Searcharxiv⌕ Search

arXiv subjects

Yongyi Mao

Publications and source records attributed to Yongyi Mao.

At least 19 recordsLinked to original sources

Fundamental Limits of Statistical Watermarking for LLMs with Fixed and Optimizable Common Randomness

Watermarking has emerged as an important technique for distinguishing AI-generated text from human-written text. However, most existing work is primarily empirical, relies on ad hoc watermarking designs, and lacks a fundamental understanding of the performance limits. In this paper, we propose a unified theoretical framework for statistical watermarking of large language models (LLMs), casting the problem as an independence test between generated text and common randomness and jointly designing the watermarking embedder and the detector. Our objective is to maximize detection power while controlling the worst-case false positive rate (FPR) and limiting distortion to text quality. A central focus of our framework is the design of the common randomness, which plays a critical role in both watermark embedding and detection. We characterize the fundamental limits of statistical watermarking under both fixed and optimizable common randomness. Our results show that optimizing the common randomness does not improve the fundamental limits: the best achievable performance is the same as that with fixed common randomness. Nevertheless, common randomness design remains highly important because it can substantially simplify detector design. In particular, with optimizable common randomness, a fixed, deterministic, and model-agnostic detector can achieve optimal performance. We further study a practical setting with fixed uniform common randomness and a fixed deterministic detector, and characterize its performance gap relative to the optimizable case, thereby demonstrating the practical value of optimizable common randomness.

cs.IT↗

On the Recoverability of Private Information Unlearning in Large Language Models

Large language models (LLMs) can memorize sensitive information, raising serious privacy concerns. Machine unlearning offers a potential solution to remove such information, but it remains unclear whether existing methods truly erase it or merely hide it within the model. A key challenge is quantifying the persistence of sensitive data under a unified evaluation framework. To address this, we construct a synthetic dataset containing fake private information and propose a white-box auditing framework to systematically assess whether claimed-forgotten information is genuinely removed. Using this framework, we evaluate five existing unlearning methods and find that a simple "inverse greedy" decoding -- selecting the least likely token at each step -- can recover supposedly forgotten private information. Our results reveal that current unlearning approaches often fail to fully eliminate sensitive information, highlighting the need for more reliable methods to ensure privacy in deployed LLMs.

cs.LG↗

On the Cost and Benefit of Chain of Thought: A Learning-Theoretic Perspective

We develop a learning-theoretic framework for understanding Chain of Thought (CoT). We model CoT as the interaction between an answer map and a chain rule that generates intermediate questions autoregressively, and define the reasoning risk of a hypothesis under this interaction. Our first result is a tight canonical decomposition of this risk into two terms with opposing roles: an oracle-trajectory risk (OTR), which captures the benefit of CoT and reduces to a target-domain risk in a domain adaptation problem, and a trajectory-mismatch risk (TMR), which captures the cost of CoT through error accumulation along mismatched reasoning trajectories. We then show that this cost is unavoidable without structure: if any one of the loss, the hypothesis answer map, or the chain rule lacks stability, the TMR can be arbitrarily large even when the OTR is zero and the hypothesis is uniformly close to the ground truth. Conversely, under stability, we prove a tight upper bound on the TMR governed by an exact amplification factor that identifies bounded, linear, and exponential error-growth regimes. Together, these results give a precise theory of when CoT helps, when it hurts, and what controls the transition between the two.

cs.LG↗

Parallel Recursive LSTM

Transformers have become the dominant architecture for sequence modeling by using self-attention to enable expressive and highly parallel processing. However, the resulting quadratic time and memory costs limit efficiency in long-context settings. Recurrent models such as LSTMs provide explicit nonlinear state updates and strong state-tracking capabilities, yet their strictly sequential computation limits parallelism. We introduce the Parallel Recursive LSTM (PR-LSTM), a hierarchical recurrent architecture that replaces left-to-right recurrence with recursive nonlinear state composition over a balanced computation tree. Tokens are first mapped independently to latent states, which are then recursively merged by a learned gated composition block. This structure uses the reduction pattern underlying parallel scans as a fixed execution schedule, rather than assuming an associative recurrence. As a result, PR-LSTM retains nonlinear gated state representations while reducing recurrent parallel depth from linear to logarithmic. Empirically, PR-LSTM achieves strong sequence-length generalization on formal-language benchmarks, solving more tasks than standard RNN, LSTM, and Transformer baselines, while avoiding the quadratic scaling of attention. These results suggest that recurrent computation can be reorganized hierarchically to expose parallelism without restricting the transition dynamics to linear or associative forms.

