Language Models (LMs) have shown substantial improvements in handling factual knowledge, yet their capability to consistently represent temporal facts, which are valid only within specific timeframes, remains underexplored. To investigate this, we introduce TimeStress, a novel dataset comprising 521K statements on 2003 of the most popular temporal facts in Wikidata. Each statement contextualizes a fact with correct and incorrect dates across three precisions (Day, Month, Year). This setup allows us to evaluate LMs' ability to discern between correct and incorrect temporal statements based on their probability of being generated. We assess 18 LMs across various architectures using two metrics: the win rate, indicating how often correct dates outperform incorrect ones, and robustness, reflecting consistent performance across all dates. Our findings reveal that while some LMs achieve a win rate exceeding 80\%, robustness remains low, with the best model achieving only 6\%. Furthermore, robust knowledge at one date precision does not reliably transfer to others, highlighting a significant generalization gap. These results underscore the struggle of LMs to maintain a consistent temporal representation, supporting their limitations as reliable sources of temporal knowledge. We provide all data and code for further research.

Large Language Models (LLMs) exhibit impressive reasoning abilities, yet their reliance on structured step-by-step processing reveals a critical limitation. While human cognition fluidly adapts between intuitive, heuristic (System 1) and analytical, deliberative (System 2) reasoning depending on the context, LLMs lack this dynamic flexibility. This rigidity can lead to brittle and unreliable performance when faced with tasks that deviate from their trained patterns. To address this, we create a dataset of 2,000 samples with valid System 1 and System 2 answers, explicitly align LLMs with these reasoning styles, and evaluate their performance across reasoning benchmarks. Our results reveal an accuracy-efficiency trade-off: System 2-aligned models excel in arithmetic and symbolic reasoning, while System 1-aligned models perform better in commonsense tasks. A mechanistic analysis of model responses shows that System 1 models employ more definitive answers, whereas System 2 models demonstrate greater uncertainty. Interpolating between these extremes produces a monotonic transition in reasoning accuracy, preserving coherence. This work challenges the assumption that step-by-step reasoning is always optimal and highlights the need for adapting reasoning strategies based on task demands.

Retrieval-augmented generation (RAG) enhances Large Language Models (LLMs) with relevant and up-to-date knowledge, improving their ability to answer knowledge-intensive questions. It has been shown to enhance both generation quality and trustworthiness. While numerous works have focused on improving retrieval, generation, and evaluation, the role of reward models in reinforcement learning for optimizing RAG remains underexplored. In this paper, we introduce \textbf{RAG-Reward}, a framework designed to develop reward models to enable \textit{hallucination-free, comprehensive, reliable, and efficient RAG}. We define four key metrics to assess generation quality and develop an automated benchmarking pipeline to evaluate the outputs of multiple LLMs across a variety of RAG scenarios. Using \textbf{RAG-Reward}, we train reward models and apply {reinforcement learning with human feedback (RLHF)} to improve LLMs' effectiveness in RAG. Experimental results demonstrate that our reward model achieves state-of-the-art performance in automatic benchmarking and aligns closely with human evaluations. Furthermore, the improved generation quality of the trained policy model highlights the feasibility and efficiency of using RLHF to enhance RAG outputs.

We propose a novel paradigm for automatic fact-checking that leverages frame semantics to enhance the structured understanding of claims and guide the process of fact-checking them. To support this, we introduce a pilot dataset of real-world claims extracted from PolitiFact, specifically annotated for large-scale structured data. This dataset underpins two case studies: the first investigates voting-related claims using the Vote semantic frame, while the second explores various semantic frames based on data sources from the Organisation for Economic Co-operation and Development (OECD). Our findings demonstrate the effectiveness of frame semantics in improving evidence retrieval and explainability for fact-checking. Finally, we conducted a survey of frames evoked in fact-checked claims, identifying high-impact frames to guide future work in this direction.

Various benchmarks have been proposed to assess the performance of large language models (LLMs) in different coding scenarios. We refer to them as code-related benchmarks. However, there are no systematic guidelines by which such a benchmark should be developed to ensure its quality, reliability, and reproducibility. We propose How2Bench, which is comprised of a 55-criteria checklist as a set of guidelines to govern the development of code-related benchmarks comprehensively. Using HOW2BENCH, we profiled 274 benchmarks released within the past decade and found concerning issues. Nearly 70% of the benchmarks did not take measures for data quality assurance; over 10% did not even open source or only partially open source. Many highly cited benchmarks have loopholes, including duplicated samples, incorrect reference codes/tests/prompts, and unremoved sensitive/confidential information. Finally, we conducted a human study involving 49 participants, which revealed significant gaps in awareness of the importance of data quality, reproducibility, and transparency.

Recent studies have raised concerns about the reliability of current mathematical benchmarks, highlighting issues such as simplistic design and potential data contamination. Therefore, creating a reliable benchmark that effectively evaluates the genuine capabilities of large language models (LLMs) in mathematical reasoning remains a significant challenge. To address this, we propose RV-Bench, a framework for Benchmarking LLMs via Random Variables in mathematical reasoning. Specifically, the background content of a random variable question (RV question) mirrors the original problem in existing benchmarks, but the variable combinations are randomized, making it "unseen" by the LLMs. Models must completely understand the question pattern of the original problem to correctly answer RV questions with various variable values. As a result, the LLM's genuine capability in mathematical reasoning is reflected by its accuracy and robustness on RV-Bench. We conducted extensive experiments on over 30 representative LLMs across more than 1000 RV questions. Our findings suggest that LLMs exhibit an imbalance in proficiency between encountered and "unseen" data domains. Proficiency generalization across similar mathematical reasoning tasks is verified to be limited by accuracy and robustness, but it can still be enhanced through test-time scaling.

