Steer AGI

As large language models (LLMs) become increasingly capable, ensuring that their self-generated explanations are faithful to their internal decision-making process is critical for safety and oversight. In this work, we conduct a comprehensive counterfactual faithfulness analysis across 62 models from 8 families, encompassing both pretrained and instruction-tuned variants and significantly extending prior studies of counterfactual tests. We introduce phi-CCT, a simplified variant of the Correlational Counterfactual Test, which avoids the need for token probabilities while explaining most of the variance of the original test. Our findings reveal clear scaling trends: larger models are consistently more faithful on our metrics. However, when comparing instruction-tuned and human-imitated explanations, we find that observed differences in faithfulness can often be attributed to explanation verbosity, leading to shifts along the true-positive/false-positive Pareto frontier. While instruction-tuning and prompting can influence this trade-off, we find limited evidence that they fundamentally expand the frontier of explanatory faithfulness beyond what is achievable with pretrained models of comparable size. Our analysis highlights the nuanced relationship between instruction-tuning, verbosity, and the faithful representation of model decision processes.

With the rapid development of artificial intelligence, intelligent decision-making techniques have gradually surpassed human levels in various human-machine competitions, especially in complex multi-agent cooperative task scenarios. Multi-agent cooperative decision-making involves multiple agents working together to complete established tasks and achieve specific objectives. These techniques are widely applicable in real-world scenarios such as autonomous driving, drone navigation, disaster rescue, and simulated military confrontations. This paper begins with a comprehensive survey of the leading simulation environments and platforms used for multi-agent cooperative decision-making. Specifically, we provide an in-depth analysis for these simulation environments from various perspectives, including task formats, reward allocation, and the underlying technologies employed. Subsequently, we provide a comprehensive overview of the mainstream intelligent decision-making approaches, algorithms and models for multi-agent systems (MAS). Theseapproaches can be broadly categorized into five types: rule-based (primarily fuzzy logic), game theory-based, evolutionary algorithms-based, deep multi-agent reinforcement learning (MARL)-based, and large language models(LLMs)reasoning-based. Given the significant advantages of MARL andLLMs-baseddecision-making methods over the traditional rule, game theory, and evolutionary algorithms, this paper focuses on these multi-agent methods utilizing MARL and LLMs-based techniques. We provide an in-depth discussion of these approaches, highlighting their methodology taxonomies, advantages, and drawbacks. Further, several prominent research directions in the future and potential challenges of multi-agent cooperative decision-making are also detailed.

Large Language Models (LLMs) have achieved remarkable success across diverse tasks, largely driven by well-designed prompts. However, crafting and selecting such prompts often requires considerable human effort, significantly limiting its scalability. To mitigate this, recent studies have explored automated prompt optimization as a promising solution. Despite these efforts, existing methods still face critical challenges in robustness, efficiency, and generalization. To systematically address these challenges, we first conduct an empirical analysis to identify the limitations of current reflection-based prompt optimization paradigm. Building on these insights, we propose 7 innovative approaches inspired by traditional deep learning paradigms for prompt optimization (DLPO), seamlessly integrating these concepts into text-based gradient optimization. Through these advancements, we progressively tackle the aforementioned challenges and validate our methods through extensive experimentation. We hope our study not only provides valuable guidance for future research but also offers a comprehensive understanding of the challenges and potential solutions in prompt optimization. Our code is available at https://github.com/sfasfaffa/DLPO.

Classifier-free guidance has become a staple for conditional generation with denoising diffusion models. However, a comprehensive understanding of classifier-free guidance is still missing. In this work, we carry out an empirical study to provide a fresh perspective on classifier-free guidance. Concretely, instead of solely focusing on classifier-free guidance, we trace back to the root, i.e., classifier guidance, pinpoint the key assumption for the derivation, and conduct a systematic study to understand the role of the classifier. We find that both classifier guidance and classifier-free guidance achieve conditional generation by pushing the denoising diffusion trajectories away from decision boundaries, i.e., areas where conditional information is usually entangled and is hard to learn. Based on this classifier-centric understanding, we propose a generic postprocessing step built upon flow-matching to shrink the gap between the learned distribution for a pre-trained denoising diffusion model and the real data distribution, majorly around the decision boundaries. Experiments on various datasets verify the effectiveness of the proposed approach.

Minibatch optimal transport coupling straightens paths in unconditional flow matching. This leads to computationally less demanding inference as fewer integration steps and less complex numerical solvers can be employed when numerically solving an ordinary differential equation at test time. However, in the conditional setting, minibatch optimal transport falls short. This is because the default optimal transport mapping disregards conditions, resulting in a conditionally skewed prior distribution during training. In contrast, at test time, we have no access to the skewed prior, and instead sample from the full, unbiased prior distribution. This gap between training and testing leads to a subpar performance. To bridge this gap, we propose conditional optimal transport C^2OT that adds a conditional weighting term in the cost matrix when computing the optimal transport assignment. Experiments demonstrate that this simple fix works with both discrete and continuous conditions in 8gaussians-to-moons, CIFAR-10, ImageNet-32x32, and ImageNet-256x256. Our method performs better overall compared to the existing baselines across different function evaluation budgets. Code is available at https://hkchengrex.github.io/C2OT

We describe a concrete device roadmap towards a fault-tolerant quantum computing architecture based on noise-resilient, topologically protected Majorana-based qubits. Our roadmap encompasses four generations of devices: a single-qubit device that enables a measurement-based qubit benchmarking protocol; a two-qubit device that uses measurement-based braiding to perform single-qubit Clifford operations; an eight-qubit device that can be used to show an improvement of a two-qubit operation when performed on logical qubits rather than directly on physical qubits; and a topological qubit array supporting lattice surgery demonstrations on two logical qubits. Devices that enable this path require a superconductor-semiconductor heterostructure that supports a topological phase, quantum dots and coupling between those quantum dots that can create the appropriate loops for interferometric measurements, and a microwave readout system that can perform fast, low-error single-shot measurements. We describe the key design components of these qubit devices, along with the associated protocols for demonstrations of single-qubit benchmarking, Clifford gate execution, quantum error detection, and quantum error correction, which differ greatly from those in more conventional qubits. Finally, we comment on implications and advantages of this architecture for utility-scale quantum computation.

Knowledge distillation (KD) has become a prevalent technique for compressing large language models (LLMs). Existing KD methods are constrained by the need for identical tokenizers (i.e., vocabularies) between teacher and student models, limiting their versatility in handling LLMs of different architecture families. In this paper, we introduce the Multi-Level Optimal Transport (MultiLevelOT), a novel approach that advances the optimal transport for universal cross-tokenizer knowledge distillation. Our method aligns the logit distributions of the teacher and the student at both token and sequence levels using diverse cost matrices, eliminating the need for dimensional or token-by-token correspondence. At the token level, MultiLevelOT integrates both global and local information by jointly optimizing all tokens within a sequence to enhance robustness. At the sequence level, we efficiently capture complex distribution structures of logits via the Sinkhorn distance, which approximates the Wasserstein distance for divergence measures. Extensive experiments on tasks such as extractive QA, generative QA, and summarization demonstrate that the MultiLevelOT outperforms state-of-the-art cross-tokenizer KD methods under various settings. Our approach is robust to different student and teacher models across model families, architectures, and parameter sizes. Codes and models are available at https://github.com/2018cx/Multi-Level-OT.

The rapid advancement of large language models (LLMs) has led to significant improvements in their capabilities, but also to increased concerns about their alignment with human values and intentions. Current alignment strategies, including adaptive training and inference-time methods, have demonstrated potential in this area. However, these approaches still struggle to balance deployment complexity and capability across various tasks and difficulties. In this work, we introduce the Streaming Distribution Induce Aligner (Stream Aligner), a novel alignment paradigm that combines efficiency with enhanced performance in various tasks throughout the generation process. Stream Aligner achieves dynamic sentence-level correction by using a small model to learn the preferences of the suffix sentence, iteratively correcting the suffix sentence output by the upstream model, and then using the corrected sentence to replace the suffix sentence in subsequent generations. Compared to Aligner, our experiments demonstrate that Stream Aligner reduces reliance on the capabilities of additional models, enhances the reasoning abilities of LLMs, and decreases latency during user interaction. Specifically, Stream Aligner-2B model has achieved an improvement of 76.1% in helpfulness, 36.0% in harmlessness on the tested Llama2-70B-chat model, and Stream Aligner-8B has achieved an improvement of 3.5% on the math ability of the tested Llama3-70B-Instruct model.

