Grapheme-to-phoneme (G2P) conversion for Persian presents unique challenges due to its complex phonological features, particularly homographs and Ezafe, which exist in formal and informal language contexts. This paper introduces an intermediate language specifically designed for Persian language processing that addresses these challenges through a multi-faceted approach. Our methodology combines two key components: Large Language Model (LLM) prompting techniques and a specialized sequence-to-sequence machine transliteration architecture. We developed and implemented a systematic approach for constructing a comprehensive lexical database for homographs with multiple pronunciations disambiguation often termed polyphones, utilizing formal concept analysis for semantic differentiation. We train our model using two distinct datasets: the LLM-generated dataset for formal and informal Persian and the B-Plus podcasts for informal language variants. The experimental results demonstrate superior performance compared to existing state-of-the-art approaches, particularly in handling the complexities of Persian phoneme conversion. Our model significantly improves Phoneme Error Rate (PER) metrics, establishing a new benchmark for Persian G2P conversion accuracy. This work contributes to the growing research in low-resource language processing and provides a robust solution for Persian text-to-speech systems and demonstrating its applicability beyond Persian. Specifically, the approach can extend to languages with rich homographic phenomena such as Chinese and Arabic
The common assumption in on-device AI is that GPUs, with their superior parallel processing, always provide the best performance for large language model (LLM) inference. In this work, we challenge this notion by empirically demonstrating that, under certain conditions, CPUs can outperform GPUs for LLM inference on mobile devices. Using a 1-billion-parameter LLM deployed via llama.cpp on the iPhone 15 Pro, we show that a CPU-only configuration (two threads, F16 precision) achieves 17 tokens per second, surpassing the 12.8 tokens per second obtained with GPU acceleration. We analyze the architectural factors driving this counterintuitive result, revealing that GPU memory transfer overhead and CPU thread optimization play a critical role. Furthermore, we explore the impact of thread oversubscription, quantization strategies, and hardware constraints, providing new insights into efficient on-device AI execution. Our findings challenge conventional GPU-first thinking, highlighting the untapped potential of optimized CPU inference and paving the way for smarter deployment strategies in mobile AI. However, fully explaining the observed CPU advantage remains difficult due to limited access to low-level profiling tools on iOS.
Multimodal large language models (MLLMs) demonstrate impressive performance on scientific reasoning tasks (e.g., ScienceQA). However, most existing benchmarks focus narrowly on the accuracy of the final answer while ignoring other metrics. In particular, when applying MLLMs to educational contexts, the goal is not only correctness but also the ability to teach. In this paper, we propose a framework that evaluates MLLMs as science tutors using a comprehensive educational rubric and a simulated student model that judges the teaching performance of the tutors. Given a list of candidate MLLM science tutors, we use rubric-based student judgments to produce a range of tutor performance scores, identifying both strong and weak tutors. Using the training section of the ScienceQA dataset, we then construct a data set of pairwise comparisons between the outputs of strong and weak tutors. This enables us to apply multiple preference optimization methods to fine-tune an underperforming tutor model (Qwen2-VL-2B) into more effective ones. Our results also show that strong problem-solving skills do not guarantee high-quality tutoring and that performance optimization-guided refinements can yield more educationally aligned tutor models. This approach opens avenues for building MLLMs that serve not only as problem solvers, but as genuinely helpful educational assistants.
We introduce MLE-Dojo, a Gym-style framework for systematically reinforcement learning, evaluating, and improving autonomous large language model (LLM) agents in iterative machine learning engineering (MLE) workflows. Unlike existing benchmarks that primarily rely on static datasets or single-attempt evaluations, MLE-Dojo provides an interactive environment enabling agents to iteratively experiment, debug, and refine solutions through structured feedback loops. Built upon 200+ real-world Kaggle challenges, MLE-Dojo covers diverse, open-ended MLE tasks carefully curated to reflect realistic engineering scenarios such as data processing, architecture search, hyperparameter tuning, and code debugging. Its fully executable environment supports comprehensive agent training via both supervised fine-tuning and reinforcement learning, facilitating iterative experimentation, realistic data sampling, and real-time outcome verification. Extensive evaluations of eight frontier LLMs reveal that while current models achieve meaningful iterative improvements, they still exhibit significant limitations in autonomously generating long-horizon solutions and efficiently resolving complex errors. Furthermore, MLE-Dojo's flexible and extensible architecture seamlessly integrates diverse data sources, tools, and evaluation protocols, uniquely enabling model-based agent tuning and promoting interoperability, scalability, and reproducibility. We open-source our framework and benchmarks to foster community-driven innovation towards next-generation MLE agents.
