论文检索

Xue Tan, Hao Luan, Mingyu Luo, Xiaoyan Sun, Ping Chen, Jun Dai

Retrieval-Augmented Generation (RAG) enriches the input to LLMs by retrieving information from the relevant knowledge database, enabling them to produce responses that are more accurate and contextually appropriate. It is worth noting that the knowledge database, being sourced from publicly available channels such as Wikipedia, inevitably introduces a new attack surface. RAG poisoning attack involves injecting malicious texts into the knowledge database, ultimately leading to the generation of the attacker’s target response (also called poisoned response). However, there are currently limited methods available for detecting such poisoning attacks. We aim to bridge the gap in this work by introducing RevPRAG, a flexible and automated detection pipeline that leverages the activations of LLMs for poisoned response detection. Our investigation uncovers distinct patterns in LLMs’ activations when generating poisoned responses versus correct responses. Our results on multiple benchmarks and RAG architectures show our approach can achieve a 98% true positive rate, while maintaining a false positive rate close to 1%.

Cheng Yan, Feng Zhao, Ruilin Zhao, Hong Zhang

Knowledge graphs are dynamic structures that continuously evolve as new entities emerge, often accompanied by only a handful of associated triples. Current knowledge graph reasoning methods struggle in these few-shot scenarios due to their reliance on extensive structural information.To address this limitation, we introduce ENGRAM, a novel approach that enables inductive reasoning on few-shot KGs by innovatively enriching the semantics from both textual and structural perspectives. Our key innovation lies in designing a task-aware language model that activates the language model’s in-context learning ability for structured KG tasks, effectively bridging the gap between unstructured natural language and structured tasks. Unlike prior methods that inefficiently employ classification over exhaustive candidate sets, we recast knowledge graph reasoning from a generative perspective, allowing for direct computation of inference results without iterative enumeration. Additionally, we propose a distant neighborhood awareness strategy to enrich the sparse structural features of few-shot entities.Our experimental findings indicate that our method not only achieves state-of-the-art performance in few-shot scenarios. The tunable parameters of our model are approximately 1% of those in previous language model-based methods, and the inference time has been reduced to 1/10 of that required by previous methods.

Yunhai Hu, Yilun Zhao, Chen Zhao, Arman Cohan

We introduce MCTS-RAG, a novel approach that enhances the reasoning capabilities of small language models on knowledge-intensive tasks by leveraging retrieval-augmented generation (RAG) to provide relevant context and Monte Carlo Tree Search (MCTS) to refine reasoning paths. MCTS-RAG dynamically integrates retrieval and reasoning through an iterative decision-making process. Unlike standard RAG methods, which typically retrieve information independently from reasoning and thus integrate knowledge suboptimally, or conventional MCTS reasoning, which depends solely on internal model knowledge without external facts, MCTS-RAG combines structured reasoning with adaptive retrieval. This integrated approach enhances decision-making, reduces hallucinations, and ensures improved factual accuracy and response consistency. The experimental results on multiple reasoning and knowledge-intensive datasets datasets (ComplexWebQA, GPQA, and FoolMeTwice) show that our method enables small-scale LMs to achieve performance comparable to frontier LLMs like GPT-4o by effectively scaling inference-time compute, setting a new standard for reasoning in small-scale models.

Meinan Liu, Yunfang Dong, Xixian Liao, Bonnie Webber

Previous work has shown that simple mask-filling can provide useful information about the discourse informativeness of syntactic structures. Dong et al. (2024) first adopted this approach to investigating preposing constructions. The problem with single token mask fillers was that they were, by and large, ambiguous. We address the issue by adapting the approach of Kalinsky et al. (2023) to support the prediction of multi-token connectives in masked positions. Our first experiment demonstrates that this multi-token mask-filling approach substantially outperforms the previously considered single-token approach in recognizing implicit discourse relations. Our second experiment corroborates previous findings, providing additional empirical support for the role of preposed syntactic constituents in signaling discourse coherence. Overall, our study extends existing mask-filling methods to a new discourse-level task and reinforces the linguistic hypothesis concerning the discourse informativeness of preposed structures.

