Lin Liu

Hanbo Huang, Yiran Zhang, Hao Zheng et al.

Large language model (LLM) watermarking has shown promise in detecting AI-generated content and mitigating misuse, with prior work claiming robustness against paraphrasing and text editing. In this paper, we argue that existing evaluations are not sufficiently adversarial, obscuring critical vulnerabilities and overstating the security. To address this, we introduce the adaptive robustness radius, a formal metric that quantifies the worst-case resilience of watermarks against adaptive adversaries. By lifting the paraphrase space into a KL-divergence ball, we approximate this radius and theoretically demonstrate that optimizing the attack context and model parameters can significantly reduce the approximate radius, making watermarks highly vulnerable to paraphrase attacks. Leveraging this insight, we propose RLCracker, a reinforcement learning (RL)-based adaptive attack that erases watermark signals with limited watermarked examples and limited access to the detector. Despite weak supervision, it empowers a 3B model to achieve 98.5% removal success with minimal semantic shift on 1,500-token Unigram-marked texts after training on only 100 short samples. This performance dramatically exceeds 6.75% by GPT-4o and generalizes across five model sizes over ten watermarking schemes. Our code is available at https://github.com/OTT0-OTO/RLCracker.

Mingming Li, Huimu Wang, Zuxu Chen et al.

Generative retrieval introduces a groundbreaking paradigm to document retrieval by directly generating the identifier of a pertinent document in response to a specific query. This paradigm has demonstrated considerable benefits and potential, particularly in representation and generalization capabilities, within the context of large language models. However, it faces significant challenges in E-commerce search scenarios, including the complexity of generating detailed item titles from brief queries, the presence of noise in item titles with weak language order, issues with long-tail queries, and the interpretability of results. To address these challenges, we have developed an innovative framework for E-commerce search, called generative retrieval with preference optimization. This framework is designed to effectively learn and align an autoregressive model with target data, subsequently generating the final item through constraint-based beam search. By employing multi-span identifiers to represent raw item titles and transforming the task of generating titles from queries into the task of generating multi-span identifiers from queries, we aim to simplify the generation process. The framework further aligns with human preferences using click data and employs a constrained search method to identify key spans for retrieving the final item, thereby enhancing result interpretability. Our extensive experiments show that this framework achieves competitive performance on a real-world dataset, and online A/B tests demonstrate the superiority and effectiveness in improving conversion gains.

Zhirong Huang, Shichao Zhang, Debo Cheng et al.

In recommender systems, popularity and conformity biases undermine recommender effectiveness by disproportionately favouring popular items, leading to their over-representation in recommendation lists and causing an unbalanced distribution of user-item historical data. We construct a causal graph to address both biases and describe the abstract data generation mechanism. Then, we use it as a guide to develop a novel Debiased Contrastive Learning framework for Mitigating Dual Biases, called DCLMDB. In DCLMDB, both popularity bias and conformity bias are handled in the model training process by contrastive learning to ensure that user choices and recommended items are not unduly influenced by conformity and popularity. Extensive experiments on two real-world datasets, Movielens-10M and Netflix, show that DCLMDB can effectively reduce the dual biases, as well as significantly enhance the accuracy and diversity of recommendations.

Hanbo Huang, Yihan Li, Bowen Jiang et al.

Privacy-sensitive users require deploying large language models (LLMs) within their own infrastructure (on-premises) to safeguard private data and enable customization. However, vulnerabilities in local environments can lead to unauthorized access and potential model theft. To address this, prior research on small models has explored securing only the output layer within hardware-secured devices to balance model confidentiality and customization. Yet this approach fails to protect LLMs effectively. In this paper, we discover that (1) query-based distillation attacks targeting the secured top layer can produce a functionally equivalent replica of the victim model; (2) securing the same number of layers, bottom layers before a transition layer provide stronger protection against distillation attacks than top layers, with comparable effects on customization performance; and (3) the number of secured layers creates a trade-off between protection and customization flexibility. Based on these insights, we propose SOLID, a novel deployment framework that secures a few bottom layers in a secure environment and introduces an efficient metric to optimize the trade-off by determining the ideal number of hidden layers. Extensive experiments on five models (1.3B to 70B parameters) demonstrate that SOLID outperforms baselines, achieving a better balance between protection and downstream customization.

Debo Cheng, Ziqi Xu, Jiuyong Li et al.

Querying causal effects from time-series data is important across various fields, including healthcare, economics, climate science, and epidemiology. However, this task becomes complex in the existence of time-varying latent confounders, which affect both treatment and outcome variables over time and can introduce bias in causal effect estimation. Traditional instrumental variable (IV) methods are limited in addressing such complexities due to the need for predefined IVs or strong assumptions that do not hold in dynamic settings. To tackle these issues, we develop a novel Time-varying Conditional Instrumental Variables (CIV) for Debiasing causal effect estimation, referred to as TDCIV. TDCIV leverages Long Short-Term Memory (LSTM) and Variational Autoencoder (VAE) models to disentangle and learn the representations of time-varying CIV and its conditioning set from proxy variables without prior knowledge. Under the assumptions of the Markov property and availability of proxy variables, we theoretically establish the validity of these learned representations for addressing the biases from time-varying latent confounders, thus enabling accurate causal effect estimation. Our proposed TDCIV is the first to effectively learn time-varying CIV and its associated conditioning set without relying on domain-specific knowledge.