Plagiarism involves using another person's work or concepts without proper attribution, presenting them as original creations. With the growing amount of data communicated in regional languages such as Marathi -- one of India's regional languages -- it is crucial to design robust plagiarism detection systems tailored for low-resource languages. Language models like Bidirectional Encoder Representations from Transformers (BERT) have demonstrated exceptional capability in text representation and feature extraction, making them essential tools for semantic analysis and plagiarism detection. However, the application of BERT for low-resource languages remains under-explored, particularly in the context of plagiarism detection. This paper presents a method to enhance the accuracy of plagiarism detection for Marathi texts using BERT sentence embeddings in conjunction with Term Frequency-Inverse Document Frequency (TF-IDF) feature representation. This approach effectively captures statistical, semantic, and syntactic aspects of text features through a weighted voting ensemble of machine learning models.

Decoding speech from non-invasive brain signals, such as electroencephalography (EEG), has the potential to advance brain-computer interfaces (BCIs), with applications in silent communication and assistive technologies for individuals with speech impairments. However, EEG-based speech decoding faces major challenges, such as noisy data, limited datasets, and poor performance on complex tasks like speech perception. This study attempts to address these challenges by employing variational autoencoders (VAEs) for EEG data augmentation to improve data quality and applying a state-of-the-art (SOTA) sequence-to-sequence deep learning architecture, originally successful in electromyography (EMG) tasks, to EEG-based speech decoding. Additionally, we adapt this architecture for word classification tasks. Using the Brennan dataset, which contains EEG recordings of subjects listening to narrated speech, we preprocess the data and evaluate both classification and sequence-to-sequence models for EEG-to-words/sentences tasks. Our experiments show that VAEs have the potential to reconstruct artificial EEG data for augmentation. Meanwhile, our sequence-to-sequence model achieves more promising performance in generating sentences compared to our classification model, though both remain challenging tasks. These findings lay the groundwork for future research on EEG speech perception decoding, with possible extensions to speech production tasks such as silent or imagined speech.

Large language models (LLMs) are increasingly deployed in everyday applications, demanding robust general reasoning capabilities and diverse reasoning skillset. However, current LLM reasoning benchmarks predominantly focus on mathematical and coding abilities, leaving a gap in evaluating broader reasoning proficiencies. One particular exception is the BIG-Bench dataset, which has served as a crucial benchmark for evaluating the general reasoning capabilities of LLMs, thanks to its diverse set of challenging tasks that allowed for a comprehensive assessment of general reasoning across various skills within a unified framework. However, recent advances in LLMs have led to saturation on BIG-Bench, and its harder version BIG-Bench Hard (BBH). State-of-the-art models achieve near-perfect scores on many tasks in BBH, thus diminishing its utility. To address this limitation, we introduce BIG-Bench Extra Hard (BBEH), a new benchmark designed to push the boundaries of LLM reasoning evaluation. BBEH replaces each task in BBH with a novel task that probes a similar reasoning capability but exhibits significantly increased difficulty. We evaluate various models on BBEH and observe a (harmonic) average accuracy of 9.8\% for the best general-purpose model and 44.8\% for the best reasoning-specialized model, indicating substantial room for improvement and highlighting the ongoing challenge of achieving robust general reasoning in LLMs. We release BBEH publicly at: https://github.com/google-deepmind/bbeh.

Scaling reasoning capabilities beyond traditional domains such as math and coding is hindered by the lack of diverse and high-quality questions. To overcome this limitation, we introduce a scalable approach for generating diverse and challenging reasoning questions, accompanied by reference answers. We present NaturalReasoning, a comprehensive dataset comprising 2.8 million questions that span multiple domains, including STEM fields (e.g., Physics, Computer Science), Economics, Social Sciences, and more. We demonstrate the utility of the questions in NaturalReasoning through knowledge distillation experiments which show that NaturalReasoning can effectively elicit and transfer reasoning capabilities from a strong teacher model. Furthermore, we demonstrate that NaturalReasoning is also effective for unsupervised self-training using external reward models or self-rewarding.

We present Step-Video-T2V, a state-of-the-art text-to-video pre-trained model with 30B parameters and the ability to generate videos up to 204 frames in length. A deep compression Variational Autoencoder, Video-VAE, is designed for video generation tasks, achieving 16x16 spatial and 8x temporal compression ratios, while maintaining exceptional video reconstruction quality. User prompts are encoded using two bilingual text encoders to handle both English and Chinese. A DiT with 3D full attention is trained using Flow Matching and is employed to denoise input noise into latent frames. A video-based DPO approach, Video-DPO, is applied to reduce artifacts and improve the visual quality of the generated videos. We also detail our training strategies and share key observations and insights. Step-Video-T2V's performance is evaluated on a novel video generation benchmark, Step-Video-T2V-Eval, demonstrating its state-of-the-art text-to-video quality when compared with both open-source and commercial engines. Additionally, we discuss the limitations of current diffusion-based model paradigm and outline future directions for video foundation models. We make both Step-Video-T2V and Step-Video-T2V-Eval available at https://github.com/stepfun-ai/Step-Video-T2V. The online version can be accessed from https://yuewen.cn/videos as well. Our goal is to accelerate the innovation of video foundation models and empower video content creators.

We introduce ReStyle3D, a novel framework for scene-level appearance transfer from a single style image to a real-world scene represented by multiple views. The method combines explicit semantic correspondences with multi-view consistency to achieve precise and coherent stylization. Unlike conventional stylization methods that apply a reference style globally, ReStyle3D uses open-vocabulary segmentation to establish dense, instance-level correspondences between the style and real-world images. This ensures that each object is stylized with semantically matched textures. It first transfers the style to a single view using a training-free semantic-attention mechanism in a diffusion model. It then lifts the stylization to additional views via a learned warp-and-refine network guided by monocular depth and pixel-wise correspondences. Experiments show that ReStyle3D consistently outperforms prior methods in structure preservation, perceptual style similarity, and multi-view coherence. User studies further validate its ability to produce photo-realistic, semantically faithful results. Our code, pretrained models, and dataset will be publicly released, to support new applications in interior design, virtual staging, and 3D-consistent stylization.

We present PRINCIPLE-BASED PROMPTING, a simple but effective multi-agent prompting strategy for text classification. It first asks multiple LLM agents to independently generate candidate principles based on analysis of demonstration samples with or without labels, consolidates them into final principles via a finalizer agent, and then sends them to a classifier agent to perform downstream classification tasks. Extensive experiments on binary and multi-class classification datasets with different sizes of LLMs show that our approach not only achieves substantial performance gains (1.55% - 19.37%) over zero-shot prompting on macro-F1 score but also outperforms other strong baselines (CoT and stepback prompting). Principles generated by our approach help LLMs perform better on classification tasks than human crafted principles on two private datasets. Our multi-agent PRINCIPLE-BASED PROMPTING approach also shows on-par or better performance compared to demonstration-based few-shot prompting approaches, yet with substantially lower inference costs. Ablation studies show that label information and the multi-agent cooperative LLM framework play an important role in generating high-quality principles to facilitate downstream classification tasks.

Automated mammography screening plays an important role in early breast cancer detection. However, current machine learning models, developed on some training datasets, may exhibit performance degradation and bias when deployed in real-world settings. In this paper, we analyze the performance of high-performing AI models on two mammography datasets-the Emory Breast Imaging Dataset (EMBED) and the RSNA 2022 challenge dataset. Specifically, we evaluate how these models perform across different subgroups, defined by six attributes, to detect potential biases using a range of classification metrics. Our analysis identifies certain subgroups that demonstrate notable underperformance, highlighting the need for ongoing monitoring of these subgroups' performance. To address this, we adopt a monitoring method designed to detect performance drifts over time. Upon identifying a drift, this method issues an alert, which can enable timely interventions. This approach not only provides a tool for tracking the performance but also helps ensure that AI models continue to perform effectively across diverse populations.