Sikun Guo, Amir Hassan Shariatmadari, Peng Wang, Albert Huang, Aidong Zhang

Advancements in Large Language Models (LLMs) have opened new opportunities for scientific discovery by assisting researchers in generating novel hypotheses and ideas. In this process, a major challenge is how to optimally and efficiently utilize LLMs’ parametric knowledge obtained from their pretraining process. Inspired by Generative Adversarial Networks (GANs), we propose inference time adversarial learning (termed InfAL), implemented through multi-LLM-agent interactions, to enhance research ideation. This approach optimizes the utilization of LLMs’ parametric knowledge without requiring additional model training, making adversarial learning efficient and context-driven. To evaluate the quality of generated ideas, we propose a relative quality ranking metric as a scalable alternative to human evaluation. Our results show that InfAL significantly improves idea generation, with GPT-4o achieving a 21% increase in novelty and a 322% increase in feasibility, demonstrating its transformative potential for driving innovation in scientific research.

Zhibo Xu, Zhu JianHao, Jingwen Xu, Changze Lv, Zhenghua Wang, Zisu Huang, Xiaohua Wang, Muling Wu, Qi Qian, Xiaoqing Zheng 等

The primary goal of traditional federated learning is to protect data privacy by enabling distributed edge devices to collaboratively train a shared global model while keeping raw data decentralized at local clients. The rise of large language models (LLMs) has introduced new challenges in distributed systems, as their substantial computational requirements and the need for specialized expertise raise critical concerns about protecting intellectual property (IP). This highlights the need for a federated learning approach that can safeguard both sensitive data and proprietary models. To tackle this challenge, we propose FedQSN, a federated learning approach that leverages random masking to obscure a subnetwork of model parameters and applies quantization to the remaining parameters. Consequently, the server transmits only a privacy-preserving proxy of the global model to clients during each communication round, thus enhancing the model’s confidentiality. Experimental results across various models and tasks demonstrate that our approach not only maintains strong model performance in federated learning settings but also achieves enhanced protection of model parameters compared to baseline methods.

Yiyang Feng, Yichen Wang, Shaobo Cui, Boi Faltings, Mina Lee, Jiawei Zhou

Large Language Models (LLMs) have demonstrated impressive capabilities in reasoning, positioning them as promising tools for supporting human problem-solving. However, what happens when their performance is affected by *misinformation*, i.e., incorrect inputs introduced by users due to oversights or gaps in knowledge? Such misinformation is prevalent in real-world interactions with LLMs, yet how it propagates within LLMs’ reasoning process remains underexplored. Focusing on mathematical reasoning, we present a comprehensive analysis of how misinformation affects intermediate reasoning steps and final answers. We also examine how effectively LLMs can correct misinformation when explicitly instructed to do so. Even with explicit instructions, LLMs succeed less than half the time in rectifyingmisinformation, despite possessing correct internal knowledge, leading to significant accuracy drops (10.02% – 72.20%), and the degradation holds with thinking models (4.30% – 19.97%). Further analysis shows that applying factual corrections early in the reasoning process most effectively reduces misinformation propagation, and fine-tuning on synthesized data with early-stage corrections significantly improves reasoning factuality. Our work offers a practical approach to mitigating misinformation propagation.

Prerna Agarwal, Srikanta Bedathur

Existing low-resource Knowledge Graph Question Answering (KGQA) methods rely heavily on Large Language Models (LLMs) for semantic parsing of natural language question to its corresponding logical form (LF) such as SPARQL, S-Expression, etc. However, LLMs becomes bottleneck for practical applications due to: (i) its high computational resource requirements; (2) limited knowledge of LLM about different LFs; (3) unavailability of low-resource annotated data for new KGs and settings. This motivates us to design a KGQA framework that can operate in a zero-shot setting without the need for additional resources. In this paper, we propose (NS-KGQA): a zero-shot neuro-symbolic approach based on neural KG embeddings that have demonstrated their ability to effectively model KG structure without the need of additional data. We extract a link-prediction based symbolic question subgraph. We then propose a Symbolic Resolver that uses Dual KG Embeddings combined with a symbolic approach to resolve the symbolic question subgraph. Our extensive experiments on Complex KGQA benchmarks such as KQA Pro demonstrate the effectiveness of our approach. NS-KGQA outperforms all other LLM-based zero-shot baselines by 26% (avg).

