With the rise of LLMs, a large number of Model Context Protocol (MCP) services have emerged since the end of 2024. However, the effectiveness and efficiency of MCP servers have not been well studied. To study these questions, we propose an evaluation framework, called MCPBench. We selected several widely used MCP server and conducted an experimental evaluation on their accuracy, time, and token usage. Our experiments showed that the most effective MCP, Bing Web Search, achieved an accuracy of 64%. Importantly, we found that the accuracy of MCP servers can be substantially enhanced by involving declarative interface. This research paves the way for further investigations into optimized MCP implementations, ultimately leading to better AI-driven applications and data retrieval solutions.
We introduce AI University (AI-U), a flexible framework for AI-driven course content delivery that adapts to instructors' teaching styles. At its core, AI-U fine-tunes a large language model (LLM) with retrieval-augmented generation (RAG) to generate instructor-aligned responses from lecture videos, notes, and textbooks. Using a graduate-level finite-element-method (FEM) course as a case study, we present a scalable pipeline to systematically construct training data, fine-tune an open-source LLM with Low-Rank Adaptation (LoRA), and optimize its responses through RAG-based synthesis. Our evaluation - combining cosine similarity, LLM-based assessment, and expert review - demonstrates strong alignment with course materials. We also have developed a prototype web application, available at https://my-ai-university.com, that enhances traceability by linking AI-generated responses to specific sections of the relevant course material and time-stamped instances of the open-access video lectures. Our expert model is found to have greater cosine similarity with a reference on 86% of test cases. An LLM judge also found our expert model to outperform the base Llama 3.2 model approximately four times out of five. AI-U offers a scalable approach to AI-assisted education, paving the way for broader adoption in higher education. Here, our framework has been presented in the setting of a class on FEM - a subject that is central to training PhD and Master students in engineering science. However, this setting is a particular instance of a broader context: fine-tuning LLMs to research content in science.
Configuring computational fluid dynamics (CFD) simulations typically demands extensive domain expertise, limiting broader access. Although large language models (LLMs) have advanced scientific computing, their use in automating CFD workflows is underdeveloped. We introduce a novel approach centered on domain-specific LLM adaptation. By fine-tuning Qwen2.5-7B-Instruct on NL2FOAM, our custom dataset of 28716 natural language-to-OpenFOAM configuration pairs with chain-of-thought (CoT) annotations, we enable direct translation from natural language descriptions to executable CFD setups. A multi-agent framework orchestrates the process, autonomously verifying inputs, generating configurations, running simulations, and correcting errors. Evaluation on a benchmark of 21 diverse flow cases demonstrates state-of-the-art performance, achieving 88.7% solution accuracy and 82.6% first-attempt success rate. This significantly outperforms larger general-purpose models like Qwen2.5-72B-Instruct, DeepSeek-R1, and Llama3.3-70B-Instruct, while also requiring fewer correction iterations and maintaining high computational efficiency. The results highlight the critical role of domain-specific adaptation in deploying LLM assistants for complex engineering workflows. Our code and fine-tuned model have been deposited at https://github.com/YYgroup/AutoCFD.
Are we running out of learning signal? Predicting the next word in an existing text has turned out to be a powerful signal, at least at scale. But there are signs that we are running out of this resource. In recent months, interaction between learner and feedback-giver has come into focus, both for "alignment" (with a reward model judging the quality of instruction following attempts) and for improving "reasoning" (process- and outcome-based verifiers judging reasoning steps). In this paper, we explore to what extent synthetic interaction in what we call Dialogue Games -- goal-directed and rule-governed activities driven predominantly by verbal actions -- can provide a learning signal, and how this signal can be used. We introduce an environment for producing such interaction data (with the help of a Large Language Model as counterpart to the learner model), both offline and online. We investigate the effects of supervised fine-tuning on this data, as well as reinforcement learning setups such as DPO, and GRPO; showing that all of these approaches achieve some improvements in in-domain games, but only GRPO demonstrates the ability to generalise to out-of-domain games as well as retain competitive performance in reference-based tasks. We release the framework and the baseline training setups in the hope that this can foster research in this promising new direction.