cs.LG↗

Fundamental Trade-Offs in Multi-Bit Watermarking of Stochastic Processes

We study multi-bit watermarking for data generated by stochastic processes, where a hidden message is embedded during sampling and must be decodable by an authorized detector that possesses side information unavailable to unauthorized observers. In high-stakes deployments, a practical watermark must simultaneously control false alarms, preserve generation quality without distorting the output distribution, and support reliable multi-bit decoding. Satisfying all three goals at once inevitably creates fundamental trade-offs. We formulate watermark embedding as a distributional information-embedding problem and watermark detection as a multiple-hypothesis testing problem under distortion and rate constraints, leading to four fundamental metrics: false-alarm probability, detection error probability, distortion, and information rate. Within this information-theoretic framework, we derive matched converse and achievability bounds that characterize the optimal trade-offs and provide scheme-agnostic benchmarks for any watermarking method. For stationary ergodic stochastic processes, we further obtain matched asymptotic limits and connect them to the finite-sample regime. Finally, we present a reference watermarking construction satisfying our assumptions and empirically illustrating the predicted trade-offs.

cs.IT↗

Theoretically Grounded Framework for LLM Watermarking: A Distribution-Adaptive Approach

Watermarking has emerged as a crucial method to distinguish AI-generated text from human-created text. Current watermarking approaches often lack formal optimality guarantees or address the scheme and detector design separately. In this paper, we introduce a novel, unified theoretical framework for watermarking Large Language Models (LLMs) that jointly optimizes both the watermarking scheme and detector. Our approach aims to maximize detection performance while maintaining control over the worst-case false positive rate (FPR) and distortion on text quality. We derive closed-form optimal solutions for this joint design and characterize the fundamental trade-off between watermark detectability and distortion. Notably, we reveal that the optimal watermarking schemes should be adaptive to the LLM's generative distribution. Building on our theoretical insights, we propose a distortion-free, distribution-adaptive watermarking algorithm (DAWA) that leverages a surrogate model for model-agnosticism and efficiency. Experiments on Llama2-13B and Mistral-8$\times$7B models confirm the effectiveness of our approach, particularly at ultra-low FPRs. Our code is available at https://github.com/yepengliu/DAWA.

cs.CR↗

On the Hardness of Unsupervised Domain Adaptation: Optimal Learners and Information-Theoretic Perspective

This paper studies the hardness of unsupervised domain adaptation (UDA) under covariate shift. We model the uncertainty that the learner faces by a distribution $π$ in the ground-truth triples $(p, q, f)$ -- which we call a UDA class -- where $(p, q)$ is the source -- target distribution pair and $f$ is the classifier. We define the performance of a learner as the overall target domain risk, averaged over the randomness of the ground-truth triple. This formulation couples the source distribution, the target distribution and the classifier in the ground truth, and deviates from the classical worst-case analyses, which pessimistically emphasize the impact of hard but rare UDA instances. In this formulation, we precisely characterize the optimal learner. The performance of the optimal learner then allows us to define the learning difficulty for the UDA class and for the observed sample. To quantify this difficulty, we introduce an information-theoretic quantity -- Posterior Target Label Uncertainty (PTLU) -- along with its empirical estimate (EPTLU) from the sample , which capture the uncertainty in the prediction for the target domain. Briefly, PTLU is the entropy of the predicted label in the target domain under the posterior distribution of ground-truth classifier given the observed source and target samples. By proving that such a quantity serves to lower-bound the risk of any learner, we suggest that these quantities can be used as proxies for evaluating the hardness of UDA learning. We provide several examples to demonstrate the advantage of PTLU, relative to the existing measures, in evaluating the difficulty of UDA learning.

stat.ML↗

Adversarial Defenses via Vector Quantization

Adversarial attacks pose significant challenges to the robustness of modern deep neural networks in computer vision, and defending these networks against adversarial attacks has attracted intense research efforts. Among various defense strategies, preprocessing-based defenses are practically appealing since there is no need to train the network under protection. However, such approaches typically do not achieve comparable robustness as other methods such as adversarial training. In this paper, we propose a novel framework for preprocessing-based defenses, where a vector quantizer is used as a preprocessor. This framework, inspired by and extended from Randomized Discretization (RandDisc), is theoretically principled by rate-distortion theory: indeed, RandDisc may be viewed as a scalar quantizer, and rate-distortion theory suggests that such quantization schemes are inferior to vector quantization. In our framework, the preprocessing vector quantizer treats the input image as a collection of patches and finds a set of representative patches based on the patch distributions; each original patch is then modified according to the representative patches close to it. We present two lightweight defenses in this framework, referred to as patched RandDisc (pRD) and sliding-window RandDisc (swRD), where the patches are disjoint in the former and overlapping in the latter. We show that vector-quantization-based defenses have certifiable robust accuracy and that pRD and swRD demonstrate state-of-the-art performances, surpassing RandDisc by a large margin. Notably, the proposed defenses possess the obfuscated gradients property. Our experiments however show that pRD and swRD remain effective under the STE and EOT attacks, which are designed specifically for defenses with gradient obfuscation. ...