Post-training quantization is a key technique for reducing the memory and inference latency of large language models by quantizing weights and activations without requiring retraining. However, existing methods either (1) fail to account for the varying importance of hidden features to the end loss or, when incorporating end loss, (2) neglect the critical interactions between model weights. To address these limitations, we propose GuidedQuant, a novel quantization approach that integrates gradient information from the end loss into the quantization objective while preserving cross-weight dependencies within output channels. GuidedQuant consistently boosts the performance of state-of-the-art quantization methods across weight-only scalar, weight-only vector, and weight-and-activation quantization. Additionally, we introduce a novel non-uniform scalar quantization algorithm, which is guaranteed to monotonically decrease the quantization objective value, and outperforms existing methods in this category. We release the code at https://github.com/snu-mllab/GuidedQuant.
Micro-expressions (MEs) are crucial psychological responses with significant potential for affective computing. However, current automatic micro-expression recognition (MER) research primarily focuses on discrete emotion classification, neglecting a convincing analysis of the subtle dynamic movements and inherent emotional cues. The rapid progress in multimodal large language models (MLLMs), known for their strong multimodal comprehension and language generation abilities, offers new possibilities. MLLMs have shown success in various vision-language tasks, indicating their potential to understand MEs comprehensively, including both fine-grained motion patterns and underlying emotional semantics. Nevertheless, challenges remain due to the subtle intensity and short duration of MEs, as existing MLLMs are not designed to capture such delicate frame-level facial dynamics. In this paper, we propose a novel Micro-Expression Large Language Model (MELLM), which incorporates a subtle facial motion perception strategy with the strong inference capabilities of MLLMs, representing the first exploration of MLLMs in the domain of ME analysis. Specifically, to explicitly guide the MLLM toward motion-sensitive regions, we construct an interpretable motion-enhanced color map by fusing onset-apex optical flow dynamics with the corresponding grayscale onset frame as the model input. Additionally, specialized fine-tuning strategies are incorporated to further enhance the model's visual perception of MEs. Furthermore, we construct an instruction-description dataset based on Facial Action Coding System (FACS) annotations and emotion labels to train our MELLM. Comprehensive evaluations across multiple benchmark datasets demonstrate that our model exhibits superior robustness and generalization capabilities in ME understanding (MEU). Code is available at https://github.com/zyzhangUstc/MELLM.
This paper proposes batch augmentation with unimodal fine-tuning to detect the fetus's organs from ultrasound images and associated clinical textual information. We also prescribe pre-training initial layers with investigated medical data before the multimodal training. At first, we apply a transferred initialization with the unimodal image portion of the dataset with batch augmentation. This step adjusts the initial layer weights for medical data. Then, we apply neural networks (NNs) with fine-tuned initial layers to images in batches with batch augmentation to obtain features. We also extract information from descriptions of images. We combine this information with features obtained from images to train the head layer. We write a dataloader script to load the multimodal data and use existing unimodal image augmentation techniques with batch augmentation for the multimodal data. The dataloader brings a new random augmentation for each batch to get a good generalization. We investigate the FPU23 ultrasound and UPMC Food-101 multimodal datasets. The multimodal large language model (LLM) with the proposed training provides the best results among the investigated methods. We receive near state-of-the-art (SOTA) performance on the UPMC Food-101 dataset. We share the scripts of the proposed method with traditional counterparts at the following repository: github.com/dipuk0506/multimodal
In this paper, we present Optimized Prompt-based Unified System (OPUS), a framework that utilizes a Large Language Model (LLM) to control Pan-Tilt-Zoom (PTZ) cameras, providing contextual understanding of natural environments. To achieve this goal, the OPUS system improves cost-effectiveness by generating keywords from a high-level camera control API and transferring knowledge from larger closed-source language models to smaller ones through Supervised Fine-Tuning (SFT) on synthetic data. This enables efficient edge deployment while maintaining performance comparable to larger models like GPT-4. OPUS enhances environmental awareness by converting data from multiple cameras into textual descriptions for language models, eliminating the need for specialized sensory tokens. In benchmark testing, our approach significantly outperformed both traditional language model techniques and more complex prompting methods, achieving a 35% improvement over advanced techniques and a 20% higher task accuracy compared to closed-source models like Gemini Pro. The system demonstrates OPUS's capability to simplify PTZ camera operations through an intuitive natural language interface. This approach eliminates the need for explicit programming and provides a conversational method for interacting with camera systems, representing a significant advancement in how users can control and utilize PTZ camera technology.
