Large language model (LLM) routing has emerged as a crucial strategy for balancing computational costs with performance by dynamically assigning queries to the most appropriate model based on query complexity. Despite recent advances showing that preference-data-based routers can outperform traditional methods, current evaluation benchmarks remain limited. They largely focus on general model capabilities while overlooking task-specific behaviors and critical concerns such as privacy, safety, and potential backdoor vulnerabilities introduced through preference data. In response, we propose the DSC benchmark: Diverse, Simple, and Categorized, an evaluation framework that categorizes router performance across a broad spectrum of query types, including coding, translation, mathematics, human instructions, general knowledge, and LLM jailbreaking. Additionally, it integrates privacy and safety assessments to reveal hidden risks. Our experiments on three preference-based routers and two commercial counterparts demonstrate that while these systems improve efficiency, they often make suboptimal, category-driven decisions. For instance, a BERT-based router directs all coding and mathematics queries to the most powerful LLM even when simpler models would suffice, while routing jailbreaking attempts to weaker models, thereby elevating safety risks.
Error detection (ED) in tabular data is crucial yet challenging due to diverse error types and the need for contextual understanding. Traditional ED methods often rely heavily on manual criteria and labels, making them labor-intensive. Large language models (LLM) can minimize human effort but struggle with errors requiring a comprehensive understanding of data context. In this paper, we propose ZeroED, a novel hybrid zero-shot error detection framework, which combines LLM reasoning ability with the manual label-based ED pipeline. ZeroED operates in four steps, i.e., feature representation, error labeling, training data construction, and detector training. Initially, to enhance error distinction, ZeroED generates rich data representations using error reason-aware binary features, pre-trained embeddings, and statistical features. Then, ZeroED employs LLM to label errors holistically through in-context learning, guided by a two-step reasoning process for detailed error detection guidelines. To reduce token costs, LLMs are applied only to representative data selected via clustering-based sampling. High-quality training data is constructed through in-cluster label propagation and LLM augmentation with verification. Finally, a classifier is trained to detect all errors. Extensive experiments on seven public datasets demonstrate that, ZeroED substantially outperforms state-of-the-art methods by a maximum 30% improvement in F1 score and up to 90% token cost reduction.
Disaggregated inference has become an essential framework that separates the prefill (P) and decode (D) stages in large language model inference to improve throughput. However, the KV cache transfer faces significant delays between prefill and decode nodes. The block-wise calling method and discontinuous KV cache memory allocation increase the number of calls to the transmission kernel. Additionally, existing frameworks often fix the roles of P and D nodes, leading to computational imbalances. In this paper, we propose FlowKV, a novel disaggregated inference framework, which reduces the average transmission latency of KV cache by 96%, from 0.944s to 0.053s, almost eliminating the transfer time relative to the total request latency by optimizing the KV cache transfer. FlowKV introduces the Load-Aware Scheduler for balanced request scheduling and flexible PD node allocation. This design maximizes hardware resource utilization, achieving peak system throughput across various scenarios, including normal, computational imbalance, and extreme overload conditions. Experimental results demonstrate that FlowKV significantly accelerates inference by 15.2%-48.9% on LongBench dataset compared to the baseline and supports applications with heterogeneous GPUs.
Large Language Models (LLMs) have recently emerged as powerful tools in cybersecurity, offering advanced capabilities in malware detection, generation, and real-time monitoring. Numerous studies have explored their application in cybersecurity, demonstrating their effectiveness in identifying novel malware variants, analyzing malicious code structures, and enhancing automated threat analysis. Several transformer-based architectures and LLM-driven models have been proposed to improve malware analysis, leveraging semantic and structural insights to recognize malicious intent more accurately. This study presents a comprehensive review of LLM-based approaches in malware code analysis, summarizing recent advancements, trends, and methodologies. We examine notable scholarly works to map the research landscape, identify key challenges, and highlight emerging innovations in LLM-driven cybersecurity. Additionally, we emphasize the role of static analysis in malware detection, introduce notable datasets and specialized LLM models, and discuss essential datasets supporting automated malware research. This study serves as a valuable resource for researchers and cybersecurity professionals, offering insights into LLM-powered malware detection and defence strategies while outlining future directions for strengthening cybersecurity resilience.
Computing the full Hessian matrix -- the matrix of second-order derivatives for an entire Large Language Model (LLM) is infeasible due to its sheer size. In this technical report, we aim to provide a comprehensive guide on how to accurately compute at least a small portion of the Hessian for LLMs using PyTorch autograd library. We also demonstrate how to compute the full diagonal of the Hessian matrix using multiple samples of vector-Hessian Products (HVPs). We hope that both this guide and the accompanying GitHub code will be valuable resources for practitioners and researchers interested in better understanding the behavior and structure of the Hessian in LLMs.
