arXiv subjects

Liang Pang

Publications and source records attributed to Liang Pang.

At least 37 records · Page 2Linked to original sources

arXiv2026

D-Models and E-Models: Diversity-Stability Trade-offs in the Sampling Behavior of Large Language Models

The predictive probability of the next token (P_token) in large language models (LLMs) is inextricably linked to the probability of relevance for the next piece of information, the purchase probability of the next product, and the execution probability of the next action-all of which fall under the scope of the task-level target distribution (P_task). While LLMs are known to generate samples that approximate real-world distributions, whether their fine-grained sampling probabilities faithfully align with task requirements remains an open question. Through controlled distribution-sampling simulations, we uncover a striking dichotomy in LLM behavior, distinguishing two model types: D-models (e.g. Qwen-2.5), whose P_token exhibits large step-to-step variability and poor alignment with P_task; and E-models (e.g. Mistral-Small), whose P_token is more stable and better aligned with P_task. We further evaluate these two model types in downstream tasks such as code generation and recommendation, revealing systematic trade-offs between diversity and stability that shape task outcomes. Finally, we analyze the internal properties of both model families to probe their underlying mechanisms. These findings offer foundational insights into the probabilistic sampling behavior of LLMs and provide practical guidance on when to favor D- versus E-models. For web-scale applications, including recommendation, search, and conversational agents, our results inform model selection and configuration to balance diversity with reliability under real-world uncertainty, providing a better level of interpretation.

cs.CL↗

arXiv2026

Less is More: Compact Clue Selection for Efficient Retrieval-Augmented Generation Reasoning

Current RAG retrievers are designed primarily for human readers, emphasizing complete, readable, and coherent paragraphs. However, Large Language Models (LLMs) benefit more from precise, compact, and well-structured input, which enhances reasoning quality and efficiency. Existing methods rely on reranking or summarization to identify key sentences, but may introduce semantic breaks and unfaithfulness. Thus, efficiently extracting and organizing answer-relevant clues from large-scale documents while reducing LLM reasoning costs remains challenging in RAG systems. Inspired by Occam's razor, we frame LLM-centric retrieval as MinMax optimization: maximizing the extraction of potential clues and reranking them for well-organization, while minimizing reasoning costs by truncating to the smallest sufficient set of clues. In this paper, we propose CompSelect, a compact clue selection mechanism for LLM-centric RAG, consisting of a clue extractor, a reranker, and a truncator. (1) The clue extractor first uses answer-containing sentences as fine-tuning targets, aiming to extract sufficient potential clues; (2) The reranker is trained to prioritize effective clues based on real LLM feedback; (3) The truncator uses the truncated text containing the minimum sufficient clues for answering the question as fine-tuning targets, thereby enabling efficient RAG reasoning. Experiments on three QA datasets demonstrate that CompSelect improves performance while reducing both total and online latency compared to a range of baseline methods. Further analysis also confirms its robustness to unreliable retrieval and generalization across different scenarios.

cs.CL↗

arXiv2026

GeoVLMath: Enhancing Geometry Reasoning in Vision-Language Models via Cross-Modal Reward for Auxiliary Line Creation

Auxiliary lines are essential for solving complex geometric problems but remain challenging for large vision-language models (LVLMs). Recent attempts construct auxiliary lines via code-driven rendering, a strategy that relies on accurate and executable code generation to produce visual renderings of the auxiliary lines for subsequent reasoning. However, in complex solid geometry settings, such a strong dependence on precise specifications substantially restricts the robustness of this strategy. Alternatively, we turn to a simpler and more stable solution, representing auxiliary-line constructions as structured textual descriptions. To bridge the gap between textual descriptions and spatial structure, we propose a reinforcement learning framework that enhances diagram-text alignment. The core is a cross-modal reward model that evaluates how well the generated auxiliary-line description matches the ground-truth auxiliary-line diagram. The reward signal drives a GRPO-based RL stage to yield informative auxiliary-line descriptions for the reasoning. To support the training and evaluation, we develop a scalable data pipeline and construct AuxSolidMath, a dataset of 3,018 real-exam geometry problems with paired diagrams and aligned textual fields. Based on this framework, we derive GeoVLMath, an LVLM for solving complex solid geometry.

cs.CV↗

arXiv2026

The Evolution of Thought: Tracking LLM Overthinking via Reasoning Dynamics Analysis

