting system observability records (logs) from compromised OSs has gained significant traction in recent times, with several note-worthy approaches proposed. Unfortunately, none of the proposed approaches achieve high performance with tiny log protection delays. They also leverage risky environments for protection (\eg many use general-purpose hypervisors or TrustZone, which have large TCB and attack surfaces). HitchHiker is an attempt to rectify this problem. The system is designed to ensure (a) in-memory protection of batched logs within a short and configurable real-time deadline by efficient hardware permission switching, and (b) an end-to-end high-assurance environment built upon hardware protection primitives with debloating strategies for secure log protection, persistence, and management. Security evaluations and validations show that HitchHiker reduces log protection delay by 93.3--99.3% compared to the state-of-the-art, while reducing TCB by 9.4--26.9X. Performance evaluations show HitchHiker incurs a geometric mean of less than 6% overhead on diverse real-world programs, improving on the state-of-the-art approach by 61.9--77.5%.

cs.CR↗

arXiv2024

Effective Prompt Extraction from Language Models

The text generated by large language models is commonly controlled by prompting, where a prompt prepended to a user's query guides the model's output. The prompts used by companies to guide their models are often treated as secrets, to be hidden from the user making the query. They have even been treated as commodities to be bought and sold on marketplaces. However, anecdotal reports have shown adversarial users employing prompt extraction attacks to recover these prompts. In this paper, we present a framework for systematically measuring the effectiveness of these attacks. In experiments with 3 different sources of prompts and 11 underlying large language models, we find that simple text-based attacks can in fact reveal prompts with high probability. Our framework determines with high precision whether an extracted prompt is the actual secret prompt, rather than a model hallucination. Prompt extraction from real systems such as Claude 3 and ChatGPT further suggest that system prompts can be revealed by an adversary despite existing defenses in place.

cs.CL↗

arXiv2024

Forcing Diffuse Distributions out of Language Models

Despite being trained specifically to follow user instructions, today's instructiontuned language models perform poorly when instructed to produce random outputs. For example, when prompted to pick a number uniformly between one and ten Llama-2-13B-chat disproportionately favors the number five, and when tasked with picking a first name at random, Mistral-7B-Instruct chooses Avery 40 times more often than we would expect based on the U.S. population. When these language models are used for real-world tasks where diversity of outputs is crucial, such as language model assisted dataset construction, their inability to produce diffuse distributions over valid choices is a major hurdle. In this work, we propose a fine-tuning method that encourages language models to output distributions that are diffuse over valid outcomes. The methods we introduce generalize across a variety of tasks and distributions and make large language models practical for synthetic dataset generation with little human intervention.

cs.CL↗

arXiv2024

A simple hybrid linear and non-linear interpolation finite element for adaptive cracking elements method

Cracking Elements Method (CEM) is a numerical tool to simulate quasi-brittle fractures, which does not need remeshing, nodal enrichment, or complicated crack tracking strategy. The cracking elements used in the CEM can be considered as a special type of finite element implemented in the standard finite element frameworks. One disadvantage of CEM is that it uses nonlinear interpolation of the displacement field (Q8 or T6 elements), introducing more nodes and consequent computing efforts than the cases with elements using linear interpolation of the displacement field. Aiming at solving this problem, we propose a simple hybrid linear and non-linear interpolation finite element for adaptive cracking elements method in this work. A simple strategy is proposed for treating the elements with $p$ edge nodes $p\in\left[0,n\right]$ and $n$ being the edge number of the element. Only a few codes are needed. Then, by only adding edge and center nodes on the elements experiencing cracking and keeping linear interpolation of the displacement field for the elements outside the cracking domain, the number of total nodes was reduced almost to half of the case using the conventional cracking elements. Numerical investigations prove that the new approach inherits all the advantages of CEM with greatly improved computing efficiency.

cs.CE↗

arXiv2024

FoleyCrafter: Bring Silent Videos to Life with Lifelike and Synchronized Sounds

We study Neural Foley, the automatic generation of high-quality sound effects synchronizing with videos, enabling an immersive audio-visual experience. Despite its wide range of applications, existing approaches encounter limitations when it comes to simultaneously synthesizing high-quality and video-aligned (i.e.,, semantic relevant and temporal synchronized) sounds. To overcome these limitations, we propose FoleyCrafter, a novel framework that leverages a pre-trained text-to-audio model to ensure high-quality audio generation. FoleyCrafter comprises two key components: the semantic adapter for semantic alignment and the temporal controller for precise audio-video synchronization. The semantic adapter utilizes parallel cross-attention layers to condition audio generation on video features, producing realistic sound effects that are semantically relevant to the visual content. Meanwhile, the temporal controller incorporates an onset detector and a timestampbased adapter to achieve precise audio-video alignment. One notable advantage of FoleyCrafter is its compatibility with text prompts, enabling the use of text descriptions to achieve controllable and diverse video-to-audio generation according to user intents. We conduct extensive quantitative and qualitative experiments on standard benchmarks to verify the effectiveness of FoleyCrafter. Models and codes are available at https://github.com/open-mmlab/FoleyCrafter.