Conventional Vision-Language Models(VLMs) typically utilize a fixed number of vision tokens, regardless of task complexity. This one-size-fits-all strategy introduces notable inefficiencies: using excessive tokens leads to unnecessary computational overhead in simpler tasks, whereas insufficient tokens compromise fine-grained visual comprehension in more complex contexts. To overcome these limitations, we present TokenFLEX, an innovative and adaptable vision-language framework that encodes images into a variable number of tokens for efficient integration with a Large Language Model (LLM). Our approach is underpinned by two pivotal innovations. Firstly, we present a novel training paradigm that enhances performance across varying numbers of vision tokens by stochastically modulating token counts during training. Secondly, we design a lightweight vision token projector incorporating an adaptive pooling layer and SwiGLU, allowing for flexible downsampling of vision tokens and adaptive selection of features tailored to specific token counts. Comprehensive experiments reveal that TokenFLEX consistently outperforms its fixed-token counterparts, achieving notable performance gains across various token counts enhancements of 1.6%, 1.0%, and 0.4% with 64, 144, and 256 tokens, respectively averaged over eight vision-language benchmarks. These results underscore TokenFLEX's remarkable flexibility while maintaining high-performance vision-language understanding.
Retrieval-Augmented Generation (RAG) has emerged as a widely adopted approach for knowledge integration during large language model (LLM) inference in recent years. However, current RAG implementations face challenges in effectively addressing noise, repetition and redundancy in retrieved content, primarily due to their limited ability to exploit fine-grained inter-document relationships. To address these limitations, we propose an \textbf{E}fficient \textbf{D}ynamic \textbf{C}lustering-based document \textbf{C}ompression framework (\textbf{EDC\textsuperscript{2}-RAG}) that effectively utilizes latent inter-document relationships while simultaneously removing irrelevant information and redundant content. We validate our approach, built upon GPT-3.5, on widely used knowledge-QA and hallucination-detected datasets. The results show that this method achieves consistent performance improvements across various scenarios and experimental settings, demonstrating strong robustness and applicability. Our code and datasets can be found at https://github.com/Tsinghua-dhy/EDC-2-RAG.
Large Language Models (LLMs) have impressive capabilities in text understanding and zero-shot reasoning. However, delays in knowledge updates may cause them to reason incorrectly or produce harmful results. Knowledge Graphs (KGs) provide rich and reliable contextual information for the reasoning process of LLMs by structurally organizing and connecting a wide range of entities and relations. Existing KG-based LLM reasoning methods only inject KGs' knowledge into prompts in a textual form, ignoring its structural information. Moreover, they mostly rely on close-source models or open-source models with large parameters, which poses challenges to high resource consumption. To address this, we propose a novel Lightweight and efficient Prompt learning-ReasOning Framework for KGQA (LightPROF), which leverages the full potential of LLMs to tackle complex reasoning tasks in a parameter-efficient manner. Specifically, LightPROF follows a "Retrieve-Embed-Reason process", first accurately, and stably retrieving the corresponding reasoning graph from the KG through retrieval module. Next, through a Transformer-based Knowledge Adapter, it finely extracts and integrates factual and structural information from the KG, then maps this information to the LLM's token embedding space, creating an LLM-friendly prompt to be used by the LLM for the final reasoning. Additionally, LightPROF only requires training Knowledge Adapter and can be compatible with any open-source LLM. Extensive experiments on two public KGQA benchmarks demonstrate that LightPROF achieves superior performance with small-scale LLMs. Furthermore, LightPROF shows significant advantages in terms of input token count and reasoning time.
E-participation platforms can be an important asset for governments in increasing trust and fostering democratic societies. By engaging non-governmental and private institutions, domain experts, and even the general public, policymakers can make informed and inclusive decisions. Drawing on the Media Richness Theory and applying the Design Science Research method, we explore how a chatbot can be designed to improve the effectiveness of the policy-making process of existing citizen involvement platforms. Leveraging the Have Your Say platform, which solicits feedback on European Commission initiatives and regulations, a Large Language Model based chatbot, called AskThePublic is created, providing policymakers, journalists, researchers, and interested citizens with a convenient channel to explore and engage with public input. By conducting 11 semistructured interviews, the results show that the participants value the interactive and structured responses as well as enhanced language capabilities, thus increasing their likelihood of engaging with AskThePublic over the existing platform. An outlook for future iterations is provided and discussed with regard to the perspectives of the different stakeholders.