We propose an automated framework for quantum circuit design by integrating large-language models (LLMs) with evolutionary optimization to overcome the rigidity, scalability limitations, and expert dependence of traditional ones in variational quantum algorithms. Our approach (FunSearch) autonomously discovers hardware-efficient ans\"atze with new features of scalability and system-size-independent number of variational parameters entirely from scratch. Demonstrations on the Ising and XY spin chains with n = 9 qubits yield circuits containing 4 parameters, achieving near-exact energy extrapolation across system sizes. Implementations on quantum hardware (Zuchongzhi chip) validate practicality, where two-qubit quantum gate noises can be effectively mitigated via zero-noise extrapolations for a spin chain system as large as 20 sites. This framework bridges algorithmic design and experimental constraints, complementing contemporary quantum architecture search frameworks to advance scalable quantum simulations.
Educators evaluate student knowledge using knowledge component (KC) models that map assessment questions to KCs. Still, designing KC models for large question banks remains an insurmountable challenge for instructors who need to analyze each question by hand. The growing use of Generative AI in education is expected only to aggravate this chronic deficiency of expert-designed KC models, as course engineers designing KCs struggle to keep up with the pace at which questions are generated. In this work, we propose KCluster, a novel KC discovery algorithm based on identifying clusters of congruent questions according to a new similarity metric induced by a large language model (LLM). We demonstrate in three datasets that an LLM can create an effective metric of question similarity, which a clustering algorithm can use to create KC models from questions with minimal human effort. Combining the strengths of LLM and clustering, KCluster generates descriptive KC labels and discovers KC models that predict student performance better than the best expert-designed models available. In anticipation of future work, we illustrate how KCluster can reveal insights into difficult KCs and suggest improvements to instruction.
As the Large-Language-Model-driven (LLM-driven) Artificial Intelligence (AI) bots became popular, people realized their strong potential in Task-Oriented Dialogue (TOD). However, bots relying wholly on LLMs are unreliable in their knowledge, and whether they can finally produce a correct result for the task is not guaranteed. The collaboration among these agents also remains a challenge, since the necessary information to convey is unclear, and the information transfer is by prompts -- unreliable, and malicious knowledge is easy to inject. With the help of logic programming tools such as Answer Set Programming (ASP), conversational agents can be built safely and reliably, and communication among the agents made more efficient and secure. We proposed an Administrator-Assistant Dual-Agent paradigm, where the two ASP-driven bots share the same knowledge base and complete their tasks independently, while the information can be passed by a Collaborative Rule Set (CRS). The knowledge and information conveyed are encapsulated and invisible to the users, ensuring the security of information transmission. We have constructed AutoManager, a dual-agent system for managing the drive-through window of a fast-food restaurant such as Taco Bell in the US. In AutoManager, the assistant bot takes the customer's order while the administrator bot manages the menu and food supply. We evaluated our AutoManager and compared it with the real-world Taco Bell Drive-Thru AI Order Taker, and the results show that our method is more reliable.
In November 2022, Europe and the world by and large were stunned by the birth of a new large language model : ChatGPT. Ever since then, both academic and populist discussions have taken place in various public spheres such as LinkedIn and X(formerly known as Twitter) with the view to both understand the tool and its benefits for the society. The views of real actors in professional spaces, especially in regulated industries such as finance and law have been largely missing. We aim to begin to close this gap by presenting results from an empirical investigation conducted through interviews with professional actors in the Fintech industry. The paper asks the question, how and to what extent are large language models in general and ChatGPT in particular being adopted and used in the Fintech industry? The results show that while the fintech experts we spoke with see a potential in using large language models in the future, a lot of questions marks remain concerning how they are policed and therefore might be adopted in a regulated industry such as Fintech. This paper aims to add to the existing academic discussing around large language models, with a contribution to our understanding of professional viewpoints.
