Compute scaling for language model (LM) pretraining has outpaced the growth of human-written texts, leading to concerns that data will become the bottleneck to LM scaling. To continue scaling pretraining in this data-constrained regime, we propose that explicitly modeling and inferring the latent thoughts that underlie the text generation process can significantly improve pretraining data efficiency. Intuitively, our approach views web text as the compressed final outcome of a verbose human thought process and that the latent thoughts contain important contextual knowledge and reasoning steps that are critical to data-efficient learning. We empirically demonstrate the effectiveness of our approach through data-constrained continued pretraining for math. We first show that synthetic data approaches to inferring latent thoughts significantly improve data efficiency, outperforming training on the same amount of raw data (5.7\% $\rightarrow$ 25.4\% on MATH). Furthermore, we demonstrate latent thought inference without a strong teacher, where an LM bootstraps its own performance by using an EM algorithm to iteratively improve the capability of the trained LM and the quality of thought-augmented pretraining data. We show that a 1B LM can bootstrap its performance across at least three iterations and significantly outperform baselines trained on raw data, with increasing gains from additional inference compute when performing the E-step. The gains from inference scaling and EM iterations suggest new opportunities for scaling data-constrained pretraining.
Mathematical error detection in educational settings presents a significant challenge for Multimodal Large Language Models (MLLMs), requiring a sophisticated understanding of both visual and textual mathematical content along with complex reasoning capabilities. Though effective in mathematical problem-solving, MLLMs often struggle with the nuanced task of identifying and categorizing student errors in multimodal mathematical contexts. Therefore, we introduce MathAgent, a novel Mixture-of-Math-Agent framework designed specifically to address these challenges. Our approach decomposes error detection into three phases, each handled by a specialized agent: an image-text consistency validator, a visual semantic interpreter, and an integrative error analyzer. This architecture enables more accurate processing of mathematical content by explicitly modeling relationships between multimodal problems and student solution steps. We evaluate MathAgent on real-world educational data, demonstrating approximately 5% higher accuracy in error step identification and 3% improvement in error categorization compared to baseline models. Besides, MathAgent has been successfully deployed in an educational platform that has served over one million K-12 students, achieving nearly 90% student satisfaction while generating significant cost savings by reducing manual error detection.
Difficult problems, which often result in long reasoning traces, are widely recognized as key factors for enhancing the performance of reasoning models. However, such high-challenge problems are scarce, limiting the size of available datasets. In this paper, we propose a simple method to decouple the reliance on problem difficulty. First, we empirically demonstrate that reasoning length, rather than problem difficulty, primarily influences the performance of trained models. Second, we identify a scaling law on reasoning length, showing that model performance increases in a log-linear fashion as the reasoning data length grows. Finally, we introduce a straightforward technique to generate reasoning data of arbitrary length, and show that synthesized data is effective for training reasoning models. After fine-tuning the Qwen2.5-32B-Instruct language model on our Long1K dataset, we present our model, Long1K-32B, which achieves remarkable performance with only 1,000 training samples, achieving 95.6\% accuracy on MATH, and 71.1\% on GPQA outperforming DeepSeek-R1-Distill-Qwen-32B. The model, code, and dataset are all open-sourced, available at https://huggingface.co/ZTss/LONG1.
Automatic math correction aims to check students' solutions to mathematical problems via artificial intelligence technologies. Most existing studies focus on judging the final answer at the problem level, while they ignore detailed feedback on each step in a math problem-solving process, which requires abilities of semantic understanding and reasoning. In this paper, we propose a reinforcement learning (RL)-based method to boost large language model (LLM) for step-level automatic math correction, named StepAMC. Particularly, we convert the step-level automatic math correction within the text classification task into an RL problem to enhance the reasoning capabilities of LLMs. Then, we design a space-constrained policy network to improve the stability of RL. Then, we introduce a fine-grained reward network to convert the binary human feedback into a continuous value. We conduct extensive experiments over two benchmark datasets and the results show that our model outperforms the eleven strong baselines.