Test-time scaling via explicit reasoning trajectories significantly boosts large language model (LLM) performance but often triggers overthinking. To explore this, we analyze reasoning through two lenses: Reasoning Length Dynamics, which reveals a compensatory trade-off between thinking and answer content length that eventually leads to thinking redundancy, and Reasoning Semantic Dynamics, which identifies semantic convergence and repetitive oscillations. These dynamics uncover an instance-specific Reasoning Completion Point (RCP), beyond which computation continues without further performance gain. Since the RCP varies across instances, we propose a Reasoning Completion Point Detector (RCPD), an inference-time early-exit method that identifies the RCP by monitoring the rank dynamics of termination tokens (e.g., ). Across AIME and GPQA benchmarks using Qwen3 and DeepSeek-R1, RCPD reduces token usage by up to 44% while preserving accuracy, offering a principled approach to efficient test-time scaling.

cs.CL↗

arXiv2026

Circular Reasoning: Understanding Self-Reinforcing Loops in Large Reasoning Models

Despite the success of test-time scaling, Large Reasoning Models (LRMs) frequently encounter repetitive loops that lead to computational waste and inference failure. In this paper, we identify a distinct failure mode termed Circular Reasoning. Unlike traditional model degeneration, this phenomenon manifests as a self-reinforcing trap where generated content acts as a logical premise for its own recurrence, compelling the reiteration of preceding text. To systematically analyze this phenomenon, we introduce LoopBench, a dataset designed to capture two distinct loop typologies: numerical loops and statement loops. Mechanistically, we characterize circular reasoning as a state collapse exhibiting distinct boundaries, where semantic repetition precedes textual repetition. We reveal that reasoning impasses trigger the loop onset, which subsequently persists as an inescapable cycle driven by a self-reinforcing V-shaped attention mechanism. Guided by these findings, we employ the Cumulative Sum (CUSUM) algorithm to capture these precursors for early loop prediction. Experiments across diverse LRMs validate its accuracy and elucidate the stability of long-chain reasoning.

cs.AI↗

arXiv2026

Projecting Out the Malice: A Global Subspace Approach to LLM Detoxification

Large language models (LLMs) exhibit exceptional performance but pose inherent risks of generating toxic content, restricting their safe deployment. While traditional methods (e.g., alignment) adjust output preferences, they fail to eliminate underlying toxic regions in parameters, leaving models vulnerable to adversarial attacks. Prior mechanistic studies characterize toxic regions as "toxic vectors" or "layer-wise subspaces", yet our analysis identifies critical limitations: i) Removed toxic vectors can be reconstructed via linear combinations of non-toxic vectors, demanding targeting of entire toxic subspace; ii) Contrastive objective over limited samples inject noise into layer-wise subspaces, hindering stable extraction. These highlight the challenge of identifying robust toxic subspace and removing them. Therefore, we propose GLOSS (GLobal tOxic Subspace Suppression), a lightweight method that mitigates toxicity by identifying and eliminating this global subspace from FFN parameters. Experiments on LLMs (e.g., Qwen3) show GLOSS achieves SOTA detoxification while preserving general capabilities without requiring large-scale retraining. WARNING: This paper contains context which is toxic in nature.

cs.LG↗

arXiv2025

Large Language Model Sourcing: A Survey

Due to the black-box nature of large language models (LLMs) and the realism of their generated content, issues such as hallucinations, bias, unfairness, and copyright infringement have become significant. In this context, sourcing information from multiple perspectives is essential. This survey presents a systematic investigation organized around four interrelated dimensions: Model Sourcing, Model Structure Sourcing, Training Data Sourcing, and External Data Sourcing. Moreover, a unified dual-paradigm taxonomy is proposed that classifies existing sourcing methods into prior-based (proactive traceability embedding) and posterior-based (retrospective inference) approaches. Traceability across these dimensions enhances the transparency, accountability, and trustworthiness of LLMs deployment in real-world applications.

cs.CL↗

arXiv2025

Privacy-Preserving Reasoning with Knowledge-Distilled Parametric Retrieval Augmented Generation