Quantum machine learning offers a transformative approach to solving complex problems, but the inherent noise hinders its practical implementation in near-term quantum devices. This obstacle makes it difficult to understand the generalizability of quantum circuit models. Designing robust quantum machine learning models under noise requires a principled understanding of complexity and generalization, extending beyond classical capacity measures. This study investigates the generalization properties of parameterized quantum machine learning models under the influence of noise. We present a data-dependent generalization bound grounded in the quantum Fisher information matrix. We leverage statistical learning theory to relate the parameter space volumes and training sizes to estimate the generalization capability of the trained model. We provide a structured characterization of complexity in quantum models by integrating local parameter neighborhoods and effective dimensions defined through quantum Fisher information matrix eigenvalues. We also analyze the tightness of the bound and discuss the tradeoff between model expressiveness and generalization performance.

There is a widespread and longstanding belief that machine learning models are biased towards the majority (or negative) class when learning from imbalanced data, leading them to neglect or ignore the minority (or positive) class. In this study, we show that this belief is not necessarily correct for decision trees, and that their bias can actually be in the opposite direction. Motivated by a recent simulation study that suggested that decision trees can be biased towards the minority class, our paper aims to reconcile the conflict between that study and decades of other works. First, we critically evaluate past literature on this problem, finding that failing to consider the data generating process has led to incorrect conclusions about the bias in decision trees. We then prove that, under specific conditions related to the predictors, decision trees fit to purity and trained on a dataset with only one positive case are biased towards the minority class. Finally, we demonstrate that splits in a decision tree are also biased when there is more than one positive case. Our findings have implications on the use of popular tree-based models, such as random forests.

Retrieval Augmented Generation (RAG) improves correctness of Question Answering (QA) and addresses hallucinations in Large Language Models (LLMs), yet greatly increase computational costs. Besides, RAG is not always needed as may introduce irrelevant information. Recent adaptive retrieval methods integrate LLMs' intrinsic knowledge with external information appealing to LLM self-knowledge, but they often neglect efficiency evaluations and comparisons with uncertainty estimation techniques. We bridge this gap by conducting a comprehensive analysis of 35 adaptive retrieval methods, including 8 recent approaches and 27 uncertainty estimation techniques, across 6 datasets using 10 metrics for QA performance, self-knowledge, and efficiency. Our findings show that uncertainty estimation techniques often outperform complex pipelines in terms of efficiency and self-knowledge, while maintaining comparable QA performance.

Offline meta-reinforcement learning aims to equip agents with the ability to rapidly adapt to new tasks by training on data from a set of different tasks. Context-based approaches utilize a history of state-action-reward transitions -- referred to as the context -- to infer representations of the current task, and then condition the agent, i.e., the policy and value function, on the task representations. Intuitively, the better the task representations capture the underlying tasks, the better the agent can generalize to new tasks. Unfortunately, context-based approaches suffer from distribution mismatch, as the context in the offline data does not match the context at test time, limiting their ability to generalize to the test tasks. This leads to the task representations overfitting to the offline training data. Intuitively, the task representations should be independent of the behavior policy used to collect the offline data. To address this issue, we approximately minimize the mutual information between the distribution over the task representations and behavior policy by maximizing the entropy of behavior policy conditioned on the task representations. We validate our approach in MuJoCo environments, showing that compared to baselines, our task representations more faithfully represent the underlying tasks, leading to outperforming prior methods in both in-distribution and out-of-distribution tasks.

The ML community is rapidly exploring techniques for prompting language models (LMs) and for stacking them into pipelines that solve complex tasks. Unfortunately, existing LM pipelines are typically implemented using hard-coded "prompt templates", i.e. lengthy strings discovered via trial and error. Toward a more systematic approach for developing and optimizing LM pipelines, we introduce DSPy, a programming model that abstracts LM pipelines as text transformation graphs, i.e. imperative computational graphs where LMs are invoked through declarative modules. DSPy modules are parameterized, meaning they can learn (by creating and collecting demonstrations) how to apply compositions of prompting, finetuning, augmentation, and reasoning techniques. We design a compiler that will optimize any DSPy pipeline to maximize a given metric. We conduct two case studies, showing that succinct DSPy programs can express and optimize sophisticated LM pipelines that reason about math word problems, tackle multi-hop retrieval, answer complex questions, and control agent loops. Within minutes of compiling, a few lines of DSPy allow GPT-3.5 and llama2-13b-chat to self-bootstrap pipelines that outperform standard few-shot prompting (generally by over 25% and 65%, respectively) and pipelines with expert-created demonstrations (by up to 5-46% and 16-40%, respectively). On top of that, DSPy programs compiled to open and relatively small LMs like 770M-parameter T5 and llama2-13b-chat are competitive with approaches that rely on expert-written prompt chains for proprietary GPT-3.5. DSPy is available at https://github.com/stanfordnlp/dspy

Generative Artificial Intelligence (GenAI) systems are increasingly being deployed across diverse industries and research domains. Developers and end-users interact with these systems through the use of prompting and prompt engineering. Although prompt engineering is a widely adopted and extensively researched area, it suffers from conflicting terminology and a fragmented ontological understanding of what constitutes an effective prompt due to its relatively recent emergence. We establish a structured understanding of prompt engineering by assembling a taxonomy of prompting techniques and analyzing their applications. We present a detailed vocabulary of 33 vocabulary terms, a taxonomy of 58 LLM prompting techniques, and 40 techniques for other modalities. Additionally, we provide best practices and guidelines for prompt engineering, including advice for prompting state-of-the-art (SOTA) LLMs such as ChatGPT. We further present a meta-analysis of the entire literature on natural language prefix-prompting. As a culmination of these efforts, this paper presents the most comprehensive survey on prompt engineering to date.

Large Language Models (LLMs) are increasingly used by software engineers for code generation. However, limitations of LLMs such as irrelevant or incorrect code have highlighted the need for prompt programming (or prompt engineering) where engineers apply specific prompt techniques (e.g., chain-of-thought or input-output examples) to improve the generated code. Despite this, the impact of different prompt techniques -- and their combinations -- on code generation remains underexplored. In this study, we introduce CodePromptEval, a dataset of 7072 prompts designed to evaluate five prompt techniques (few-shot, persona, chain-of-thought, function signature, list of packages) and their effect on the correctness, similarity, and quality of complete functions generated by three LLMs (GPT-4o, Llama3, and Mistral). Our findings show that while certain prompt techniques significantly influence the generated code, combining multiple techniques does not necessarily improve the outcome. Additionally, we observed a trade-off between correctness and quality when using prompt techniques. Our dataset and replication package enable future research on improving LLM-generated code and evaluating new prompt techniques.

Recent advances in decision-making policies have led to significant progress in fields such as autonomous driving and robotics. However, testing these policies remains crucial with the existence of critical scenarios that may threaten their reliability. Despite ongoing research, challenges such as low testing efficiency and limited diversity persist due to the complexity of the decision-making policies and their environments. To address these challenges, this paper proposes an adaptable Large Language Model (LLM)-driven online testing framework to explore critical and diverse testing scenarios for decision-making policies. Specifically, we design a "generate-test-feedback" pipeline with templated prompt engineering to harness the world knowledge and reasoning abilities of LLMs. Additionally, a multi-scale scenario generation strategy is proposed to address the limitations of LLMs in making fine-grained adjustments, further enhancing testing efficiency. Finally, the proposed LLM-driven method is evaluated on five widely recognized benchmarks, and the experimental results demonstrate that our method significantly outperforms baseline methods in uncovering both critical and diverse scenarios. These findings suggest that LLM-driven methods hold significant promise for advancing the testing of decision-making policies.