Progress in scientific discovery is rarely the result of a single "Eureka" moment, but is rather the product of hundreds of scientists incrementally working together toward a common goal. While existing agent workflows are capable of producing research autonomously, they do so in isolation, without the ability to continuously improve upon prior research results. To address these challenges, we introduce AgentRxiv-a framework that lets LLM agent laboratories upload and retrieve reports from a shared preprint server in order to collaborate, share insights, and iteratively build on each other's research. We task agent laboratories to develop new reasoning and prompting techniques and find that agents with access to their prior research achieve higher performance improvements compared to agents operating in isolation (11.4% relative improvement over baseline on MATH-500). We find that the best performing strategy generalizes to benchmarks in other domains (improving on average by 3.3%). Multiple agent laboratories sharing research through AgentRxiv are able to work together towards a common goal, progressing more rapidly than isolated laboratories, achieving higher overall accuracy (13.7% relative improvement over baseline on MATH-500). These findings suggest that autonomous agents may play a role in designing future AI systems alongside humans. We hope that AgentRxiv allows agents to collaborate toward research goals and enables researchers to accelerate discovery.
Large Language Models (LLMs) have significantly advanced various fields, particularly coding, mathematical reasoning, and logical problem solving. However, a critical question remains: Do these mathematical reasoning abilities persist when LLMs are presented with culturally adapted math problems? Specifically, how do LLMs perform when faced with math problems embedded in cultural contexts that have no significant representation in main stream web-scale AI training data? To explore this, we generated six synthetic cultural datasets from GSM8K, a widely used benchmark for assessing LLMs' mathematical reasoning skills. While preserving the mathematical logic and numerical values of the original GSM8K test set, we modify cultural elements such as personal names, food items, place names, etc. These culturally adapted datasets provide a more reliable framework for evaluating LLMs' mathematical reasoning under shifting cultural contexts. Our findings reveal that LLMs struggle with math problems when cultural references change, even though the underlying mathematical structure remains constant. Smaller models exhibit greater performance drops compared to larger models. Interestingly, our results also suggest that cultural familiarity can enhance mathematical reasoning. Even models with no explicit mathematical training but exposure to relevant cultural contexts sometimes outperform larger, mathematically proficient models on culturally embedded math problems. This study highlights the impact of cultural context on the mathematical reasoning abilities of LLMs, underscoring the need for more diverse and representative training data to improve robustness in real-world applications. The benchmark data sets and script for reproducing the results are available at https://github.com/akarim23131/Lost_in_Cultural_Translation
Improving the training efficiency remains one of the most significant challenges in large-scale reinforcement learning. In this paper, we investigate how the model's context length and the complexity of the training dataset influence the training process of R1-like models. Our experiments reveal three key insights: (1) adopting longer context lengths may not necessarily result in better performance; (2) selecting an appropriate context length helps mitigate entropy collapse; and (3) appropriately controlling the model's context length and curating training data based on input prompt length can effectively improve RL training efficiency, achieving better performance with shorter thinking length. Inspired by these insights, we propose FastCuRL, a curriculum reinforcement learning framework with the progressive context extension strategy, and successfully accelerate the training process of RL models. Experimental results demonstrate that FastCuRL-1.5B-Preview surpasses DeepScaleR-1.5B-Preview across all five benchmarks while only utilizing 50\% of training steps. Furthermore, all training stages for FastCuRL-1.5B-Preview are completed using a single node with 8 GPUs.
Large language models (LLMs) and Vision language models (VLMs) have been able to perform various forms of reasoning tasks in a wide range of scenarios, but are they truly engaging in task abstraction and rule-based reasoning beyond mere memorization and pattern matching? To answer this question, we propose a novel experimental approach, Misleading Fine-Tuning (MisFT), to examine whether LLMs/VLMs perform abstract reasoning by altering their original understanding of fundamental rules. In particular, by constructing a dataset with math expressions that contradict correct operation principles, we fine-tune the model to learn those contradictory rules and assess its generalization ability on different test domains. Through a series of experiments, we find that current LLMs/VLMs are capable of effectively applying contradictory rules to solve practical math word problems and math expressions represented by images, implying the presence of an internal mechanism that abstracts before reasoning.
