Training tool-augmented LLMs has emerged as a promising approach to enhancing language models’ capabilities for complex tasks. The current supervised fine-tuning paradigm relies on constructing extensive domain-specific datasets to train models. However, this approach often struggles to generalize effectively to unfamiliar or intricate tool-use scenarios. Recently, reinforcement learning (RL) paradigm can endow LLMs with superior reasoning and generalization abilities. In this work, we address a key question: Can the pure RL be used to effectively elicit a model’s intrinsic reasoning capabilities and enhance the tool-agnostic generalization? We propose a dynamic generalization-guided reward design for rule-based RL, which progressively shifts rewards from exploratory to exploitative tool-use patterns. Based on this design, we introduce the Tool-Zero series models. These models are trained to enable LLMs to autonomously utilize general tools by directly scaling up RL from Zero models (i.e., base models without post-training). Experimental results demonstrate that our models achieve over 7% performance improvement compared to both SFT and RL-with-SFT models under the same experimental settings. These gains are consistently replicated across cross-dataset and intra-dataset evaluations, validating the effectiveness and robustness of our methods.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
The growing capabilities of large language models in natural language understanding significantly strengthen existing agentic systems. To power performant on-device mobile agents for better data privacy, we introduce DroidCall, the first training and testing dataset for accurate Android Intent invocation. With a highly flexible and reusable data generation pipeline, we constructed 10k samples in DroidCall. Given a task instruction in natural language, small language models such as Qwen2.5-3B and Gemma2-2B fine-tuned with DroidCall can approach or even surpass the capabilities of GPT-4o for accurate Android intent invocation. We also provide an end-to-end Android app equipped with these fine-tuned models to demonstrate the Android intent invocation process. The code and dataset are available at https://github.com/UbiquitousLearning/DroidCall
The performance of pre-trained Large Language Models (LLMs) is often sensitive to nuances in prompt templates, requiring careful prompt engineering, adding costs in terms of computing and human effort. In this study, we present experiments encompassing multiple LLMs variants of varying sizes aimed at probing their preference with different prompts. Through experiments on Question Answering, we show prompt preference consistency across LLMs of different sizes. We also show that this consistency extends to other tasks, such as Natural Language Inference. Utilizing this consistency, we propose a method to use a smaller model to select effective prompt templates for a larger model. We show that our method substantially reduces the cost of prompt engineering while consistently matching performance with optimal prompts among candidates. More importantly, our experiment shows the efficacy of our strategy across fourteen LLMs and its applicability to a broad range of NLP tasks, highlighting its robustness.
EventRelBench: A Comprehensive Benchmark for Evaluating Event Relation Understanding in Large Language Models
PDF ↗Understanding event relationships is critical for tasks such as narrative comprehension, information extraction, and reasoning in natural language processing. Despite the remarkable advancements of large language models (LLMs) across diverse NLP tasks, current studies have not systematically evaluated their ability to capture the complex of event relations. To this end, we aim to assess LLMs on event relationship extraction (ERE) by designing the benchmark EventRelBench. EventRelBench comprises 35K diverse event relation questions covering four key categories—coreference, temporal, causal, and supersub relations. These questions are provided at two levels of granularity: document-level and sentence-level. Extensive experiments on different sizes and types of LLMs show that existing LLMs still fall short in accurately extracting and understanding event relationships. To address this gap, we introduce EventRelInst, a 48K instruction fine‐tuning dataset in the event relation extraction domain. Experimental results not only highlight the shortcomings of current general-purpose LLMs in extracting event relationships but also demonstrate the effectiveness of EventRelInst. Both EventRelBench and EventRelBench will be publicly available.
