Cheng Qian

Yujia Qin, Shengding Hu, Yankai Lin et al. · tsinghua

Humans possess an extraordinary ability to create and utilize tools, allowing them to overcome physical limitations and explore new frontiers. With the advent of foundation models, AI systems have the potential to be equally adept in tool use as humans. This paradigm, i.e., tool learning with foundation models, combines the strengths of specialized tools and foundation models to achieve enhanced accuracy, efficiency, and automation in problem-solving. Despite its immense potential, there is still a lack of a comprehensive understanding of key challenges, opportunities, and future endeavors in this field. To this end, we present a systematic investigation of tool learning in this paper. We first introduce the background of tool learning, including its cognitive origins, the paradigm shift of foundation models, and the complementary roles of tools and models. Then we recapitulate existing tool learning research into tool-augmented and tool-oriented learning. We formulate a general tool learning framework: starting from understanding the user instruction, models should learn to decompose a complex task into several subtasks, dynamically adjust their plan through reasoning, and effectively conquer each sub-task by selecting appropriate tools. We also discuss how to train models for improved tool-use capabilities and facilitate the generalization in tool learning. Considering the lack of a systematic tool learning evaluation in prior works, we experiment with 18 representative tools and show the potential of current foundation models in skillfully utilizing tools. Finally, we discuss several open problems that require further investigation for tool learning. In general, we hope this paper could inspire future research in integrating tools with foundation models.

Cheng Qian, Chenyan Xiong, Zhenghao Liu et al. · tsinghua

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.

Yujia Qin, Cheng Qian, Jing Yi et al. · tencent-ai, tsinghua

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.

Hyeonjeong Ha, Jeonghwan Kim, Cheng Qian et al.

Memory-augmented large language models extend reasoning beyond a fixed context window by maintaining long-term memory across interactions. However, existing memory systems often collapse stable user facts, episodic events, and behavioral rules into a shared space, allowing functionally distinct memories to be retrieved and used as interchangeable evidence. We identify this failure mode as heterogeneous memory contamination, where context-specific events become overgeneralized claims, or semantically relevant but functionally incompatible memories mislead generation. To this end, we introduce MemGuard, a type-aware memory framework that preserves functional memory boundaries during memory construction and retrieval. It assigns each memory an explicit functional role at write time, maintains relations across type-isolated memories, and selectively composes evidence only from necessary memory types, reducing contamination from irrelevant or functionally incompatible evidence. Across hallucination and long-horizon conversation benchmarks, MemGuard improves memory reliability by up to 28.27% while retrieving up to 5.8x fewer memory tokens than prior methods. These results suggest that reliable long-term reasoning depends on principled organization and selective use of heterogeneous memory.

Cheng Qian, Xinran Zhao, Sherry Tongshuang Wu · tsinghua

Large language models (LLMs) acquire extensive knowledge during pre-training, known as their parametric knowledge. However, in order to remain up-to-date and align with human instructions, LLMs inevitably require external knowledge during their interactions with users. This raises a crucial question: How will LLMs respond when external knowledge interferes with their parametric knowledge? To investigate this question, we propose a framework that systematically elicits LLM parametric knowledge and introduces external knowledge. Specifically, we uncover the impacts by constructing a parametric knowledge graph to reveal the different knowledge structures of LLMs, and introduce external knowledge through distractors of varying degrees, methods, positions, and formats. Our experiments on both black-box and open-source models demonstrate that LLMs tend to produce responses that deviate from their parametric knowledge, particularly when they encounter direct conflicts or confounding changes of information within detailed contexts. We also find that while LLMs are sensitive to the veracity of external knowledge, they can still be distracted by unrelated information. These findings highlight the risk of hallucination when integrating external knowledge, even indirectly, during interactions with current LLMs. All the data and results are publicly available.

Jiayu Liu, Cheng Qian, Zhenhailong Wang et al.

Planning for real-world problems by language models often involves both world and user constraints, which may not be fully specified upfront and are progressively disclosed through interaction. However, existing benchmarks still underexplore adaptive planning under such progressively revealed dual constraints. To address this gap, we introduce AdaPlanBench, a dynamic interactive benchmark for evaluating whether Large Language Model (LLM) agents can adaptively plan and re-plan under progressively revealed world and user constraints. AdaPlanBench is built on 307 household tasks, with a scalable constraint construction pipeline that augments each task with dual constraints. At runtime, agents interact with the environment in a multi-turn protocol where hidden constraints are revealed only when the agent proposes a plan that violates them, requiring iterative plan revision under accumulating feedback. This makes planning challenging, as agents must infer and track constraints from feedback while re-planning effectively. Experiments on ten leading LLMs show that adaptive planning under dual constraints remains challenging, with the best model reaching only 67.75% accuracy. We further observe that performance degrades as more constraints accumulate, with user constraints posing a particularly large challenge and failures often stemming from weaker physical grounding and reduced effectiveness. These results establish AdaPlanBench as a testbed for dual-constrained interactive planning and highlight the challenge of reliable adaptation to dynamically revealed constraints in LLM agents.