Since the success of GPT, large language models (LLMs) have been revolutionizing machine learning and have initiated the so-called LLM prompting paradigm. In the era of LLMs, people train a single general-purpose LLM and provide the LLM with different prompts to perform different tasks. However, such empirical success largely lacks theoretical understanding. Here, we present the first theoretical study on the LLM prompting paradigm to the best of our knowledge. In this work, we show that prompting is in fact Turing-complete: there exists a finite-size Transformer such that for any computable function, there exists a corresponding prompt following which the Transformer computes the function. Furthermore, we show that even though we use only a single finite-size Transformer, it can still achieve nearly the same complexity bounds as that of the class of all unbounded-size Transformers. Overall, our result reveals that prompting can enable a single finite-size Transformer to be efficiently universal, which establishes a theoretical underpinning for prompt engineering in practice.

We propose a novel persona-driven data synthesis methodology that leverages various perspectives within a large language model (LLM) to create diverse synthetic data. To fully exploit this methodology at scale, we introduce Persona Hub -- a collection of 1 billion diverse personas automatically curated from web data. These 1 billion personas (~13% of the world's total population), acting as distributed carriers of world knowledge, can tap into almost every perspective encapsulated within the LLM, thereby facilitating the creation of diverse synthetic data at scale for various scenarios. By showcasing Persona Hub's use cases in synthesizing high-quality mathematical and logical reasoning problems, instructions (i.e., user prompts), knowledge-rich texts, game NPCs and tools (functions) at scale, we demonstrate persona-driven data synthesis is versatile, scalable, flexible, and easy to use, potentially driving a paradigm shift in synthetic data creation and applications in practice, which may have a profound impact on LLM research and development.

The rapid development of autoregressive Large Language Models (LLMs) has significantly improved the quality of generated texts, necessitating reliable machine-generated text detectors. A huge number of detectors and collections with AI fragments have emerged, and several detection methods even showed recognition quality up to 99.9% according to the target metrics in such collections. However, the quality of such detectors tends to drop dramatically in the wild, posing a question: Are detectors actually highly trustworthy or do their high benchmark scores come from the poor quality of evaluation datasets? In this paper, we emphasise the need for robust and qualitative methods for evaluating generated data to be secure against bias and low generalising ability of future model. We present a systematic review of datasets from competitions dedicated to AI-generated content detection and propose methods for evaluating the quality of datasets containing AI-generated fragments. In addition, we discuss the possibility of using high-quality generated data to achieve two goals: improving the training of detection models and improving the training datasets themselves. Our contribution aims to facilitate a better understanding of the dynamics between human and machine text, which will ultimately support the integrity of information in an increasingly automated world. The code is available at https://github.com/Advacheck-OU/ai-dataset-analysing.

Sequential models have become increasingly popular in powering personalized recommendation systems over the past several years. These approaches traditionally model a user's actions on a website as a sequence to predict the user's next action. While theoretically simplistic, these models are quite challenging to deploy in production, commonly requiring streaming infrastructure to reflect the latest user activity and potentially managing mutable data for encoding a user's hidden state. Here we introduce PinnerFormer, a user representation trained to predict a user's future long-term engagement using a sequential model of a user's recent actions. Unlike prior approaches, we adapt our modeling to a batch infrastructure via our new dense all-action loss, modeling long-term future actions instead of next action prediction. We show that by doing so, we significantly close the gap between batch user embeddings that are generated once a day and realtime user embeddings generated whenever a user takes an action. We describe our design decisions via extensive offline experimentation and ablations and validate the efficacy of our approach in A/B experiments showing substantial improvements in Pinterest's user retention and engagement when comparing PinnerFormer against our previous user representation. PinnerFormer is deployed in production as of Fall 2021.

In this work, we aim to develop an MLLM that understands and solves questions by learning to create each intermediate step of the reasoning involved till the final answer. To this end, we propose Collective Monte Carlo Tree Search (CoMCTS), a new learning-to-reason method for MLLMs, which introduces the concept of collective learning into ``tree search'' for effective and efficient reasoning-path searching and learning. The core idea of CoMCTS is to leverage collective knowledge from multiple models to collaboratively conjecture, search and identify effective reasoning paths toward correct answers via four iterative operations including Expansion, Simulation and Error Positioning, Backpropagation, and Selection. Using CoMCTS, we construct Mulberry-260k, a multimodal dataset with a tree of rich, explicit and well-defined reasoning nodes for each question. With Mulberry-260k, we perform collective SFT to train our model, Mulberry, a series of MLLMs with o1-like step-by-step Reasoning and Reflection capabilities. Extensive experiments demonstrate the superiority of our proposed methods on various benchmarks. Code will be available at https://github.com/HJYao00/Mulberry

Who controls the development of Artificial General Intelligence (AGI) might matter less than how we handle the fight for control itself. We formalize this "steering wheel problem" as humanity's greatest near-term existential risk may stem not from misaligned AGI, but from the dynamics of competing to develop it. Just as a car crash can occur from passengers fighting over the wheel before reaching any destination, catastrophic outcomes could arise from development competition long before AGI exists. While technical alignment research focuses on ensuring safe arrival, we show how coordination failures during development could drive us off the cliff first. We present a game theoretic framework modeling AGI development dynamics and prove conditions for sustainable cooperative equilibria. Drawing from nuclear control while accounting for AGI's unique characteristics, we propose concrete mechanisms including pre-registration, shared technical infrastructure, and automated deterrence to stabilize cooperation. Our key insight is that AGI creates network effects in safety: shared investments become more valuable as participation grows, enabling mechanism designs where cooperation dominates defection. This work bridges formal methodology and policy frameworks, providing foundations for practical governance of AGI competition risks.

Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task. While typically task-agnostic in architecture, this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples. By contrast, humans can generally perform a new language task from only a few examples or from simple instructions - something which current NLP systems still largely struggle to do. Here we show that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even reaching competitiveness with prior state-of-the-art fine-tuning approaches. Specifically, we train GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, and test its performance in the few-shot setting. For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model. GPT-3 achieves strong performance on many NLP datasets, including translation, question-answering, and cloze tasks, as well as several tasks that require on-the-fly reasoning or domain adaptation, such as unscrambling words, using a novel word in a sentence, or performing 3-digit arithmetic. At the same time, we also identify some datasets where GPT-3's few-shot learning still struggles, as well as some datasets where GPT-3 faces methodological issues related to training on large web corpora. Finally, we find that GPT-3 can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans. We discuss broader societal impacts of this finding and of GPT-3 in general.

Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, yet their performance is highly dependent on the prompting strategy and model scale. While reinforcement learning and fine-tuning have been deployed to boost reasoning, these approaches incur substantial computational and data overhead. In this work, we introduce Adaptive Graph of Thoughts (AGoT), a dynamic, graph-based inference framework that enhances LLM reasoning solely at test time. Rather than relying on fixed-step methods like Chain of Thought (CoT) or Tree of Thoughts (ToT), AGoT recursively decomposes complex queries into structured subproblems, forming an dynamic directed acyclic graph (DAG) of interdependent reasoning steps. By selectively expanding only those subproblems that require further analysis, AGoT unifies the strengths of chain, tree, and graph paradigms into a cohesive framework that allocates computation where it is most needed. We validate our approach on diverse benchmarks spanning multi-hop retrieval, scientific reasoning, and mathematical problem-solving, achieving up to 46.2% improvement on scientific reasoning tasks (GPQA) - comparable to gains achieved through computationally intensive reinforcement learning approaches and outperforming state-of-the-art iterative approaches. These results suggest that dynamic decomposition and structured recursion offer a scalable, cost-effective alternative to post-training modifications, paving the way for more robust, general-purpose reasoning in LLMs.