cs.CV↗

arXiv2024

Zero-Shot Audio Captioning Using Soft and Hard Prompts

In traditional audio captioning methods, a model is usually trained in a fully supervised manner using a human-annotated dataset containing audio-text pairs and then evaluated on the test sets from the same dataset. Such methods have two limitations. First, these methods are often data-hungry and require time-consuming and expensive human annotations to obtain audio-text pairs. Second, these models often suffer from performance degradation in cross-domain scenarios, i.e., when the input audio comes from a different domain than the training set, which, however, has received little attention. We propose an effective audio captioning method based on the contrastive language-audio pre-training (CLAP) model to address these issues. Our proposed method requires only textual data for training, enabling the model to generate text from the textual feature in the cross-modal semantic space.In the inference stage, the model generates the descriptive text for the given audio from the audio feature by leveraging the audio-text alignment from CLAP.We devise two strategies to mitigate the discrepancy between text and audio embeddings: a mixed-augmentation-based soft prompt and a retrieval-based acoustic-aware hard prompt. These approaches are designed to enhance the generalization performance of our proposed model, facilitating the model to generate captions more robustly and accurately. Extensive experiments on AudioCaps and Clotho benchmarks show the effectiveness of our proposed method, which outperforms other zero-shot audio captioning approaches for in-domain scenarios and outperforms the compared methods for cross-domain scenarios, underscoring the generalization ability of our method.

cs.SD↗

arXiv2024

DORY: Deliberative Prompt Recovery for LLM

Prompt recovery in large language models (LLMs) is crucial for understanding how LLMs work and addressing concerns regarding privacy, copyright, etc. The trend towards inference-only APIs complicates this task by restricting access to essential outputs for recovery. To tackle this challenge, we extract prompt-related information from limited outputs and identify a strong(negative) correlation between output probability-based uncertainty and the success of prompt recovery. This finding led to the development of Deliberative PrOmpt RecoverY (DORY), our novel approach that leverages uncertainty to recover prompts accurately. DORY involves reconstructing drafts from outputs, refining these with hints, and filtering out noise based on uncertainty. Our evaluation across diverse LLMs and prompt benchmarks shows that DORY outperforms existing baselines, improving performance by approximately 10.82% and establishing a new state-of-the-art record in prompt recovery tasks. Significantly, DORY operates using a single LLM without any external resources or model, offering a cost-effective, user-friendly prompt recovery solution.

cs.CL↗

arXiv2024

Creating cyclo-N$_5$$^{+}$ cation and assembling N$_5$$^{+}$N$_5$$^{-}$ salt via electronegativity co-matching in tailored ionic compounds

The recent discovery of crystalline pentazolates marks a major advance in polynitrogen science and raises prospects of making the long-touted potent propellant N$_5$$^{+}$N$_5$$^{-}$ salt. However, despite the synthesis of cyclo-N$_5$$^{-}$ anion in pentazolates, counter cation cyclo-N$_5$$^{+}$ remains elusive due to the strong oxidizing power of pentazole ion; moreover, pure N$_5$$^{+}$N$_5$$^{-}$ salt is known to be unstable. Here, we devise a new strategy for making rare cyclo-N$_5$$^{+}$ cation and assembling the long-sought N$_5$$^{+}$N$_5$$^{-}$ salt in tailored ionic compounds, wherein the negative/positive host ions act as oxidizing/reducing agents to form cyclo-N$_5$$^{+}$/N$_5$$^{-}$ species. This strategy is implemented via an advanced computational crystal structure search, which identifies XN$_5$N$_5$F (X = Li, Na, K) compounds that stabilize at high pressures and remain viable at ambient pressure-temperature conditions based on \textit{ab initio} molecular dynamics simulations. This finding opens an avenue for creating and stabilizing N$_5$$^{+}$N$_5$$^{-}$ salt assembly in ionic compounds, where cyclo-N$_5$ species are oxidized/reduced via co-matching with host ions of high/low electronegativity. The present results demonstrate novel polynitrogen chemistry, and these findings offer new insights and prospects in the design and synthesis of diverse chemical species that exhibit unusual charge states, bonding structures, and superior functionality.

cond-mat.mtrl-sci↗

arXiv2024

Low-Overhead Defect-Adaptive Surface Code with Bandage-Like Super-Stabilizers

To make practical quantum algorithms work, large-scale quantum processors protected by error-correcting codes are required to resist noise and ensure reliable computational outcomes. However, a major challenge arises from defects in processor fabrication, as well as occasional losses or cosmic rays during the computing process, all of which can lead to qubit malfunctions and disrupt error-correcting codes' normal operations. In this context, we introduce an automatic adapter to implement the surface code on defective lattices. Unlike previous approaches, this adapter leverages newly proposed bandage-like super-stabilizers to save more qubits when defects are clustered, thus enhancing the code distance and reducing super-stabilizer weight. For instance, in comparison with earlier methods, with a code size of 27 and a random defect rate of 2\%, the disabled qubits decrease by $1/3$, and the average preserved code distance increases by 63\%. This demonstrates a significant reduction in overhead when handling defects using our approach, and this advantage amplifies with increasing processor size and defect rates. Our work presents a low-overhead, automated solution to the challenge of adapting the surface code to defects, an essential step towards scaling up the construction of large-scale quantum computers for practical applications.

