Accurately quantifying a large language model's (LLM) predictive uncertainty is crucial for judging the reliability of its answers. While most existing research focuses on short, directly answerable questions with closed-form outputs (e.g., multiple-choice), involving intermediate reasoning steps in LLM responses is increasingly important. This added complexity complicates uncertainty quantification (UQ) because the probabilities assigned to answer tokens are conditioned on a vast space of preceding reasoning tokens. Direct marginalization is infeasible, and the dependency inflates probability estimates, causing overconfidence in UQ. To address this, we propose UQAC, an efficient method that narrows the reasoning space to a tractable size for marginalization. UQAC iteratively constructs an "attention chain" of tokens deemed "semantically crucial" to the final answer via a backtracking procedure. Starting from the answer tokens, it uses attention weights to identify the most influential predecessors, then iterates this process until reaching the input tokens. Similarity filtering and probability thresholding further refine the resulting chain, allowing us to approximate the marginal probabilities of the answer tokens, which serve as the LLM's confidence. We validate UQAC on multiple reasoning benchmarks with advanced open-source LLMs, demonstrating that it consistently delivers reliable UQ estimates with high computational efficiency.
Large video-language models (VLMs) have demonstrated promising progress in various video understanding tasks. However, their effectiveness in long-form video analysis is constrained by limited context windows. Traditional approaches, such as uniform frame sampling, often inevitably allocate resources to irrelevant content, diminishing their effectiveness in real-world scenarios. In this paper, we introduce BOLT, a method to BOost Large VLMs without additional Training through a comprehensive study of frame selection strategies. First, to enable a more realistic evaluation of VLMs in long-form video understanding, we propose a multi-source retrieval evaluation setting. Our findings reveal that uniform sampling performs poorly in noisy contexts, underscoring the importance of selecting the right frames. Second, we explore several frame selection strategies based on query-frame similarity and analyze their effectiveness at inference time. Our results show that inverse transform sampling yields the most significant performance improvement, increasing accuracy on the Video-MME benchmark from 53.8% to 56.1% and MLVU benchmark from 58.9% to 63.4%. Our code is available at https://github.com/sming256/BOLT.
Lip Reading, or Visual Automatic Speech Recognition (V-ASR), is a complex task requiring the interpretation of spoken language exclusively from visual cues, primarily lip movements and facial expressions. This task is especially challenging due to the absence of auditory information and the inherent ambiguity when visually distinguishing phonemes that have overlapping visemes where different phonemes appear identical on the lips. Current methods typically attempt to predict words or characters directly from these visual cues, but this approach frequently encounters high error rates due to coarticulation effects and viseme ambiguity. We propose a novel two-stage, phoneme-centric framework for Visual Automatic Speech Recognition (V-ASR) that addresses these longstanding challenges. First, our model predicts a compact sequence of phonemes from visual inputs using a Video Transformer with a CTC head, thereby reducing the task complexity and achieving robust speaker invariance. This phoneme output then serves as the input to a fine-tuned Large Language Model (LLM), which reconstructs coherent words and sentences by leveraging broader linguistic context. Unlike existing methods that either predict words directly-often faltering on visually similar phonemes-or rely on large-scale multimodal pre-training, our approach explicitly encodes intermediate linguistic structure while remaining highly data efficient. We demonstrate state-of-the-art performance on two challenging datasets, LRS2 and LRS3, where our method achieves significant reductions in Word Error Rate (WER) achieving a SOTA WER of 18.7 on LRS3 despite using 99.4% less labelled data than the next best approach.
Adapting Large Language Models (LLMs) to downstream tasks using Reinforcement Learning (RL) has proven to be an effective approach. However, LLMs do not inherently define the structure of an agent for RL training, particularly in terms of defining the action space. This paper studies learning a compact latent action space to enhance the controllability and exploration of RL for LLMs. We propose Controlling Large Language Models with Latent Actions (CoLA), a framework that integrates a latent action space into pre-trained LLMs. We apply CoLA to the Llama-3.1-8B model. Our experiments demonstrate that, compared to RL with token-level actions, CoLA's latent action enables greater semantic diversity in text generation. For enhancing downstream tasks, we show that CoLA with RL achieves a score of 42.4 on the math500 benchmark, surpassing the baseline score of 38.2, and reaches 68.2 when augmented with a Monte Carlo Tree Search variant. Furthermore, CoLA with RL consistently improves performance on agent-based tasks without degrading the pre-trained LLM's capabilities, unlike the baseline. Finally, CoLA reduces computation time by half in tasks involving enhanced thinking prompts for LLMs by RL. These results highlight CoLA's potential to advance RL-based adaptation of LLMs for downstream applications.
