Traditional benchmarks struggle to evaluate increasingly sophisticated language models in multilingual and culturally diverse contexts. To address this gap, we introduce MMLU-ProX, a comprehensive multilingual benchmark covering 13 typologically diverse languages with approximately 11,829 questions per language. Building on the challenging reasoning-focused design of MMLU-Pro, our framework employs a semi-automatic translation process: translations generated by state-of-the-art large language models (LLMs) are rigorously evaluated by expert annotators to ensure conceptual accuracy, terminological consistency, and cultural relevance. We comprehensively evaluate 25 state-of-the-art LLMs using 5-shot chain-of-thought (CoT) and zero-shot prompting strategies, analyzing their performance across linguistic and cultural boundaries. Our experiments reveal consistent performance degradation from high-resource languages to lower-resource ones, with the best models achieving over 70% accuracy on English but dropping to around 40% for languages like Swahili, highlighting persistent gaps in multilingual capabilities despite recent advances. MMLU-ProX is an ongoing project; we are expanding our benchmark by incorporating additional languages and evaluating more language models to provide a more comprehensive assessment of multilingual capabilities.
Reinforcement Learning (RL) algorithms for safety alignment of Large Language Models (LLMs), such as Direct Preference Optimization (DPO), encounter the challenge of distribution shift. Current approaches typically address this issue through online sampling from the target policy, which requires significant computational resources. In this paper, we hypothesize that during off-policy training, while the ranking order of output generated by policy changes, their overall distribution remains relatively stable. This stability allows the transformation of the sampling process from the target policy into a re-ranking of preference data. Building on this hypothesis, We propose a new framework that leverages the model's intrinsic safety judgment capability to extract reward signals, which are then used to calculate label confidence for preferences reordering. Extensive experimental results and theoretical analysis demonstrate that the proposed method effectively addresses the distribution shift issue, remarkably enhancing the safety performance while reducing about 300x computational overheads.
As large language models continue to scale, computational costs and resource consumption have emerged as significant challenges. While existing sparsification methods like pruning reduce computational overhead, they risk losing model knowledge through parameter removal. This paper proposes DSMoE (Dynamic Sparse Mixture-of-Experts), a novel approach that achieves sparsification by partitioning pre-trained FFN layers into computational blocks. We implement adaptive expert routing using sigmoid activation and straight-through estimators, enabling tokens to flexibly access different aspects of model knowledge based on input complexity. Additionally, we introduce a sparsity loss term to balance performance and computational efficiency. Extensive experiments on LLaMA models demonstrate that under equivalent computational constraints, DSMoE achieves superior performance compared to existing pruning and MoE approaches across language modeling and downstream tasks, particularly excelling in generation tasks. Analysis reveals that DSMoE learns distinctive layerwise activation patterns, providing new insights for future MoE architecture design.
With the emerging trend of GPT models, we have established a framework called AutoML-GPT that integrates a comprehensive set of tools and libraries. This framework grants users access to a wide range of data preprocessing techniques, feature engineering methods, and model selection algorithms. Through a conversational interface, users can specify their requirements, constraints, and evaluation metrics. Throughout the process, AutoML-GPT employs advanced techniques for hyperparameter optimization and model selection, ensuring that the resulting model achieves optimal performance. The system effectively manages the complexity of the machine learning pipeline, guiding users towards the best choices without requiring deep domain knowledge. Through our experimental results on diverse datasets, we have demonstrated that AutoML-GPT significantly reduces the time and effort required for machine learning tasks. Its ability to leverage the vast knowledge encoded in large language models enables it to provide valuable insights, identify potential pitfalls, and suggest effective solutions to common challenges faced during model training.
Current image generation and editing methods primarily process textual prompts as direct inputs without reasoning about visual composition and explicit operations. We present Generation Chain-of-Thought (GoT), a novel paradigm that enables generation and editing through an explicit language reasoning process before outputting images. This approach transforms conventional text-to-image generation and editing into a reasoning-guided framework that analyzes semantic relationships and spatial arrangements. We define the formulation of GoT and construct large-scale GoT datasets containing over 9M samples with detailed reasoning chains capturing semantic-spatial relationships. To leverage the advantages of GoT, we implement a unified framework that integrates Qwen2.5-VL for reasoning chain generation with an end-to-end diffusion model enhanced by our novel Semantic-Spatial Guidance Module. Experiments show our GoT framework achieves excellent performance on both generation and editing tasks, with significant improvements over baselines. Additionally, our approach enables interactive visual generation, allowing users to explicitly modify reasoning steps for precise image adjustments. GoT pioneers a new direction for reasoning-driven visual generation and editing, producing images that better align with human intent. To facilitate future research, we make our datasets, code, and pretrained models publicly available at https://github.com/rongyaofang/GoT.