Rares Dolga, Lucas Maystre, Tudor Berariu, David Barber

Subword tokenization methods like Byte Pair Encoding (BPE) are widely used in large language models due to their balance of vocabulary compactness and representational power. However, they suffer from inefficiencies in representing rare words and require large embedding matrices. Character-level models address these issues but introduce performance bottlenecks, particularly in Transformer-based architectures. Recent hierarchical models attempt to merge the benefits of both paradigms by grouping characters into patches, but existing patching strategies either rely on whitespace—limiting applicability to certain languages—or require auxiliary models that introduce new dependencies. In this paper, we propose a dynamic character grouping method that leverages the structure of existing BPE tokenization without requiring additional models. By appending explicit end-of-patch markers to BPE tokens and introducing a second-level BPE compression stage to control patch granularity, our method offers efficient, flexible, and language-agnostic representations. Empirical results demonstrate that our approach matches or exceeds the performance of dynamic entropy- and whitespace-based patching strategies, while maintaining a compact vocabulary.

Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, Chao Huang

Retrieval-Augmented Generation (RAG) systems enhance large language models (LLMs) by integrating external knowledge sources, enabling more accurate and contextually relevant responses tailored to user needs. However, existing RAG systems have significant limitations, including reliance on flat data representations and inadequate contextual awareness, which can lead to fragmented answers that fail to capture complex interdependencies. To address these challenges, we propose LightRAG, a novel framework that incorporates graph structures into text indexing and retrieval processes. This innovative approach employs a dual-level retrieval system that enhances comprehensive information retrieval from both low- and high-level knowledge discovery. Additionally, the integration of graph structures with vector representations facilitates efficient retrieval of related entities and their relationships, significantly improving response times while maintaining contextual relevance. This capability is further enhanced by an incremental update algorithm that ensures the timely integration of new data, allowing the system to remain effective and responsive in rapidly changing data environments. Extensive experimental validation demonstrates considerable improvements in retrieval accuracy and efficiency compared to existing approaches. We have made our LightRAG framework open source and anonymously available at the link: https://anonymous.4open.science/r/LightRAG-2BEE.

Yanhong Li, Zixuan Lan, Jiawei Zhou

Large language models (LLMs) and their multimodal variants can now process visual inputs, including images of text. This raises an intriguing question: Can we compress textual inputs by feeding them as images to reduce token usage while preserving performance?In this paper, we show that *visual text representations* are a practical and surprisingly effective form of input compression for decoder LLMs. We exploit this idea by rendering long text inputs as a single image and providing it directly to the model. This approach dramatically reduces the number of decoder tokens required, offering a new form of input compression. Through experiments on two distinct benchmarks — RULER (long-context retrieval) and CNN/DailyMail (document summarization) — we demonstrate that this text-as-image method yields substantial token savings *without degrading task performance*.

Jian Jornbowrl Wang, Xiaofei Xie, Qiang Hu, Shangqing Liu, Yi Li

Code Large Language Models (Code LLMs) have opened a new era in programming with their impressive capabilities. However, recent research has revealed critical limitations in their ability to reason about runtime behavior and understand the actual functionality of programs, which poses significant challenges for their post-training and practical deployment. Specifically, Code LLMs encounter two principal issues: (1) a lack of proficiency in reasoning about program execution behavior, as they struggle to interpret what programs actually do during runtime, and (2) inconsistent and fragmented representation of semantic information, such as execution traces, across existing methods, which hinders their ability to generalize and reason effectively. These challenges underscore the necessity for more systematic approaches to enhance the reasoning capabilities of Code LLMs. To address these issues, we introduce a generic framework to support integrating semantic information (e.g., execution trace) to code task-relevant prompts, and conduct a comprehensive study to explore the role of semantic information in enhancing the reasoning ability of Code LLMs accordingly. Specifically, we focus on investigating the usefulness of trace-based semantic information in boosting supervised fine-tuning(SFT) and post-phase inference of Code LLMs. The experimental results surprisingly disagree with previous works and demonstrate that semantic information has limited usefulness for SFT and test time scaling of Code LLM.