Comprehending 3D environments is vital for intelligent systems in domains like robotics and autonomous navigation. Voxel grids offer a structured representation of 3D space, but extracting high-level semantic meaning remains challenging. This paper proposes a novel approach utilizing a Vision-Language Model (VLM) to extract "voxel semantics"-object identity, color, and location-from voxel data. Critically, instead of employing complex 3D networks, our method processes the voxel space by systematically slicing it along a primary axis (e.g., the Z-axis, analogous to CT scan slices). These 2D slices are then formatted and sequentially fed into the image encoder of a standard VLM. The model learns to aggregate information across slices and correlate spatial patterns with semantic concepts provided by the language component. This slice-based strategy aims to leverage the power of pre-trained 2D VLMs for efficient 3D semantic understanding directly from voxel representations.
Accurate and interpretable detection of AI-generated images is essential for mitigating risks associated with AI misuse. However, the substantial domain gap among generative models makes it challenging to develop a generalizable forgery detection model. Moreover, since every pixel in an AI-generated image is synthesized, traditional saliency-based forgery explanation methods are not well suited for this task. To address these challenges, we propose modeling AI-generated image detection and explanation as a Forgery Detection and Reasoning task (FDR-Task), leveraging vision-language models (VLMs) to provide accurate detection through structured and reliable reasoning over forgery attributes. To facilitate this task, we introduce the Multi-Modal Forgery Reasoning dataset (MMFR-Dataset), a large-scale dataset containing 100K images across 10 generative models, with 10 types of forgery reasoning annotations, enabling comprehensive evaluation of FDR-Task. Additionally, we propose FakeReasoning, a forgery detection and reasoning framework with two key components. First, Forgery-Aligned Contrastive Learning enhances VLMs' understanding of forgery-related semantics through both cross-modal and intra-modal contrastive learning between images and forgery attribute reasoning. Second, a Classification Probability Mapper bridges the optimization gap between forgery detection and language modeling by mapping the output logits of VLMs to calibrated binary classification probabilities. Experiments across multiple generative models demonstrate that FakeReasoning not only achieves robust generalization but also outperforms state-of-the-art methods on both detection and reasoning tasks.
Video understanding models often struggle with high computational requirements, extensive parameter counts, and slow inference speed, making them inefficient for practical use. To tackle these challenges, we propose Mobile-VideoGPT, an efficient multimodal framework designed to operate with fewer than a billion parameters. Unlike traditional video large multimodal models (LMMs), Mobile-VideoGPT consists of lightweight dual visual encoders, efficient projectors, and a small language model (SLM), enabling real-time throughput. To further improve efficiency, we present an Attention-Based Frame Scoring mechanism to select the key-frames, along with an efficient token projector that prunes redundant visual tokens and preserves essential contextual cues. We evaluate our model across well-established six video understanding benchmarks (e.g., MVBench, EgoSchema, NextQA, and PercepTest). Our results show that Mobile-VideoGPT-0.5B can generate up to 46 tokens per second while outperforming existing state-of-the-art 0.5B-parameter models by 6 points on average with 40% fewer parameters and more than 2x higher throughput. Our code and models are publicly available at: https://github.com/Amshaker/Mobile-VideoGPT.
Prior work has shown that large language models (LLMs) can predict human attitudes based on other attitudes, but this work has largely focused on predictions from highly similar and interrelated attitudes. In contrast, human attitudes are often strongly associated even across disparate and dissimilar topics. Using a novel dataset of human responses toward diverse attitude statements, we found that a frontier language model (GPT-4o) was able to recreate the pairwise correlations among individual attitudes and to predict individuals' attitudes from one another. Crucially, in an advance over prior work, we tested GPT-4o's ability to predict in the absence of surface-similarity between attitudes, finding that while surface similarity improves prediction accuracy, the model was still highly-capable of generating meaningful social inferences between dissimilar attitudes. Altogether, our findings indicate that LLMs capture crucial aspects of the deeper, latent structure of human belief systems.