Detecting organized political campaigns is of paramount importance in fighting against disinformation on social media. Existing approaches for the identification of such organized actions employ techniques mostly from network science, graph machine learning and natural language processing. Their ultimate goal is to analyze the relationships and interactions (e.g. re-posting) among users and the textual similarities of their posts. Despite their effectiveness in recognizing astroturf campaigns, these methods face significant challenges, notably the class imbalance in available training datasets. To mitigate this issue, recent methods usually resort to data augmentation or increasing the number of positive samples, which may not always be feasible or sufficient in real-world settings. Following a different path, in this paper, we propose a novel framework for identifying astroturf campaigns based solely on large language models (LLMs), introducing a Balanced Retrieval-Augmented Generation (Balanced RAG) component. Our approach first gives both textual information concerning the posts (in our case tweets) and the user interactions of the social network as input to a language model. Then, through prompt engineering and the proposed Balanced RAG method, it effectively detects coordinated disinformation campaigns on X (Twitter). The proposed framework does not require any training or fine-tuning of the language model. Instead, by strategically harnessing the strengths of prompt engineering and Balanced RAG, it facilitates LLMs to overcome the effects of class imbalance and effectively identify coordinated political campaigns. The experimental results demonstrate that by incorporating the proposed prompt engineering and Balanced RAG methods, our framework outperforms the traditional graph-based baselines, achieving 2x-3x improvements in terms of precision, recall and F1 scores.

Medical language models (MLMs) have become pivotal in advancing medical natural language processing. However, prior models that rely on pre-training or supervised fine-tuning often exhibit low data efficiency and limited practicality in real-world clinical applications. While OpenAI's o1 highlights test-time scaling in mathematics, attempts to replicate this approach in medicine typically distill responses from GPT-series models to open-source models, focusing primarily on multiple-choice tasks. This strategy, though straightforward, neglects critical concerns like data privacy and realistic deployment in clinical settings. In this work, we present a deployable, small-scale medical reasoning system, MedS3, designed for long-chain reasoning in clinical tasks using a self-evolution paradigm. Starting with a seed dataset of around 8,000 instances spanning five domains and 16 datasets, we prompt a base policy model to perform Monte Carlo Tree Search (MCTS) to construct rule-verifiable reasoning chains. Each reasoning step is assigned an evolution rollout value, allowing verified trajectories to train the policy model and the process reward model (PRM). During inference, the policy model generates multiple responses, and the reward model selects the one with a newly proposed PRM-guided Vote-Sum (P-VS) strategy. Experiments on eleven evaluation datasets demonstrate that MedS3 outperforms not only the prior strongest medical model by 6.59, but also 32B-level general reasoning models by 8.71 points. Code and data are available at https://github.com/pixas/MedSSS.

Large language models (LLMs) generally utilize a consistent data distribution throughout the pretraining process. However, as the model's capability improves, it is intuitive that its data preferences dynamically change, indicating the need for pretraining with different data at various training stages. To achieve it, we propose the Perplexity Difference (PD) based Preference Curriculum learning (PDPC) framework, which always perceives and uses the data preferred by LLMs to train and boost them. First, we introduce the PD metric to quantify the difference in how challenging a sample is for weak versus strong models. Samples with high PD are more challenging for weak models to learn and are more suitable to be arranged in the later stage of pretraining. Second, we propose the preference function to approximate and predict the data preference of the LLM at any training step, so as to complete the arrangement of the dataset offline and ensure continuous training without interruption. Experimental results on 1.3B and 3B models demonstrate that PDPC significantly surpasses baselines. Notably, the 3B model trained on 1T tokens achieves an increased average accuracy of over 8.1% across MMLU and CMMLU.

Large Language Models (LLMs) exhibit strong general language capabilities. However, fine-tuning these models on domain-specific tasks often leads to catastrophic forgetting, where the model overwrites or loses essential knowledge acquired during pretraining. This phenomenon significantly limits the broader applicability of LLMs. To address this challenge, we propose a novel approach to compute the element-wise importance of model parameters crucial for preserving general knowledge during fine-tuning. Our method utilizes a dual-objective optimization strategy: (1) regularization loss based on element-wise parameter importance, which constrains the updates to parameters crucial for general knowledge; (2) cross-entropy loss to adapt to domain-specific tasks. Additionally, we introduce layer-wise coefficients to account for the varying contributions of different layers, dynamically balancing the dual-objective optimization. Extensive experiments on scientific, medical, and physical tasks using GPT-J and LLaMA-3 demonstrate that our approach mitigates catastrophic forgetting while enhancing model adaptability. Compared to previous methods, our solution is approximately 20 times faster and requires only 10-15% of the storage, highlighting the practical efficiency. The code will be released.