Pixel grounding, encompassing tasks such as Referring Expression Segmentation (RES), has garnered considerable attention due to its immense potential for bridging the gap between vision and language modalities. However, advancements in this domain are currently constrained by limitations inherent in existing datasets, including limited object categories, insufficient textual diversity, and a scarcity of high-quality annotations. To mitigate these limitations, we introduce GroundingSuite, which comprises: (1) an automated data annotation framework leveraging multiple Vision-Language Model (VLM) agents; (2) a large-scale training dataset encompassing 9.56 million diverse referring expressions and their corresponding segmentations; and (3) a meticulously curated evaluation benchmark consisting of 3,800 images. The GroundingSuite training dataset facilitates substantial performance improvements, enabling models trained on it to achieve state-of-the-art results. Specifically, a cIoU of 68.9 on gRefCOCO and a gIoU of 55.3 on RefCOCOm. Moreover, the GroundingSuite annotation framework demonstrates superior efficiency compared to the current leading data annotation method, i.e., $4.5 \times$ faster than the GLaMM.
Analyzing animal behavior is crucial in advancing neuroscience, yet quantifying and deciphering its intricate dynamics remains a significant challenge. Traditional machine vision approaches, despite their ability to detect spontaneous behaviors, fall short due to limited interpretability and reliance on manual labeling, which restricts the exploration of the full behavioral spectrum. Here, we introduce MouseGPT, a Vision-Language Model (VLM) that integrates visual cues with natural language to revolutionize mouse behavior analysis. Built upon our first-of-its-kind dataset - incorporating pose dynamics and open-vocabulary behavioral annotations across over 42 million frames of diverse psychiatric conditions - MouseGPT provides a novel, context-rich method for comprehensive behavior interpretation. Our holistic analysis framework enables detailed behavior profiling, clustering, and novel behavior discovery, offering deep insights without the need for labor - intensive manual annotation. Evaluations reveal that MouseGPT surpasses existing models in precision, adaptability, and descriptive richness, positioning it as a transformative tool for ethology and for unraveling complex behavioral dynamics in animal models.
This work concerns the path-star task, a minimal example of searching over a graph. The graph, $G$, is star-shaped with $D$ arms radiating from a start node, $s$. A language model (LM) is given $G$, $s$, and a target node $t$, which ends one of the arms and is tasked with generating the arm containing $t$. The minimal nature of this task means only a single choice needs to be made: which of the $D$ arms contains $t$? Decoder-only LMs fail to solve this elementary task above $1/D$ chance due to a learned shortcut that absorbs training supervision. We show how this pathology is caused by excess supervision and we present a series of solutions demonstrating that the task is solvable via decoder-only LMs. We find that the task's minimal nature causes its difficulty, as it prevents task decomposition. Our solutions provide insight into the pathology and its implications for LMs trained via next-token prediction.
Existing Image Restoration (IR) studies typically focus on task-specific or universal modes individually, relying on the mode selection of users and lacking the cooperation between multiple task-specific/universal restoration modes. This leads to insufficient interaction for unprofessional users and limits their restoration capability for complicated real-world applications. In this work, we present HybridAgent, intending to incorporate multiple restoration modes into a unified image restoration model and achieve intelligent and efficient user interaction through our proposed hybrid agents. Concretely, we propose the hybrid rule of fast, slow, and feedback restoration agents. Here, the slow restoration agent optimizes the powerful multimodal large language model (MLLM) with our proposed instruction-tuning dataset to identify degradations within images with ambiguous user prompts and invokes proper restoration tools accordingly. The fast restoration agent is designed based on a lightweight large language model (LLM) via in-context learning to understand the user prompts with simple and clear requirements, which can obviate the unnecessary time/resource costs of MLLM. Moreover, we introduce the mixed distortion removal mode for our HybridAgents, which is crucial but not concerned in previous agent-based works. It can effectively prevent the error propagation of step-by-step image restoration and largely improve the efficiency of the agent system. We validate the effectiveness of HybridAgent with both synthetic and real-world IR tasks.
