arXiv subjects

Geng Liu

Publications and source records attributed to Geng Liu.

At least 19 recordsLinked to original sources

arXiv2026

Conditions for Global Optimality in Quantum Arimoto-Blahut Algorithms

Generalized Arimoto--Blahut (AB) algorithms are widely used in information theory and quantum optimization, but monotonic objective decrease and numerical stabilization do not guarantee global optimality. We establish necessary and sufficient conditions for the global optimality of full-rank AB fixed points for convex differentiable objectives under linear constraints. An AB fixed point is globally optimal if and only if the AB update direction and the objective gradient differ by an element of the constraint normal space at that point. The same compatibility condition characterizes agreement between individual AB and mirror-descent (MD) updates, whereas pathwise equivalence requires it along the entire common trajectory. Thus, an AB algorithm may follow a trajectory different from MD and still reach the global optimum. We also develop a posteriori optimality certificates based on feasible directional derivatives, including finite-difference upper bounds on the objective gap that require only objective evaluations. For channel relative entropy between dephasing and depolarizing channels, we analytically identify the global minimizer as the unique full-rank AB fixed point, although pathwise equivalence fails. Numerical experiments illustrate convergence of AB and MD along different trajectories and validate the certificates. By contrast, an amplitude-damping example shows that a monotone AB iteration can stabilize at a suboptimal fixed point, whose nonoptimality is detected by the finite-difference certificate. These results provide structural and computable criteria for assessing global optimality in quantum AB algorithms.

cs.IT↗

arXiv2026

Denser $\neq$ Better: Limits of On-Policy Self-Distillation for Continual Post-Training

Continual post-training enables foundation models to acquire new knowledge while preserving existing capabilities. Recent work suggests that on-policy learning can mitigate forgetting, with on-policy self-distillation emerging as a particularly attractive approach. In this work, we revisit this optimistic view through self-distillation policy optimization (SDPO). Our experiments show that SDPO can accelerate in-domain specialization when teacher signals are stable and well aligned, but it struggles to generalize to out-of-distribution scenarios. In continual post-training, SDPO exhibits stronger forgetting and can even collapse, whereas on-policy reinforcement learning methods such as GRPO adapt more conservatively and better preserve prior capabilities. Further analyses reveal that denser self-distillation induces larger drift in both parameter space and response space, and can amplify high-frequency formatting artifacts through a self-reinforcing teacher--student loop. These findings suggest that on-policy data alone is insufficient for continual learning. Dense self-distillation can accelerate specialization when teacher targets are stable and token-level supervision is reliable, but it should not be treated as a default stabilizer for continual post-training. Our code is available at https://github.com/Moenupa/SDPO-CL.

cs.LG↗

arXiv2026

Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation

Multi-turn conversation has emerged as a predominant interaction paradigm for Large Language Models (LLMs). Users often employ follow-up questions to refine their intent, expecting LLMs to adapt dynamically. However, recent research reveals that LLMs suffer a substantial performance drop in multi-turn settings compared to single-turn interactions with fully specified instructions, a phenomenon termed ``Lost in Conversation'' (LiC). While this prior work attributes LiC to model unreliability, we argue that the root cause lies in an intent alignment gap rather than intrinsic capability deficits. In this paper, we first demonstrate that LiC is not a failure of model capability but rather a breakdown in interaction between users and LLMs. We theoretically show that scaling model size or improving training alone cannot resolve this gap, as it arises from structural ambiguity in conversational context rather than representational limitations. To address this, we propose to decouple intent understanding from task execution through a Mediator-Assistant architecture. By utilizing an experience-driven Mediator to explicate user inputs into explicit, well-structured instructions based on historical interaction patterns, our approach effectively bridges the gap between vague user intent and model interpretation. Experimental results demonstrate that this method significantly mitigates performance degradation in multi-turn conversations across diverse LLMs.

cs.CL↗

arXiv2026

A Posteriori Certification Framework for Generalized Quantum Arimoto-Blahut Algorithms