We present EXAONE Deep series, which exhibits superior capabilities in various reasoning tasks, including math and coding benchmarks. We train our models mainly on the reasoning-specialized dataset that incorporates long streams of thought processes. Evaluation results show that our smaller models, EXAONE Deep 2.4B and 7.8B, outperform other models of comparable size, while the largest model, EXAONE Deep 32B, demonstrates competitive performance against leading open-weight models. All EXAONE Deep models are openly available for research purposes and can be downloaded from https://huggingface.co/LGAI-EXAONE
Large language models (LLMs) have demonstrated remarkable capabilities in various natural language processing tasks. However, achieving strong performance in specialized domains like mathematical reasoning and non-English languages often requires extensive training on massive datasets. This paper investigates a contrasting approach: strategic fine-tuning on a small, high-quality, bilingual (English-French) dataset to enhance both the reasoning capabilities and French language proficiency of a large language model. Rather than relying on scale, we explore the hypothesis that targeted data curation and optimized training can achieve competitive, or even superior, performance. We demonstrate, through targeted supervised fine-tuning (SFT) on only 2,000 carefully selected samples, significant improvements in mathematical reasoning. Specifically, Pensez 7B exhibits an increase in accuracy of the base model up to 20% on the AIME25 and a 12% increase on a French MATH level 5 benchmark. These results challenge the prevailing assumption that massive datasets are aprerequisite for strong reasoning performance in LLMs, highlighting the potential of strategic data curation and optimized fine-tuning for enhancing both specialized skills and multilingual capabilities. Our findings have implications for the efficient development of high-performing, multilingual LLMs, especially in resource-constrained scenarios.
Recent breakthroughs in solving reasoning, math and coding problems with Large Language Models (LLMs) have been enabled by investing substantial computation budgets at inference time. Therefore, inference speed is one of the most critical properties of LLM architectures, and there is a growing need for LLMs that are efficient and fast at inference. Recently, LLMs built on the xLSTM architecture have emerged as a powerful alternative to Transformers, offering linear compute scaling with sequence length and constant memory usage, both highly desirable properties for efficient inference. However, such xLSTM-based LLMs have yet to be scaled to larger models and assessed and compared with respect to inference speed and efficiency. In this work, we introduce xLSTM 7B, a 7-billion-parameter LLM that combines xLSTM's architectural benefits with targeted optimizations for fast and efficient inference. Our experiments demonstrate that xLSTM 7B achieves performance on downstream tasks comparable to other similar-sized LLMs, while providing significantly faster inference speeds and greater efficiency compared to Llama- and Mamba-based LLMs. These results establish xLSTM 7B as the fastest and most efficient 7B LLM, offering a solution for tasks that require large amounts of test-time computation. Our work highlights xLSTM's potential as a foundational architecture for methods building on heavy use of LLM inference. Our model weights, model code and training code are open-source.
Reasoning and strategic behavior in social interactions is a hallmark of intelligence. This form of reasoning is significantly more sophisticated than isolated planning or reasoning tasks in static settings (e.g., math problem solving). In this paper, we present Strategic Planning, Interaction, and Negotiation (SPIN-Bench), a new multi-domain evaluation designed to measure the intelligence of strategic planning and social reasoning. While many existing benchmarks focus on narrow planning or single-agent reasoning, SPIN-Bench combines classical PDDL tasks, competitive board games, cooperative card games, and multi-agent negotiation scenarios in one unified framework. The framework includes both a benchmark as well as an arena to simulate and evaluate the variety of social settings to test reasoning and strategic behavior of AI agents. We formulate the benchmark SPIN-Bench by systematically varying action spaces, state complexity, and the number of interacting agents to simulate a variety of social settings where success depends on not only methodical and step-wise decision making, but also conceptual inference of other (adversarial or cooperative) participants. Our experiments reveal that while contemporary LLMs handle basic fact retrieval and short-range planning reasonably well, they encounter significant performance bottlenecks in tasks requiring deep multi-hop reasoning over large state spaces and socially adept coordination under uncertainty. We envision SPIN-Bench as a catalyst for future research on robust multi-agent planning, social reasoning, and human--AI teaming. Project Website: https://spinbench.github.io/
We propose a new algorithm for fine-tuning large language models using reinforcement learning. Tapered Off-Policy REINFORCE (TOPR) uses an asymmetric, tapered variant of importance sampling to speed up learning while maintaining stable learning dynamics, even without the use of KL regularization. TOPR can be applied in a fully offline fashion, allows the handling of positive and negative examples in a unified framework, and benefits from the implementational simplicity that is typical of Monte Carlo algorithms. We demonstrate the effectiveness of our approach with a series of experiments on the GSM8K and MATH reasoning benchmarks, finding performance gains for training both a model for solution generation and as a generative verifier. We show that properly leveraging positive and negative examples alike in the off-policy regime simultaneously increases test-time accuracy and training data efficiency, all the while avoiding the ``wasted inference'' that comes with discarding negative examples. We find that this advantage persists over multiple iterations of training and can be amplified by dataset curation techniques, enabling us to match 70B-parameter model performance with 8B language models. As a corollary to this work, we find that REINFORCE's baseline parameter plays an important and unexpected role in defining dataset composition in the presence of negative examples, and is consequently critical in driving off-policy performance.