cs.LG↗

Distributional Information Embedding: A Framework for Multi-bit Watermarking

This paper introduces a novel problem, distributional information embedding, motivated by the practical demands of multi-bit watermarking for large language models (LLMs). Unlike traditional information embedding, which embeds information into a pre-existing host signal, LLM watermarking actively controls the text generation process--adjusting the token distribution--to embed a detectable signal. We develop an information-theoretic framework to analyze this distributional information embedding problem, characterizing the fundamental trade-offs among three critical performance metrics: text quality, detectability, and information rate. In the asymptotic regime, we demonstrate that the maximum achievable rate with vanishing error corresponds to the entropy of the LLM's output distribution and increases with higher allowable distortion. We also characterize the optimal watermarking scheme to achieve this rate. Extending the analysis to the finite-token case with non-i.i.d. tokens, we identify schemes that maximize detection probability while adhering to constraints on false alarm and distortion.

cs.CR↗

Generalization in Federated Learning: A Conditional Mutual Information Framework

Federated learning (FL) is a widely adopted privacy-preserving distributed learning framework, yet its generalization performance remains less explored compared to centralized learning. In FL, the generalization error consists of two components: the out-of-sample gap, which measures the gap between the empirical and true risk for participating clients, and the participation gap, which quantifies the risk difference between participating and non-participating clients. In this work, we apply an information-theoretic analysis via the conditional mutual information (CMI) framework to study FL's two-level generalization. Beyond the traditional supersample-based CMI framework, we introduce a superclient construction to accommodate the two-level generalization setting in FL. We derive multiple CMI-based bounds, including hypothesis-based CMI bounds, illustrating how privacy constraints in FL can imply generalization guarantees. Furthermore, we propose fast-rate evaluated CMI bounds that recover the best-known convergence rate for two-level FL generalization in the small empirical risk regime. For specific FL model aggregation strategies and structured loss functions, we refine our bounds to achieve improved convergence rates with respect to the number of participating clients. Empirical evaluations confirm that our evaluated CMI bounds are non-vacuous and accurately capture the generalization behavior of FL algorithms.

stat.ML↗

Generalization Bounds via Conditional $f$-Information

In this work, we introduce novel information-theoretic generalization bounds using the conditional $f$-information framework, an extension of the traditional conditional mutual information (MI) framework. We provide a generic approach to derive generalization bounds via $f$-information in the supersample setting, applicable to both bounded and unbounded loss functions. Unlike previous MI-based bounds, our proof strategy does not rely on upper bounding the cumulant-generating function (CGF) in the variational formula of MI. Instead, we set the CGF or its upper bound to zero by carefully selecting the measurable function invoked in the variational formula. Although some of our techniques are partially inspired by recent advances in the coin-betting framework (e.g., Jang et al. (2023)), our results are independent of any previous findings from regret guarantees of online gambling algorithms. Additionally, our newly derived MI-based bound recovers many previous results and improves our understanding of their potential limitations. Finally, we empirically compare various $f$-information measures for generalization, demonstrating the improvement of our new bounds over the previous bounds.

stat.ML↗

On $f$-Divergence Principled Domain Adaptation: An Improved Framework

Unsupervised domain adaptation (UDA) plays a crucial role in addressing distribution shifts in machine learning. In this work, we improve the theoretical foundations of UDA proposed in Acuna et al. (2021) by refining their $f$-divergence-based discrepancy and additionally introducing a new measure, $f$-domain discrepancy ($f$-DD). By removing the absolute value function and incorporating a scaling parameter, $f$-DD obtains novel target error and sample complexity bounds, allowing us to recover previous KL-based results and bridging the gap between algorithms and theory presented in Acuna et al. (2021). Using a localization technique, we also develop a fast-rate generalization bound. Empirical results demonstrate the superior performance of $f$-DD-based learning algorithms over previous works in popular UDA benchmarks.

stat.ML↗

Two Facets of SDE Under an Information-Theoretic Lens: Generalization of SGD via Training Trajectories and via Terminal States

Stochastic differential equations (SDEs) have been shown recently to characterize well the dynamics of training machine learning models with SGD. When the generalization error of the SDE approximation closely aligns with that of SGD in expectation, it provides two opportunities for understanding better the generalization behaviour of SGD through its SDE approximation. Firstly, viewing SGD as full-batch gradient descent with Gaussian gradient noise allows us to obtain trajectory-based generalization bound using the information-theoretic bound from Xu and Raginsky [2017]. Secondly, assuming mild conditions, we estimate the steady-state weight distribution of SDE and use information-theoretic bounds from Xu and Raginsky [2017] and Negrea et al. [2019] to establish terminal-state-based generalization bounds. Our proposed bounds have some advantages, notably the trajectory-based bound outperforms results in Wang and Mao [2022], and the terminal-state-based bound exhibits a fast decay rate comparable to stability-based bounds.

