NeurIPS 2025 E2LM Competition : Early Training Evaluation of Language Models

Existing benchmarks have proven effective for assessing the performance of fully trained large language models. However, we find striking differences in the early training stages of small models, where benchmarks often fail to provide meaningful or discriminative signals. To explore how these differences arise, this competition tackles the challenge of designing scientific knowledge evaluation tasks specifically tailored for measuring early training progress of language models. Participants are invited to develop novel evaluation methodologies or adapt existing benchmarks to better capture performance differences among language models. To support this effort, we provide three pre-trained small models (0.5B, 1B, and 3B parameters), along with intermediate checkpoints sampled during training up to 200B tokens. All experiments and development work can be run on widely available free cloud-based GPU platforms, making participation accessible to researchers with limited computational resources. Submissions will be evaluated based on three criteria: the quality of the performance signal they produce, the consistency of model rankings at 1 trillion tokens of training, and their relevance to the scientific knowledge domain. By promoting the design of tailored evaluation strategies for early training, this competition aims to attract a broad range of participants from various disciplines, including those who may not be machine learning experts or have access to dedicated GPU resources. Ultimately, this initiative seeks to make foundational LLM research more systematic and benchmark-informed from the earliest phases of model development.

Navigating loss manifolds via rigid body dynamics: A promising avenue for robustness and generalisation

Training large neural networks through gradient-based optimization requires navigating high-dimensional loss landscapes, which often exhibit pathological geometry, leading to undesirable training dynamics. In particular, poor generalization frequently results from convergence to sharp minima that are highly sensitive to input perturbations, causing the model to overfit the training data while failing to generalize to unseen examples. Furthermore, these optimization procedures typically display strong dependence on the fine structure of the loss landscape, leading to unstable training dynamics, due to the fractal-like nature of the loss surface. In this work, we propose an alternative optimizer that simultaneously reduces this dependence, and avoids sharp minima, thereby improving generalization. This is achieved by simulating the motion of the center of a ball rolling on the loss landscape. The degree to which our optimizer departs from the standard gradient descent is controlled by a hyperparameter, representing the radius of the ball. Changing this hyperparameter allows for probing the loss landscape at different scales, making it a valuable tool for understanding its geometry.

Divide by Question, Conquer by Agent: SPLIT-RAG with Question-Driven Graph Partitioning

Retrieval-Augmented Generation (RAG) systems empower large language models (LLMs) with external knowledge, yet struggle with efficiency-accuracy trade-offs when scaling to large knowledge graphs. Existing approaches often rely on monolithic graph retrieval, incurring unnecessary latency for simple queries and fragmented reasoning for complex multi-hop questions. To address these challenges, this paper propose SPLIT-RAG, a multi-agent RAG framework that addresses these limitations with question-driven semantic graph partitioning and collaborative subgraph retrieval. The innovative framework first create Semantic Partitioning of Linked Information, then use the Type-Specialized knowledge base to achieve Multi-Agent RAG. The attribute-aware graph segmentation manages to divide knowledge graphs into semantically coherent subgraphs, ensuring subgraphs align with different query types, while lightweight LLM agents are assigned to partitioned subgraphs, and only relevant partitions are activated during retrieval, thus reduce search space while enhancing efficiency. Finally, a hierarchical merging module resolves inconsistencies across subgraph-derived answers through logical verifications. Extensive experimental validation demonstrates considerable improvements compared to existing approaches.

Accurate and Diverse LLM Mathematical Reasoning via Automated PRM-Guided GFlowNets

Achieving both accuracy and diverse reasoning remains challenging for Large Language Models (LLMs) in complex domains like mathematics. A key bottleneck is evaluating intermediate reasoning steps to guide generation without costly human annotations. To address this, we first introduce a novel Process Reward Model (PRM) trained automatically using Monte Carlo Tree Search coupled with a similarity-based data augmentation technique, effectively capturing step-level reasoning quality. Leveraging this PRM, we then adapt Generative Flow Networks (GFlowNets) to operate at the reasoning step level. Unlike traditional reinforcement learning focused on maximizing a single reward, GFlowNets naturally sample diverse, high-quality solutions proportional to their rewards, as measured by our PRM. Empirical evaluation shows strong improvements in both accuracy and solution diversity on challenging mathematical benchmarks (e.g., +2.59% absolute accuracy on MATH Level 5 for Llama3.2-3B), with effective generalization to unseen datasets (+9.4% absolute on SAT MATH). Our work demonstrates the potential of PRM-guided, step-level GFlowNets for developing more robust and versatile mathematical reasoning in LLMs.

KG-IRAG: A Knowledge Graph-Based Iterative Retrieval-Augmented Generation Framework for Temporal Reasoning

Graph Retrieval-Augmented Generation (GraphRAG) has proven highly effective in enhancing the performance of Large Language Models (LLMs) on tasks that require external knowledge. By leveraging Knowledge Graphs (KGs), GraphRAG improves information retrieval for complex reasoning tasks, providing more precise and comprehensive retrieval and generating more accurate responses to QAs. However, most RAG methods fall short in addressing multi-step reasoning, particularly when both information extraction and inference are necessary. To address this limitation, this paper presents Knowledge Graph-Based Iterative Retrieval-Augmented Generation (KG-IRAG), a novel framework that integrates KGs with iterative reasoning to improve LLMs' ability to handle queries involving temporal and logical dependencies. Through iterative retrieval steps, KG-IRAG incrementally gathers relevant data from external KGs, enabling step-by-step reasoning. The proposed approach is particularly suited for scenarios where reasoning is required alongside dynamic temporal data extraction, such as determining optimal travel times based on weather conditions or traffic patterns. Experimental results show that KG-IRAG improves accuracy in complex reasoning tasks by effectively integrating external knowledge with iterative, logic-based retrieval. Additionally, three new datasets: weatherQA-Irish, weatherQA-Sydney, and trafficQA-TFNSW, are formed to evaluate KG-IRAG's performance, demonstrating its potential beyond traditional RAG applications.