Mete Ismayilzada, Antonio Laverghetta Jr., Simone A. Luchini, Reet Patel, Antoine Bosselut, Lonneke van der Plas, Roger E. Beaty

While Large Language Models (LLMs) have demonstrated impressive performance across natural language generation tasks, their ability to generate truly creative content—characterized by novelty, diversity, surprise, and quality—remains limited. Existing methods for enhancing LLM creativity often focus narrowly on diversity or specific tasks, failing to address creativity’s multifaceted nature in a generalizable way. In this work, we propose Creative Preference Optimization (CrPO), a novel alignment method that injects signals from multiple creativity dimensions into the preference optimization objective in a modular fashion. We train and evaluate creativity-augmented versions of several models using CrPO and MuCE, a new large-scale human preference dataset spanning over 200,000 human-generated responses and ratings from more than 30 psychological creativity assessments. Our models outperform strong baselines, including GPT-4o, on both automated and human evaluations, producing more novel, diverse, and surprising generations while maintaining high output quality. Additional evaluations on NoveltyBench further confirm the generalizability of our approach. Together, our results demonstrate that directly optimizing for creativity within preference frameworks is a promising direction for advancing the creative capabilities of LLMs without compromising output quality.

Hoang Phan, Victor Li, Qi Lei

Large language models (LLMs) have revolutionized natural language processing with their ability to generate coherent and contextually relevant text. However, their deployment raises significant concerns about the potential for generating harmful or inappropriate content. In this paper, we introduce Progressive Self-Reflection, a novel inference-time technique that empowers LLMs to self-monitor and correct their outputs dynamically. Experimental results demonstrate that applying our proposed method to Llama-3.1-8B-Instruct reduces the attack success rate from 77.47% to 5.86%, to Llama-3.1-8B base from 89.70% to 5.56%, and to Qwen2.5-7B-Instruct from 44.44% to 3.84%, without additional training. Furthermore, our method maintains their original performance across diverse tasks, including summarization, general knowledge, reasoning, and mathematics. Our approach acts as a test-time scaling method, where additional self-reflection rounds enhance safety at the cost of inference overhead. To balance safety with computational efficiency, we introduce a lightweight self-reflection predictor that estimates the optimal number of reflection rounds based on input complexity. This adaptive mechanism prevents unnecessary self-assessment on benign inputs while ensuring thorough evaluation when encountering potentially harmful content. Our findings suggest that Progressive Self-Reflection serves as a scalable test-time approach, enhancing LLM safety by dynamically allocating computational resources in proportion to the input’s risk profile.

Davood Wadi, Marc Fredette

Scientific evaluation of Large Language Models is an important topic that quantifies any degree of progress we make with new models. Even though current LLMs show high level of accuracy on benchmark datasets, the single-sample approach to evaluating them is not sufficient as it ignores high entropy of LLM responses. We introduce a Monte-Carlo evaluation framework for evaluating LLMs that follows behavioral science methodologies and provides statistical guarantees for estimates of performance. We test our framework on multiple LLMs to see if they are susceptible to cognitive biases. We find significant effect of prompts that induce cognitive biases in LLMs, raising questions about their reliability in social sciences and business. We also see higher susceptibility of newer and larger LLMs to cognitive biases, which shows a development towards more human-like and less rational LLM responses. We conclude by calling for the use of Monte-Carlo sampling as opposed to pass@1 for the broader LLM evaluations.

Sihan Yang, Chenhang Cui, Zihao Zhao, Yiyang Zhou, Weilong Yan, Ying Wei, Huaxiu Yao

The rapid advancements in Large Language Models (LLMs) and Large Visual-Language Models (LVLMs) have opened up new opportunities for integrating visual and linguistic modalities. Yet, challenges remain in aligning these modalities effectively, causing issues such as hallucinations, where generated outputs are not grounded in the visual input, and safety concerns in the application of LVLMs across various domains. Existing alignment methods, such as instruction tuning and preference tuning, often rely on external datasets, human annotations, or complex post-processing, which limit scalability and introduce additional costs. To address these challenges, we propose a novel approach that generates the debiased self-judgment score, a self-evaluation metric created internally by the model without relying on external resources. This enables the model to autonomously improve alignment. Our method enhances both decoding strategies and preference tuning processes, resulting in improved alignment, reduced hallucinations, and enhanced safety. Empirical results show that our approach significantly outperforms traditional methods, offering a more effective solution for aligning LVLMs.

