TextArena is an open-source collection of competitive text-based games for training and evaluation of agentic behavior in Large Language Models (LLMs). It spans 57+ unique environments (including single-player, two-player, and multi-player setups) and allows for easy evaluation of model capabilities via an online-play system (against humans and other submitted models) with real-time TrueSkill scores. Traditional benchmarks rarely assess dynamic social skills such as negotiation, theory of mind, and deception, creating a gap that TextArena addresses. Designed with research, community and extensibility in mind, TextArena emphasizes ease of adding new games, adapting the framework, testing models, playing against the models, and training models. Detailed documentation of environments, games, leaderboard, and examples are available on https://github.com/LeonGuertler/TextArena and https://www.textarena.ai/.

Task arithmetic refers to editing the pre-trained model by adding a weighted sum of task vectors, each of which is the weight update from the pre-trained model to fine-tuned models for certain tasks. This approach recently gained attention as a computationally efficient inference method for model editing, e.g., multi-task learning, forgetting, and out-of-domain generalization capabilities. However, the theoretical understanding of why task vectors can execute various conceptual operations remains limited, due to the highly non-convexity of training Transformer-based models. To the best of our knowledge, this paper provides the first theoretical characterization of the generalization guarantees of task vector methods on nonlinear Transformers. We consider a conceptual learning setting, where each task is a binary classification problem based on a discriminative pattern. We theoretically prove the effectiveness of task addition in simultaneously learning a set of irrelevant or aligned tasks, as well as the success of task negation in unlearning one task from irrelevant or contradictory tasks. Moreover, we prove the proper selection of linear coefficients for task arithmetic to achieve guaranteed generalization to out-of-domain tasks. All of our theoretical results hold for both dense-weight parameters and their low-rank approximations. Although established in a conceptual setting, our theoretical findings were validated on a practical machine unlearning task using the large language model Phi-1.5 (1.3B).

Large language models (LLMs) have demonstrated impressive capabilities and are receiving increasing attention to enhance their reasoning through scaling test--time compute. However, their application in open--ended, knowledge--intensive, complex reasoning scenarios is still limited. Reasoning--oriented methods struggle to generalize to open--ended scenarios due to implicit assumptions of complete world knowledge. Meanwhile, knowledge--augmented reasoning (KAR) methods fail to address two core challenges: 1) error propagation, where errors in early steps cascade through the chain, and 2) verification bottleneck, where the explore--exploit tradeoff arises in multi--branch decision processes. To overcome these limitations, we introduce ARise, a novel framework that integrates risk assessment of intermediate reasoning states with dynamic retrieval--augmented generation (RAG) within a Monte Carlo tree search paradigm. This approach enables effective construction and optimization of reasoning plans across multiple maintained hypothesis branches. Experimental results show that ARise significantly outperforms the state--of--the--art KAR methods by up to 23.10%, and the latest RAG-equipped large reasoning models by up to 25.37%. Our project page is at https://opencausalab.github.io/ARise.

Automated fact-checking (AFC) still falters on claims that are time-sensitive, entity-ambiguous, or buried beneath noisy search-engine results. We present PASS-FC, a Progressive and Adaptive Search Scheme for Fact Checking. Each atomic claim is first grounded with a precise time span and disambiguated entity descriptors. An adaptive search loop then issues structured queries, filters domains through credible-source selection, and expands queries cross-lingually; when necessary, a lightweight reflection routine restarts the loop. Experiments on six benchmark--covering general knowledge, scientific literature, real-world events, and ten languages--show that PASS-FC consistently outperforms prior systems, even those powered by larger backbone LLMs. On the multilingual X-FACT set, performance of different languages partially correlates with typological closeness to English, and forcing the model to reason in low-resource languages degrades accuracy. Ablations highlight the importance of temporal grounding and the adaptive search scheme, while detailed analysis shows that cross-lingual retrieval contributes genuinely new evidence. Code and full results will be released to facilitate further research.