The generalized quantum Arimoto--Blahut (QAB) algorithm is a powerful derivative-free iterative method in quantum information theory. A key obstacle to its broader use is that existing convergence guarantees typically rely on analytical conditions that are either overly restrictive or difficult to verify for concrete problems. We address this issue by introducing an a posteriori certification viewpoint: instead of requiring fully a priori verifiable assumptions, we provide convergence and error guarantees that can be validated directly from the iterates produced by the algorithm. Specifically, we prove a generalized global convergence theorem showing that, under convexity and a substantially weaker numerically verifiable condition, the QAB iteration converges to the global minimizer. This theorem yields a practical certification procedure: by checking explicit inequalities along the computed trajectory, one can certify global optimality and bound the suboptimality of the obtained value. As an application, we develop a certified iterative scheme for computing the quantum relative entropy of channels, a fundamental measure of distinguishability in quantum dynamics. This quantity is notoriously challenging to evaluate numerically: gradient-based methods are impeded by the complexity of matrix functions such as square roots and logarithms, while recent semidefinite programming approaches can become computationally and memory intensive at high precision. Our method avoids these bottlenecks by combining the QAB iteration with a posteriori certification, yielding an efficient and scalable algorithm. Numerical experiments demonstrate rapid convergence and improved scalability and adaptivity compared with SDP-based approaches.

quant-ph↗

arXiv2025

Evaluating Reliability Asymmetries in Chinese Factual Search and AI Answers

Search engines and AI-powered systems increasingly mediate access to factual information, yet their reliability remains difficult to evaluate in realistic information-seeking settings. We study this problem in the Chinese web ecosystem by constructing a query-based fact-checking dataset from real Chinese search logs and comparing nine systems across traditional search engines, standalone large language models, and search-integrated AI Overviews. Focusing on factual Chinese-language factual Yes/No questions, we evaluate whether systems provide correct, incorrect, or uncertain decisions against evidence-derived ground truth. We find that systems are similarly accurate when they provide definitive answers, but differ sharply in how often they do so. Conditional accuracy ranges from 73.2% to 78.9%, yet search engines answer definitively on over 83% of queries, while Qwen-Max does so on fewer than half. We also find a consistent polarity gap: all systems perform better on yes-labeled queries than on no-labeled queries. We also use Baidu Index data to identify Chinese provinces with higher health-related search attention, which may indicate greater potential exposure to misinformation. Overall, our results show that reliability depends not only on whether systems are correct when they answer, but also on how often they answer, how they handle negative claims, and where information demand may increase exposure risks.

cs.IR↗

arXiv2025

Probing Social Identity Bias in Chinese LLMs with Gendered Pronouns and Social Groups

Large language models (LLMs) are increasingly deployed in user-facing applications, raising concerns that they may reflect and amplify social biases. We investigate social identity biases in Chinese LLMs using Mandarin-specific prompts across ten representative models. Our evaluation compares ingroup ("We") and outgroup ("They") framings across 240 social groups salient in the Chinese context, using a two-tiered measurement framework that assesses both sentiment and toxicity. The prompt design explicitly accounts for linguistic properties of Mandarin, including the distinction between the default gender-neutral plural pronoun and its explicitly feminine counterpart, enabling a controlled comparison of social identity framing effects. Across models, we observe systematic ingroup-outgroup asymmetries, although their expression differs across measurement dimensions. In particular, instruction tuning often reduces sentiment asymmetries, while toxicity gaps remain more persistent. Moreover, the feminine-marked plural pronoun is associated with higher toxicity than the default gender-neutral plural in several models. Our study introduces a language-aware evaluation framework for Chinese LLMs and shows that (i) social identity biases previously documented in English also manifest in Chinese and that (ii) Mandarin-specific linguistic structure can reveal bias patterns that are not directly observable in English-only settings.