Hard Negatives, Hard Lessons: Revisiting Training Data Quality for Robust Information Retrieval with LLMs
PDF ↗Training robust retrieval and reranker models typically relies on large-scale retrieval datasets; for example, the BGE collection contains 1.6 million query-passage pairs sourced from various data sources. However, we find that certain datasets can negatively impact model effectiveness — pruning 8 out of 15 datasets from the BGE collection, reduces the training set size by 2.35×, surprisingly increases nDCG@10 on BEIR by 1.0 point. This motivates a deeper examination of training data quality, with a particular focus on “false negatives”, where relevant passages are incorrectly labeled as irrelevant. We utilize LLMs as a simple, cost-effective approach to identify and relabel false negatives in training datasets. Experimental results show that relabeling false negatives as true positives improves both E5 (base) and Qwen2.5-7B retrieval models by 0.7-1.4 points on BEIR and by 1.7-1.8 points at nDCG@10 on zero-shot AIR-Bench evaluation. Similar gains are observed for rerankers fine-tuned on the relabeled data, such as Qwen2.5-3B on BEIR. The reliability of LLMs to identify false negatives is supported by human annotation results. Our training dataset and code are publicly available.
Large Language Models (LLMs) have demonstrated significant potential in specialized tasks, including medical problem-solving. However, most studies predominantly focus on English-language contexts. This study introduces a novel benchmark dataset based on Polish medical licensing and specialization exams (LEK, LDEK, PES). The dataset, sourced from publicly available materials provided by the Medical Examination Center and the Chief Medical Chamber, includes Polish medical exam questions, along with a subset of parallel Polish-English corpora professionally translated for foreign candidates. By structuring a benchmark from these exam questions, we evaluate state-of-the-art LLMs, spanning general-purpose, domain-specific, and Polish-specific models, and compare their performance with that of human medical students and doctors. Our analysis shows that while models like GPT-4o achieve near-human performance, challenges persist in cross-lingual translation and domain-specific understanding. These findings highlight disparities in model performance across languages and medical specialties, emphasizing the limitations and ethical considerations of deploying LLMs in clinical practice.
Well-designed prompts are crucial for enhancing Large language models’ (LLMs) reasoning capabilities while aligning their outputs with task requirements across diverse domains. However, manually designed prompts require expertise and iterative experimentation. While existing prompt optimization methods aim to automate this process, they rely heavily on external references such as ground truth or by humans, limiting their applicability in real-world scenarios where such data is unavailable or costly to obtain. To address this, we propose Self-Supervised Prompt Optimization (SPO), a cost-efficient framework that discovers effective prompts for both closed and open-ended tasks without requiring external reference. Motivated by the observations that prompt quality manifests directly in LLM outputs and LLMs can effectively assess adherence to task requirements, we derive evaluation and optimization signals purely from output comparisons. Specifically, SPO selects superior prompts through pairwise output comparisons evaluated by an LLM evaluator, followed by an LLM optimizer that aligns outputs with task requirements. Extensive experiments demonstrate that SPO outperforms state-of-the-art prompt optimization methods, achieving comparable or superior results with significantly lower costs (e.g., 1.1% to 5.6% of existing methods) and fewer samples (e.g., three samples).
Unveiling Multimodal Processing: Exploring Activation Patterns in Multimodal LLMs for Interpretability and Efficiency
PDF ↗Recent Multimodal Large Language Models (MLLMs) have achieved remarkable advancements, yet their internal mechanisms for concurrently processing diverse modalities like text, image, and audio remain largely opaque. In this paper, we propose a methodology to convert dense MLLMs into fine-grained Mixture-of-Experts (MoE) architectures. This allows us to visually investigate their multimodal activation patterns through expert activation frequency heatmaps. Conducting comprehensive experiments on representative MLLMs, we analyze the similarities and differences in internal neuron activations when handling distinct modalities. Specifically, we examine the distribution of high-frequency activated experts, the distinct roles of high-frequency (e.g., fundamental logic) and low-frequency (e.g., domain-specific concepts) multimodal shared experts, and the prevalence and localization of modality-specific experts. Furthermore, we explore leveraging these discovered activation discrepancies to guide sparse activation and model pruning. Experimental results demonstrate that our approach substantially outperforms random expert pruning and can achieve comparable or even superior performance to the original unpruned models while utilizing significantly fewer active parameters. Our work not only sheds light on the multimodal processing mechanisms within MLLMs but also provides a practical pathway toward developing more interpretable and efficient multimodal systems.