Operations Research (OR) has been widely applied in various fields such as resource allocation, production planning, and supply chain management. However, addressing real-world OR problems requires OR experts to perform mathematical modeling and programmers to develop solution algorithms. This traditional method, heavily reliant on experts, is costly and has long development cycles, severely limiting the widespread adoption of OR techniques. Few have considered using Artificial Intelligence (AI) to replace professionals to achieve fully automated solutions for OR problems. We propose OR-LLM-Agent, the first AI agent that enables end-to-end automation for solving real-world OR problems. OR-LLM-Agent leverages the Chain-of-Thought (CoT) reasoning capabilities of Large Language Models (LLMs) to translate natural language problem descriptions into formal mathematical models and automatically generate Gurobi solver code. In OR-LLM-Agent, OR-CodeAgent is designed to automate code execution and repair within a sandbox environment, facilitating the derivation of the final solution. Due to the lack of dedicated benchmark datasets for evaluating the automated solving of OR problems, we construct a benchmark dataset comprising 83 real-world OR problems described in natural language. We conduct comparative experiments with state-of-the-art (SOTA) reasoning LLMs, including GPT-o3-mini, DeepSeek-R1, and Gemini 2.0 Flash Thinking. The OR-LLM-Agent achieved the highest pass rate of 100% and the highest solution accuracy of 85%, demonstrating the feasibility of automated OR problem-solving. Data and code have been publicly available at https://github.com/bwz96sco/or_llm_agent.
Knowing gene regulatory networks (GRNs) is important for understanding various biological mechanisms. In this paper, we present a method, QWENDY, that uses single-cell gene expression data measured at four time points to infer GRNs. Based on a linear gene expression model, it solves the transformation for the covariance matrices. Unlike its predecessor WENDY, QWENDY avoids solving a non-convex optimization problem and produces a unique solution. Then we enhance QWENDY by the transformer neural networks and large language models to obtain two variants: TEQWENDY and LEQWENDY. We test the performance of these methods on two experimental data sets and two synthetic data sets. TEQWENDY has the best overall performance, while QWENDY ranks the first on experimental data sets.
We introduce Tempest, a multi-turn adversarial framework that models the gradual erosion of Large Language Model (LLM) safety through a tree search perspective. Unlike single-turn jailbreaks that rely on one meticulously engineered prompt, Tempest expands the conversation at each turn in a breadth-first fashion, branching out multiple adversarial prompts that exploit partial compliance from previous responses. By tracking these incremental policy leaks and re-injecting them into subsequent queries, Tempest reveals how minor concessions can accumulate into fully disallowed outputs. Evaluations on the JailbreakBench dataset show that Tempest achieves a 100% success rate on GPT-3.5-turbo and 97% on GPT-4 in a single multi-turn run, using fewer queries than baselines such as Crescendo or GOAT. This tree search methodology offers an in-depth view of how model safeguards degrade over successive dialogue turns, underscoring the urgency of robust multi-turn testing procedures for language models.
We propose PRISM, a novel framework designed to overcome the limitations of 2D-based Preference-Based Reinforcement Learning (PBRL) by unifying 3D point cloud modeling and future-aware preference refinement. At its core, PRISM adopts a 3D Point Cloud-Language Model (3D-PC-LLM) to mitigate occlusion and viewpoint biases, ensuring more stable and spatially consistent preference signals. Additionally, PRISM leverages Chain-of-Thought (CoT) reasoning to incorporate long-horizon considerations, thereby preventing the short-sighted feedback often seen in static preference comparisons. In contrast to conventional PBRL techniques, this integration of 3D perception and future-oriented reasoning leads to significant gains in preference agreement rates, faster policy convergence, and robust generalization across unseen robotic environments. Our empirical results, spanning tasks such as robotic manipulation and autonomous navigation, highlight PRISM's potential for real-world applications where precise spatial understanding and reliable long-term decision-making are critical. By bridging 3D geometric awareness with CoT-driven preference modeling, PRISM establishes a comprehensive foundation for scalable, human-aligned reinforcement learning.
Vision-and-Language Navigation (VLN) in continuous environments requires agents to interpret natural language instructions while navigating unconstrained 3D spaces. Existing VLN-CE frameworks rely on a two-stage approach: a waypoint predictor to generate waypoints and a navigator to execute movements. However, current waypoint predictors struggle with spatial awareness, while navigators lack historical reasoning and backtracking capabilities, limiting adaptability. We propose a zero-shot VLN-CE framework integrating an enhanced waypoint predictor with a Multi-modal Large Language Model (MLLM)-based navigator. Our predictor employs a stronger vision encoder, masked cross-attention fusion, and an occupancy-aware loss for better waypoint quality. The navigator incorporates history-aware reasoning and adaptive path planning with backtracking, improving robustness. Experiments on R2R-CE and MP3D benchmarks show our method achieves state-of-the-art (SOTA) performance in zero-shot settings, demonstrating competitive results compared to fully supervised methods. Real-world validation on Turtlebot 4 further highlights its adaptability.