cs.CL↗

arXiv2025

Towards an Automated Framework to Audit Youth Safety on TikTok

This paper investigates the effectiveness of TikTok's enforcement mechanisms for limiting the exposure of harmful content to youth accounts. We collect over 7000 videos, classify them as harmful vs not-harmful, and then simulate interactions using age-specific sockpuppet accounts through both passive and active engagement strategies. We also evaluate the performance of large language (LLMs) and vision-language models (VLMs) in detecting harmful content, identifying key challenges in precision and scalability. Preliminary results show minimal differences in content exposure between adult and youth accounts, raising concerns about the platform's age-based moderation. These findings suggest that the platform needs to strengthen youth safety measures and improve transparency in content moderation.

cs.CY↗

arXiv2025

Evaluating Chinese Large Language Models: The Influence of Persona Assignment on Stereotypes and Safeguards

Recent research has highlighted that assigning specific personas to large language models (LLMs) can significantly increase harmful content generation. However, limited attention has been given to persona-driven toxicity in non-Western contexts, particularly in Chinese-based LLMs. In this paper, we perform a large-scale, cross-model analysis of refusal behavior and persona-driven toxicity amplification across four Chinese LLMs, leveraging a comprehensive dataset of over 1,400,000 generated texts. We identify significant disparities in persona-driven refusal behavior, including systematic gender differences in refusal triggering across the evaluated Chinese LLMs. Furthermore, we provide quantitative evidence of persona-driven toxicity amplification with respect to model default baselines. We show that this amplification--whose magnitude varies substantially across models--is driven by interactions across several factors, involving persona conditioning, prompting strategy, target social group, and model-specific safety mechanisms. Leveraging model-specific regression analyses, we systematically characterize how persona categories, target social groups, and prompt templates independently and jointly shape both refusal behavior and output toxicity. As a complementary case study, we further explore an iterative, evaluator-guided mitigation strategy based on model feedback with an external LLM evaluator, demonstrating that highly toxic outputs can be substantially reduced without costly model retraining. Overall, our findings highlight the importance of culturally contextualized safety evaluations for Chinese-language LLMs and provide a structured framework for assessing persona-induced risks and exploratory mitigation strategies in LLM-generated content.

cs.CY↗

arXiv2025

Semi-parametric Memory Consolidation: Towards Brain-like Deep Continual Learning

Humans and most animals inherently possess a distinctive capacity to continually acquire novel experiences and accumulate worldly knowledge over time. This ability, termed continual learning, is also critical for deep neural networks (DNNs) to adapt to the dynamically evolving world in open environments. However, DNNs notoriously suffer from catastrophic forgetting of previously learned knowledge when trained on sequential tasks. In this work, inspired by the interactive human memory and learning system, we propose a novel biomimetic continual learning framework that integrates semi-parametric memory and the wake-sleep consolidation mechanism. For the first time, our method enables deep neural networks to retain high performance on novel tasks while maintaining prior knowledge in real-world challenging continual learning scenarios, e.g., class-incremental learning on ImageNet. This study demonstrates that emulating biological intelligence provides a promising path to enable deep neural networks with continual learning capabilities.

cs.LG↗

arXiv2025

Analyzing the Safety of Japanese Large Language Models in Stereotype-Triggering Prompts

In recent years, Large Language Models have attracted growing interest for their significant potential, though concerns have rapidly emerged regarding unsafe behaviors stemming from inherent stereotypes and biases. Most research on stereotypes in LLMs has primarily relied on indirect evaluation setups, in which models are prompted to select between pairs of sentences associated with particular social groups. Recently, direct evaluation methods have emerged, examining open-ended model responses to overcome limitations of previous approaches, such as annotator biases. Most existing studies have focused on English-centric LLMs, whereas research on non-English models, particularly Japanese, remains sparse, despite the growing development and adoption of these models. This study examines the safety of Japanese LLMs when responding to stereotype-triggering prompts in direct setups. We constructed 3,612 prompts by combining 301 social group terms, categorized by age, gender, and other attributes, with 12 stereotype-inducing templates in Japanese. Responses were analyzed from three foundational models trained respectively on Japanese, English, and Chinese language. Our findings reveal that LLM-jp, a Japanese native model, exhibits the lowest refusal rate and is more likely to generate toxic and negative responses compared to other models. Additionally, prompt format significantly influence the output of all models, and the generated responses include exaggerated reactions toward specific social groups, varying across models. These findings underscore the insufficient ethical safety mechanisms in Japanese LLMs and demonstrate that even high-accuracy models can produce biased outputs when processing Japanese-language prompts. We advocate for improving safety mechanisms and bias mitigation strategies in Japanese LLMs, contributing to ongoing discussions on AI ethics beyond linguistic boundaries.