Collaborative assistants, or chatbots, are data-driven decision support systems that enable natural interaction for task completion. While they can meet critical needs in modern society, concerns about their reliability and trustworthiness persist. In particular, Large Language Model (LLM)-based chatbots like ChatGPT, Gemini, and DeepSeek are becoming more accessible. However, such chatbots have limitations, including their inability to explain response generation, the risk of generating problematic content, the lack of standardized testing for reliability, and the need for deep AI expertise and extended development times. These issues make chatbots unsuitable for trust-sensitive applications like elections or healthcare. To address these concerns, we introduce SafeChat, a general architecture for building safe and trustworthy chatbots, with a focus on information retrieval use cases. Key features of SafeChat include: (a) safety, with a domain-agnostic design where responses are grounded and traceable to approved sources (provenance), and 'do-not-respond' strategies to prevent harmful answers; (b) usability, with automatic extractive summarization of long responses, traceable to their sources, and automated trust assessments to communicate expected chatbot behavior, such as sentiment; and (c) fast, scalable development, including a CSV-driven workflow, automated testing, and integration with various devices. We implemented SafeChat in an executable framework using the open-source chatbot platform Rasa. A case study demonstrates its application in building ElectionBot-SC, a chatbot designed to safely disseminate official election information. SafeChat is being used in many domains, validating its potential, and is available at: https://github.com/ai4society/trustworthy-chatbot.
Electronic health record (EHR) foundation models have been an area ripe for exploration with their improved performance in various medical tasks. Despite the rapid advances, there exists a fundamental limitation: Processing unseen medical codes out of the vocabulary. This problem limits the generality of EHR foundation models and the integration of models trained with different vocabularies. To deal with this problem, we propose MedRep for EHR foundation models based on the observational medical outcome partnership (OMOP) common data model (CDM), providing the integrated medical concept representations and the basic data augmentation strategy for patient trajectories. For concept representation learning, we enrich the information of each concept with a minimal definition through large language model (LLM) prompts and enhance the text-based representations through graph ontology of OMOP vocabulary. Trajectory augmentation randomly replaces selected concepts with other similar concepts that have closely related representations to let the model practice with the concepts out-of-vocabulary. Finally, we demonstrate that EHR foundation models trained with MedRep better maintain the prediction performance in external datasets. Our code implementation is publicly available at https://github.com/kicarussays/MedRep.
Large Language Models (LLMs) have become indispensable across academia, industry, and daily applications, yet current evaluation methods struggle to keep pace with their rapid development. One core challenge of evaluation in the large language model (LLM) era is the generalization issue: how to infer a model's near-unbounded abilities from inevitably bounded benchmarks. We address this challenge by proposing Model Utilization Index (MUI), a mechanism interpretability enhanced metric that complements traditional performance scores. MUI quantifies the effort a model expends on a task, defined as the proportion of activated neurons or features during inference. Intuitively, a truly capable model should achieve higher performance with lower effort. Extensive experiments across popular LLMs reveal a consistent inverse logarithmic relationship between MUI and performance, which we formulate as the Utility Law. From this law we derive four practical corollaries that (i) guide training diagnostics, (ii) expose data contamination issue, (iii) enable fairer model comparisons, and (iv) design model-specific dataset diversity. Our code can be found at https://github.com/ALEX-nlp/MUI-Eva.
Multi-modal Large Language Models (MLLMs) integrate visual and linguistic reasoning to address complex tasks such as image captioning and visual question answering. While MLLMs demonstrate remarkable versatility, MLLMs appears limited performance on special applications. But tuning MLLMs for downstream tasks encounters two key challenges: Task-Expert Specialization, where distribution shifts between pre-training and target datasets constrain target performance, and Open-World Stabilization, where catastrophic forgetting erases the model general knowledge. In this work, we systematically review recent advancements in MLLM tuning methodologies, classifying them into three paradigms: (I) Selective Tuning, (II) Additive Tuning, and (III) Reparameterization Tuning. Furthermore, we benchmark these tuning strategies across popular MLLM architectures and diverse downstream tasks to establish standardized evaluation analysis and systematic tuning principles. Finally, we highlight several open challenges in this domain and propose future research directions. To facilitate ongoing progress in this rapidly evolving field, we provide a public repository that continuously tracks developments: https://github.com/WenkeHuang/Awesome-MLLM-Tuning.