In this work, we introduce BurTorch, a compact high-performance framework designed to optimize Deep Learning (DL) training on single-node workstations through an exceptionally efficient CPU-based backpropagation (Rumelhart et al., 1986; Linnainmaa, 1970) implementation. Although modern DL frameworks rely on compilerlike optimizations internally, BurTorch takes a different path. It adopts a minimalist design and demonstrates that, in these circumstances, classical compiled programming languages can play a significant role in DL research. By eliminating the overhead of large frameworks and making efficient implementation choices, BurTorch achieves orders-of-magnitude improvements in performance and memory efficiency when computing $\nabla f(x)$ on a CPU. BurTorch features a compact codebase designed to achieve two key goals simultaneously. First, it provides a user experience similar to script-based programming environments. Second, it dramatically minimizes runtime overheads. In large DL frameworks, the primary source of memory overhead for relatively small computation graphs $f(x)$ is due to feature-heavy implementations. We benchmarked BurTorch against widely used DL frameworks in their execution modes: JAX (Bradbury et al., 2018), PyTorch (Paszke et al., 2019), TensorFlow (Abadi et al., 2016); and several standalone libraries: Autograd (Maclaurin et al., 2015), Micrograd (Karpathy, 2020), Apple MLX (Hannun et al., 2023). For small compute graphs, BurTorch outperforms best-practice solutions by up to $\times 2000$ in runtime and reduces memory consumption by up to $\times 3500$. For a miniaturized GPT-3 model (Brown et al., 2020), BurTorch achieves up to a $\times 20$ speedup and reduces memory up to $\times 80$ compared to PyTorch.
Time-series analysis is critical for a wide range of fields such as healthcare, finance, transportation, and energy, among many others. The practical applications often involve analyzing time-series data alongside contextual information in the form of natural language to support informed decisions. However, current time-series models are limited in their ability to perform reasoning that involves both time-series and their textual content. In this work, we address this gap by introducing \textit{Chat-TS}, a large language model (LLM) based framework, designed to support reasoning over time series and textual data. Unlike traditional models, Chat-TS integrates time-series tokens into LLMs' vocabulary, enhancing its reasoning ability over both modalities without compromising the core natural language capabilities, enabling practical analysis and reasoning across modalities. To support learning and evaluation in this setup, we contribute new datasets: the \textit{TS Instruct Training Dataset} which pairs diverse time-series data with relevant text instructions and responses for instruction tuning, the \textit{TS Instruct Question and Answer (QA) Gold Dataset} which provides multiple-choice questions designed to evaluate multimodal reasoning, and a \textit{TS Instruct Quantitative Probing Set} which contains a small subset of the TS Instruct QA tasks alongside math and decision-making questions for LLM evaluation. We designed a training strategy to preserve the inherent reasoning capabilities of LLMs while augmenting them for time-series reasoning. Experiments show that Chat-TS achieves state-of-the-art performance in multi-modal reasoning tasks by maintaining strong natural language proficiency while improving time-series reasoning. ~\footnote{To ensure replicability and facilitate future research, all models, datasets, and code will be available at [\texttt{Github-URL}].}
This paper introduces Light-R1, an open-source suite for training long reasoning models using reproducible and cost-effective methodology. Given the proprietary nature of data used in the DeepSeek-R1 series, we develop an alternative approach leveraging exclusively public data and models. Our curriculum training progressively increases data difficulty, combined with multi-staged post-training. Our Light-R1-32B model, trained from Qwen2.5-32B-Instruct, outperforms DeepSeek-R1-Distill-Qwen-32B in math reasoning. Experimental results show that this curriculum approach becomes more effective when distinct, diverse datasets are available for different training stages: fine-tuning DeepSeek-R1-Distilled models (pre-tuned by DeepSeek team on proprietary data) with 3,000 challenging examples from our curriculum dataset yielded state-of-the-art 7B and 14B models, while the 32B model, Light-R1-32B-DS performed comparably to QwQ-32B and DeepSeek-R1. Furthermore, we extend our work by applying GRPO on long reasoning models. Our final Light-R1-14B-DS achieves SOTA performance among 14B models in math, with AIME24 \& 25 scores of 74.0 and 60.2 respectively, surpassing many 32B models and DeepSeek-R1-Distill-Llama-70B. Despite math-focused training, Light-R1-14B-DS demonstrates strong cross-domain generalization. Light-R1 represents a significant advancement in making sophisticated reasoning models more accessible and implementable in real-world applications. Our models, training data and code have been made available at https://github.com/Qihoo360/Light-R1.