cs.CL↗

arXiv2024

Comparing diversity, negativity, and stereotypes in Chinese-language AI technologies: an investigation of Baidu, Ernie and Qwen

Large Language Models (LLMs) and search engines have the potential to perpetuate biases and stereotypes by amplifying existing prejudices in their training data and algorithmic processes, thereby influencing public perception and decision-making. While most work has focused on Western-centric AI technologies, we study Chinese-based tools by investigating social biases embedded in the major Chinese search engine, Baidu, and two leading LLMs, Ernie and Qwen. Leveraging a dataset of 240 social groups across 13 categories describing Chinese society, we collect over 30k views encoded in the aforementioned tools by prompting them for candidate words describing such groups. We find that language models exhibit a larger variety of embedded views compared to the search engine, although Baidu and Qwen generate negative content more often than Ernie. We also find a moderate prevalence of stereotypes embedded in the language models, many of which potentially promote offensive and derogatory views. Our work highlights the importance of promoting fairness and inclusivity in AI technologies with a global perspective.

cs.CY↗

arXiv2024

A comparison of online search engine autocompletion in Google and Baidu

Warning: This paper contains content that may be offensive or upsetting. Online search engine auto-completions make it faster for users to search and access information. However, they also have the potential to reinforce and promote stereotypes and negative opinions about a variety of social groups. We study the characteristics of search auto-completions in two different linguistic and cultural contexts: Baidu and Google. We find differences between the two search engines in the way they suppress or modify original queries, and we highlight a concerning presence of negative suggestions across all social groups. Our study highlights the need for more refined, culturally sensitive moderation strategies in current language technologies.

cs.CY↗

arXiv2023

Generalized quantum Arimoto-Blahut algorithm and its application to quantum information bottleneck

We generalize the quantum Arimoto-Blahut algorithm by Ramakrishnan et al. (IEEE Trans. IT, 67, 946 (2021)) to a function defined over a set of density matrices with linear constraints so that our algorithm can be applied to optimizations of quantum operations. This algorithm has wider applicability. Hence, we apply our algorithm to the quantum information bottleneck with three quantum systems, which can be used for quantum learning. We numerically compare our obtained algorithm with the existing algorithm by Grimsmo and Still (Phys. Rev. A, 94, 012338 (2016)). Our numerical analysis shows that our algorithm is better than their algorithm.

quant-ph↗

arXiv2023

Quantum sequential scattering model for quantum state learning

Learning probability distribution is an essential framework in classical learning theory. As a counterpart, quantum state learning has spurred the exploration of quantum machine learning theory. However, as dimensionality increases, learning a high-dimensional unknown quantum state via conventional quantum neural network approaches remains challenging due to trainability issues. In this work, we devise the quantum sequential scattering model (QSSM), inspired by the classical diffusion model, to overcome this scalability issue. Training of our model could effectively circumvent the vanishing gradient problem to a large class of high-dimensional target states possessing polynomial-scaled Schmidt ranks. Theoretical analysis and numerical experiments provide evidence for our model's effectiveness in learning both physical and algorithmic meaningful quantum states and show an out-performance beating the conventional approaches in training speed and learning accuracy. Our work has indicated that an increasing entanglement, a property of quantum states, in the target states, necessitates a larger scaled model, which could reduce our model's learning performance and efficiency.