Large language model (LLM) inference serving systems are essential to various LLM-based applications. As demand for LLM services continues to grow, scaling these systems to handle high request rates while meeting latency Service-Level Objectives (SLOs), referred to as effective throughput, becomes critical. However, existing systems often struggle to improve effective throughput, primarily due to a significant decline in Time To First Token (TTFT) SLO attainment. We identify two major causes of this bottleneck: (1) memory-intensive KV cache that limits batch size expansion under GPU memory constraints, and (2) rigid batch composition enforced by the default First-Come-First-Serve scheduling policy. In this paper, we introduce Apt-Serve, a scalable framework designed to enhance effective throughput in LLM inference serving. Apt-Serve features a new hybrid cache scheme that combines KV cache with a memory-efficient hidden cache for reusable input hidden state vectors, allowing large batch sizes and improving request concurrency. Based on the hybrid cache, Apt-Serve employs an adaptive runtime scheduling mechanism that dynamically optimizes batch composition. We formally define the adaptive scheduling optimization problem and propose an efficient algorithm with theoretical guarantees. Extensive evaluations on three real-world datasets and LLMs ranging from 13B to 66B parameters demonstrate that Apt-Serve achieves up to 8.8x improvement in effective throughput compared to the state-of-the-art inference serving systems.
Every time an Interactive Storytelling (IS) system gets a player input, it is facing the world-update problem. Classical approaches to this problem consist in mapping that input to known preprogrammed actions, what can severely constrain the free will of the player. When the expected experience has a strong focus on improvisation, like in Role-playing Games (RPGs), this problem is critical. In this paper we present PAYADOR, a different approach that focuses on predicting the outcomes of the actions instead of representing the actions themselves. To implement this approach, we ground a Large Language Model to a minimal representation of the fictional world, obtaining promising results. We make this contribution open-source, so it can be adapted and used for other related research on unleashing the co-creativity power of RPGs.
Modern Large Language Model (LLM) systems typically rely on Retrieval Augmented Generation (RAG) which aims to gather context that is useful for response generation. These RAG systems typically optimize strictly towards retrieving context that is maximally relevant to the query. However, conventional theory suggests that retrieval systems which seek to maximize context relevance without any additional explicit criteria can create information bottlenecks. We reaffirm this finding in the modern age of LLM's by showing that in standard RAG pipelines, maximizing for context relevance alone can degrade downstream response quality. In response, we show evaluations of existing RAG methods which account for both context relevance and answer quality. These evaluations introduce a novel finding that existing RAG systems scale poorly with inference time compute usage when considering our combined metric. We introduce "RErank BEyond reLevance (REBEL)", which enables RAG systems to scale with inference-time compute via injection of multi-criteria optimization using Chain-of-Thought prompting (and optionally Multi-Turn dialogue). Ultimately, this enables a new performance/speed tradeoff curve, where RAG systems are able to achieve both higher relevance of retrieved contexts and superior answer quality as inference time increases. Code for the implementation of our method in llama-index can be found at the following PR: https://github.com/run-llama/llama_index/pull/17590. Code for running experiments using this llama-index implementation can be found at https://github.com/microsoft/REBEL.
This paper presents Thanos, a novel weight-pruning algorithm designed to reduce the memory footprint and enhance the computational efficiency of large language models (LLMs) by removing redundant weights while maintaining accuracy. Thanos introduces a block-wise pruning strategy with adaptive masks that dynamically adjust to weight importance, enabling flexible sparsity patterns and structured formats, such as $n:m$ sparsity, optimized for hardware acceleration. Experimental evaluations demonstrate that Thanos achieves state-of-the-art performance in structured pruning and outperforms existing methods in unstructured pruning. By providing an efficient and adaptable approach to model compression, Thanos offers a practical solution for deploying large models in resource-constrained environments.