The current RAG system requires uploading plaintext documents to the cloud, risking private data leakage. Parametric RAG (PRAG) encodes documents as LoRA parameters within LLMs, offering a possible way to reduce exposure of raw content. However, it still faces two issues: (1) PRAG demands synthesizing QA pairs and fine-tuning LLM for each individual document to create its corresponding LoRA, leading to unacceptable inference latency. (2) The performance of PRAG relies solely on synthetic QA data while lacking internal alignment with standard RAG, resulting in poor generalization on out-of-distribution(OOD) inputs. Therefore, achieving high-efficiency parameterization while maintaining RAG-level performance remains a critical challenge for privacy-preserving reasoning. In this paper, we propose DistilledPRAG, a generalizable knowledge-distilled parametric RAG model aligned with standard RAG in document structure and parameter activation. We first synthesize QA pairs from single and multi-documents to enhance cross-document reasoning. Then, we mask the plaintext documents with a special token and translate them to LoRA via a parameter generator, maintaining the standard RAG document structure. Finally, guided by synthetic QA data, we train the parameter generator to match standard RAG's hidden states and output logits, enabling RAG-style reasoning without original documents. Experiments on four QA datasets show that DistilledPRAG outperforms baselines in accuracy and generalizes well on OOD data.

cs.CL↗

arXiv2025

Following the Autoregressive Nature of LLM Embeddings via Compression and Alignment

A new trend uses LLMs as dense text encoders via contrastive learning. However, since LLM embeddings predict the probability distribution of the next token, they are inherently generative and distributive, conflicting with contrastive learning, which requires embeddings to capture full-text semantics and align via cosine similarity. This discrepancy hinders the full utilization of LLMs' pre-training capabilities, resulting in inefficient learning. In response to this issue, we propose AutoRegEmbed, a new contrastive learning method built on embedding conditional probability distributions, which integrates two core tasks: information compression and conditional distribution alignment. The information compression task encodes text into the embedding space, ensuring that the embedding vectors capture global semantics. The conditional distribution alignment task focuses on aligning text embeddings with positive samples embeddings by leveraging the conditional distribution of embeddings while simultaneously reducing the likelihood of generating negative samples from text embeddings, thereby achieving embedding alignment and uniformity. Experimental results demonstrate that our method significantly outperforms traditional contrastive learning approaches and achieves performance comparable to state-of-the-art models when using the same amount of data.

cs.CL↗

arXiv2025

Rowen: Adaptive Retrieval-Augmented Generation for Hallucination Mitigation in LLMs

Hallucinations present a significant challenge for large language models (LLMs). The utilization of parametric knowledge in generating factual content is constrained by the limited knowledge of LLMs, potentially resulting in internal hallucinations. While incorporating external information can help fill knowledge gaps, it also introduces the risk of irrelevant information, thereby increasing the likelihood of external hallucinations. To balance the use of parametric knowledge within LLMs and external information, in this study, we present Rowen, a novel framework that enhances LLMs with an adaptive retrieval augmentation process tailored to address hallucinated outputs. Rowen introduces a consistency-based hallucination detection module, which assesses the model's uncertainty regarding the input query by evaluating the semantic inconsistencies in various responses generated across different languages or models. When high uncertainties in the responses are detected, Rowen activates the retrieval of external information to rectify the model outputs. Through comprehensive empirical experiments, we demonstrate that Rowen surpasses the current state-of-the-art in both detecting and mitigating hallucinated content within the outputs of LLMs.

cs.CL↗

arXiv2025

On the Diminishing Returns of Complex Robust RAG Training in the Era of Powerful LLMs

Retrieval-augmented generation (RAG) systems traditionally employ sophisticated training strategies to enhance robustness against retrieval noise. In this work, we investigate a critical question: does the benefit of these complex robust training methods diminish as language models become more powerful? Through systematic evaluation across multiple model scales and question-answering datasets, our analysis reveals a consistent trend: \emph{the marginal robustness benefit of sophisticated training strategies decreases substantially as model capacity increases.} While smaller models show significant performance improvements from complex document selection and adversarial objectives, more capable models achieve comparable or even superior performance with simpler training approaches. Further investigation demonstrates that stronger models naturally exhibit better confidence calibration, cross-dataset generalization capability, and more effective attention patterns, even under simple training regimes. These findings suggest that as foundation models evolve, the engineering effort invested in complex robust training may yield diminishing returns, indicating that simplified RAG pipelines could suffice for powerful models while maintaining competitive performance.

cs.CL↗

arXiv2025

Can Synthetic Query Rewrites Capture User Intent Better than Humans in Retrieval-Augmented Generation?

Multi-turn RAG systems often face queries with colloquial omissions and ambiguous references, posing significant challenges for effective retrieval and generation. Traditional query rewriting relies on human annotators to clarify queries, but due to limitations in annotators' expressive ability and depth of understanding, manually rewritten queries often diverge from those needed in real-world RAG systems, resulting in a gap between user intent and system response. We observe that high-quality synthetic queries can better bridge this gap, achieving superior performance in both retrieval and generation compared to human rewrites. This raises an interesting question: Can rewriting models trained on synthetic queries better capture user intent than human annotators? In this paper, we propose SynRewrite, a synthetic data-driven query rewriting model to generate high-quality synthetic rewrites more aligned with user intent. To construct training data, we prompt GPT-4o with dialogue history, current queries, positive documents, and answers to synthesize high-quality rewrites. A Flan-T5 model is then finetuned on this dataset to map dialogue history and queries to synthetic rewrites. Finally, we further enhance the rewriter using the generator's feedback through the DPO algorithm to boost end-task performance. Experiments on TopiOCQA and QRECC datasets show that SynRewrite consistently outperforms human rewrites in both retrieval and generation tasks. Our results demonstrate that synthetic rewrites can serve as a scalable and effective alternative to human annotations.

cs.IR↗

arXiv2025

A Factuality and Diversity Reconciled Decoding Method for Knowledge-Grounded Dialogue Generation

Grounding external knowledge can enhance the factuality of responses in dialogue generation. However, excessive emphasis on it might result in the lack of engaging and diverse expressions. Through the introduction of randomness in sampling, current approaches can increase the diversity. Nevertheless, such sampling method could undermine the factuality in dialogue generation. In this study, to discover a solution for advancing creativity without relying on questionable randomness and to subtly reconcile the factuality and diversity within the source-grounded paradigm, a novel method named DoGe is proposed. DoGe can dynamically alternate between the utilization of internal parameter knowledge and external source knowledge based on the model's factual confidence. Extensive experiments on three widely-used datasets show that DoGe can not only enhance response diversity but also maintain factuality, and it significantly surpasses other various decoding strategy baselines.

cs.CL↗

arXiv2025

Reverse Physician-AI Relationship: Full-process Clinical Diagnosis Driven by a Large Language Model

Full-process clinical diagnosis in the real world encompasses the entire diagnostic workflow that begins with only an ambiguous chief complaint. While artificial intelligence (AI), particularly large language models (LLMs), is transforming clinical diagnosis, its role remains largely as an assistant to physicians. This AI-assisted working pattern makes AI can only answer specific medical questions at certain parts within the diagnostic process, but lack the ability to drive the entire diagnostic process starting from an ambiguous complaint, which still relies heavily on human physicians. This gap limits AI's ability to fully reduce physicians' workload and enhance diagnostic efficiency. To address this, we propose a paradigm shift that reverses the relationship between physicians and AI: repositioning AI as the primary director, with physicians serving as its assistants. So we present DxDirector-7B, an LLM endowed with advanced deep thinking capabilities, enabling it to drive the full-process diagnosis with minimal physician involvement. Furthermore, DxDirector-7B establishes a robust accountability framework for misdiagnoses, delineating responsibility between AI and human physicians. In evaluations across rare, complex, and real-world cases under full-process diagnosis setting, DxDirector-7B not only achieves significant superior diagnostic accuracy but also substantially reduces physician workload than state-of-the-art medical LLMs as well as general-purpose LLMs. Fine-grained analyses across multiple clinical departments and tasks validate its efficacy, with expert evaluations indicating its potential to serve as a viable substitute for medical specialists. These findings mark a new era where AI, traditionally a physicians' assistant, now drives the entire diagnostic process to drastically reduce physicians' workload, indicating an efficient and accurate diagnostic solution.

cs.AI↗

arXiv2025

Generative Ghost: Investigating Ranking Bias Hidden in AI-Generated Videos

With the rapid development of AI-generated content (AIGC), the creation of high-quality AI-generated videos has become faster and easier, resulting in the Internet being flooded with all kinds of video content. However, the impact of these videos on the content ecosystem remains largely unexplored. Video information retrieval remains a fundamental approach for accessing video content. Building on the observation that retrieval models often favor AI-generated content in ad-hoc and image retrieval tasks, we investigate whether similar biases emerge in the context of challenging video retrieval, where temporal and visual factors may further influence model behavior. To explore this, we first construct a comprehensive benchmark dataset containing both real and AI-generated videos, along with a set of fair and rigorous metrics to assess bias. This benchmark consists of 13,000 videos generated by two state-of-the-art open-source video generation models. We meticulously design a suite of rigorous metrics to accurately measure this preference, accounting for potential biases arising from the limited frame rate and suboptimal quality of AIGC videos. We then applied three off-the-shelf video retrieval models to perform retrieval tasks on this hybrid dataset. Our findings reveal a clear preference for AI-generated videos in retrieval. Further investigation shows that incorporating AI-generated videos into the training set of retrieval models exacerbates this bias. Unlike the preference observed in image modalities, we find that video retrieval bias arises from both unseen visual and temporal information, making the root causes of video bias a complex interplay of these two factors. To mitigate this bias, we fine-tune the retrieval models using a contrastive learning approach. The results of this study highlight the potential implications of AI-generated videos on retrieval systems.

cs.IR↗

arXiv2025

Exploring the Escalation of Source Bias in User, Data, and Recommender System Feedback Loop

Recommender systems are essential for information access, allowing users to present their content for recommendation. With the rise of large language models (LLMs), AI-generated content (AIGC), primarily in the form of text, has become a central part of the content ecosystem. As AIGC becomes increasingly prevalent, it is important to understand how it affects the performance and dynamics of recommender systems. To this end, we construct an environment that incorporates AIGC to explore its short-term impact. The results from popular sequential recommendation models reveal that AIGC are ranked higher in the recommender system, reflecting the phenomenon of source bias. To further explore the long-term impact of AIGC, we introduce a feedback loop with realistic simulators. The results show that the model's preference for AIGC increases as the user clicks on AIGC rises and the model trains on simulated click data. This leads to two issues: In the short term, bias toward AIGC encourages LLM-based content creation, increasing AIGC content, and causing unfair traffic distribution. From a long-term perspective, our experiments also show that when AIGC dominates the content ecosystem after a feedback loop, it can lead to a decline in recommendation performance. To address these issues, we propose a debiasing method based on L1-loss optimization to maintain long-term content ecosystem balance. In a real-world environment with AIGC generated by mainstream LLMs, our method ensures a balance between AIGC and human-generated content in the ecosystem. The code and dataset are available at https://github.com/Yuqi-Zhou/Rec_SourceBias.

cs.IR↗

arXiv2025

A Comprehensive Survey in LLM(-Agent) Full Stack Safety: Data, Training and Deployment

The remarkable success of Large Language Models (LLMs) has illuminated a promising pathway toward achieving Artificial General Intelligence for both academic and industrial communities, owing to their unprecedented performance across various applications. As LLMs continue to gain prominence in both research and commercial domains, their security and safety implications have become a growing concern, not only for researchers and corporations but also for every nation. Currently, existing surveys on LLM safety primarily focus on specific stages of the LLM lifecycle, e.g., deployment phase or fine-tuning phase, lacking a comprehensive understanding of the entire "lifechain" of LLMs. To address this gap, this paper introduces, for the first time, the concept of "full-stack" safety to systematically consider safety issues throughout the entire process of LLM training, deployment, and eventual commercialization. Compared to the off-the-shelf LLM safety surveys, our work demonstrates several distinctive advantages: (I) Comprehensive Perspective. We define the complete LLM lifecycle as encompassing data preparation, pre-training, post-training, deployment and final commercialization. To our knowledge, this represents the first safety survey to encompass the entire lifecycle of LLMs. (II) Extensive Literature Support. Our research is grounded in an exhaustive review of over 800+ papers, ensuring comprehensive coverage and systematic organization of security issues within a more holistic understanding. (III) Unique Insights. Through systematic literature analysis, we have developed reliable roadmaps and perspectives for each chapter. Our work identifies promising research directions, including safety in data generation, alignment techniques, model editing, and LLM-based agent systems. These insights provide valuable guidance for researchers pursuing future work in this field.

cs.CR↗

arXiv2025

ToolCoder: A Systematic Code-Empowered Tool Learning Framework for Large Language Models

Tool learning has emerged as a crucial capability for large language models (LLMs) to solve complex real-world tasks through interaction with external tools. Existing approaches face significant challenges, including reliance on hand-crafted prompts, difficulty in multi-step planning, and lack of precise error diagnosis and reflection mechanisms. We propose ToolCoder, a novel framework that reformulates tool learning as a code generation task. Inspired by software engineering principles, ToolCoder transforms natural language queries into structured Python function scaffold and systematically breaks down tasks with descriptive comments, enabling LLMs to leverage coding paradigms for complex reasoning and planning. It then generates and executes function implementations to obtain final responses. Additionally, ToolCoder stores successfully executed functions in a repository to promote code reuse, while leveraging error traceback mechanisms for systematic debugging, optimizing both execution efficiency and robustness. Experiments demonstrate that ToolCoder achieves superior performance in task completion accuracy and execution reliability compared to existing approaches, establishing the effectiveness of code-centric approaches in tool learning.

cs.CL↗