Cheng Qian, Jiayu Liu, Heng Ji

Understanding what a user believes and intends is central to building effective agent assistants. This ability is often evaluated through Theory-of-Mind (ToM) tasks, where success requires reasoning from the user's perspective. However, many existing approaches address ToM with complex pipelines that model behavior indirectly, without explicitly reconstructing the user's mental state. This misses the core structure of the problem: users act based on their beliefs, which are updated through observations of the environment; beliefs and intentions jointly determine actions, which in turn change the environment; and social reasoning often requires nested beliefs about what others believe or intend. We propose UserHarness, a simple framework that reframes ToM reasoning as explicit user-mind reconstruction. UserHarness decomposes the user's mental state, its relation to the external environment, and the actions that follow from it, enabling agents to track what the user observes, believes, intends, and does. Across five benchmarks, UserHarness reaches up to 95.94% macro accuracy, improving over existing inference methods by more than 15% relative and over the strongest prompt-only harness by about 20% relative. These results suggest that robust user understanding requires reasoning from the roots of the user's mind, positioning user harnessing as a promising foundation for more adaptive future assistants.

Tong Liu, Cheng Qian, Matej Cief et al.

Tool-calling is a central component of modern large language model (LLM) agents, equipping them with skills beyond their parametric knowledge. This paper studies tool-calling along two complementary axes: effectiveness, i.e., how this capability is measured, and efficiency, i.e., how it is learned. On effectiveness, we systematically analyze tool-calling evaluation pipelines and show that results can be highly sensitive to seemingly minor, often undocumented implementation choices including the random seed, system prompt, multi-turn template construction, and how prior interaction/reasoning history is carried forward. These choices can lead to substantial differences in reported performance, especially in multi-turn settings where without rigorous standardization, leaderboard rankings are unreliable. On efficiency, we examine standard reinforcement learning (RL) for tool-calling and identify two sources of computational waste: (i) during rollouts, many prompts produce no learning signal, and (ii) during policy updates, optimization incurs high computational cost. Guided by these findings, we introduce two techniques that accelerate RL-based tool-calling training, achieving substantial wall-clock speedup without degrading performance.

Jiateng Liu, Bingxuan Li, Zhenhailong Wang et al.

We dream of AI agents that can read arbitrary designs and construct real-world objects from reusable building blocks. As a first step toward this vision, we study whether multimodal large language models (MLLMs) possess the visual grounding and spatial reasoning capabilities required for brick assembly. We formulate brick assembly as a sequential decision-making problem, where each step involves two subtasks: brick selection, identifying the target brick from candidate components, and brick pose estimation, predicting where and how the selected brick should be placed. To support this study, we introduce BC-Bench (Brick Construction Benchmark), the first benchmark for evaluating MLLMs on assembly with diverse bricks. Experiments show that current state-of-the-art MLLMs remain far from reliable builders, struggling with fine-grained brick selection and failing at precise pose estimation. To bridge this gap, we propose Brick-Composer, a learning framework that equips MLLMs with assembly skills through three complementary signals: Human Design Sparks, which provide affordance-rich construction demonstrations; World Feedback, which grounds predicted actions in visual and physical consequences; and Synthetic Experience, which scales learning beyond existing object designs. Brick-Composer improves brick selection accuracy by over three times, substantially reduces pose estimation errors, and raises strict step-level assembly success from less than 1% to around 15%. After training, a Qwen-3-8B can correctly compose up to 42% of the steps for a complete object, suggesting that MLLMs can acquire assembly capabilities through targeted, physically grounded learning.

Yilong Wang, Cheng Qian, Edward Johns

Deformable object manipulation (DOM) is challenging due to high-dimensional, partially observable states that evolve through long-horizon, topology-changing interactions with multiple valid manipulation modes. We introduce Instant-Fold, an in-context imitation learning framework for DOM. Given a single human demonstration, our policy infers and executes diverse manipulation modes directly from the demonstration, including variations in spatial execution and ordering, without requiring gradient updates. Our approach first learns deformation-aware visual representations via temporal contrastive pretraining, after which a flow-matching transformer policy conditioned on the demonstration predicts actions to execute the intended manipulation mode. Trained entirely in simulation, Instant-Fold generalizes across diverse folding modes and transfers zero-shot to real-world settings without additional data collection or finetuning. Videos are available at https://instant-fold.github.io.

Cheng Qian, Nicholas D. Sidiropoulos, Kejun Huang et al.

Reconstructing a signal from squared linear (rank-one quadratic) measurements is a challenging problem with important applications in optics and imaging, where it is known as phase retrieval. This paper proposes two new phase retrieval algorithms based on non-convex quadratically constrained quadratic programming (QCQP) formulations, and a recently proposed approximation technique dubbed feasible point pursuit (FPP). The first is designed for uniformly distributed bounded measurement errors, such as those arising from high-rate quantization (B-FPP). The second is designed for Gaussian measurement errors, using a least squares criterion (LS-FPP). Their performance is measured against state-of-the-art algorithms and the Cramér-Rao bound (CRB), which is also derived here. Simulations show that LS-FPP outperforms the state-of-art and operates close to the CRB. Compact CRB expressions, properties, and insights are obtained by explicitly computing the CRB in various special cases -- including when the signal of interest admits a sparse parametrization, using harmonic retrieval as an example.

Cheng Qian, Yulun Zhang, Varun Bhatt et al. · cmu

We use the Quality Diversity (QD) algorithm with Neural Cellular Automata (NCA) to automatically evaluate Multi-Agent Path Finding (MAPF) algorithms by generating diverse maps. Previously, researchers typically evaluate MAPF algorithms on a set of specific, human-designed maps at their initial stage of algorithm design. However, such fixed maps may not cover all scenarios, and algorithms may overfit to the small set of maps. To seek further improvements, systematic evaluations on a diverse suite of maps are needed. In this work, we propose Quality-Diversity Multi-Agent Path Finding Performance EvaluatoR (QD-MAPPER), a general framework that takes advantage of the QD algorithm to comprehensively understand the performance of MAPF algorithms by generating maps with patterns, be able to make fair comparisons between two MAPF algorithms, providing further information on the selection between two algorithms and on the design of the algorithms. Empirically, we employ this technique to evaluate and compare the behavior of different types of MAPF algorithms, including search-based, priority-based, rule-based, and learning-based algorithms. Through both single-algorithm experiments and comparisons between algorithms, researchers can identify patterns that each MAPF algorithm excels and detect disparities in runtime or success rates between different algorithms.

Yaxuan Li, Yuxin Zuo, Bingxiang He et al.

On-policy distillation (OPD) has become a core technique in the post-training of large language models, yet its training dynamics remain poorly understood. This paper provides a systematic investigation of OPD dynamics and mechanisms. We first identify that two conditions govern whether OPD succeeds or fails: (i) the student and teacher should share compatible thinking patterns; and (ii) even with consistent thinking patterns and higher scores, the teacher must offer genuinely new capabilities beyond what the student has seen during training. We validate these findings through weak-to-strong reverse distillation, showing that same-family 1.5B and 7B teachers are distributionally indistinguishable from the student's perspective. Probing into the token-level mechanism, we show that successful OPD is characterized by progressive alignment on high-probability tokens at student-visited states, a small shared token set that concentrates most of the probability mass (97%-99%). We further propose two practical strategies to recover failing OPD: off-policy cold start and teacher-aligned prompt selection. Finally, we show that OPD's apparent free lunch of dense token-level reward comes at a cost, raising the question of whether OPD can scale to long-horizon distillation.

Rakshith S Srinivasa, Cheng Qian, Brandon Theodorou et al.

The ongoing pandemic has highlighted the importance of reliable and efficient clinical trials in healthcare. Trial sites, where the trials are conducted, are chosen mainly based on feasibility in terms of medical expertise and access to a large group of patients. More recently, the issue of diversity and inclusion in clinical trials is gaining importance. Different patient groups may experience the effects of a medical drug/ treatment differently and hence need to be included in the clinical trials. These groups could be based on ethnicity, co-morbidities, age, or economic factors. Thus, designing a method for trial site selection that accounts for both feasibility and diversity is a crucial and urgent goal. In this paper, we formulate this problem as a ranking problem with fairness constraints. Using principles of fairness in machine learning, we learn a model that maps a clinical trial description to a ranked list of potential trial sites. Unlike existing fairness frameworks, the group membership of each trial site is non-binary: each trial site may have access to patients from multiple groups. We propose fairness criteria based on demographic parity to address such a multi-group membership scenario. We test our method on 480 real-world clinical trials and show that our model results in a list of potential trial sites that provides access to a diverse set of patients while also ensuing a high number of enrolled patients.