Large Language Models (LLMs) have become indispensable across academia, industry, and daily applications, yet current evaluation methods struggle to keep pace with their rapid development. One core challenge of evaluation in the large language model (LLM) era is the generalization issue: how to infer a model's near-unbounded abilities from inevitably bounded benchmarks. We address this challenge by proposing Model Utilization Index (MUI), a mechanism interpretability enhanced metric that complements traditional performance scores. MUI quantifies the effort a model expends on a task, defined as the proportion of activated neurons or features during inference. Intuitively, a truly capable model should achieve higher performance with lower effort. Extensive experiments across popular LLMs reveal a consistent inverse logarithmic relationship between MUI and performance, which we formulate as the Utility Law. From this law we derive four practical corollaries that (i) guide training diagnostics, (ii) expose data contamination issue, (iii) enable fairer model comparisons, and (iv) design model-specific dataset diversity. Our code can be found at https://github.com/ALEX-nlp/MUI-Eva.

The integration of tool learning with Large Language Models (LLMs) has expanded their capabilities in handling complex tasks by leveraging external tools. However, existing benchmarks for tool learning inadequately address critical real-world personalized scenarios, particularly those requiring multi-hop reasoning and inductive knowledge adaptation in dynamic environments. To bridge this gap, we introduce FamilyTool, a novel benchmark grounded in a family-based knowledge graph (KG) that simulates personalized, multi-hop tool use scenarios. FamilyTool, including base and extended datasets, challenges LLMs with queries spanning from 1 to 4 relational hops (e.g., inferring familial connections and preferences) and 2 to 6 hops respectively, and incorporates an inductive KG setting where models must adapt to unseen user preferences and relationships without re-training, a common limitation in prior approaches that compromises generalization. We further propose KGETool: a simple KG-augmented evaluation pipeline to systematically assess LLMs' tool use ability in these settings. Experiments reveal significant performance gaps in state-of-the-art LLMs, with accuracy dropping sharply as hop complexity increases and inductive scenarios exposing severe generalization deficits. These findings underscore the limitations of current LLMs in handling personalized, evolving real-world contexts and highlight the urgent need for advancements in tool-learning frameworks. FamilyTool serves as a critical resource for evaluating and advancing LLM agents' reasoning, adaptability, and scalability in complex, dynamic environments. Code and dataset are available at \href{https://github.com/yxzwang/FamilyTool}{https://github.com/yxzwang/FamilyTool}.

The rapid growth in the parameter size of Large Language Models (LLMs) has spurred the development of Parameter-Efficient Fine-Tuning (PEFT) methods to mitigate the substantial computational costs of fine-tuning. Among these, Fisher Induced Sparse uncHanging (FISH) Mask is a selection-based PEFT technique that identifies a critical subset of pre-trained parameters using approximate Fisher information. While addition-based and reparameterization-based PEFT methods like LoRA and Adapter already fine-tune only a small number of parameters, the newly introduced parameters within these methods themselves present an opportunity for further optimization. Selectively fine-tuning only the most impactful among these new parameters could further reduce resource consumption while maintaining, or even improving, fine-tuning effectiveness. In this paper, we propose \textbf{FISH-Tuning}, a novel approach that incorporates FISH Mask into such PEFT methods, including LoRA, Adapter, and their variants. By leveraging Fisher information to identify and update only the most significant parameters within these added or reparameterized components, FISH-Tuning aims to achieve superior performance without increasing training time or inference latency compared to the vanilla PEFT methods. Experimental results across various datasets and pre-trained models demonstrate that FISH-Tuning consistently outperforms the vanilla PEFT methods when using the same proportion of trainable parameters. Code is available at https://anonymous.4open.science/r/FISH-Tuning-6F7C.

