Large language models (LLMs) are increasingly deployed in multilingual, real-world applications with user inputs -- naturally introducing typographical errors (typos). Yet most benchmarks assume clean input, leaving the robustness of LLMs to typos across languages largely underexplored. To address this gap, we introduce MulTypo, a multilingual typo generation algorithm that simulates human-like errors based on language-specific keyboard layouts and typing behavior. We evaluate 18 open-source LLMs across three model families and five downstream tasks spanning language inference, multi-choice question answering, mathematical reasoning, and machine translation tasks. Our results show that typos consistently degrade performance, particularly in generative tasks and those requiring reasoning -- while the natural language inference task is comparatively more robust. Instruction tuning improves clean-input performance but may increase brittleness under noise. We also observe language-dependent robustness: high-resource languages are generally more robust than low-resource ones, and translation from English is more robust than translation into English. Our findings underscore the need for noise-aware training and multilingual robustness evaluation. We make our code and data publicly available.

Large reasoning models (LRMs) boosted by Reinforcement Learning from Verifier Reward (RLVR) have shown great power in problem solving, yet they often cause overthinking: excessive, meandering reasoning that inflates computational cost. Prior designs of penalization in RLVR manage to reduce token consumption while often harming model performance, which arises from the oversimplicity of token-level supervision. In this paper, we argue that the granularity of supervision plays a crucial role in balancing efficiency and accuracy, and propose Group Relative Segment Penalization (GRSP), a step-level method to regularize reasoning. Since preliminary analyses show that reasoning segments are strongly correlated with token consumption and model performance, we design a length-aware weighting mechanism across segment clusters. Extensive experiments demonstrate that GRSP achieves superior token efficiency without heavily compromising accuracy, especially the advantages with harder problems. Moreover, GRSP stabilizes RL training and scales effectively across model sizes.

Estimating model performance without labels is an important goal for understanding how NLP models generalize. While prior work has proposed measures based on dataset similarity or predicted correctness, it remains unclear when these estimates produce reliable performance rankings across domains. In this paper, we analyze the factors that affect ranking reliability using a two-step evaluation setup with four base classifiers and several large language models as error predictors. Experiments on the GeoOLID and Amazon Reviews datasets, spanning 15 domains, show that large language model-based error predictors produce stronger and more consistent rank correlations with true accuracy than drift-based or zero-shot baselines. Our analysis reveals two key findings: ranking is more reliable when performance differences across domains are larger, and when the error model's predictions align with the base model's true failure patterns. These results clarify when performance estimation methods can be trusted and provide guidance for their use in cross-domain model evaluation.

Large language models (LLMs) have advanced rapidly in recent years (Brown et al., 2020; Touvron et al., 2023), demonstrating remarkable progress in complex reasoning (Guo et al., 2025), fluent text generation, and even automated proof discovery (Yu et al., 2025). These advances have spurred growing adoption of LLMs across education, data science, and research, where they are increasingly used for tutoring, problem explanation, data analysis, and hypothesis formulation (Wu et al., 2021; Polu and Sutskever, 2020; Khan et al., 2023; Gao et al., 2023). However, despite their broad deployment in quantitative domains, the field of statistics, which forms the foundation of modern data-driven science, has received little attention in LLM evaluation. Statistics differs fundamentally from other quantitative disciplines. Rather than focusing on symbolic manipulation or fixed-form computation, it emphasizes reasoning under uncertainty, connecting probability theory, inference, regression, Bayesian analysis, multivariate methods, and asymptotic theory into a unified framework. Yet existing large-scale LLM evaluations rarely cover these competencies: statistical problems account for less than 3% of recent reasoning benchmarks (Paster et al., 2025), and when included, they are typically treated as isolated probability puzzles without structured categorization or coverage of inferential reasoning (Gao et al., 2024). This gap makes it impossible to rigorously assess whether LLMs can function as capable statisticians or support data-driven scientific discovery. To bridge this critical gap, we introduce StatEval, the first large-scale benchmark dedicated to evaluating large language models on statistical reasoning. With nearly 20,000 meticulously curated problems, StatEval covers the entire spectrum of statistics, from basic undergraduate exercises to advanced research-level challenges, captures the full 2 breadth and depth of the discipline, as illustrated in Figure 1.

Knowin Figure 1: For an Embodied Agent, using physical tools is crucial in many tasks. Phys-ToolBench (Bottom) systematically evaluates the understanding of physical tools of multimodal LLMs. The benchmark is designed with three progressive levels of difficulty and employs a Visual Question Answering (VQA) format. Notice that in the actual benchmark, tools in the images are numerically labeled, and images here are for illustrative purposes only. The ability to use, understand, and create tools is a hallmark of human intelligence, enabling sophisticated interaction with the physical world. For any general-purpose intelligent agent to achieve true versatility, it must also master these fundamental skills. While modern Multimodal Large Language Models (MLLMs) leverage their extensive common knowledge for high-level planning in embodied AI and in downstream Vision-Language-Action (VLA) models, the extent of their true understanding of physical tools remains unquantified. To bridge this gap, we present PhysT oolBench, the first benchmark dedicated to evaluating the comprehension of physical tools by MLLMs. Our benchmark is structured as a Visual Question Answering (VQA) dataset comprising over 1,000 image-text pairs. Our comprehensive evaluation of 32 MLLMs--spanning proprietary, open-source, specialized embodied, and backbones in VLAs--reveals a significant deficiency in the tool understanding.

Modern conversational agents like ChatGPT and Alexa+ rely on predefined policies specifying metadata, response styles, and tool-usage rules. As these LLM-based systems expand to support diverse business and user queries, such policies, often implemented as in-context prompts, are becoming increasingly complex and lengthy, making faithful adherence difficult and imposing large fixed computational costs. With the rise of multimodal agents, policies that govern visual and multimodal behaviors are critical but remain understudied. Prior prompt-compression work mainly shortens task templates and demonstrations, while existing policy-alignment studies focus only on text-based safety rules. We introduce Multimodal Policy Internalization (MPI), a new task that internalizes reasoning-intensive multimodal policies into model parameters, enabling stronger policy-following without including the policy during inference. MPI poses unique data and algorithmic challenges. We build two datasets spanning synthetic and real-world decision-making and tool-using tasks and propose TriMPI, a three-stage training framework. TriMPI first injects policy knowledge via continual pretraining, then performs supervised finetuning, and finally applies PolicyRollout, a GRPO-style reinforcement learning extension that augments rollouts with policy-aware responses for grounded exploration. TriMPI achieves notable gains in end-to-end accuracy, generalization, and robustness to forgetting. As the first work on multimodal policy internalization, we provide datasets, training recipes, and comprehensive evaluations to foster future research. Project page: https://mikewangwzhl.github.io/TriMPI.

Despite the remarkable progress in neural models, their ability to generalize, a cornerstone for applications like logical reasoning, remains a critical challenge. We delineate two fundamental aspects of this ability: compositionality, the capacity to abstract atomic logical rules underlying complex inferences, and recursiveness, the aptitude to build intricate representations through iterative application of inference rules. In the literature, these two aspects are often confounded together under the umbrella term of generalization. To sharpen this distinction, we investigated the logical generalization capabilities of pre-trained large language models (LLMs) using the syllogistic fragment as a benchmark for natural language reasoning. Though simple, this fragment provides a foundational yet expressive subset of formal logic that supports controlled evaluation of essential reasoning abilities. Our findings reveal a significant disparity: while LLMs demonstrate reasonable proficiency in recursiveness, they struggle with compositionality. To overcome these limitations and establish a reliable logical prover, we propose a hybrid architecture integrating symbolic reasoning with neural computation. This synergistic interaction enables robust and efficient inference, neural components accelerate processing, while symbolic reasoning ensures completeness. Our experiments show that high efficiency is preserved even with relatively small neural components. As part of our proposed methodology, this analysis gives a rationale and highlights the potential of hybrid models to effectively address key generalization barriers in neural reasoning systems.