The exponential growth in Large Language Model (LLM) deployment has intensified the need for efficient model compression techniques to reduce computational and memory costs. While pruning and quantization have shown promise, their combined potential remains largely unexplored. In this paper, we examine joint compression and how strategically combining pruning and quantization could yield superior performance-to-compression ratios compared to single-method approaches. Recognizing the challenges in accurately assessing LLM performance, we address key limitations of previous evaluation frameworks and introduce the Semantic Retention Compression Rate (SrCr), a novel metric that quantifies the trade-off between model compression and semantic preservation, facilitating the optimization of pruning-quantization configurations. Experiments demonstrate that our recommended combination achieves, on average, a 20% performance increase compared to an equivalent quantization-only model at the same theoretical compression rate.
With the growing use of large language models (LLMs) hosted on cloud platforms to offer inference services, privacy concerns about the potential leakage of sensitive information are escalating. Secure multi-party computation (MPC) is a promising solution to protect the privacy in LLM inference. However, MPC requires frequent inter-server communication, causing high performance overhead. Inspired by the prevalent activation sparsity of LLMs, where most neuron are not activated after non-linear activation functions, we propose an efficient private inference system, Comet. This system employs an accurate and fast predictor to predict the sparsity distribution of activation function output. Additionally, we introduce a new private inference protocol. It efficiently and securely avoids computations involving zero values by exploiting the spatial locality of the predicted sparse distribution. While this computation-avoidance approach impacts the spatiotemporal continuity of KV cache entries, we address this challenge with a low-communication overhead cache refilling strategy that merges miss requests and incorporates a prefetching mechanism. Finally, we evaluate Comet on four common LLMs and compare it with six state-of-the-art private inference systems. Comet achieves a 1.87x-2.63x speedup and a 1.94x-2.64x communication reduction.
Retrieval-augmented generation (RAG) systems combine large language models (LLMs) with external knowledge retrieval, making them highly effective for knowledge-intensive tasks. A crucial but often under-explored component of these systems is the reranker. Since irrelevant documents in RAG systems can mislead the generator, the reranker plays a vital role in refining retrieved documents to enhance generation quality and explainability. However, it is challenging to determine the appropriate number of documents ($k$) that the reranker should select: too few may result in missing critical information, while too many introduce noise and inefficiencies. Although recent studies have explored LLM-based rerankers, they primarily leverage internal model knowledge and overlook the rich supervisory signals that LLMs can provide, such as using response quality as feedback for optimizing reranking decisions. In this paper, we propose DynamicRAG, a novel RAG framework where the reranker dynamically adjusts both the order and number of retrieved documents based on the query. We model the reranker as an agent optimized through reinforcement learning (RL), using rewards derived from LLM output quality. Across seven knowledge-intensive datasets, DynamicRAG demonstrates superior performance, achieving state-of-the-art results among models of same parameter sizes. The model, data and code are available at https://github.com/GasolSun36/DynamicRAG.
We present gg-bench, a collection of game environments designed to evaluate general reasoning capabilities in language models. Unlike most static benchmarks, gg-bench is a data generating process where new evaluation instances can be generated at will. In particular, gg-bench is synthetically generated by (1) using a large language model (LLM) to generate natural language descriptions of novel games, (2) using the LLM to implement each game in code as a Gym environment, and (3) training reinforcement learning (RL) agents via self-play on the generated games. We evaluate language models by their winrate against these RL agents by prompting models with the game description, current board state, and a list of valid moves, after which models output the moves they wish to take. gg-bench is challenging: state-of-the-art LLMs such as GPT-4o and Claude 3.7 Sonnet achieve winrates of 7-9% on gg-bench using in-context learning, while reasoning models such as o1, o3-mini and DeepSeek-R1 achieve average winrates of 31-36%. We release the generated games, data generation process, and evaluation code in order to support future modeling work and expansion of our benchmark.
As Test-Time Scaling emerges as an active research focus in the large language model community, advanced post-training methods increasingly emphasize extending chain-of-thought (CoT) generation length, thereby enhancing reasoning capabilities to approach Deepseek R1-like reasoning models. However, recent studies reveal that reasoning models (even Qwen3) consistently exhibit excessive thought redundancy in CoT generation. This overthinking problem stems from conventional outcome-reward reinforcement learning's systematic neglect in regulating intermediate reasoning steps. This paper proposes Serial-Group Decaying-Reward Policy Optimization (namely S-GRPO), a novel reinforcement learning method that empowers models with the capability to determine the sufficiency of reasoning steps, subsequently triggering early exit of CoT generation. Specifically, unlike GRPO, which samples multiple possible completions (parallel group) in parallel, we select multiple temporal positions in the generation of one CoT to allow the model to exit thinking and instead generate answers (serial group), respectively. For the correct answers in a serial group, we assign rewards that decay according to positions, with lower rewards towards the later ones, thereby reinforcing the model's behavior to generate higher-quality answers at earlier phases with earlier exits of thinking. Empirical evaluations demonstrate compatibility with state-of-the-art reasoning models, including Qwen3 and Deepseek-distill models, achieving 35.4% ~ 61.1\% sequence length reduction with 0.72% ~ 6.08% accuracy improvements across GSM8K, AIME 2024, AMC 2023, MATH-500, and GPQA Diamond benchmarks.
We present MiMo-7B, a large language model born for reasoning tasks, with optimization across both pre-training and post-training stages. During pre-training, we enhance the data preprocessing pipeline and employ a three-stage data mixing strategy to strengthen the base model's reasoning potential. MiMo-7B-Base is pre-trained on 25 trillion tokens, with additional Multi-Token Prediction objective for enhanced performance and accelerated inference speed. During post-training, we curate a dataset of 130K verifiable mathematics and programming problems for reinforcement learning, integrating a test-difficulty-driven code-reward scheme to alleviate sparse-reward issues and employing strategic data resampling to stabilize training. Extensive evaluations show that MiMo-7B-Base possesses exceptional reasoning potential, outperforming even much larger 32B models. The final RL-tuned model, MiMo-7B-RL, achieves superior performance on mathematics, code and general reasoning tasks, surpassing the performance of OpenAI o1-mini. The model checkpoints are available at https://github.com/xiaomimimo/MiMo.
Leveraging large language model (LLM) based agents to simulate human social behaviors has recently gained significant attention. In this paper, we introduce a novel social simulator called YuLan-OneSim. Compared to previous works, YuLan-OneSim distinguishes itself in five key aspects: (1) Code-free scenario construction: Users can simply describe and refine their simulation scenarios through natural language interactions with our simulator. All simulation code is automatically generated, significantly reducing the need for programming expertise. (2) Comprehensive default scenarios: We implement 50 default simulation scenarios spanning 8 domains, including economics, sociology, politics, psychology, organization, demographics, law, and communication, broadening access for a diverse range of social researchers. (3) Evolvable simulation: Our simulator is capable of receiving external feedback and automatically fine-tuning the backbone LLMs, significantly enhancing the simulation quality. (4) Large-scale simulation: By developing a fully responsive agent framework and a distributed simulation architecture, our simulator can handle up to 100,000 agents, ensuring more stable and reliable simulation results. (5) AI social researcher: Leveraging the above features, we develop an AI social researcher. Users only need to propose a research topic, and the AI researcher will automatically analyze the input, construct simulation environments, summarize results, generate technical reports, review and refine the reports--completing the social science research loop. To demonstrate the advantages of YuLan-OneSim, we conduct experiments to evaluate the quality of the automatically generated scenarios, the reliability, efficiency, and scalability of the simulation process, as well as the performance of the AI social researcher.
Social media user profiling through content analysis is crucial for tasks like misinformation detection, engagement prediction, hate speech monitoring, and user behavior modeling. However, existing profiling techniques, including tweet summarization, attribute-based profiling, and latent representation learning, face significant limitations: they often lack transferability, produce non-interpretable features, require large labeled datasets, or rely on rigid predefined categories that limit adaptability. We introduce a novel large language model (LLM)-based approach that leverages domain-defining statements, which serve as key characteristics outlining the important pillars of a domain as foundations for profiling. Our two-stage method first employs semi-supervised filtering with a domain-specific knowledge base, then generates both abstractive (synthesized descriptions) and extractive (representative tweet selections) user profiles. By harnessing LLMs' inherent knowledge with minimal human validation, our approach is adaptable across domains while reducing the need for large labeled datasets. Our method generates interpretable natural language user profiles, condensing extensive user data into a scale that unlocks LLMs' reasoning and knowledge capabilities for downstream social network tasks. We contribute a Persian political Twitter (X) dataset and an LLM-based evaluation framework with human validation. Experimental results show our method significantly outperforms state-of-the-art LLM-based and traditional methods by 9.8%, demonstrating its effectiveness in creating flexible, adaptable, and interpretable user profiles.