The recent DeepSeek-R1 has showcased the emergence of reasoning capabilities in LLMs through reinforcement learning (RL) with rule-based rewards. Despite its success in language models, its application in multi-modal domains, particularly in graphic user interface (GUI) agent tasks, remains under-explored. To address this issue, we propose UI-R1, the first framework to explore how rule-based RL can enhance the reasoning capabilities of multimodal large language models (MLLMs) for GUI action prediction tasks. Specifically, UI-R1 introduces a novel rule-based action reward, enabling model optimization via policy-based algorithms such as Group Relative Policy Optimization (GRPO). For efficient training, we curate a small yet high-quality dataset of 136 challenging tasks, encompassing five common action types on mobile devices. Experimental results demonstrate that our proposed UI-R1-3B achieves significant improvements over the base model (i.e. Qwen2.5-VL-3B) on both in-domain (ID) and out-of-domain (OOD) tasks, with average accuracy gains of 22.1% on ScreenSpot, 6.0% on ScreenSpot-Pro, and 12.7% on ANDROIDCONTROL. Furthermore, UI-R1-3B delivers competitive performance compared to larger models (e.g., OS-Atlas-7B) trained via supervised fine-tuning (SFT) on 76K samples. We additionally develop an optimized version, UI-R1-E-3B, which significantly improves both grounding efficiency and accuracy. These results underscore the potential of rule-based reinforcement learning to advance GUI understanding and control, paving the way for future research in this domain. Code website: https://github.com/lll6gg/UI-R1.

Malicious URL detection and webpage classification are critical tasks in cybersecurity and information management. In recent years, extensive research has explored using BERT or similar language models to replace traditional machine learning methods for detecting malicious URLs and classifying webpages. While previous studies show promising results, they often apply existing language models to these tasks without accounting for the inherent differences in domain data (e.g., URLs being loosely structured and semantically sparse compared to text), leaving room for performance improvement. Furthermore, current approaches focus on single tasks and have not been tested in multi-task scenarios. To address these challenges, we propose urlBERT, a pre-trained URL encoder leveraging Transformer to encode foundational knowledge from billions of unlabeled URLs. To achieve it, we propose to use 5 unsupervised pretraining tasks to capture multi-level information of URL lexical, syntax, and semantics, and generate contrastive and adversarial representations. Furthermore, to avoid inter-pre-training competition and interference, we proposed a grouped sequential learning method to ensure effective training across multi-tasks. Finally, we leverage a two-stage fine-tuning approach to improve the training stability and efficiency of the task model. To assess the multitasking potential of urlBERT, we fine-tune the task model in both single-task and multi-task modes. The former creates a classification model for a single task, while the latter builds a classification model capable of handling multiple tasks. We evaluate urlBERT on three downstream tasks: phishing URL detection, advertising URL detection, and webpage classification. The results demonstrate that urlBERT outperforms standard pre-trained models, and its multi-task mode is capable of addressing the real-world demands of multitasking.

The ability of a model to learn a task depends strongly on both the task difficulty and the model size. We aim to understand how task difficulty relates to the minimum number of parameters required for learning specific tasks in small transformer models. Our study focuses on the ListOps dataset, which consists of nested mathematical operations. We gradually increase task difficulty by introducing new operations or combinations of operations into the training data. We observe that sum modulo n is the hardest to learn. Curiously, when combined with other operations such as maximum and median, the sum operation becomes easier to learn and requires fewer parameters. We show that joint training not only improves performance but also leads to qualitatively different model behavior. We show evidence that models trained only on SUM might be memorizing and fail to capture the number structure in the embeddings. In contrast, models trained on a mixture of SUM and other operations exhibit number-like representations in the embedding space, and a strong ability to distinguish parity. Furthermore, the SUM-only model relies more heavily on its feedforward layers, while the jointly trained model activates the attention mechanism more. Finally, we show that learning pure SUM can be induced in models below the learning threshold of pure SUM, by pretraining them on MAX+MED. Our findings indicate that emergent abilities in language models depend not only on model size, but also the training curriculum.