Test-time scaling is a promising new approach to language modeling that uses extra test-time compute to improve performance. Recently, OpenAI's o1 model showed this capability but did not publicly share its methodology, leading to many replication efforts. We seek the simplest approach to achieve test-time scaling and strong reasoning performance. First, we curate a small dataset s1K of 1,000 questions paired with reasoning traces relying on three criteria we validate through ablations: difficulty, diversity, and quality. Second, we develop budget forcing to control test-time compute by forcefully terminating the model's thinking process or lengthening it by appending "Wait" multiple times to the model's generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps. After supervised finetuning the Qwen2.5-32B-Instruct language model on s1K and equipping it with budget forcing, our model s1-32B exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24). Further, scaling s1-32B with budget forcing allows extrapolating beyond its performance without test-time intervention: from 50% to 57% on AIME24. Our model, data, and code are open-source at https://github.com/simplescaling/s1

This paper introduces an approach for training o1-like RAG models that retrieve and reason over relevant information step by step before generating the final answer. Conventional RAG methods usually perform a single retrieval step before the generation process, which limits their effectiveness in addressing complex queries due to imperfect retrieval results. In contrast, our proposed method, CoRAG (Chain-of-Retrieval Augmented Generation), allows the model to dynamically reformulate the query based on the evolving state. To train CoRAG effectively, we utilize rejection sampling to automatically generate intermediate retrieval chains, thereby augmenting existing RAG datasets that only provide the correct final answer. At test time, we propose various decoding strategies to scale the model's test-time compute by controlling the length and number of sampled retrieval chains. Experimental results across multiple benchmarks validate the efficacy of CoRAG, particularly in multi-hop question answering tasks, where we observe more than 10 points improvement in EM score compared to strong baselines. On the KILT benchmark, CoRAG establishes a new state-of-the-art performance across a diverse range of knowledge-intensive tasks. Furthermore, we offer comprehensive analyses to understand the scaling behavior of CoRAG, laying the groundwork for future research aimed at developing factual and grounded foundation models.

Language has long been conceived as an essential tool for human reasoning. The breakthrough of Large Language Models (LLMs) has sparked significant research interest in leveraging these models to tackle complex reasoning tasks. Researchers have moved beyond simple autoregressive token generation by introducing the concept of "thought" -- a sequence of tokens representing intermediate steps in the reasoning process. This innovative paradigm enables LLMs' to mimic complex human reasoning processes, such as tree search and reflective thinking. Recently, an emerging trend of learning to reason has applied reinforcement learning (RL) to train LLMs to master reasoning processes. This approach enables the automatic generation of high-quality reasoning trajectories through trial-and-error search algorithms, significantly expanding LLMs' reasoning capacity by providing substantially more training data. Furthermore, recent studies demonstrate that encouraging LLMs to "think" with more tokens during test-time inference can further significantly boost reasoning accuracy. Therefore, the train-time and test-time scaling combined to show a new research frontier -- a path toward Large Reasoning Model. The introduction of OpenAI's o1 series marks a significant milestone in this research direction. In this survey, we present a comprehensive review of recent progress in LLM reasoning. We begin by introducing the foundational background of LLMs and then explore the key technical components driving the development of large reasoning models, with a focus on automated data construction, learning-to-reason techniques, and test-time scaling. We also analyze popular open-source projects at building large reasoning models, and conclude with open challenges and future research directions.

Machine learning, the foundation of modern artificial intelligence, has driven innovations that have fundamentally transformed the world. Yet, behind advancements lies a complex and often tedious process requiring labor and compute intensive iteration and experimentation. Engineers and scientists developing machine learning models spend much of their time on trial-and-error tasks instead of conceptualizing innovative solutions or research hypotheses. To address this challenge, we introduce AI-Driven Exploration (AIDE), a machine learning engineering agent powered by large language models (LLMs). AIDE frames machine learning engineering as a code optimization problem, and formulates trial-and-error as a tree search in the space of potential solutions. By strategically reusing and refining promising solutions, AIDE effectively trades computational resources for enhanced performance, achieving state-of-the-art results on multiple machine learning engineering benchmarks, including our Kaggle evaluations, OpenAI MLE-Bench and METRs RE-Bench.

Large language models (LLMs) have demonstrated remarkable capabilities in various complex tasks, yet they still suffer from hallucinations. Introducing external knowledge, such as knowledge graph, can enhance the LLMs' ability to provide factual answers. LLMs have the ability to interactively explore knowledge graphs. However, most approaches have been affected by insufficient internal knowledge excavation in LLMs, limited generation of trustworthy knowledge reasoning paths, and a vague integration between internal and external knowledge. Therefore, we propose KnowPath, a knowledge-enhanced large model framework driven by the collaboration of internal and external knowledge. It relies on the internal knowledge of the LLM to guide the exploration of interpretable directed subgraphs in external knowledge graphs, better integrating the two knowledge sources for more accurate reasoning. Extensive experiments on multiple real-world datasets confirm the superiority of KnowPath.

This paper introduces CodeQUEST, a novel framework leveraging Large Language Models (LLMs) to iteratively evaluate and enhance code quality across multiple dimensions, including readability, maintainability, efficiency, and security. The framework is divided into two main components: an Evaluator that assesses code quality across ten dimensions, providing both quantitative scores and qualitative summaries, and an Optimizer that iteratively improves the code based on the Evaluator's feedback. Our study demonstrates that CodeQUEST can effectively and robustly evaluate code quality, with its assessments aligning closely with established code quality metrics. Through a series of experiments using a curated dataset of Python and JavaScript examples, CodeQUEST demonstrated significant improvements in code quality, achieving a mean relative percentage improvement of 52.6%. The framework's evaluations were validated against a set of proxy metrics comprising of Pylint Score, Radon Maintainability Index, and Bandit output logs, showing a meaningful correlation. This highlights the potential of LLMs in automating code quality evaluation and improvement processes, presenting a significant advancement toward enhancing software development practices. The code implementation of the framework is available at: https://github.com/jpmorganchase/CodeQuest.

Human-in-the-loop (HITL) frameworks are increasingly recognized for their potential to improve annotation accuracy in emotion estimation systems by combining machine predictions with human expertise. This study focuses on integrating a high-performing image-based emotion model into a HITL annotation framework to evaluate the collaborative potential of human-machine interaction and identify the psychological and practical factors critical to successful collaboration. Specifically, we investigate how varying model reliability and cognitive framing influence human trust, cognitive load, and annotation behavior in HITL systems. We demonstrate that model reliability and psychological framing significantly impact annotators' trust, engagement, and consistency, offering insights into optimizing HITL frameworks. Through three experimental scenarios with 29 participants--baseline model reliability (S1), fabricated errors (S2), and cognitive bias introduced by negative framing (S3)--we analyzed behavioral and qualitative data. Reliable predictions in S1 yielded high trust and annotation consistency, while unreliable outputs in S2 led to increased critical evaluations but also heightened frustration and response variability. Negative framing in S3 revealed how cognitive bias influenced participants to perceive the model as more relatable and accurate, despite misinformation regarding its reliability. These findings highlight the importance of both reliable machine outputs and psychological factors in shaping effective human-machine collaboration. By leveraging the strengths of both human oversight and automated systems, this study establishes a scalable HITL framework for emotion annotation and lays the foundation for broader applications in adaptive learning and human-computer interaction.

Computational models offer powerful tools for formalising psychological theories, making them both testable and applicable in digital contexts. However, they remain little used in the study of motivation within psychology. We focus on the "need for competence", postulated as a key basic human need within Self-Determination Theory (SDT) -- arguably the most influential psychological framework for studying intrinsic motivation (IM). The need for competence is treated as a single construct across SDT texts. Yet, recent research has identified multiple, ambiguously defined facets of competence in SDT. We propose that these inconsistencies may be alleviated by drawing on computational models from the field of artificial intelligence, specifically from the domain of reinforcement learning (RL). By aligning the aforementioned facets of competence -- effectance, skill use, task performance, and capacity growth -- with existing RL formalisms, we provide a foundation for advancing competence-related theory in SDT and motivational psychology more broadly. The formalisms reveal underlying preconditions that SDT fails to make explicit, demonstrating how computational models can improve our understanding of IM. Additionally, our work can support a cycle of theory development by inspiring new computational models formalising aspects of the theory, which can then be tested empirically to refine the theory. While our research lays a promising foundation, empirical studies of these models in both humans and machines are needed, inviting collaboration across disciplines.

