High-quality pre-training data is crucial for large language models (LLMs), yet existing score-based selection approaches often fail to balance quality and diversity. Empirical studies have revealed a non-monotonic relationship between dataset scores and downstream performance: simply selecting the top-scored data often leads to performance degradation due to the collapse of correlated evaluation dimensions, which sacrifices diversity. To address this, we propose **O**rthogonal **Di**versity-Aware **S**election (**ODiS**), a novel framework to preserve both quality and diversity by decomposing correlated evaluation metrics into orthogonal feature dimensions. First, ODiS evaluates data from multiple dimensions, covering language quality, knowledge quality, and comprehension difficulty. The resulting multi-dimensional scores are then decorrelated via Principal Component Analysis (PCA), yielding orthogonal evaluation dimensions. For each dimension, a RoBERTa-based scorer is trained to regress the data onto PCA-projected scores, enabling scalable inference on large corpora. Finally, ODiS constructs the training dataset by selecting the top-scored data from each orthogonal dimension, thereby ensuring both quality and diversity. Empirical results show that ODiS-selected data exhibit less than 2% inter-dimension overlap, confirming the orthogonality between dimensions. More importantly, models trained on ODiS-selected data significantly outperform baselines across multiple downstream benchmarks, underscoring the necessity of orthogonal, diversity-aware data selection for LLMs.

Data curation for large language models (LLMs) has been extensively studied for both the pre-training and post-training stages, yet the intermediate mid-training phase—bridging general pretraining and task-specific fine-tuning—remains critically underexplored. Most existing mid-training data for state-of-the-art models remains proprietary. Meanwhile, the Hugging Face platform hosts a huge amount of community-contributed datasets, yet no systematic effort has been made to refine this rich ecosystem into a unified high-quality corpus. We present HuggingBrew, a large-scale mid-training corpus constructed through a principled, multi-step curation pipeline applied to the massive publicly accessible Hugging Face datasets. The pipeline operates at two complementary granularities: dataset-level operations act on each dataset independently, while corpus-level operations are applied jointly across all retained corpus. We construct three progressively curated variants—\textsf{Draught}, \textsf{Craft}, and \textsf{Reserve}—containing 1.6T, 450B, and 40B tokens, respectively. Continual pretraining experiments under a 100B-token budget show that series yields substantial improvements over baselines across mathematical reasoning, general knowledge, and code generation, validating the untapped value of community-contributed data for mid-training.

Causal self-attention provides positional information to Transformer decoders. Prior work has shown that stacks of causal self-attention layers alone induce a positional bias in attention scores toward earlier tokens. However, this differs from the bias toward later tokens typically observed in Transformer decoders, known as recency bias. We address this discrepancy by analyzing the interaction between causal self-attention and other architectural components. We show that stacked causal self-attention layers combined with LayerNorm induce recency bias. Furthermore, we examine the effects of residual connections and the distribution of input token embeddings on this bias. Our results provide new theoretical insights into how positional information interacts with architectural components and suggest directions for improving positional encoding strategies.

Large language models (LLMs) have demonstrated strong reasoning capabilities through step-by-step chain-of-thought (CoT) reasoning. Nevertheless, at the limits of model capability, CoT often proves insufficient, and its strictly sequential nature constrains test-time scalability. A potential alternative is divide-and-conquer (DAC) reasoning, which decomposes a complex problem into subproblems to facilitate more effective exploration of the solution space. Although promising, our analysis reveals a fundamental misalignment between general-purpose post-training and DAC-style inference, which limits the model’s capacity to fully leverage this potential. To bridge this gap and fully unlock LLMs’ reasoning capabilities on the most challenging tasks, we propose an end-to-end reinforcement learning (RL) framework to enhance their DAC-style reasoning capacity. At each step, the policy decomposes a problem into a group of subproblems, solves them sequentially, and addresses the original problem conditioned on the subproblem solutions, with both decomposition and solution integrated into RL training. Under comparable training settings, our DAC-style framework endows the model with a higher performance ceiling and stronger test-time scalability, surpassing CoT by 8.6% in Pass@1 and 6.3% in Pass@32 on competition-level benchmarks. The code is available at the provided link.

The financial industry faces a substantial workload in verifying document images. Existing methods based on visual features struggle to identify fraudulent document images due to the lack of visual clues on the tampering region. This paper proposes CSIAD (Cross-Sample Image Anomaly Detection) by leveraging LLMs to identify logical inconsistencies in similar images. This novel framework accurately detects forged images with slight tampering traces and explains anomaly detection results. Furthermore, we introduce CrossCred, a new benchmark of real-world fraudulent images with fine-grained manual annotations. Experiments demonstrate that CSIAD outperforms state-of-the-art image fraud detection methods by 79.6% (F1) on CrossCred and deployed industrial solutions by 21.7% (F1) on business data. The benchmark is available at https://github.com/XMUDM/CSIAD.