D4R is a digital platform designed to assist non-technical users, particularly historians, in exploring textual documents through advanced graphical tools for text analysis and knowledge extraction. By leveraging a large language model, D4R translates natural language questions into Cypher queries, enabling the retrieval of data from a Neo4J database. A user-friendly graphical interface allows for intuitive interaction, enabling users to navigate and analyse complex relational data extracted from unstructured textual documents. Originally designed to bridge the gap between AI technologies and historical research, D4R's capabilities extend to various other domains. A demonstration video and a live software demo are available.
Using large teacher models to guide the training of smaller student models has become the prevailing paradigm for efficient and effective learning. However, vocabulary mismatches between teacher and student language models pose significant challenges in language modeling, resulting in divergent token sequences and output distributions. To overcome these limitations, we propose Vocabulary-agnostic Teacher Guided Language Modeling (VocAgnoLM), a novel approach that bridges the gap caused by vocabulary mismatch through two key methods: (1) Token-level Lexical Alignment, which aligns token sequences across mismatched vocabularies, and (2) Teacher Guided Loss, which leverages the loss of teacher model to guide effective student training. We demonstrate its effectiveness in language modeling with 1B student model using various 7B teacher models with different vocabularies. Notably, with Qwen2.5-Math-Instruct, a teacher model sharing only about 6% of its vocabulary with TinyLlama, VocAgnoLM achieves a 46% performance improvement compared to naive continual pretraining. Furthermore, we demonstrate that VocAgnoLM consistently benefits from stronger teacher models, providing a robust solution to vocabulary mismatches in language modeling.
In traditional innovation practices, concept and IP generation are often iteratively integrated. Both processes demand an intricate understanding of advanced technical domain knowledge. Existing large language models (LLMs), while possessing massive pre-trained knowledge, often fall short in the innovative concept generation due to a lack of specialized knowledge necessary for the generation. To bridge this critical gap, we propose a novel knowledge finetuning (KFT) framework to endow LLM-based AI with the ability to autonomously mine, understand, and apply domain-specific knowledge and concepts for invention generation, i.e., concept and patent generation together. Our proposed PatentGPT integrates knowledge injection pre-training (KPT), domain-specific supervised finetuning (SFT), and reinforcement learning from human feedback (RLHF). Extensive evaluation shows that PatentGPT significantly outperforms the state-of-the-art models on patent-related benchmark tests. Our method not only provides new insights into data-driven innovation but also paves a new path to fine-tune LLMs for applications in the context of technology. We also discuss the managerial and policy implications of AI-generating inventions in the future.
We study last-layer outlier dimensions, i.e. dimensions that display extreme activations for the majority of inputs. We show that outlier dimensions arise in many different modern language models, and trace their function back to the heuristic of constantly predicting frequent words. We further show how a model can block this heuristic when it is not contextually appropriate, by assigning a counterbalancing weight mass to the remaining dimensions, and we investigate which model parameters boost outlier dimensions and when they arise during training. We conclude that outlier dimensions are a specialized mechanism discovered by many distinct models to implement a useful token prediction heuristic.
In recent years, large language models (LLMs) have shown significant advancements in natural language processing (NLP), with strong capa-bilities in generation, comprehension, and rea-soning. These models have found applications in education, intelligent decision-making, and gaming. However, effectively utilizing LLMs for strategic planning and decision-making in the game of Gomoku remains a challenge. This study aims to develop a Gomoku AI system based on LLMs, simulating the human learning process of playing chess. The system is de-signed to understand and apply Gomoku strat-egies and logic to make rational decisions. The research methods include enabling the model to "read the board," "understand the rules," "select strategies," and "evaluate positions," while en-hancing its abilities through self-play and rein-forcement learning. The results demonstrate that this approach significantly improves the se-lection of move positions, resolves the issue of generating illegal positions, and reduces pro-cess time through parallel position evaluation. After extensive self-play training, the model's Gomoku-playing capabilities have been notably enhanced.