Foundation models have demonstrated strong reasoning and generalization capabilities in driving-related tasks, including scene understanding, planning, and control. However, they still face challenges in hallucinations, uncertainty, and long inference latency. While existing foundation models have general knowledge of avoiding collisions, they often lack transportation-specific safety knowledge. To overcome these limitations, we introduce LetsPi, a physics-informed, dual-phase, knowledge-driven framework for safe, human-like trajectory planning. To prevent hallucinations and minimize uncertainty, this hybrid framework integrates Large Language Model (LLM) reasoning with physics-informed social force dynamics. LetsPi leverages the LLM to analyze driving scenes and historical information, providing appropriate parameters and target destinations (goals) for the social force model, which then generates the future trajectory. Moreover, the dual-phase architecture balances reasoning and computational efficiency through its Memory Collection phase and Fast Inference phase. The Memory Collection phase leverages the physics-informed LLM to process and refine planning results through reasoning, reflection, and memory modules, storing safe, high-quality driving experiences in a memory bank. Surrogate safety measures and physics-informed prompt techniques are introduced to enhance the LLM's knowledge of transportation safety and physical force, respectively. The Fast Inference phase extracts similar driving experiences as few-shot examples for new scenarios, while simplifying input-output requirements to enable rapid trajectory planning without compromising safety. Extensive experiments using the HighD dataset demonstrate that LetsPi outperforms baseline models across five safety metrics.See PDF for project Github link.
Large language model-based agents are becoming increasingly popular as a low-cost mechanism to provide personalized, conversational advice, and have demonstrated impressive capabilities in relatively simple scenarios, such as movie recommendations. But how do these agents perform in complex high-stakes domains, where domain expertise is essential and mistakes carry substantial risk? This paper investigates the effectiveness of LLM-advisors in the finance domain, focusing on three distinct challenges: (1) eliciting user preferences when users themselves may be unsure of their needs, (2) providing personalized guidance for diverse investment preferences, and (3) leveraging advisor personality to build relationships and foster trust. Via a lab-based user study with 64 participants, we show that LLM-advisors often match human advisor performance when eliciting preferences, although they can struggle to resolve conflicting user needs. When providing personalized advice, the LLM was able to positively influence user behavior, but demonstrated clear failure modes. Our results show that accurate preference elicitation is key, otherwise, the LLM-advisor has little impact, or can even direct the investor toward unsuitable assets. More worryingly, users appear insensitive to the quality of advice being given, or worse these can have an inverse relationship. Indeed, users reported a preference for and increased satisfaction as well as emotional trust with LLMs adopting an extroverted persona, even though those agents provided worse advice.
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.
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.
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.
Predicting future events stands as one of the ultimate aspirations of artificial intelligence. Recent advances in large language model (LLM)-based systems have shown remarkable potential in forecasting future events, thereby garnering significant interest in the research community. Currently, several benchmarks have been established to evaluate the forecasting capabilities by formalizing the event prediction as a retrieval-augmented generation (RAG) and reasoning task. In these benchmarks, each prediction question is answered with relevant retrieved news articles. However, because there is no consideration on whether the questions can be supported by valid or sufficient supporting rationales, some of the questions in these benchmarks may be inherently noninferable. To address this issue, we introduce a new benchmark, PROPHET, which comprises inferable forecasting questions paired with relevant news for retrieval. To ensure the inferability of the benchmark, we propose Causal Intervened Likelihood (CIL), a statistical measure that assesses inferability through causal inference. In constructing this benchmark, we first collected recent trend forecasting questions and then filtered the data using CIL, resulting in an inferable benchmark for event prediction. Through extensive experiments, we first demonstrate the validity of CIL and in-depth investigations into event prediction with the aid of CIL. Subsequently, we evaluate several representative prediction systems on PROPHET, drawing valuable insights for future directions.
Referring video object segmentation (RVOS) is a challenging task that requires the model to segment the object in a video given the language description. MeViS is a recently proposed dataset that contains motion expressions of the target objects, leading to a challenging benchmark, compared with existing RVOS benchmarks. On the other hand, for referring expression tasks, a new trend is to adopt multi-modal large language model (MLLM) to achieve better image and text alignment. In this report, we show that with a simple modification to the test time inference method on stronger MLLMs, we can lead to stronger results on MeVIS. In particular, we adopt the recent method Sa2VA, a unified model for dense grounded understanding of both images and videos. By enlarging the scope of key frames, without any further training, we can achieve the 3rd place in the 4th PVUW workshop.
Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities.