Language model agents excel in long-session planning and reasoning, but existing benchmarks primarily focus on goal-oriented tasks with explicit objectives, neglecting creative adaptation in unfamiliar environments. To address this, we introduce EscapeBench—a benchmark suite of room escape game environments designed to challenge agents with creative reasoning, unconventional tool use, and iterative problem-solving to uncover implicit goals. Our results show that current LM models, despite employing working memory and Chain-of-Thought reasoning, achieve only 15% average progress without hints, highlighting their limitations in creativity. To bridge this gap, we propose EscapeAgent, a framework designed to enhance creative reasoning through Foresight (innovative tool use) and Reflection (identifying unsolved tasks). Experiments show that EscapeAgent can execute action chains over 1,000 steps while maintaining logical coherence. It navigates and completes games with up to 40% fewer steps and hints, performs robustly across difficulty levels, and achieves higher action success rates with more efficient and innovative puzzle-solving strategies.

Large Language Models (LLMs) have shown remarkable capabilities as autonomous agents; yet existing benchmarks either focus on single-agent tasks or are confined to narrow domains, failing to capture the dynamics of multi-agent coordination and competition. In this paper, we introduce MultiAgentBench, a comprehensive benchmark designed to evaluate LLM-based multi-agent systems across diverse, interactive scenarios. Our framework measures not only task completion but also the quality of collaboration and competition using novel, milestone-based key performance indicators. Moreover, we evaluate various coordination protocols (including star, chain, tree, and graph topologies) and innovative strategies such as group discussion and cognitive planning. Notably, cognitive planning improves milestone achievement rates by 3%. Code and dataset will be made publicly available. Code and datasets are publicavailable at https://github.com/ulab-uiuc/MARBLE

Large Language Models (LLMs) excel in traditional natural language processing tasks but struggle with problems that require complex domain-specific calculations or simulations. While equipping LLMs with external tools to build LLM-based agents can enhance their capabilities, existing approaches lack the flexibility to address diverse and ever-evolving user queries in open domains. Currently, there is also no existing dataset that evaluates LLMs on open-domain knowledge that requires tools to solve. To this end, we introduce OpenAct benchmark to evaluate the open-domain task-solving capability, which is built on human expert consultation and repositories in GitHub. It comprises 339 questions spanning 7 diverse domains that need to be solved with domain-specific methods. In our experiments, even state-of-the-art LLMs and LLM-based agents demonstrate unsatisfactory success rates, underscoring the need for a novel approach.Furthermore, we present OpenAgent, a novel LLM-based agent system that can tackle evolving queries in open domains through autonomously integrating specialized tools from GitHub. OpenAgent employs 1) a hierarchical framework where specialized agents handle specific tasks and can assign tasks to inferior agents, 2) a bi-level experience learning mechanism to learn from both humans’ and its own experiences to tackle tool flaws. Experiments demonstrate its superior effectiveness and efficiency, which significantly outperforms baselines. Our data and code are open-source at https://github.com/OpenBMB/OpenAct.

As large language models (LLMs) demonstrate increasingly advanced capabilities, aligning their behaviors with human values and preferences becomes crucial for their wide adoption. While previous research focuses on general alignment to principles such as helpfulness, harmlessness, and honesty, the need to account for individual and diverse preferences has been largely overlooked, potentially undermining customized human experiences. To address this gap, we train LLMs that can “interact to align”, essentially cultivating the meta-skill of LLMs to implicitly infer the unspoken personalized preferences of the current user through multi-turn conversations, and then dynamically align their following behaviors and responses to these inferred preferences. Our approach involves establishing a diverse pool of 3,310 distinct user personas by initially creating seed examples, which are then expanded through iterative self-generation and filtering. Guided by distinct user personas, we leverage multi-LLM collaboration to develop a multi-turn preference dataset containing 3K+ multi-turn conversations in tree structures. Finally, we apply supervised fine-tuning and reinforcement learning to enhance LLMs using this dataset. For evaluation, we establish the ALOE (ALign with custOmized prEferences) benchmark, consisting of 100 carefully selected examples and well-designed metrics to measure the customized alignment performance during conversations. Experimental results demonstrate the effectiveness of our method in enabling dynamic, personalized alignment via interaction. The code and dataset will be made public.

Incorporating external context can significantly enhance the response quality of Large Language Models (LLMs). However, real-world contexts often mix relevant information with disproportionate inappropriate content, posing reliability risks. How do LLMs process and prioritize mixed context? To study this, we introduce the Poisoned Context Testbed, pairing queries with real-world contexts containing relevant and inappropriate content. Inspired by associative learning in animals, we adapt the Rescorla-Wagner (RW) model from neuroscience to quantify how competing contextual signals influence LLM outputs. Our adapted model reveals a consistent behavioral pattern: LLMs exhibit a strong tendency to incorporate information that is less prevalent in the context. This susceptibility is harmful in real-world settings, where small amounts of inappropriate content can substantially degrade response quality. Empirical evaluations on our testbed further confirm this vulnerability. To tackle this, we introduce RW-Steering, a two-stage finetuning-based approach that enables the model to internally identify and ignore inappropriate signals. Unlike prior methods that rely on extensive supervision across diverse context mixtures, RW-Steering generalizes robustly across varying proportions of inappropriate content. Experiments show that our best fine-tuned model improves response quality by 39.8% and reverses the undesirable behavior curve, establishing RW-Steering as a robust, generalizable solution for improving LLM safety in real-world use.

Hallucination is a persistent challenge in large language models (LLMs), where even with rigorous quality control, models often generate distorted facts. This paradox, in which error generation continues despite high-quality training data, calls for a deeper understanding of the underlying LLM mechanisms. To address it, we propose a novel concept: knowledge overshadowing, where model’s dominant knowledge can obscure less prominent knowledge during text generation, causing the model to fabricate inaccurate details. Building on this idea, we introduce a novel framework to quantify factual hallucinations by modeling knowledge overshadowing. Central to our approach is the log-linear law, which predicts that the rate of factual hallucination increases linearly with the logarithmic scale of (1) Knowledge Popularity, (2) Knowledge Length, and (3) Model Size. The law provides a means to preemptively quantify hallucinations, offering foresight into their occurrence even before model training or inference. Built on the overshadowing effect, we propose a new decoding strategy CoDA, to mitigate hallucinations, which notably enhances model factuality on Overshadow (27.9%), MemoTrap (13.1%) and NQ-Swap (18.3%). Our findings not only deepen understandings of the underlying mechanisms behind hallucinations but also provide actionable insights for developing more predictable and controllable language models.

Understanding alignment techniques begins with comprehending zero-shot generalization brought by instruction tuning, but little of the mechanism has been understood. Existing work has largely been confined to the task level, without considering that tasks are artificially defined and, to LLMs, merely consist of tokens and representations. To bridge this gap, we investigate zero-shot generalization from the perspective of the data itself. We first demonstrate that zero-shot generalization happens very early during instruction tuning, with loss serving as a stable indicator. Next, we investigate training data arrangement through similarity and granularity perspectives, confirming that the timing of exposure to certain training examples may greatly facilitate generalization on unseen tasks. Finally, we propose a more grounded training data arrangement framework, Test-centric Multi-turn Arrangement, and show its effectiveness in promoting continual learning and further loss reduction. For the first time, we show that zero-shot generalization during instruction tuning is a form of similarity-based generalization between training and test data at the instance level.

Positional bias in large language models hinders their ability to effectively process long inputs. A prominent example is the “lost in the middle” phenomenon, where LLMs struggle to utilize relevant information situated in the middle of the input. While prior research primarily focuses on single pieces of relevant information, real-world applications often involve multiple relevant information pieces. To bridge this gap, we present LongPiBench, a benchmark designed to assess positional bias involving multiple pieces of relevant information. It includes various tasks and input lengths. Thorough experiments are conducted with three commercial and six open-source models. These experiments reveal that while most current models are more robust against the “lost in the middle” issue, there also exist noticeable biases related to the spacing of relevant information pieces. These findings highlight the importance of evaluating and reducing positional biases for long-context LLMs.

Current Large Language Model (LLM) agents demonstrate strong reasoning and tool use capabilities, but often lack self-awareness, failing to balance these approaches effectively. This imbalance leads to **Tool Overuse**, where models unnecessarily rely on external tools for tasks solvable with parametric knowledge, increasing computational overhead. Inspired by human metacognition, we introduce **SMART** (Strategic Model-Aware Reasoning with Tools), a paradigm that enhances an agent’s self-awareness to optimize task handling and reduce tool overuse. To support this paradigm, we introduce **SMART-ER**, a dataset spanning three domains, where reasoning alternates between parametric knowledge and tool-dependent steps, with each step enriched by rationales explaining when tools are necessary. Through supervised training, we develop **SMARTAgent**, a family of models that dynamically balance parametric knowledge and tool use. Evaluations show that SMARTAgent reduces tool use by 24% while improving performance by over 37%, enabling 7B-scale models to match its 70B counterpart and GPT-4. Additionally, SMARTAgent generalizes to out-of-distribution test data like GSM8K and MINTQA, maintaining accuracy with just one-fifth the tool calls. These highlight the potential of strategic tool use to enhance reasoning, mitigate overuse, and bridge the gap between model size and performance, advancing intelligent and resource-efficient agent designs.

Hallucination is a persistent challenge in large language models (LLMs), where even with rigorous quality control, models often generate distorted facts. This paradox, in which error generation continues despite high-quality training data, calls for a deeper understanding of the underlying LLM mechanisms. To address it, we propose a novel concept: knowledge overshadowing, where model’s dominant knowledge can obscure less prominent knowledge during text generation, causing the model to fabricate inaccurate details. Building on this idea, we introduce a novel framework to quantify factual hallucinations by modeling knowledge overshadowing. Central to our approach is the log-linear law, which predicts that the rate of factual hallucination increases linearly with the logarithmic scale of (1) Knowledge Popularity, (2) Knowledge Length, and (3) Model Size. The law provides a means to preemptively quantify hallucinations, offering foresight into their occurrence even before model training or inference. Built on overshadowing effect, we propose a new decoding strategy CoDa, to mitigate hallucinations, which notably enhance model factuality on Overshadow (27.9%), MemoTrap (13.1%) and NQ-Swap (18.3%). Our findings not only deepen understandings of the underlying mechanisms behind hallucinations but also provide actionable insights for developing more predictable and controllable language models.

Recent progress in large language models (LLMs) has enabled substantial advances in solving mathematical problems. However, existing benchmarks often fail to reflect real-world complexity, which demand open-ended, interdisciplinary reasoning and integration of computational tools. To address this gap, we introduce **ModelingBench**, a novel benchmark featuring real-world-inspired, open-ended problems from math modeling competitions across diverse domains, ranging from urban traffic optimization to ecosystem resource planning. These tasks require translating natural language into formal mathematical formulations, applying appropriate tools, and producing structured, defensible reports. ModelingBench supports multiple valid solutions, capturing the ambiguity and creativity of practical modeling. To solve these challenges, we present **ModelingAgent**, a multi-agent framework that coordinates tool use, supports structured workflows, and enables iterative self-refinement to generate well-grounded, creative solutions. Empirical results show that ModelingAgent substantially outperforms strong baselines and often produces solutions indistinguishable from those of human experts. Together, our work provides a comprehensive framework for evaluating and advancing real-world problem-solving in open-ended, interdisciplinary modeling challenges. All the codes are released for future research.

Large language models (LLMs) exhibit exceptional capabilities across various tasks but also pose risks by generating harmful content. Existing safety mechanisms, while improving model safety, often lead to overly cautious behavior and fail to fully leverage LLMs’ internal cognitive processes. Inspired by humans’ reflective thinking capability, we first show that LLMs can similarly perform internal assessments about safety in their internal states. Building on this insight, we propose **SafeSwitch**, a dynamic framework that regulates unsafe outputs by utilizing the prober-based internal state monitor that actively detects harmful intentions, and activates a safety head that leads to safer and more conservative responses only when necessary. SafeSwitch reduces harmful outputs by approximately 80% on harmful queries while maintaining strong utility, reaching a Pareto optimal among several methods. Our method is also advantageous over traditional methods in offering more informative, context-aware refusals, and achieves these benefits while only tuning less than 6% of the original parameters. SafeSwitch demonstrates large language models’ capacity for self-awareness and reflection regarding safety, offering a promising approach to more nuanced and effective safety controls.

Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP) but pose risks of inadvertently exposing copyrighted or proprietary data, especially when such data is used for training but not intended for distribution. Traditional methods address these leaks only after content is generated, which can lead to the exposure of sensitive information. This study introduces a proactive approach: examining LLMs’ internal states before text generation to detect potential leaks. By using a curated dataset of copyrighted materials, we trained a neural network classifier to identify risks, allowing for early intervention by stopping the generation process or altering outputs to prevent disclosure. Integrated with a Retrieval-Augmented Generation (RAG) system, this framework ensures adherence to copyright and licensing requirements while enhancing data privacy and ethical standards. Our results show that analyzing internal states effectively mitigates the risk of copyrighted data leakage, offering a scalable solution that fits smoothly into AI workflows, ensuring compliance with copyright regulations while maintaining high-quality text generation. Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP) but pose risks of inadvertently exposing copyrighted or proprietary data, especially when such data is used for training but not intended for distribution. Traditional methods address these leaks only after content is generated, which can lead to the exposure of sensitive information. This study introduces a proactive approach: examining LLMs’ internal states before text generation to detect potential leaks. By using a curated dataset of copyrighted materials, we trained a neural network classifier to identify risks, allowing for early intervention by stopping the generation process or altering outputs to prevent disclosure. Integrated with a Retrieval-Augmented Generation (RAG) system, this framework ensures adherence to copyright and licensing requirements while enhancing data privacy and ethical standards. Our results show that analyzing internal states effectively mitigates the risk of copyrighted data leakage, offering a scalable solution that fits smoothly into AI workflows, ensuring compliance with copyright regulations while maintaining high-quality text generation. Our code can be found here: (https://anonymous.4open.science/r/Internal-states-leakage-9D6E).

In high-stakes domains such as healthcare and finance, effective decision-making demands not just accurate outcomes but transparent and explainable reasoning. However, current language models often lack the structured deliberation needed for such tasks, instead generating decisions and justifications in a disconnected, post-hoc manner. To address this, we propose DecisionFlow, a novel decision modeling framework that guides models to reason over structured representations of actions, attributes, and constraints. Rather than predicting answers directly from prompts, DecisionFlow builds a semantically grounded decision space and infers a latent utility function to evaluate trade-offs in a transparent, utility-driven manner. This process produces decisions tightly coupled with interpretable rationales reflecting the model’s reasoning. Empirical results on two high-stakes benchmarks show that DecisionFlow not only achieves up to 30% accuracy gains over strong prompting baselines but also enhances alignment in outcomes. Our work is a critical step toward integrating symbolic reasoning with LLMs, enabling more accountable, explainable, and reliable LLM decision support systems. Code and data are at https://github.com/xiusic/DecisionFlow.

Current language model-driven agents often lack mechanisms for effective user participation, which is crucial given the vagueness commonly found in user instructions. Although adept at devising strategies and performing tasks, these agents struggle with seeking clarification and grasping precise user intentions. To bridge this gap, we introduce Intention-in-Interaction (IN3), a novel benchmark designed to inspect users’ implicit intentions through explicit queries. Next, we propose the incorporation of model experts as the upstream in agent designs to enhance user-agent interaction. Employing IN3, we empirically train Mistral-Interact, a powerful model that proactively assesses task vagueness, inquires about user intentions, and refines them into actionable goals before starting downstream agent task execution. Integrating it into the XAgent framework, we comprehensively evaluate the enhanced agent system regarding user instruction understanding and execution, revealing that our approach notably excels at identifying vague user tasks, recovering and summarizing critical missing information, setting precise and necessary agent execution goals, and minimizing redundant tool usage, thus boosting overall efficiency.

Large Language Models (LLMs) have demonstrated remarkable progress in utilizing tools, but their closed-source nature and high inference costs pose limitations on their adaptability, necessitating a valid method that leverages smaller, open-sourced models. In this paper, we introduce Toolink, a comprehensive framework that performs task-solving by first creating a toolkit and then integrating the planning and calling of tools through a chain-of-solving (CoS) approach. We first validate the efficacy of Toolink in harnessing the model’s creativity and CoS ability on ChatGPT. Subsequently, we curate CoS-GPT, a chain-of-solving dataset designed for tool-using, and finetune the LLaMA-7B model. It results in LLaMA-CoS, a powerful open-source model with advanced tool-planning and tool-calling capabilities. Evaluation of diverse tasks from BIG-bench demonstrates its CoS ability matches that of ChatGPT while its performance surpasses the chain-of-thought approach. Further studies highlight the generalization of LLaMA-CoS to unseen tasks and showcase its capability in using toolkits not explicitly tailored for the target task, affirming its robustness in real-world scenarios. All codes and data are released.

Continual pre-training is the paradigm where pre-trained language models (PLMs) continually acquire fresh knowledge from growing data and gradually get upgraded. Before an upgraded PLM is released, we may have tuned the original PLM for various tasks and stored the adapted weights. However, when tuning the upgraded PLM, these outdated adapted weights will typically be ignored and discarded, causing a potential waste of resources. We bring this issue to the forefront and contend that proper algorithms for recycling outdated adapted weights should be developed. To this end, we formulate the task of recyclable tuning for continual pre-training. In pilot studies, we find that after continual pre-training, the upgraded PLM remains compatible with the outdated adapted weights to some extent. Motivated by this finding, we analyze the connection between continually pre-trained PLMs from two novel aspects, i.e., mode connectivity, and functional similarity. Based on the corresponding findings, we propose both an initialization-based method and a distillation-based method for our task. We demonstrate their feasibility in improving the convergence and performance for tuning the upgraded PLM. We also show that both methods can be combined to achieve better performance.

Large Language Models (LLMs) have made significant progress in utilizing tools, but their ability is limited by API availability and the instability of implicit reasoning, particularly when both planning and execution are involved. To overcome these limitations, we propose CREATOR, a novel framework that enables LLMs to create their own tools using documentation and code realization. CREATOR disentangles abstract tool creation and concrete decision execution, resulting in improved performance. We evaluate CREATOR on MATH and TabMWP benchmarks, respectively consisting of challenging math competition problems and diverse tabular contents. Remarkably, CREATOR outperforms existing chain-of-thought, program-of-thought, and tool-using baselines. Additionally, we introduce the Creation Challenge dataset, featuring 2K diverse questions, to emphasize the necessity and benefits of LLMs’ tool creation ability. Further research demonstrates that leveraging LLMs as tool creators facilitates knowledge transfer, and LLMs exhibit varying levels of tool creation abilities, enabling them to adapt to diverse situations. The tool creation ability revolutionizes the LLM’s problem-solving paradigm, driving us closer to the next frontier of artificial intelligence.

Recent years have witnessed the prevalent application of pre-trained language models (PLMs) in NLP. From the perspective of parameter space, PLMs provide generic initialization, starting from which high-performance minima could be found. Although plenty of works have studied how to effectively and efficiently adapt PLMs to high-performance minima, little is known about the connection of various minima reached under different adaptation configurations. In this paper, we investigate the geometric connections of different minima through the lens of mode connectivity, which measures whether two minima can be connected with a low-loss path. We conduct empirical analyses to investigate three questions: (1) how could hyperparameters, specific tuning methods, and training data affect PLM’s mode connectivity? (2) How does mode connectivity change during pre-training? (3) How does the PLM’s task knowledge change along the path connecting two minima? In general, exploring the mode connectivity of PLMs conduces to understanding the geometric connection of different minima, which may help us fathom the inner workings of PLM downstream adaptation. The codes are publicly available at https://github.com/thunlp/Mode-Connectivity-PLM.

Out of Scope (OOS) detection in Conversational AI solutions enables a chatbot to handle a conversation gracefully when it is unable to make sense of the end-user query. Accurately tagging a query as out-of-domain is particularly hard in scenarios when the chatbot is not equipped to handle a topic which has semantic overlap with an existing topic it is trained on. We propose a simple yet effective OOS detection method that outperforms standard OOS detection methods in a real-world deployment of virtual assistants. We discuss the various design and deployment considerations for a cloud platform solution to train virtual assistants and deploy them at scale. Additionally, we propose a collection of datasets that replicates real-world scenarios and show comprehensive results in various settings using both offline and online evaluation metrics.

Current virtual assistant (VA) platforms are beholden to the limited number of languages they support. Every component, such as the tokenizer and intent classifier, is engineered for specific languages in these intricate platforms. Thus, supporting a new language in such platforms is a resource-intensive operation requiring expensive re-training and re-designing. In this paper, we propose a benchmark for evaluating language-agnostic intent classification, the most critical component of VA platforms. To ensure the benchmarking is challenging and comprehensive, we include 29 public and internal datasets across 10 low-resource languages and evaluate various training and testing settings with consideration of both accuracy and training time. The benchmarking result shows that Watson Assistant, among 7 commercial VA platforms and pre-trained multilingual language models (LMs), demonstrates close-to-best accuracy with the best accuracy-training time trade-off.