The performance of Large Language Models (LLMs) is determined by their training data. Despite the proliferation of open-weight LLMs, access to LLM training data has remained limited. Even for fully open LLMs, the scale of the data makes it all but inscrutable to the general scientific community, despite potentially containing critical data scraped from the internet. In this paper, we present the full-text indexing pipeline for the Apertus LLM training data. Leveraging Elasticsearch parallel indices and the Alps infrastructure, a state-of-the-art, highly energy-efficient arm64 supercluster, we were able to index 8.6T tokens out of 15.2T used to train the Apertus LLM family, creating both a critical LLM safety tool and effectively an offline, curated, open web search engine. Our contribution is threefold. First, we demonstrate that Elasticsearch can be successfully ported onto next-generation arm64-based infrastructure. Second, we demonstrate that full-text indexing at the scale of modern LLM training datasets and the entire open web is feasible and accessible. Finally, we demonstrate that such indices can be used to ensure previously inaccessible jailbreak-agnostic LLM safety. We hope that our findings will be useful to other teams attempting large-scale data indexing and facilitate the general transition towards greener computation.

Free-text crash narratives recorded in real-world crash databases have been shown to play a significant role in improving traffic safety. However, large-scale analyses remain difficult to implement as there are no documented tools that can batch process the unstructured, non standardized text content written by various authors with diverse experience and attention to detail. In recent years, Transformer-based pre-trained language models (PLMs), such as Bidirectional Encoder Representations from Transformers (BERT) and large language models (LLMs), have demonstrated strong capabilities across various natural language processing tasks. These models can extract explicit facts from crash narratives, but their performance declines on inference-heavy tasks in, for example, Crash Type identification, which can involve nearly 100 categories. Moreover, relying on closed LLMs through external APIs raises privacy concerns for sensitive crash data. Additionally, these black-box tools often underperform due to limited domain knowledge. Motivated by these challenges, we study whether compact open-source PLMs can support reasoning-intensive extraction from crash narratives. We target two challenging objectives: 1) identifying the Manner of Collision for a crash, and 2) Crash Type for each vehicle involved in the crash event from real-world crash narratives. To bridge domain gaps, we apply fine-tuning techniques to inject task-specific knowledge to LLMs with Low-Rank Adaption (LoRA) and BERT. Experiments on the authoritative real-world dataset Crash Investigation Sampling System (CISS) demonstrate that our fine-tuned compact models outperform strong closed LLMs, such as GPT-4o, while requiring only minimal training resources. Further analysis reveals that the fine-tuned PLMs can capture richer narrative details and even correct some mislabeled annotations in the dataset.

This work presents the first large-scale investigation into constructing a fully open bilingual large language model (LLM) for a non-English language, specifically Korean, trained predominantly on synthetic data. We introduce KORMo-10B, a 10.8B-parameter model trained from scratch on a Korean-English corpus in which 68.74% of the Korean portion is synthetic. Through systematic experimentation, we demonstrate that synthetic data, when carefully curated with balanced linguistic coverage and diverse instruction styles, does not cause instability or degradation during large-scale pretraining. Furthermore, the model achieves performance comparable to that of contemporary open-weight multilingual baselines across a wide range of reasoning, knowledge, and instruction-following benchmarks. Our experiments reveal two key findings: (1) synthetic data can reliably sustain long-horizon pretraining without model collapse, and (2) bilingual instruction tuning enables near-native reasoning and discourse coherence in Korean. By fully releasing all components including data, code, training recipes, and logs, this work establishes a transparent framework for developing synthetic data-driven fully open models (FOMs) in low-resource settings and sets a reproducible precedent for future multilingual LLM research.