Large Language Model (LLM) tools have demonstrated their potential to deliver high-quality assistance by providing instant, personalized feedback that is crucial for effective programming education. However, many of these tools operate independently from institutional Learning Management Systems, which creates a significant disconnect. This isolation limits the ability to leverage learning materials and exercise context for generating tailored, context-aware feedback. Furthermore, previous research on self-regulated learning and LLM support mainly focused on knowledge acquisition, not the development of important self-regulation skills. To address these challenges, we developed CodeRunner Agent, an LLM-based programming assistant that integrates the CodeRunner, a student-submitted code executing and automated grading plugin in Moodle. CodeRunner Agent empowers educators to customize AI-generated feedback by incorporating detailed context from lecture materials, programming questions, student answers, and execution results. Additionally, it enhances students' self-regulated learning by providing strategy-based AI responses. This integrated, context-aware, and skill-focused approach offers promising avenues for data-driven improvements in programming education.
Recently, Large Language Model (LLM)-empowered recommender systems have revolutionized personalized recommendation frameworks and attracted extensive attention. Despite the remarkable success, existing LLM-empowered RecSys have been demonstrated to be highly vulnerable to minor perturbations. To mitigate the negative impact of such vulnerabilities, one potential solution is to employ collaborative signals based on item-item co-occurrence to purify the malicious collaborative knowledge from the user's historical interactions inserted by attackers. On the other hand, due to the capabilities to expand insufficient internal knowledge of LLMs, Retrieval-Augmented Generation (RAG) techniques provide unprecedented opportunities to enhance the robustness of LLM-empowered recommender systems by introducing external collaborative knowledge. Therefore, in this paper, we propose a novel framework (RETURN) by retrieving external collaborative signals to purify the poisoned user profiles and enhance the robustness of LLM-empowered RecSys in a plug-and-play manner. Specifically, retrieval-augmented perturbation positioning is proposed to identify potential perturbations within the users' historical sequences by retrieving external knowledge from collaborative item graphs. After that, we further retrieve the collaborative knowledge to cleanse the perturbations by using either deletion or replacement strategies and introduce a robust ensemble recommendation strategy to generate final robust predictions. Extensive experiments on three real-world datasets demonstrate the effectiveness of the proposed RETURN.
The rapid advancements in large Language models (LLMs) have significantly enhanced their reasoning capabilities, driven by various strategies such as multi-agent collaboration. However, unlike the well-established performance improvements achieved through scaling data and model size, the scaling of reasoning in LLMs is more complex and can even negatively impact reasoning performance, introducing new challenges in model alignment and robustness. In this survey, we provide a comprehensive examination of scaling in LLM reasoning, categorizing it into multiple dimensions and analyzing how and to what extent different scaling strategies contribute to improving reasoning capabilities. We begin by exploring scaling in input size, which enables LLMs to process and utilize more extensive context for improved reasoning. Next, we analyze scaling in reasoning steps that improves multi-step inference and logical consistency. We then examine scaling in reasoning rounds, where iterative interactions refine reasoning outcomes. Furthermore, we discuss scaling in training-enabled reasoning, focusing on optimization through iterative model improvement. Finally, we review applications of scaling across domains and outline future directions for further advancing LLM reasoning. By synthesizing these diverse perspectives, this survey aims to provide insights into how scaling strategies fundamentally enhance the reasoning capabilities of LLMs and further guide the development of next-generation AI systems.
In this report we describe the development of Command A, a powerful large language model purpose-built to excel at real-world enterprise use cases. Command A is an agent-optimised and multilingual-capable model, with support for 23 languages of global business, and a novel hybrid architecture balancing efficiency with top of the range performance. It offers best-in-class Retrieval Augmented Generation (RAG) capabilities with grounding and tool use to automate sophisticated business processes. These abilities are achieved through a decentralised training approach, including self-refinement algorithms and model merging techniques. We also include results for Command R7B which shares capability and architectural similarities to Command A. Weights for both models have been released for research purposes. This technical report details our original training pipeline and presents an extensive evaluation of our models across a suite of enterprise-relevant tasks and public benchmarks, demonstrating excellent performance and efficiency.
We address the task of video chaptering, i.e., partitioning a long video timeline into semantic units and generating corresponding chapter titles. While relatively underexplored, automatic chaptering has the potential to enable efficient navigation and content retrieval in long-form videos. In this paper, we achieve strong chaptering performance on hour-long videos by efficiently addressing the problem in the text domain with our 'Chapter-Llama' framework. Specifically, we leverage a pretrained large language model (LLM) with large context window, and feed as input (i) speech transcripts and (ii) captions describing video frames, along with their respective timestamps. Given the inefficiency of exhaustively captioning all frames, we propose a lightweight speech-guided frame selection strategy based on speech transcript content, and experimentally demonstrate remarkable advantages. We train the LLM to output timestamps for the chapter boundaries, as well as free-form chapter titles. This simple yet powerful approach scales to processing one-hour long videos in a single forward pass. Our results demonstrate substantial improvements (e.g., 45.3 vs 26.7 F1 score) over the state of the art on the recent VidChapters-7M benchmark. To promote further research, we release our code and models at our project page.