Bao Nguyen, Binh Nguyen, Duy Nguyen, Viet Anh Nguyen

Language models, while capable of generating remarkably coherent and seemingly accurate text, can occasionally produce undesirable content including harmful or toxic outputs. In this paper, we present a new two-stage approach to detect and mitigate undesirable content generations by rectifying activations. First, we train an ensemble of layerwise classifiers to detect undesirable content using activations by minimizing a smooth surrogate of the risk-aware score. Then, for detected undesirable contents, we propose layerwise distributional steering policies that transform the attention heads. These policies are computed through principled semidefinite programming aims to minimally perturb the attention distribution while probabilistically guaranteeing the effectiveness of the editions. Empirical evaluations across multiple language models and datasets show that our method outperforms baselines in reducing the generation of undesirable output.

Aakash Sen Sharma, Debdeep Sanyal, Priyansh Srivastava, Sundar Athreya H, Shirish Karande, Mohan Kankanhalli, Murari Mandal

Large language models (LLMs) commonly risk copyright infringement by reproducing protected content verbatim or with insufficient transformative modifications, posing significant ethical, legal, and practical concerns. Current inference-time safeguards predominantly rely on restrictive refusal-based filters, often compromising the practical utility of these models. To address this, we collaborated closely with intellectual property experts to develop LAW-LM (Legally Aware Language Model), a legally-grounded framework explicitly designed to align LLM outputs with fair-use doctrine. Central to our method is FairUseDB, a carefully constructed dataset containing 18,000 expert-validated examples covering nine realistic infringement scenarios. Leveraging this dataset, we apply Direct Preference Optimization (DPO) to fine-tune open-source LLMs, encouraging them to produce legally compliant and practically useful alternatives rather than resorting to blunt refusal. Recognizing the shortcomings of traditional evaluation metrics, we propose new measures: Weighted Penalty Utility and Compliance Aware Harmonic Mean (CAH) to balance infringement risk against response utility. Extensive quantitative experiments coupled with expert evaluations confirm that LAW-LM substantially reduces problematic outputs compared to state-of-the-art approaches, while preserving real-world usability.

Jieyong Wang, Chunyao Song, Tingjian Ge

In the era of prosperity of large language models (LLMs), hallucination remains a serious issue hindering LLMs’ expansion and reliability. Predicting the presence (and absence) of certain knowledge in LLMs could aid the hallucination avoidance. However, the token-based generation mode of LLM is different from the knowledge storage structure in the form of triples, which makes it difficult to accurately evaluate the knowledge boundary of LLM. We approach this problem from a novel perspective and, for the first time, introduce the concept of token knowledge in large language models. Consequently, we propose a token knowledge dataset construction method and use the intermediate states during inference to train probes. This allows us to predict if a specific token will appear in the LLM’s generated sequence, without even generating a single token. Our approach unlocks the model’s latent potential, enhancing its accuracy in assessing token knowledge from about 60% to over 90%, with strong out-of-distribution generalization by training on just a few dozen prompts. Finally, we apply KEGT to enhance a state-of-the-art knowledge boundary detection method, achieving improved performance while reducing computational time by over 90%. Furthermore, KEGT enables prevention of hallucinations in certain cases by leveraging its guidance in the token-level knowledge semantic space. Our code is available at https://github.com/CC-2000/KEGT.

Zhuoyue Chen, Jihai Zhang, Ben Liu, Fangquan Lin, Wotao Yin

Optimizing data utilization remains a central challenge in applying Reinforcement Learning (RL) to Large Language Models (LLMs), directly impacting sample efficiency, training stability, and final model performance.Current approaches often rely on massive static datasets, leading to computational inefficiency and redundant gradient updates.In this paper, we propose ScalingRL, a data-centric RL framework that dynamically selects the most informative training samples to optimize RL for mathematical reasoning.Specifically, ScalingRL introduces the Data Effectiveness Score (DES) that quantitatively ranks prompts according to three complementary factors: problem difficulty, Chain-of-Thought complexity, and reward adaptability.Then, ScalingRL employs an adaptive curriculum scheduler that progressively adjusts the overall scale and specific mix of training prompts—balancing exploration of new, challenging data with exploitation of previously learned concepts—thereby tailoring the data distribution to the model’s current learning trajectory and performance.Experimental results demonstrate that ScalingRL achieves comparable performance to full-data training methods while requiring only 1.5K samples instead of 220K, reducing training time from 13 days to just 4 hours on 8\timesA800 GPUs.