Large Language Models (LLMs) have transformed task automation and content generation across various domains while incorporating safety filters to prevent misuse. We introduce a novel jailbreaking framework that employs distributed prompt processing combined with iterative refinements to bypass these safety measures, particularly in generating malicious code. Our architecture consists of four key modules: prompt segmentation, parallel processing, response aggregation, and LLM-based jury evaluation. Tested on 500 malicious prompts across 10 cybersecurity categories, the framework achieves a 73.2% Success Rate (SR) in generating malicious code. Notably, our comparative analysis reveals that traditional single-LLM judge evaluation overestimates SRs (93.8%) compared to our LLM jury system (73.2%), with manual verification confirming that single-judge assessments often accept incomplete implementations. Moreover, we demonstrate that our distributed architecture improves SRs by 12% over the non-distributed approach in an ablation study, highlighting both the effectiveness of distributed prompt processing and the importance of robust evaluation methodologies in assessing jailbreak attempts.
As the Information Retrieval (IR) field increasingly recognizes the importance of inclusivity, addressing the needs of low-resource languages remains a significant challenge. Transliteration between Urdu and its Romanized form, Roman Urdu, remains underexplored despite the widespread use of both scripts in South Asia. Prior work using RNNs on the Roman-Urdu-Parl dataset showed promising results but suffered from poor domain adaptability and limited evaluation. We propose a transformer-based approach using the m2m100 multilingual translation model, enhanced with masked language modeling (MLM) pretraining and fine-tuning on both Roman-Urdu-Parl and the domain-diverse Dakshina dataset. To address previous evaluation flaws, we introduce rigorous dataset splits and assess performance using BLEU, character-level BLEU, and CHRF. Our model achieves strong transliteration performance, with Char-BLEU scores of 96.37 for Urdu->Roman-Urdu and 97.44 for Roman-Urdu->Urdu. These results outperform both RNN baselines and GPT-4o Mini and demonstrate the effectiveness of multilingual transfer learning for low-resource transliteration tasks.
Large language models afford opportunities for using computers for intensive tasks, realizing research opportunities that have not been considered before. One such opportunity could be a systematic interrogation of the scientific literature. Here, we show how a large language model can be used to construct a literature review of 2699 publications associated with microphysics parametrizations in the Weather and Research Forecasting (WRF) model, with the goal of learning how they were used and their systematic biases, when simulating precipitation. The database was constructed of publications identified from Web of Science and Scopus searches. The large language model GPT-4 Turbo was used to extract information about model configurations and performance from the text of 2699 publications. Our results reveal the landscape of how nine of the most popular microphysics parameterizations have been used around the world: Lin, Ferrier, WRF Single-Moment, Goddard Cumulus Ensemble, Morrison, Thompson, and WRF Double-Moment. More studies used one-moment parameterizations before 2020 and two-moment parameterizations after 2020. Seven out of nine parameterizations tended to overestimate precipitation. However, systematic biases of parameterizations differed in various regions. Except simulations using the Lin, Ferrier, and Goddard parameterizations that tended to underestimate precipitation over almost all locations, the remaining six parameterizations tended to overestimate, particularly over China, southeast Asia, western United States, and central Africa. This method could be used by other researchers to help understand how the increasingly massive body of scientific literature can be harnessed through the power of artificial intelligence to solve their research problems.
Intrusion Detection Systems (IDS) are crucial for identifying malicious traffic, yet traditional signature-based methods struggle with zero-day attacks and high false positive rates. AI-driven packet-capture analysis offers a promising alternative. However, existing approaches rely heavily on flow-based or statistical features, limiting their ability to detect fine-grained attack patterns. This study proposes Xavier-CMAE, an enhanced Convolutional Multi-Head Attention Ensemble (CMAE) model that improves detection accuracy while reducing computational overhead. By replacing Word2Vec embeddings with a Hex2Int tokenizer and Xavier initialization, Xavier-CMAE eliminates pre-training, accelerates training, and achieves 99.971% accuracy with a 0.018% false positive rate, outperforming Word2Vec-based methods. Additionally, we introduce LLM-CMAE, which integrates pre-trained Large Language Model (LLM) tokenizers into CMAE. While LLMs enhance feature extraction, their computational cost hinders real-time detection. LLM-CMAE balances efficiency and performance, reaching 99.969% accuracy with a 0.019% false positive rate. This work advances AI-powered IDS by (1) introducing a payload-based detection framework, (2) enhancing efficiency with Xavier-CMAE, and (3) integrating LLM tokenizers for improved real-time detection.