Enforcing Consistency and Fairness in Multi-level Hierarchical Classification with a Mask-based Output Layer

Traditional Multi-level Hierarchical Classification (MLHC) classifiers often rely on backbone models with $n$ independent output layers. This structure tends to overlook the hierarchical relationships between classes, leading to inconsistent predictions that violate the underlying taxonomy. Additionally, once a backbone architecture for an MLHC classifier is selected, adapting the model to accommodate new tasks can be challenging. For example, incorporating fairness to protect sensitive attributes within a hierarchical classifier necessitates complex adjustments to maintain the class hierarchy while enforcing fairness constraints. In this paper, we extend this concept to hierarchical classification by introducing a fair, model-agnostic layer designed to enforce taxonomy and optimize specific objectives, including consistency, fairness, and exact match. Our evaluations demonstrate that the proposed layer not only improves the fairness of predictions but also enforces the taxonomy, resulting in consistent predictions and superior performance. Compared to Large Language Models (LLMs) employing in-processing de-biasing techniques and models without any bias correction, our approach achieves better outcomes in both fairness and accuracy, making it particularly valuable in sectors like e-commerce, healthcare, and education, where predictive reliability is crucial.

Leveraging Taxonomy and LLMs for Improved Multimodal Hierarchical Classification

Multi-level Hierarchical Classification (MLHC) tackles the challenge of categorizing items within a complex, multi-layered class structure. However, traditional MLHC classifiers often rely on a backbone model with independent output layers, which tend to ignore the hierarchical relationships between classes. This oversight can lead to inconsistent predictions that violate the underlying taxonomy. Leveraging Large Language Models (LLMs), we propose a novel taxonomy-embedded transitional LLM-agnostic framework for multimodality classification. The cornerstone of this advancement is the ability of models to enforce consistency across hierarchical levels. Our evaluations on the MEP-3M dataset - a multi-modal e-commerce product dataset with various hierarchical levels - demonstrated a significant performance improvement compared to conventional LLM structures.

Visual question answering: from early developments to recent advances -- a survey

Visual Question Answering (VQA) is an evolving research field aimed at enabling machines to answer questions about visual content by integrating image and language processing techniques such as feature extraction, object detection, text embedding, natural language understanding, and language generation. With the growth of multimodal data research, VQA has gained significant attention due to its broad applications, including interactive educational tools, medical image diagnosis, customer service, entertainment, and social media captioning. Additionally, VQA plays a vital role in assisting visually impaired individuals by generating descriptive content from images. This survey introduces a taxonomy of VQA architectures, categorizing them based on design choices and key components to facilitate comparative analysis and evaluation. We review major VQA approaches, focusing on deep learning-based methods, and explore the emerging field of Large Visual Language Models (LVLMs) that have demonstrated success in multimodal tasks like VQA. The paper further examines available datasets and evaluation metrics essential for measuring VQA system performance, followed by an exploration of real-world VQA applications. Finally, we highlight ongoing challenges and future directions in VQA research, presenting open questions and potential areas for further development. This survey serves as a comprehensive resource for researchers and practitioners interested in the latest advancements and future

SimpsonsVQA: Enhancing Inquiry-Based Learning with a Tailored Dataset

Visual Question Answering (VQA) has emerged as a promising area of research to develop AI-based systems for enabling interactive and immersive learning. Numerous VQA datasets have been introduced to facilitate various tasks, such as answering questions or identifying unanswerable ones. However, most of these datasets are constructed using real-world images, leaving the performance of existing models on cartoon images largely unexplored. Hence, in this paper, we present "SimpsonsVQA", a novel dataset for VQA derived from The Simpsons TV show, designed to promote inquiry-based learning. Our dataset is specifically designed to address not only the traditional VQA task but also to identify irrelevant questions related to images, as well as the reverse scenario where a user provides an answer to a question that the system must evaluate (e.g., as correct, incorrect, or ambiguous). It aims to cater to various visual applications, harnessing the visual content of "The Simpsons" to create engaging and informative interactive systems. SimpsonsVQA contains approximately 23K images, 166K QA pairs, and 500K judgments (https://simpsonsvqa.org). Our experiments show that current large vision-language models like ChatGPT4o underperform in zero-shot settings across all three tasks, highlighting the dataset's value for improving model performance on cartoon images. We anticipate that SimpsonsVQA will inspire further research, innovation, and advancements in inquiry-based learning VQA.

Maximizing the Potential of Synthetic Data: Insights from Random Matrix Theory

Synthetic data has gained attention for training large language models, but poor-quality data can harm performance (see, e.g., Shumailov et al. (2023); Seddik et al. (2024)). A potential solution is data pruning, which retains only high-quality data based on a score function (human or machine feedback). Previous work Feng et al. (2024) analyzed models trained on synthetic data as sample size increases. We extend this by using random matrix theory to derive the performance of a binary classifier trained on a mix of real and pruned synthetic data in a high dimensional setting. Our findings identify conditions where synthetic data could improve performance, focusing on the quality of the generative model and verification strategy. We also show a smooth phase transition in synthetic label noise, contrasting with prior sharp behavior in infinite sample limits. Experiments with toy models and large language models validate our theoretical results.