Large Language Models (LLMs) exhibit impressive capabilities but also present risks such as biased content generation and privacy issues. One of the current alignment techniques includes principle-driven integration, but it faces challenges arising from the imprecision of manually crafted rules and inadequate risk perception in models without safety training. To address these, we introduce Guide-Align, a two-stage approach. Initially, a safety-trained model identifies potential risks and formulates specific guidelines for various inputs, establishing a comprehensive library of guidelines and a model for input-guidelines retrieval. Subsequently, the retrieval model correlates new inputs with relevant guidelines, which guide LLMs in response generation to ensure safe and high-quality outputs, thereby aligning with human values. An additional optional stage involves fine-tuning a model with well-aligned datasets generated through the process implemented in the second stage.Our method customizes guidelines to accommodate diverse inputs, thereby enhancing the fine-grainedness and comprehensiveness of the guideline library. Furthermore, it incorporates safety expertise from a safety-trained LLM through a lightweight retrieval model.We evaluate our approach on three benchmarks, demonstrating significant improvements in LLM security and quality. Notably, our fine-tuned model, Labrador, even at 13 billion parameters, outperforms GPT-3.5-turbo and surpasses GPT-4 in alignment capabilities.

Many natural language processing (NLP) tasks rely on labeled data to train machine learning models with high performance. However, data annotation is time-consuming and expensive, especially when the task involves a large amount of data or requires specialized domains. Recently, GPT-3.5 series models have demonstrated remarkable few-shot and zero-shot ability across various NLP tasks. In this paper, we first claim that large language models (LLMs), such as GPT-3.5, can serve as an excellent crowdsourced annotator when provided with sufficient guidance and demonstrated examples. Accordingly, we propose AnnoLLM, an annotation system powered by LLMs, which adopts a two-step approach, explain-then-annotate. Concretely, we first prompt LLMs to provide explanations for why the specific ground truth answer/label was assigned for a given example. Then, we construct the few-shot chain-of-thought prompt with the self-generated explanation and employ it to annotate the unlabeled data with LLMs. Our experiment results on three tasks, including user input and keyword relevance assessment, BoolQ, and WiC, demonstrate that AnnoLLM surpasses or performs on par with crowdsourced annotators. Furthermore, we build the first conversation-based information retrieval dataset employing AnnoLLM. This dataset is designed to facilitate the development of retrieval models capable of retrieving pertinent documents for conversational text. Human evaluation has validated the dataset’s high quality.

Large language models (LLMs) have demonstrated impressive reasoning capabilities, yet there is ongoing debate about these abilities and the potential data contamination problem recently. This paper aims to evaluate the reasoning capacities of LLMs, specifically in solving recent competition-level programming problems in Codeforces, which are expert-crafted and unique, requiring deep understanding and robust reasoning skills. We first provide a comprehensive evaluation of GPT-4’s perceived zero-shot performance on this task, considering various aspects such as problems’ release time, difficulties, and types of errors encountered. Surprisingly, the perceived performance of GPT-4 has experienced a cliff like decline in problems after September 2021 consistently across all the difficulties and types of problems, which shows the potential data contamination, as well as the challenges for any existing LLM to solve unseen complex reasoning problems. We further explore various approaches such as fine-tuning, Chain-of-Thought prompting and problem description simplification. Unfortunately, none of them is able to consistently mitigate the challenges. Through our work, we emphasize the importance of this excellent data source for assessing the genuine reasoning capabilities of LLMs, and foster the development of LLMs with stronger reasoning abilities and better generalization in the future.

Most existing pre-trained language representation models (PLMs) are sub-optimal in sentiment analysis tasks, as they capture the sentiment information from word-level while under-considering sentence-level information. In this paper, we propose SentiWSP, a novel Sentiment-aware pre-trained language model with combined Word-level and Sentence-level Pre-training tasks.The word level pre-training task detects replaced sentiment words, via a generator-discriminator framework, to enhance the PLM’s knowledge about sentiment words.The sentence level pre-training task further strengthens the discriminator via a contrastive learning framework, with similar sentences as negative samples, to encode sentiments in a sentence.Extensive experimental results show that SentiWSP achieves new state-of-the-art performance on various sentence-level and aspect-level sentiment classification benchmarks. We have made our code and model publicly available at https://github.com/XMUDM/SentiWSP.