Currently, instruction-based image editing methods have made significant progress by leveraging the powerful cross-modal understanding capabilities of vision language models (VLMs). However, they still face challenges in three key areas: 1) complex scenarios; 2) semantic consistency; and 3) fine-grained editing. To address these issues, we propose FireEdit, an innovative Fine-grained Instruction-based image editing framework that exploits a REgion-aware VLM. FireEdit is designed to accurately comprehend user instructions and ensure effective control over the editing process. Specifically, we enhance the fine-grained visual perception capabilities of the VLM by introducing additional region tokens. Relying solely on the output of the LLM to guide the diffusion model may lead to suboptimal editing results. Therefore, we propose a Time-Aware Target Injection module and a Hybrid Visual Cross Attention module. The former dynamically adjusts the guidance strength at various denoising stages by integrating timestep embeddings with the text embeddings. The latter enhances visual details for image editing, thereby preserving semantic consistency between the edited result and the source image. By combining the VLM enhanced with fine-grained region tokens and the time-dependent diffusion model, FireEdit demonstrates significant advantages in comprehending editing instructions and maintaining high semantic consistency. Extensive experiments indicate that our approach surpasses the state-of-the-art instruction-based image editing methods. Our project is available at https://zjgans.github.io/fireedit.github.io.
We study how large language models can be used in combination with evolutionary computation techniques to automatically discover optimization algorithms for the design of photonic structures. Building on the Large Language Model Evolutionary Algorithm (LLaMEA) framework, we introduce structured prompt engineering tailored to multilayer photonic problems such as Bragg mirror, ellipsometry inverse analysis, and solar cell antireflection coatings. We systematically explore multiple evolutionary strategies, including (1+1), (1+5), (2+10), and others, to balance exploration and exploitation. Our experiments show that LLM-generated algorithms, generated using small-scale problem instances, can match or surpass established methods like quasi-oppositional differential evolution on large-scale realistic real-world problem instances. Notably, LLaMEA's self-debugging mutation loop, augmented by automatically extracted problem-specific insights, achieves strong anytime performance and reliable convergence across diverse problem scales. This work demonstrates the feasibility of domain-focused LLM prompts and evolutionary approaches in solving optical design tasks, paving the way for rapid, automated photonic inverse design.
The AM-DeepSeek-R1-Distilled is a large-scale dataset with thinking traces for general reasoning tasks, composed of high-quality and challenging reasoning problems. These problems are collected from a multitude of open-source datasets, subjected to semantic deduplication and meticulous cleaning to eliminate test set contamination. All responses within the dataset are distilled from reasoning models (predominantly DeepSeek-R1) and have undergone rigorous verification procedures. Mathematical problems are validated by checking against reference answers, code problems are verified using test cases, and other tasks are evaluated with the aid of a reward model. The AM-Distill-Qwen-32B model, which was trained through only simple Supervised Fine-Tuning (SFT) using this batch of data, outperformed the DeepSeek-R1-Distill-Qwen-32B model on four benchmarks: AIME2024, MATH-500, GPQA-Diamond, and LiveCodeBench. Additionally, the AM-Distill-Qwen-72B model surpassed the DeepSeek-R1-Distill-Llama-70B model on all benchmarks as well. We are releasing these 1.4 million problems and their corresponding responses to the research community with the objective of fostering the development of powerful reasoning-oriented Large Language Models (LLMs). The dataset was published in \href{https://huggingface.co/datasets/a-m-team/AM-DeepSeek-R1-Distilled-1.4M}{https://huggingface.co/datasets/a-m-team/AM-DeepSeek-R1-Distilled-1.4M}.