quant-ph↗

arXiv2024

fMRI Exploration of Visual Quality Assessment

Despite significant strides in visual quality assessment, the neural mechanisms underlying visual quality perception remain insufficiently explored. This study employed fMRI to examine brain activity during image quality assessment and identify differences in human processing of images with varying quality. Fourteen healthy participants underwent tasks assessing both image quality and content classification while undergoing functional MRI scans. The collected behavioral data was statistically analyzed, and univariate and functional connectivity analyses were conducted on the imaging data. The findings revealed that quality assessment is a more complex task than content classification, involving enhanced activation in high-level cognitive brain regions for fine-grained visual analysis. Moreover, the research showed the brain's adaptability to different visual inputs, adopting different strategies depending on the input's quality. In response to high-quality images, the brain primarily uses specialized visual areas for precise analysis, whereas with low-quality images, it recruits additional resources including higher-order visual cortices and related cognitive and attentional networks to decode and recognize complex, ambiguous signals effectively. This study pioneers the intersection of neuroscience and image quality research, providing empirical evidence through fMRI linking image quality to neural processing. It contributes novel insights into the human visual system's response to diverse image qualities, thereby paving the way for advancements in objective image quality assessment algorithms.

cs.MM↗

arXiv2024

DiffBody: Human Body Restoration by Imagining with Generative Diffusion Prior

Human body restoration plays a vital role in various applications related to the human body. Despite recent advances in general image restoration using generative models, their performance in human body restoration remains mediocre, often resulting in foreground and background blending, over-smoothing surface textures, missing accessories, and distorted limbs. Addressing these challenges, we propose a novel approach by constructing a human body-aware diffusion model that leverages domain-specific knowledge to enhance performance. Specifically, we employ a pretrained body attention module to guide the diffusion model's focus on the foreground, addressing issues caused by blending between the subject and background. We also demonstrate the value of revisiting the language modality of the diffusion model in restoration tasks by seamlessly incorporating text prompt to improve the quality of surface texture and additional clothing and accessories details. Additionally, we introduce a diffusion sampler tailored for fine-grained human body parts, utilizing local semantic information to rectify limb distortions. Lastly, we collect a comprehensive dataset for benchmarking and advancing the field of human body restoration. Extensive experimental validation showcases the superiority of our approach, both quantitatively and qualitatively, over existing methods.

cs.CV↗

arXiv2024

PIA: Your Personalized Image Animator via Plug-and-Play Modules in Text-to-Image Models

Recent advancements in personalized text-to-image (T2I) models have revolutionized content creation, empowering non-experts to generate stunning images with unique styles. While promising, adding realistic motions into these personalized images by text poses significant challenges in preserving distinct styles, high-fidelity details, and achieving motion controllability by text. In this paper, we present PIA, a Personalized Image Animator that excels in aligning with condition images, achieving motion controllability by text, and the compatibility with various personalized T2I models without specific tuning. To achieve these goals, PIA builds upon a base T2I model with well-trained temporal alignment layers, allowing for the seamless transformation of any personalized T2I model into an image animation model. A key component of PIA is the introduction of the condition module, which utilizes the condition frame and inter-frame affinity as input to transfer appearance information guided by the affinity hint for individual frame synthesis in the latent space. This design mitigates the challenges of appearance-related image alignment within and allows for a stronger focus on aligning with motion-related guidance.

cs.CV↗

arXiv2024

RECOST: External Knowledge Guided Data-efficient Instruction Tuning

In the current landscape of large language models (LLMs), the process of instruction tuning serves as an essential step. Considering the high computing power overhead, data-efficient instruction tuning was proposed to reduce the training data size in this process, aiming at selecting high-quality instructional data. Nevertheless, we argue that most current data-efficient instruction-tuning methods are highly dependent on the quality of the original instruction-tuning dataset. When it comes to datasets synthesized by LLMs, a common scenario in this field, dirty samples will even be selected with a higher probability than other samples. To address these challenges, we utilized external knowledge (relevant examples or paragraphs) to evaluate those samples synthesized by LLMs with an in-context-based relative predictive entropy. Based on the new metric, we proposed a framework, dubbed as \textbf{RECOST}, which integrates external-knowledge-base re-ranking and diversity-consistent sampling into a single pipeline. Through extensive experiments on several synthetic datasets (Alpaca and Alpaca-gpt4), we demonstrate the effectiveness of our method and achieve even better results with only \textbf{1\%} of the full dataset.

cs.CL↗