cs.LG↗

On robust overfitting: adversarial training induced distribution matters

Adversarial training may be regarded as standard training with a modified loss function. But its generalization error appears much larger than standard training under standard loss. This phenomenon, known as robust overfitting, has attracted significant research attention and remains largely as a mystery. In this paper, we first show empirically that robust overfitting correlates with the increasing generalization difficulty of the perturbation-induced distributions along the trajectory of adversarial training (specifically PGD-based adversarial training). We then provide a novel upper bound for generalization error with respect to the perturbation-induced distributions, in which a notion of the perturbation operator, referred to "local dispersion", plays an important role. Experimental results are presented to validate the usefulness of the bound and various additional insights are provided.

cs.LG↗

Narrowing the Gap between Supervised and Unsupervised Sentence Representation Learning with Large Language Model

Sentence Representation Learning (SRL) is a fundamental task in Natural Language Processing (NLP), with the Contrastive Learning of Sentence Embeddings (CSE) being the mainstream technique due to its superior performance. An intriguing phenomenon in CSE is the significant performance gap between supervised and unsupervised methods, with their only difference lying in the training data. Previous works attribute this performance gap to differences in two representation properties (alignment and uniformity). However, since alignment and uniformity only measure the results, they fail to answer "What aspects of the training data contribute to the performance gap?" and "How can the performance gap be narrowed?", In this paper, we conduct empirical experiments to answer these "What" and "How" questions. We first answer the "What" question by thoroughly comparing the behavior of supervised and unsupervised CSE during their respective training processes. From the comparison, we identify the similarity pattern as a key factor to the performance gap, and introduce a metric, called Relative Fitting Difficulty (RFD), to measure the complexity of the similarity pattern. Then, based on the insights gained from the "What" question, we tackle the "How" question by increasing the pattern complexity of the training data. We achieve this by leveraging the In-Context Learning (ICL) capability of the Large Language Model (LLM) to generate data that simulates complex patterns. By utilizing the hierarchical patterns in the LLM-generated data, we effectively narrow the gap between supervised and unsupervised CSE. We release our codes and appendix at https://github.com/BDBC-KG-NLP/NGCSE.

cs.CL↗

Sample-Conditioned Hypothesis Stability Sharpens Information-Theoretic Generalization Bounds

We present new information-theoretic generalization guarantees through the a novel construction of the "neighboring-hypothesis" matrix and a new family of stability notions termed sample-conditioned hypothesis (SCH) stability. Our approach yields sharper bounds that improve upon previous information-theoretic bounds in various learning scenarios. Notably, these bounds address the limitations of existing information-theoretic bounds in the context of stochastic convex optimization (SCO) problems, as explored in the recent work by Haghifam et al. (2023).

stat.ML↗

Anaphor Assisted Document-Level Relation Extraction

Document-level relation extraction (DocRE) involves identifying relations between entities distributed in multiple sentences within a document. Existing methods focus on building a heterogeneous document graph to model the internal structure of an entity and the external interaction between entities. However, there are two drawbacks in existing methods. On one hand, anaphor plays an important role in reasoning to identify relations between entities but is ignored by these methods. On the other hand, these methods achieve cross-sentence entity interactions implicitly by utilizing a document or sentences as intermediate nodes. Such an approach has difficulties in learning fine-grained interactions between entities across different sentences, resulting in sub-optimal performance. To address these issues, we propose an Anaphor-Assisted (AA) framework for DocRE tasks. Experimental results on the widely-used datasets demonstrate that our model achieves a new state-of-the-art performance.

cs.CL↗

Adversarial Word Dilution as Text Data Augmentation in Low-Resource Regime

Data augmentation is widely used in text classification, especially in the low-resource regime where a few examples for each class are available during training. Despite the success, generating data augmentations as hard positive examples that may increase their effectiveness is under-explored. This paper proposes an Adversarial Word Dilution (AWD) method that can generate hard positive examples as text data augmentations to train the low-resource text classification model efficiently. Our idea of augmenting the text data is to dilute the embedding of strong positive words by weighted mixing with unknown-word embedding, making the augmented inputs hard to be recognized as positive by the classification model. We adversarially learn the dilution weights through a constrained min-max optimization process with the guidance of the labels. Empirical studies on three benchmark datasets show that AWD can generate more effective data augmentations and outperform the state-of-the-art text data augmentation methods. The additional analysis demonstrates that the data augmentations generated by AWD are interpretable and can flexibly extend to new examples without further training.

cs.CL↗