quant-ph↗

arXiv2023

TDG: Text-guided Domain Generalization

Domain generalization (DG) attempts to generalize a model trained on single or multiple source domains to the unseen target domain. Benefiting from the success of Visual-and-Language Pre-trained models in recent years, we argue that it is crucial for domain generalization by introducing extra text information. In this paper, we develop a novel Text-guided Domain Generalization (TDG) paradigm for domain generalization, which includes three following aspects. Specifically, we first devise an automatic words generation method to extend the description of current domains with novel domain-relevant words. Then, we embed the generated domain information into the text feature space, by the proposed prompt learning-based text feature generation method, which shares a common representation space with the image feature. Finally, we utilize both input image features and generated text features to train a specially designed classifier that generalizes well on unseen target domains, while the image encoder is also updated under the supervision of gradients back propagated from the classifier. Our experimental results show that the techniques incorporated by TDG contribute to the performance in an easy implementation manner. Experimental results on several domain generalization benchmarks show that our proposed framework achieves superior performance by effectively leveraging generated text information in domain generalization.

cs.CV↗

arXiv2023

MDVSC -- Wireless Model Division Video Semantic Communication

In this paper, we propose a new wireless video communication scheme to achieve high-efficiency video transmission over noisy channels. It exploits the idea of model division multiple access (MDMA) and extracts common semantic features across video frames. Besides, deep joint source-channel coding (JSCC) is applied to overcome the distortion caused by noisy channels. The proposed framework is collected under the name model division video semantic communication (MDVSC). In particular, temporal relative video frames are first transformed into a latent space for computing complexity reduction and data redistribution. Accordingly, a novel entropy-based variable length coding is developed further to compress semantic information under the communication bandwidth cost limitation. The whole MDVSC is an end-to-end learnable system. It can be formulated as an optimization problem whose goal is to minimize end-to-end transmission distortion under restricted communication resources. Across standard video source test sequences, test results show that the MDVSC outperforms traditional wireless video coding schemes generally under perceptual quality metrics and has the ability to control code length precisely.

cs.MM↗

arXiv2023

ITA-ELECTION-2022: A multi-platform dataset of social media conversations around the 2022 Italian general election

Online social media play a major role in shaping public discourse and opinion, especially during political events. We present the first public multi-platform dataset of Italian-language political conversations, focused on the 2022 Italian general election taking place on September 25th. Leveraging public APIs and a keyword-based search, we collected millions of posts published by users, pages and groups on Facebook, Instagram and Twitter, along with metadata of TikTok and YouTube videos shared on these platforms, over a period of four months. We augmented the dataset with a collection of political ads sponsored on Meta platforms, and a list of social media handles associated with political representatives. Our data resource will allow researchers and academics to further our understanding of the role of social media in the democratic process.

cs.SI↗

arXiv2023

A Specific Task-oriented Semantic Image Communication System for substation patrol inspection

Intelligent inspection robots are widely used in substation patrol inspection, which can help check potential safety hazards by patrolling the substation and sending back scene images. However, when patrolling some marginal areas with weak signal, the scene images cannot be sucessfully transmissted to be used for hidden danger elimination, which greatly reduces the quality of robots'daily work. To solve such problem, a Specific Task-oriented Semantic Communication System for Imag-STSCI is designed, which involves the semantic features extraction, transmission, restoration and enhancement to get clearer images sent by intelligent robots under weak signals. Inspired by that only some specific details of the image are needed in such substation patrol inspection task, we proposed a new paradigm of semantic enhancement in such specific task to ensure the clarity of key semantic information when facing a lower bit rate or a low signal-to-noise ratio situation. Across the reality-based simulation, experiments show our STSCI can generally surpass traditional image-compression-based and channel-codingbased or other semantic communication system in the substation patrol inspection task with a lower bit rate even under a low signal-to-noise ratio situation.

cs.CV↗