Speculative decoding accelerates large language model (LLM) inference by using a smaller draft model to propose tokens, which are then verified by a larger target model. However, selecting an optimal speculation length is critical for maximizing speedup while minimizing wasted computation. We introduce \textit{GammaTune} and \textit{GammaTune+}, training-free adaptive algorithms that dynamically adjust speculation length based on token acceptance rates using a heuristic-based switching mechanism. Evaluated on SpecBench across multiple tasks and model pairs, our method outperforms other heuristic-based approaches and fixed-length speculative decoding, achieving an average speedup of 15\% ($\pm$5\%) with \textit{GammaTune} and 16\% ($\pm$3\%) with \textit{GammaTune+}, while reducing performance variance. This makes \textit{GammaTune} a robust and efficient solution for real-world deployment.
We study how to use naturally available user feedback, such as clicks, to optimize large language model (LLM) pipelines for generating personalized sentences using prompts. Naive approaches, which estimate the policy gradient in the prompt space, suffer either from variance caused by the large action space of prompts or bias caused by inaccurate reward predictions. To circumvent these challenges, we propose a novel kernel-based off-policy gradient method, which estimates the policy gradient by leveraging similarity among generated sentences, substantially reducing variance while suppressing the bias. Empirical results on our newly established suite of benchmarks demonstrate the effectiveness of the proposed approach in generating personalized descriptions for movie recommendations, particularly when the number of candidate prompts is large.
In machine learning, Neural Architecture Search (NAS) requires domain knowledge of model design and a large amount of trial-and-error to achieve promising performance. Meanwhile, evolutionary algorithms have traditionally relied on fixed rules and pre-defined building blocks. The Large Language Model (LLM)-Guided Evolution (GE) framework transformed this approach by incorporating LLMs to directly modify model source code for image classification algorithms on CIFAR data and intelligently guide mutations and crossovers. A key element of LLM-GE is the "Evolution of Thought" (EoT) technique, which establishes feedback loops, allowing LLMs to refine their decisions iteratively based on how previous operations performed. In this study, we perform NAS for object detection by improving LLM-GE to modify the architecture of You Only Look Once (YOLO) models to enhance performance on the KITTI dataset. Our approach intelligently adjusts the design and settings of YOLO to find the optimal algorithms against objective such as detection accuracy and speed. We show that LLM-GE produced variants with significant performance improvements, such as an increase in Mean Average Precision from 92.5% to 94.5%. This result highlights the flexibility and effectiveness of LLM-GE on real-world challenges, offering a novel paradigm for automated machine learning that combines LLM-driven reasoning with evolutionary strategies.
With the rapid development of 3D printing, the demand for personalized and customized production on the manufacturing line is steadily increasing. Efficient merging of printing workpieces can significantly enhance the processing efficiency of the production line. Addressing the challenge, a Large Language Model (LLM)-driven method is established in this paper for the autonomous merging of 3D printing work orders, integrated with a memory-augmented learning strategy. In industrial scenarios, both device and order features are modeled into LLM-readable natural language prompt templates, and develop an order-device matching tool along with a merging interference checking module. By incorporating a self-memory learning strategy, an intelligent agent for autonomous order merging is constructed, resulting in improved accuracy and precision in order allocation. The proposed method effectively leverages the strengths of LLMs in industrial applications while reducing hallucination.
Training large language models (LLMs) presents numerous challenges, including gradient instability and loss spikes. These phenomena can lead to catastrophic divergence, requiring costly checkpoint restoration and data batch skipping. Traditional gradient clipping techniques, such as constant or norm-based methods, fail to address these issues effectively due to their reliance on fixed thresholds or heuristics, leading to inefficient learning and requiring frequent manual intervention. In this work, we propose ZClip, an adaptive gradient clipping algorithm that dynamically adjusts the clipping threshold based on statistical properties of gradient norms over time. Unlike prior reactive strategies, ZClip proactively adapts to training dynamics without making any prior assumptions on the scale and the temporal evolution of gradient norms. At its core, it leverages z-score-based anomaly detection to identify and mitigate large gradient spikes, preventing malignant loss spikes while not interfering with convergence otherwise. Our code is available at: https://github.com/bluorion-com/ZClip.
Accurate and verifiable large language model (LLM) simulations of human research subjects promise an accessible data source for understanding human behavior and training new AI systems. However, results to date have been limited, and few social scientists have adopted this method. In this position paper, we argue that the promise of LLM social simulations can be achieved by addressing five tractable challenges. We ground our argument in a review of empirical comparisons between LLMs and human research subjects, commentaries on the topic, and related work. We identify promising directions, including context-rich prompting and fine-tuning with social science datasets. We believe that LLM social simulations can already be used for pilot and exploratory studies, and more widespread use may soon be possible with rapidly advancing LLM capabilities. Researchers should prioritize developing conceptual models and iterative evaluations to make the best use of new AI systems.