Research ideation is crucial for scientific progress, but the exponential increase in scientific literature makes it challenging to stay updated and identify impactful directions. Recent developments in large language models(LLMs) offer a promising avenue to automate this process. However, existing methods for idea generation either trivially prompt LLMs or expose LLMs to extensive literature without indicating useful information. Inspired by human research processes, we propose a Chain-of-Ideas (CoI) agent, an LLM-based agent that organizes relevant literature in a chain structure to effectively mirror the progressive development in a research domain. This organization helps LLMs better grasp current advancements, thereby improving ideation capabilities. Further, we present Idea Arena, a protocol for evaluating idea-generation methods from different perspectives, which aligns closely with the preferences of human researchers. Experiments show that CoI agent consistently outperforms existing methods and matches human quality in idea generation. Moreover, CoI agent is budget-friendly, requiring only $0.50 to generate a candidate idea and its experimental design.
NAP2: A Benchmark for Naturalness and Privacy-Preserving Text Rewriting by Learning from Human
PDF ↗The widespread use of cloud-based Large Language Models (LLMs) has heightened concerns over user privacy, as sensitive information may be inadvertently exposed during interactions with these services. To protect privacy before sending sensitive data to those models, we suggest sanitizing sensitive text using two common strategies used by humans: i) deleting sensitive expressions, and ii) obscuring sensitive details by abstracting them. To explore the issues and develop a tool for text rewriting, we curate the first corpus, coined , through both crowdsourcing and the use of large language models (LLMs). Compared to the prior works based on differential privacy, which lead to a sharp drop in information utility and unnatural texts, the human-inspired approaches result in more natural rewrites and offer an improved balance between privacy protection and data utility, as demonstrated by our extensive experiments.
Most of the existing work focuses on enabling LLMs to leverage legal rules (, law articles) to tackle complex legal reasoning tasks, but ignores their ability to understand legal rules. To better evaluate the LLMs’ capabilities on the task, in this work, we propose a new challenge task: Legal Paragraph Prediction (LPP), which aims to predict the legal paragraph given criminal facts. Moreover, to enhance the legal reasoning ability of LLMs, we propose a novel framework CLEAR, enabling LLMs to analyze legal cases with the guidance of legal rule insights. The CLEAR contains four key components, where the Legal Rules Retriever aims to retrieve legal rule knowledge, and the Rule Insights Generator is used to generate legal insights guiding the LLM’s reasoning, then the Case Analyzer analyze the case with the guidance of legal rule insights given criminal facts. Finally, the Legal Reasoner synthesizes the criminal facts, legal rule insights, and analysis results to derive the final decision. By conducting extensive experiments on a real-world dataset, experimental results validate the effectiveness of our proposed model. Our codes and dataset are available at https://anonymous.4open.science/r/CLEAR-3048.
While reasoning and multilingual capabilities in Language Models (LMs) have achieved remarkable progress in recent years, their integration into a unified paradigm—multilingual reasoning—is at a nascent stage. Multilingual reasoning requires language models to handle logical reasoning across languages while addressing misalignment, biases, and challenges in low-resource settings. This survey provides the first in-depth review of multilingual reasoning in LMs. In this survey, we provide a systematic overview of existing methods that leverage LMs for multilingual reasoning, specifically outlining the challenges, motivations, and foundational aspects of applying language models to reason across diverse languages. We provide an overview of the standard data resources used for training multilingual reasoning in LMs and the evaluation benchmarks employed to assess their multilingual capabilities. Next, we analyze various state-of-the-art methods and their performance on these benchmarks. Finally, we explore future research opportunities to improve multilingual reasoning in LMs, focusing on enhancing their ability to handle diverse languages and complex reasoning tasks.