Inspired by progress in large-scale language modeling, we apply a similar approach towards building a single generalist agent beyond the realm of text outputs. The agent, which we refer to as Gato, works as a multi-modal, multi-task, multi-embodiment generalist policy. The same network with the same weights can play Atari, caption images, chat, stack blocks with a real robot arm and much more, deciding based on its context whether to output text, joint torques, button presses, or other tokens. In this report we describe the model and the data, and document the current capabilities of Gato.

Leading AI developers and startups are increasingly deploying agentic AI systems that can plan and execute complex tasks with limited human involvement. However, there is currently no structured framework for documenting the technical components, intended uses, and safety features of agentic systems. To fill this gap, we introduce the AI Agent Index, the first public database to document information about currently deployed agentic AI systems. For each system that meets the criteria for inclusion in the index, we document the system's components (e.g., base model, reasoning implementation, tool use), application domains (e.g., computer use, software engineering), and risk management practices (e.g., evaluation results, guardrails), based on publicly available information and correspondence with developers. We find that while developers generally provide ample information regarding the capabilities and applications of agentic systems, they currently provide limited information regarding safety and risk management practices. The AI Agent Index is available online at https://aiagentindex.mit.edu/

As AI continues to advance, there is a growing demand for systems that go beyond language-based assistance and move toward intelligent agents capable of performing real-world actions. This evolution requires the transition from traditional Large Language Models (LLMs), which excel at generating textual responses, to Large Action Models (LAMs), designed for action generation and execution within dynamic environments. Enabled by agent systems, LAMs hold the potential to transform AI from passive language understanding to active task completion, marking a significant milestone in the progression toward artificial general intelligence. In this paper, we present a comprehensive framework for developing LAMs, offering a systematic approach to their creation, from inception to deployment. We begin with an overview of LAMs, highlighting their unique characteristics and delineating their differences from LLMs. Using a Windows OS-based agent as a case study, we provide a detailed, step-by-step guide on the key stages of LAM development, including data collection, model training, environment integration, grounding, and evaluation. This generalizable workflow can serve as a blueprint for creating functional LAMs in various application domains. We conclude by identifying the current limitations of LAMs and discussing directions for future research and industrial deployment, emphasizing the challenges and opportunities that lie ahead in realizing the full potential of LAMs in real-world applications. The code for the data collection process utilized in this paper is publicly available at: https://github.com/microsoft/UFO/tree/main/dataflow, and comprehensive documentation can be found at https://microsoft.github.io/UFO/dataflow/overview/.

Communication has been widely employed to enhance multi-agent collaboration. Previous research has typically assumed delay-free communication, a strong assumption that is challenging to meet in practice. However, real-world agents suffer from channel delays, receiving messages sent at different time points, termed {\it{Asynchronous Communication}}, leading to cognitive biases and breakdowns in collaboration. This paper first defines two communication delay settings in MARL and emphasizes their harm to collaboration. To handle the above delays, this paper proposes a novel framework, Communication Delay-tolerant Multi-Agent Collaboration (CoDe). At first, CoDe learns an intent representation as messages through future action inference, reflecting the stable future behavioral trends of the agents. Then, CoDe devises a dual alignment mechanism of intent and timeliness to strengthen the fusion process of asynchronous messages. In this way, agents can extract the long-term intent of others, even from delayed messages, and selectively utilize the most recent messages that are relevant to their intent. Experimental results demonstrate that CoDe outperforms baseline algorithms in three MARL benchmarks without delay and exhibits robustness under fixed and time-varying delays.

In this paper, we propose a novel approach for verifying the compliance of turn-based multi-agent reinforcement learning (TMARL) agents with complex requirements in stochastic multiplayer games. Our method overcomes the limitations of existing verification approaches, which are inadequate for dealing with TMARL agents and not scalable to large games with multiple agents. Our approach relies on tight integration of TMARL and a verification technique referred to as model checking. We demonstrate the effectiveness and scalability of our technique through experiments in different types of environments. Our experiments show that our method is suited to verify TMARL agents and scales better than naive monolithic model checking.

Composed Image Retrieval (CIR) aims to retrieve target images from candidate set using a hybrid-modality query consisting of a reference image and a relative caption that describes the user intent. Recent studies attempt to utilize Vision-Language Pre-training Models (VLPMs) with various fusion strategies for addressing the task.However, these methods typically fail to simultaneously meet two key requirements of CIR: comprehensively extracting visual information and faithfully following the user intent. In this work, we propose CIR-LVLM, a novel framework that leverages the large vision-language model (LVLM) as the powerful user intent-aware encoder to better meet these requirements. Our motivation is to explore the advanced reasoning and instruction-following capabilities of LVLM for accurately understanding and responding the user intent. Furthermore, we design a novel hybrid intent instruction module to provide explicit intent guidance at two levels: (1) The task prompt clarifies the task requirement and assists the model in discerning user intent at the task level. (2) The instance-specific soft prompt, which is adaptively selected from the learnable prompt pool, enables the model to better comprehend the user intent at the instance level compared to a universal prompt for all instances. CIR-LVLM achieves state-of-the-art performance across three prominent benchmarks with acceptable inference efficiency. We believe this study provides fundamental insights into CIR-related fields.

Searching for information on the internet and digital platforms to satisfy an information need requires effective retrieval solutions. However, such solutions are not yet available for Tetun, making it challenging to find relevant documents for text-based search queries in this language. To address these challenges, this study investigates Tetun text retrieval with a focus on the ad-hoc retrieval task. It begins by developing essential language resources -- including a list of stopwords, a stemmer, and a test collection -- which serve as foundational components for solutions tailored to Tetun text retrieval. Various strategies are then explored using both document titles and content to evaluate retrieval effectiveness. The results show that retrieving document titles, after removing hyphens and apostrophes without applying stemming, significantly improves retrieval performance compared to the baseline. Efficiency increases by 31.37%, while effectiveness achieves an average gain of 9.40% in MAP@10 and 30.35% in nDCG@10 with DFR BM25. Beyond the top-10 cutoff point, Hiemstra LM demonstrates strong performance across various retrieval strategies and evaluation metrics. Contributions of this work include the development of Labadain-Stopwords (a list of 160 Tetun stopwords), Labadain-Stemmer (a Tetun stemmer with three variants), and Labadain-Avaliad\'or (a Tetun test collection containing 59 topics, 33,550 documents, and 5,900 qrels).

Numerous retrieval models, including sparse, dense and llm-based methods, have demonstrated remarkable performance in predicting the relevance between queries and corpora. However, the preliminary effectiveness analysis experiments indicate that these models fail to achieve satisfactory performance on the majority of queries and corpora, revealing their effectiveness restricted to specific scenarios. Thus, to tackle this problem, we propose a novel Distributed Collaborative Retrieval Framework (DCRF), outperforming each single model across all queries and corpora. Specifically, the framework integrates various retrieval models into a unified system and dynamically selects the optimal results for each user's query. It can easily aggregate any retrieval model and expand to any application scenarios, illustrating its flexibility and scalability.Moreover, to reduce maintenance and training costs, we design four effective prompting strategies with large language models (LLMs) to evaluate the quality of ranks without reliance of labeled data. Extensive experiments demonstrate that proposed framework, combined with 8 efficient retrieval models, can achieve performance comparable to effective listwise methods like RankGPT and ListT5, while offering superior efficiency. Besides, DCRF surpasses all selected retrieval models on the most datasets, indicating the effectiveness of our prompting strategies on rank-oriented automatic evaluation.