As we scale to more massive machine learning models, the frequent synchronization demands inherent in data-parallel approaches create significant slowdowns, posing a critical challenge to further scaling. Recent work develops an approach (DiLoCo) that relaxes synchronization demands without compromising model quality. However, these works do not carefully analyze how DiLoCo's behavior changes with model size. In this work, we study the scaling law behavior of DiLoCo when training LLMs under a fixed compute budget. We focus on how algorithmic factors, including number of model replicas, hyperparameters, and token budget affect training in ways that can be accurately predicted via scaling laws. We find that DiLoCo scales both predictably and robustly with model size. When well-tuned, DiLoCo scales better than data-parallel training with model size, and can outperform data-parallel training even at small model sizes. Our results showcase a more general set of benefits of DiLoCo than previously documented, including increased optimal batch sizes, improved downstream generalization with scale, and improved evaluation loss for a fixed token budget.
Dance serves as a profound and universal expression of human culture, conveying emotions and stories through movements synchronized with music. Although some current works have achieved satisfactory results in the task of single-person dance generation, the field of multi-person dance generation remains relatively novel. In this work, we present a group choreography framework that leverages recent advancements in Large Language Models (LLM) by modeling the group dance generation problem as a sequence-to-sequence translation task. Our framework consists of a tokenizer that transforms continuous features into discrete tokens, and an LLM that is fine-tuned to predict motion tokens given the audio tokens. We show that by proper tokenization of input modalities and careful design of the LLM training strategies, our framework can generate realistic and diverse group dances while maintaining strong music correlation and dancer-wise consistency. Extensive experiments and evaluations demonstrate that our framework achieves state-of-the-art performance.
Sparse autoencoders (SAEs) are a popular technique for interpreting language model activations, and there is extensive recent work on improving SAE effectiveness. However, most prior work evaluates progress using unsupervised proxy metrics with unclear practical relevance. We introduce SAEBench, a comprehensive evaluation suite that measures SAE performance across seven diverse metrics, spanning interpretability, feature disentanglement and practical applications like unlearning. To enable systematic comparison, we open-source a suite of over 200 SAEs across eight recently proposed SAE architectures and training algorithms. Our evaluation reveals that gains on proxy metrics do not reliably translate to better practical performance. For instance, while Matryoshka SAEs slightly underperform on existing proxy metrics, they substantially outperform other architectures on feature disentanglement metrics; moreover, this advantage grows with SAE scale. By providing a standardized framework for measuring progress in SAE development, SAEBench enables researchers to study scaling trends and make nuanced comparisons between different SAE architectures and training methodologies. Our interactive interface enables researchers to flexibly visualize relationships between metrics across hundreds of open-source SAEs at: https://saebench.xyz
Inspired by the success of DeepSeek R1 in reasoning via reinforcement learning without human feedback, we train a 3B language model using the Countdown Game with pure reinforcement learning. Our model outperforms baselines on four of five benchmarks, demonstrating improved generalization beyond its training data. Notably, response length does not correlate with reasoning quality, and while "aha moments" emerge, they do not always yield correct answers. These findings highlight the potential of RL-only training for reasoning enhancement and suggest future work on refining reward structures to bridge emergent insights with accuracy.
This paper elucidates that current state-of-the-art Large Language Models (LLMs) are fundamentally incapable of making decisions or developing "thoughts" within the feature space due to their architectural constraints. We establish a definition of "thought" that encompasses traditional understandings of that term and adapt it for application to LLMs. We demonstrate that the architectural design and language modeling training methodology of contemporary LLMs inherently preclude them from engaging in genuine thought processes. Our primary focus is on this theoretical realization rather than practical insights derived from experimental data. Finally, we propose solutions to enable thought processes within the feature space and discuss the broader implications of these architectural modifications.
Reinforcement learning from human feedback (RLHF) has become essential for improving language model capabilities, but traditional approaches rely on the assumption that human preferences follow a transitive Bradley-Terry model. This assumption fails to capture the non-transitive nature of populational human preferences. Nash learning from human feedback (NLHF), targeting non-transitive preferences, is a problem of computing the Nash equilibrium (NE) of the two-player constant-sum game defined by the human preference. We introduce Extragradient preference optimization (EGPO), a novel algorithm for NLHF achieving last-iterate linear convergence to the NE of KL-regularized games and polynomial convergence to the NE of original games, while being robust to noise. Unlike previous approaches that rely on nested optimization, we derive an equivalent implementation using gradients of an online variant of the identity preference optimization (IPO) loss, enabling more faithful implementation for neural networks. Our empirical evaluations demonstrate EGPO's superior performance over baseline methods when training for the same number of epochs, as measured by pairwise win-rates using the ground truth preference. These results validate both the theoretical strengths and practical advantages of EGPO for language model alignment with non-transitive human preferences.