Preference optimization methods like DPO have achieved remarkable performance in LLM alignment. However, the evaluation for these methods relies on a single response and overlooks other potential outputs, which could also be generated in real-world applications within this hypothetical space. To address this issue, this paper presents a Hypothesis-based PrEference-aware AnaLysis Framework (HEAL), a novel evaluation paradigm that formulates preference alignment as a re-ranking process within hypothesis spaces. The framework incorporates two complementary metrics: ranking accuracy for evaluating ordinal consistency and preference strength correlation for assessing continuous alignment. To facilitate this framework, we develop UniHypoBench, a unified hypothesis benchmark constructed from diverse instruction-response pairs. Through extensive experiments based on HEAL, with a particular focus on the intrinsic mechanisms of preference learning, we demonstrate that current preference learning methods can effectively capture preferences provided by proxy models while simultaneously suppressing negative samples. These findings contribute to preference learning research through two significant avenues. Theoretically, we introduce hypothesis space analysis as an innovative paradigm for understanding preference alignment. Practically, HEAL offers researchers robust diagnostic tools for refining preference optimization methods, while our empirical results identify promising directions for developing more advanced alignment algorithms capable of comprehensive preference capture.
Retrieval-Augmented Generation (RAG) systems traditionally treat retrieval and generation as separate processes, requiring explicit textual queries to connect them. This separation can limit the ability of models to generalize across diverse tasks. In this work, we propose a query-free RAG system, named ImpRAG, which integrates retrieval and generation into a unified model. ImpRAG allows models to implicitly express their information needs, eliminating the need for human-specified queries. By dividing pretrained decoder-only language models into specialized layer groups, ImpRAG optimizes retrieval and generation tasks simultaneously. Our approach employs a two-stage inference process, using the same model parameters and forward pass for both retrieval and generation, thereby minimizing the disparity between retrievers and language models. Experiments on 8 knowledge-intensive tasks demonstrate that ImpRAG achieves 3.6-11.5 improvements in exact match scores on unseen tasks with diverse formats, highlighting its effectiveness in enabling models to articulate their own information needs and generalize across tasks. Our analysis underscores the importance of balancing retrieval and generation parameters and leveraging generation perplexities as retrieval training objectives for enhanced performance.
TR-MTEB: A Comprehensive Benchmark and Embedding Model Suite for Turkish Sentence Representations
PDF ↗We introduce TR-MTEB, the first large-scale, task-diverse benchmark designed to evaluate sentence embedding models for Turkish. Covering six core tasks as classification, clustering, pair classification, retrieval, bitext mining, and semantic textual similarity, TR-MTEB incorporates 26 high-quality datasets, including native and translated resources. To complement this benchmark, we construct a corpus of 34.2 million weakly supervised Turkish sentence pairs and train two Turkish-specific embedding models using contrastive pretraining and supervised fine-tuning. Evaluation results show that our models, despite being trained on limited resources, achieve competitive performance across most tasks and significantly improve upon baseline monolingual models. All datasets, models, and evaluation pipelines are publicly released to facilitate further research in Turkish natural language processing and low-resource benchmarking.
FastCuRL: Curriculum Reinforcement Learning with Stage-wise Context Scaling for Efficient Training R1-like Reasoning Models
PDF ↗Improving training efficiency continues to be one of the primary challenges in large-scale Reinforcement Learning (RL). In this paper, we investigate how context length and the complexity of training data influence the RL scaling training process of R1-distilled reasoning models, e.g., DeepSeek-R1-Distill-Qwen-1.5B.Our experimental results reveal that: text-green(1) simply controlling the context length and selecting the training data based on the input prompt length can effectively improve the training efficiency of RL scaling, achieving better performance with more concise CoT; text-blue(2) properly scaling the context length helps mitigate entropy collapse; text-redand (3) carefully choosing the context length facilitates achieving efficient LLM training and reasoning. Inspired by these insights, we propose FastCuRL, a curriculum RL framework with stage-wise context scaling to achieve efficient LLM training and reasoning. Extensive experimental results demonstrate that FastCuRL-1.5B-V3 significantly outperforms state-of-the-art reasoning models on five competition-level benchmarks and achieves 49.6% accuracy on AIME 2024. Furthermore, FastCuRL-1.5B-Preview surpasses DeepScaleR-1.5B-Preview on five benchmarks while only using a single node with 8 GPUs and a total of 50% of training steps.