Multi-domain recommendation (MDR) aims to enhance recommendation performance across various domains. However, real-world recommender systems in online platforms often need to handle dozens or even hundreds of domains, far exceeding the capabilities of traditional MDR algorithms, which typically focus on fewer than five domains. Key challenges include a substantial increase in parameter count, high maintenance costs, and intricate knowledge transfer patterns across domains. Furthermore, minor domains often suffer from data sparsity, leading to inadequate training in classical methods. To address these issues, we propose Adaptive REcommendation for All Domains with counterfactual augmentation (AREAD). AREAD employs a hierarchical structure with a limited number of expert networks at several layers, to effectively capture domain knowledge at different granularities. To adaptively capture the knowledge transfer pattern across domains, we generate and iteratively prune a hierarchical expert network selection mask for each domain during training. Additionally, counterfactual assumptions are used to augment data in minor domains, supporting their iterative mask pruning. Our experiments on two public datasets, each encompassing over twenty domains, demonstrate AREAD's effectiveness, especially in data-sparse domains. Source code is available at https://github.com/Chrissie-Law/AREAD-Multi-Domain-Recommendation.

This article describes the history of information retrieval on personal document collections.

Mixture-of-Experts (MoE) activates only a subset of experts during inference, allowing the model to maintain low inference FLOPs and latency even as the parameter count scales up. However, since MoE dynamically selects the experts, all the experts need to be loaded into VRAM. Their large parameter size still limits deployment, and offloading, which load experts into VRAM only when needed, significantly increase inference latency. To address this, we propose Mixture of Lookup Experts (MoLE), a new MoE architecture that is efficient in both communication and VRAM usage. In MoLE, the experts are Feed-Forward Networks (FFNs) during training, taking the output of the embedding layer as input. Before inference, these experts can be re-parameterized as lookup tables (LUTs) that retrieves expert outputs based on input ids, and offloaded to storage devices. Therefore, we do not need to perform expert computations during inference. Instead, we directly retrieve the expert's computation results based on input ids and load them into VRAM, and thus the resulting communication overhead is negligible. Experiments show that, with the same FLOPs and VRAM usage, MoLE achieves inference speeds comparable to dense models and significantly faster than MoE with experts offloading, while maintaining performance on par with MoE.

Large language models (LLMs) have shown remarkable potential in processing long sequences, yet efficiently serving these long-context models remains challenging due to the quadratic computational complexity of attention in the prefilling stage and the large memory footprint of the KV cache in the decoding stage. To address these issues, we introduce LServe, an efficient system that accelerates long-sequence LLM serving via hybrid sparse attention. This method unifies different hardware-friendly, structured sparsity patterns for both prefilling and decoding attention into a single framework, where computations on less important tokens are skipped block-wise. LServe demonstrates the compatibility of static and dynamic sparsity in long-context LLM attention. This design enables multiplicative speedups by combining these optimizations. Specifically, we convert half of the attention heads to nearly free streaming heads in both the prefilling and decoding stages. Additionally, we find that only a constant number of KV pages is required to preserve long-context capabilities, irrespective of context length. We then design a hierarchical KV page selection policy that dynamically prunes KV pages based on query-centric similarity. On average, LServe accelerates LLM prefilling by up to 2.9x and decoding by 1.3-2.1x over vLLM, maintaining long-context accuracy. Code is released at https://github.com/mit-han-lab/omniserve.

Understanding historical and cultural artifacts demands human expertise and advanced computational techniques, yet the process remains complex and time-intensive. While large multimodal models offer promising support, their evaluation and improvement require a standardized benchmark. To address this, we introduce TimeTravel, a benchmark of 10,250 expert-verified samples spanning 266 distinct cultures across 10 major historical regions. Designed for AI-driven analysis of manuscripts, artworks, inscriptions, and archaeological discoveries, TimeTravel provides a structured dataset and robust evaluation framework to assess AI models' capabilities in classification, interpretation, and historical comprehension. By integrating AI with historical research, TimeTravel fosters AI-powered tools for historians, archaeologists, researchers, and cultural tourists to extract valuable insights while ensuring technology contributes meaningfully to historical discovery and cultural heritage preservation. We evaluate contemporary AI models on TimeTravel, highlighting their strengths and identifying areas for improvement. Our goal is to establish AI as a reliable partner in preserving cultural heritage, ensuring that technological advancements contribute meaningfully to historical discovery. Our code is available at: \url{https://github.com/mbzuai-oryx/TimeTravel}.

Speculative sampling has emerged as an important technique for accelerating the auto-regressive generation process of large language models (LLMs) by utilizing a draft-then-verify mechanism to produce multiple tokens per forward pass. While state-of-the-art speculative sampling methods use only a single layer and a language modeling (LM) head as the draft model to achieve impressive layer compression, their efficiency gains are substantially reduced for large-vocabulary LLMs, such as Llama-3-8B with a vocabulary of 128k tokens. To address this, we present FR-Spec, a frequency-ranked speculative sampling framework that optimizes draft candidate selection through vocabulary space compression. By constraining the draft search to a frequency-prioritized token subset, our method reduces LM Head computation overhead by 75% while ensuring the equivalence of the final output distribution. Experiments across multiple datasets demonstrate an average of 1.12$\times$ speedup over the state-of-the-art speculative sampling method EAGLE-2. Code available at https://github.com/thunlp/FR-Spec.

Large language model (LLM) evaluations typically rely on aggregated metrics like accuracy or human preference, averaging across users and prompts. This averaging obscures user- and prompt-specific variations in model performance. To address this, we propose Prompt-to-Leaderboard (P2L), a method that produces leaderboards specific to a prompt. The core idea is to train an LLM taking natural language prompts as input to output a vector of Bradley-Terry coefficients which are then used to predict the human preference vote. The resulting prompt-dependent leaderboards allow for unsupervised task-specific evaluation, optimal routing of queries to models, personalization, and automated evaluation of model strengths and weaknesses. Data from Chatbot Arena suggest that P2L better captures the nuanced landscape of language model performance than the averaged leaderboard. Furthermore, our findings suggest that P2L's ability to produce prompt-specific evaluations follows a power law scaling similar to that observed in LLMs themselves. In January 2025, the router we trained based on this methodology achieved the #1 spot on the Chatbot Arena leaderboard. Our code is available on GitHub at https://github.com/lmarena/p2l.

The hypothesis that pretrained large language models (LLMs) necessitate only minimal supervision during the fine-tuning (SFT) stage (Zhou et al., 2024) has been substantiated by recent advancements in data curation and selection research. However, their stability and generalizability are compromised due to the vulnerability to experimental setups and validation protocols, falling short of surpassing random sampling (Diddee & Ippolito, 2024; Xia et al., 2024b). Built upon LLMs, multi-modal LLMs (MLLMs), combined with the sheer token volume and heightened heterogeneity of data sources, amplify both the significance and complexity of data selection. To harvest multi-modal instructional data in a robust and efficient manner, we re-define the granularity of the quality metric by decomposing it into 14 vision-language-related capabilities, and introduce multi-modal rich scorers to evaluate the capabilities of each data candidate. To promote diversity, in light of the inherent objective of the alignment stage, we take interaction style as diversity indicator and use a multi-modal rich styler to identify data instruction patterns. In doing so, our multi-modal rich scorers and styler (mmSSR) guarantee that high-scoring information is conveyed to users in diversified forms. Free from embedding-based clustering or greedy sampling, mmSSR efficiently scales to millions of data with varying budget constraints, supports customization for general or specific capability acquisition, and facilitates training-free generalization to new domains for curation. Across 10+ experimental settings, validated by 14 multi-modal benchmarks, we demonstrate consistent improvements over random sampling, baseline strategies and state-of-the-art selection methods, achieving 99.1% of full performance with only 30% of the 2.6M data.

Recent advancements in Large Language Models (LLMs) have demonstrated enhanced reasoning capabilities, evolving from Chain-of-Thought (CoT) prompting to advanced, product-oriented solutions like OpenAI o1. During our re-implementation of this model, we noticed that in multimodal tasks requiring visual input (e.g., geometry problems), Multimodal LLMs (MLLMs) struggle to maintain focus on the visual information, in other words, MLLMs suffer from a gradual decline in attention to visual information as reasoning progresses, causing text-over-relied outputs. To investigate this, we ablate image inputs during long-chain reasoning. Concretely, we truncate the reasoning process midway, then re-complete the reasoning process with the input image removed. We observe only a ~2% accuracy drop on MathVista's test-hard subset, revealing the model's textual outputs dominate the following reasoning process. Motivated by this, we propose Take-along Visual Conditioning (TVC), a strategy that shifts image input to critical reasoning stages and compresses redundant visual tokens via dynamic pruning. This methodology helps the model retain attention to the visual components throughout the reasoning. Our approach achieves state-of-the-art performance on average across five mathematical reasoning benchmarks (+3.4% vs previous sota), demonstrating the effectiveness of TVC in enhancing multimodal reasoning systems.