We study the capabilities of Large Language Models (LLM) on binary relations, a ubiquitous concept in math employed in most reasoning, math and logic benchmarks. This work focuses on equality, inequality, and inclusion, along with the properties they satisfy, such as ir/reflexivity, a/symmetry, transitivity, and logical complexity (e.g., number of reasoning ``hops''). We propose an alternative to in-context learning that trains only the representations of newly introduced tokens, namely out-of-context representation learning. This method mitigates linguistic biases already present in a model and, differently from in-context learning, does not rely on external information or illustrations. We argue out-of-context representation learning as a better alternative to in-context learning and fine-tuning to evaluate the capabilities of LLMs on logic tasks that are the building blocks of more complex reasoning benchmarks.
Vision-Language Models have made significant progress on many perception-focused tasks, however, their progress on reasoning-focused tasks seem to be limited due to the lack of high-quality and diverse training data. In this work, we aim to address the scarcity issue of reasoning-focused multimodal datasets. We propose VisualWebInstruct - a novel approach that leverages search engine to create a diverse, and high-quality dataset spanning multiple disciplines like math, physics, finance, chemistry, etc. Starting with meticulously selected 30,000 seed images, we employ Google Image search to identify websites containing similar images. We collect and process the HTMLs from over 700K unique URL sources. Through a pipeline of content extraction, filtering and synthesis, we build a dataset of approximately 900K question-answer pairs, with 40% being visual QA pairs and the rest as text QA pairs. Models fine-tuned on VisualWebInstruct demonstrate significant performance gains: (1) training from Llava-OV-mid shows 10-20% absolute point gains across benchmarks, (2) training from MAmmoTH-VL shows 5% absoluate gain. Our best model MAmmoTH-VL2 shows state-of-the-art performance within the 10B parameter class on MMMU-Pro-std (40.7%), MathVerse (42.6%), and DynaMath (55.7%). These remarkable results highlight the effectiveness of our dataset in enhancing VLMs' reasoning capabilities for complex multimodal tasks.
We consider the problem of generating valid and small prediction sets by sampling outputs (e.g., software code and natural language text) from a black-box deep generative model for a given input (e.g., textual prompt). The validity of a prediction set is determined by a user-defined binary admissibility function depending on the target application. For example, requiring at least one program in the set to pass all test cases in code generation application. To address this problem, we develop a simple and effective conformal inference algorithm referred to as Generative Prediction Sets (GPS). Given a set of calibration examples and black-box access to a deep generative model, GPS can generate prediction sets with provable guarantees. The key insight behind GPS is to exploit the inherent structure within the distribution over the minimum number of samples needed to obtain an admissible output to develop a simple conformal regression approach over the minimum number of samples. Experiments on multiple datasets for code and math word problems using different large language models demonstrate the efficacy of GPS over state-of-the-art methods.
Evaluating mathematical capabilities is critical for assessing the overall performance of large language models (LLMs). However, existing evaluation methods often focus solely on final answers, resulting in highly inaccurate and uninterpretable evaluation outcomes, as well as their failure to assess proof or open-ended problems. To address these issues, we propose a novel mathematical process evaluation agent based on Tree-of-Error, called StepMathAgent. This agent incorporates four internal core operations: logical step segmentation, step scoring, score aggregation and error tree generation, along with four external extension modules: difficulty calibration, simplicity evaluation, completeness validation and format assessment. Furthermore, we introduce StepMathBench, a benchmark comprising 1,000 step-divided process evaluation instances, derived from 200 high-quality math problems grouped by problem type, subject category and difficulty level. Experiments on StepMathBench show that our proposed StepMathAgent outperforms all state-of-the-art methods, demonstrating human-aligned evaluation preferences and broad applicability to various scenarios. Our data and code are available at https://github.com/SHU-XUN/StepMathAgent.