As Large language models (LLMs) are increasingly deployed in diverse applications, faithfully integrating evolving factual knowledge into these models remains a critical challenge. Continued pre-training on paraphrased data has shown empirical promise for enhancing knowledge acquisition. However, this approach is often costly and unreliable, as it relies on external models or manual effort for rewriting, and may inadvertently alter the factual content. In this work, we hypothesize and empirically show that an LLM’s ability to continually predict the same factual knowledge tokens given diverse paraphrased contexts is positively correlated with its capacity to extract that knowledge via question-answering. Based on this view and aiming to improve generalization to diverse paraphrased contexts, we introduce two strategies to enhance LLMs’ ability to predict the same knowledge tokens given varied contexts, thereby enhancing knowledge acquisition. First, we propose formatting-based data augmentation, which diversifies documents conveying the same knowledge by altering document formats rather than their content, thereby preserving factual integrity. Second, we adopt sharpness-aware minimization as the optimizer to better improve generalization. Extensive experiments demonstrate our methods’ effectiveness in both continued pre-training and instruction tuning, and further gains can be achieved by combining with paraphrased data. Code and data are available at https://github.com/dvlab-research/llm-knowledge-generalization.
DRBO: Mitigating the Bottleneck Effect via Dynamic Reward Balancing in Multi-reward LLM Optimization
PDF ↗In the current landscape of large language models (LLMs), many evaluation metrics have been developed and used as rewards during training to improve specific metrics. However, balancing these metrics and dynamically adjusting reward weights remains challenging, as current approaches often fail to enhance weaker metrics. To address this, we empirically propose a Dynamic Reward Balancing Optimization framework DRBO to mitigate the “bottleneck effect” by measuring performance, adjusting reward weights to prioritize weaker metrics, and optimizing the model via reinforcement learning. We apply DRBO to both single-task and multi-type task scenarios, validating its effectiveness in generation with citations and online shopping conversation tasks. The results demonstrate improved overall performance and balanced optimization across multiple metrics, effectively overcoming the diversity and complexity inherent in LLMs. Our codes are available at https://github.com/NuoJohnChen/DRBO.
Large language models (LLMs) have demonstrated exceptional performance across diverse natural language processing tasks. However, as the model size and the input sequence’s length increase, the linearly increasing key-value (KV) cache significantly degrades inference throughput. Therefore, grouped-query attention (GQA), as an alternative to multi-head attention (MHA), has been widely introduced into LLMs. In this work, we propose a cost-effective method for converting MHA into GQA with any compression ratio of KV heads. The key point of our method lies in the application of Procrustes analysis to the attention heads, which enhances the similarity among attention heads while preserving computational invariance, thereby improving the model’s post-training performance. Subsequently, we employ \mathit{L_0} regularization to prune redundant parameters. The model after pruning can be adapted to the standard GQA framework. Experimental results show that our strategy can compress up to 87.5% KV heads of LLaMA2-7B model and 75% KV heads of Sheared-LLaMA-1.3B with acceptable performance degradation. Our code is released at https://github.com/fpcsong/mha2gqa.
MAGIC: A Multi-Hop and Graph-Based Benchmark for Inter-Context Conflicts in Retrieval-Augmented Generation
PDF ↗Knowledge conflict often arises in retrieval-augmented generation (RAG) systems, where retrieved documents may be inconsistent with one another or contradict the model’s parametric knowledge.Existing benchmarks for investigating the phenomenon have notable limitations, including a narrow focus on the question answering setup, heavy reliance on entity substitution techniques, and a restricted range of conflict types. To address these issues, we propose a knowledge graph (KG)-based framework that generates varied and subtle conflicts between two similar yet distinct contexts, while ensuring interpretability through the explicit relational structure of KGs.Experimental results on our benchmark, MAGIC, provide intriguing insights into the inner workings of LLMs regarding knowledge conflict: both open-source and proprietary models struggle with conflict detection—especially when multi-hop reasoning is required—and often fail to pinpoint the exact source of contradictions.Finally, we present in-depth analyses that serve as a foundation for improving LLMs in integrating diverse, sometimes even conflicting, information.