The rapid advancement of Large Multi-modal Models (LMMs) has enabled their application in scientific problem-solving, yet their fine-grained capabilities remain under-explored. In this paper, we introduce SciVerse, a multi-modal scientific evaluation benchmark to thoroughly assess LMMs across 5,735 test instances in five distinct versions. We aim to investigate three key dimensions of LMMs: scientific knowledge comprehension, multi-modal content interpretation, and Chain-of-Thought (CoT) reasoning. To unveil whether LMMs possess sufficient scientific expertise, we first transform each problem into three versions containing different levels of knowledge required for solving, i.e., Knowledge-free, -lite, and -rich. Then, to explore how LMMs interpret multi-modal scientific content, we annotate another two versions, i.e., Vision-rich and -only, marking more question information from texts to diagrams. Comparing the results of different versions, SciVerse systematically examines the professional knowledge stock and visual perception skills of LMMs in scientific domains. In addition, to rigorously assess CoT reasoning, we propose a new scientific CoT evaluation strategy, conducting a step-wise assessment on knowledge and logical errors in model outputs. Our extensive evaluation of different LMMs on SciVerse reveals critical limitations in their scientific proficiency and provides new insights into future developments. Project page: https://sciverse-cuhk.github.io

Multimodal summarization integrating information from diverse data modalities presents a promising solution to aid the understanding of information within various processes. However, the application and advantages of multimodal summarization have not received much attention in model-based engineering (MBE), where it has become a cornerstone in the design and development of complex systems, leveraging formal models to improve understanding, validation and automation throughout the engineering lifecycle. UML and EMF diagrams in model-based engineering contain a large amount of multimodal information and intricate relational data. Hence, our study explores the application of multimodal large language models within the domain of model-based engineering to evaluate their capacity for understanding and identifying relationships, features, and functionalities embedded in UML and EMF diagrams. We aim to demonstrate the transformative potential benefits and limitations of multimodal summarization in improving productivity and accuracy in MBE practices. The proposed approach is evaluated within the context of automotive software development, while many promising state-of-art models were taken into account.

Understanding and predicting the properties of inorganic materials is crucial for accelerating advancements in materials science and driving applications in energy, electronics, and beyond. Integrating material structure data with language-based information through multi-modal large language models (LLMs) offers great potential to support these efforts by enhancing human-AI interaction. However, a key challenge lies in integrating atomic structures at full resolution into LLMs. In this work, we introduce MatterChat, a versatile structure-aware multi-modal LLM that unifies material structural data and textual inputs into a single cohesive model. MatterChat employs a bridging module to effectively align a pretrained machine learning interatomic potential with a pretrained LLM, reducing training costs and enhancing flexibility. Our results demonstrate that MatterChat significantly improves performance in material property prediction and human-AI interaction, surpassing general-purpose LLMs such as GPT-4. We also demonstrate its usefulness in applications such as more advanced scientific reasoning and step-by-step material synthesis.

As human augmentation technologies evolve, the convergence of AI, robotics, and extended reality (XR) is redefining human potential -- enhancing cognition, perception, and physical abilities. However, these advancements also introduce ethical dilemmas, security risks, and concerns over loss of control. This workshop explores both the transformative potential and the unintended consequences of augmentation technologies. Bringing together experts from HCI, neuroscience, robotics, and ethics, we will examine real-world applications, emerging risks, and governance strategies for responsible augmentation. The session will feature keynote talks and interactive discussions, addressing topics such as AI-enhanced cognition, wearable robotics, neural interfaces, and XR-driven augmentation. By fostering multidisciplinary dialogue, this workshop aims to generate actionable insights for responsible innovation, proposing ethical frameworks to balance human empowerment with risk mitigation. We invite researchers, practitioners, and industry leaders to contribute their perspectives and help shape the future of human augmentation.

With the rapid advancement of large language models (LLMs) and vision-language models (VLMs), significant progress has been made in developing open-vocabulary robotic manipulation systems. However, many existing approaches overlook the importance of object dynamics, limiting their applicability to more complex, dynamic tasks. In this work, we introduce KUDA, an open-vocabulary manipulation system that integrates dynamics learning and visual prompting through keypoints, leveraging both VLMs and learning-based neural dynamics models. Our key insight is that a keypoint-based target specification is simultaneously interpretable by VLMs and can be efficiently translated into cost functions for model-based planning. Given language instructions and visual observations, KUDA first assigns keypoints to the RGB image and queries the VLM to generate target specifications. These abstract keypoint-based representations are then converted into cost functions, which are optimized using a learned dynamics model to produce robotic trajectories. We evaluate KUDA on a range of manipulation tasks, including free-form language instructions across diverse object categories, multi-object interactions, and deformable or granular objects, demonstrating the effectiveness of our framework. The project page is available at http://kuda-dynamics.github.io.

Existing quadrupedal locomotion learning paradigms usually rely on extensive domain randomization to alleviate the sim2real gap and enhance robustness. It trains policies with a wide range of environment parameters and sensor noises to perform reliably under uncertainty. However, since optimal performance under ideal conditions often conflicts with the need to handle worst-case scenarios, there is a trade-off between optimality and robustness. This trade-off forces the learned policy to prioritize stability in diverse and challenging conditions over efficiency and accuracy in ideal ones, leading to overly conservative behaviors that sacrifice peak performance. In this paper, we propose a two-stage framework that mitigates this trade-off by integrating policy learning with imagined transitions. This framework enhances the conventional reinforcement learning (RL) approach by incorporating imagined transitions as demonstrative inputs. These imagined transitions are derived from an optimal policy and a dynamics model operating within an idealized setting. Our findings indicate that this approach significantly mitigates the domain randomization-induced negative impact of existing RL algorithms. It leads to accelerated training, reduced tracking errors within the distribution, and enhanced robustness outside the distribution.

This study proposes a new Gaussian Mixture Filter (GMF) to improve the estimation performance for the autonomous robotic radio signal source search and localization problem in unknown environments. The proposed filter is first tested with a benchmark numerical problem to validate the performance with other state-of-practice approaches such as Particle Gaussian Mixture (PGM) filters and Particle Filter (PF). Then the proposed approach is tested and compared against PF and PGM filters in real-world robotic field experiments to validate its impact for real-world robotic applications. The considered real-world scenarios have partial observability with the range-only measurement and uncertainty with the measurement model. The results show that the proposed filter can handle this partial observability effectively whilst showing improved performance compared to PF, reducing the computation requirements while demonstrating improved robustness over compared techniques.

In agricultural automation, inherent occlusion presents a major challenge for robotic harvesting. We propose a novel imitation learning-based viewpoint planning approach to actively adjust camera viewpoint and capture unobstructed images of the target crop. Traditional viewpoint planners and existing learning-based methods, depend on manually designed evaluation metrics or reward functions, often struggle to generalize to complex, unseen scenarios. Our method employs the Action Chunking with Transformer (ACT) algorithm to learn effective camera motion policies from expert demonstrations. This enables continuous six-degree-of-freedom (6-DoF) viewpoint adjustments that are smoother, more precise and reveal occluded targets. Extensive experiments in both simulated and real-world environments, featuring agricultural scenarios and a 6-DoF robot arm equipped with an RGB-D camera, demonstrate our method's superior success rate and efficiency, especially in complex occlusion conditions, as well as its ability to generalize across different crops without reprogramming. This study advances robotic harvesting by providing a practical "learn from demonstration" (LfD) solution to occlusion challenges, ultimately enhancing autonomous harvesting performance and productivity.