Chaoqi Yang, Cheng Qian, Jimeng Sun

Low-rank tensor factorization or completion is well-studied and applied in various online settings, such as online tensor factorization (where the temporal mode grows) and online tensor completion (where incomplete slices arrive gradually). However, in many real-world settings, tensors may have more complex evolving patterns: (i) one or more modes can grow; (ii) missing entries may be filled; (iii) existing tensor elements can change. Existing methods cannot support such complex scenarios. To fill the gap, this paper proposes a Generalized Online Canonical Polyadic (CP) Tensor factorization and completion framework (named GOCPT) for this general setting, where we maintain the CP structure of such dynamic tensors during the evolution. We show that existing online tensor factorization and completion setups can be unified under the GOCPT framework. Furthermore, we propose a variant, named GOCPTE, to deal with cases where historical tensor elements are unavailable (e.g., privacy protection), which achieves similar fitness as GOCPT but with much less computational cost. Experimental results demonstrate that our GOCPT can improve fitness by up to 2:8% on the JHU Covid data and 9:2% on a proprietary patient claim dataset over baselines. Our variant GOCPTE shows up to 1:2% and 5:5% fitness improvement on two datasets with about 20% speedup compared to the best model.

Bingxiang He, Zekai Qu, Zeyuan Liu et al.

Recent advances in reinforcement learning for large language models have converged on increasing complexity: multi-stage training pipelines, dynamic hyperparameter schedules, and curriculum learning strategies. This raises a fundamental question: \textbf{Is this complexity necessary?} We present \textbf{JustRL}, a minimal approach using single-stage training with fixed hyperparameters that achieves state-of-the-art performance on two 1.5B reasoning models (54.9\% and 64.3\% average accuracy across nine mathematical benchmarks) while using 2$\times$ less compute than sophisticated approaches. The same hyperparameters transfer across both models without tuning, and training exhibits smooth, monotonic improvement over 4,000+ steps without the collapses or plateaus that typically motivate interventions. Critically, ablations reveal that adding ``standard tricks'' like explicit length penalties and robust verifiers may degrade performance by collapsing exploration. These results suggest that the field may be adding complexity to solve problems that disappear with a stable, scaled-up baseline. We release our models and code to establish a simple, validated baseline for the community.

Jiayu Liu, Cheng Qian, Zhaochen Su et al.

Current evaluations of Large Language Model (LLM) agents primarily emphasize task completion, often overlooking resource efficiency and adaptability. This neglects a crucial capability: agents' ability to devise and adjust cost-optimal plans in response to changing environments. To bridge this gap, we introduce CostBench, a scalable, cost-centric benchmark designed to evaluate agents' economic reasoning and replanning abilities. Situated in the travel-planning domain, CostBench comprises tasks solvable via multiple sequences of atomic and composite tools with diverse, customizable costs. It also supports four types of dynamic blocking events, such as tool failures and cost changes, to simulate real-world unpredictability and necessitate agents to adapt in real time. Evaluating leading open-sourced and proprietary models on CostBench reveals a substantial gap in cost-aware planning: agents frequently fail to identify cost-optimal solutions in static settings, with even GPT-5 achieving less than 75% exact match rate on the hardest tasks, and performance further dropping by around 40% under dynamic conditions. By diagnosing these weaknesses, CostBench lays the groundwork for developing future agents that are both economically rational and robust.

Cheng Qian, Hyeonjeong Ha, Jiayu Liu et al.

Large multimodal models (LMMs) have rapidly advanced in perception and reasoning; however, it remains unclear whether these capabilities generalize to discovering visually grounded solutions in open-ended environments, beyond pattern recognition. In such settings, intelligence requires more than answering well-posed questions: it involves identifying how elements in a scene can be repurposed in non-obvious yet physically feasible ways. This form of creative problem-solving is central to human intelligence, but remains largely untested in current benchmarks. To evaluate this ability, we introduce MM-CreativityBench, a benchmark for affordance-grounded creative tool use in visually rich, physically constrained environments. Each instance presents a scenario image with structured views of candidate entities and their parts, enabling fine-grained, interactive evaluation of how models iteratively inspect the scene, identify relevant affordances, and compose visually and physically grounded solutions. Our experiments show that current LMMs often fall short, not due to lack of generative capability, but because they do not sustain grounded exploration. Models often overlook relevant entities, under-examine critical parts, or hallucinate attributes not grounded in the image. Motivated by this failure mode, we propose affordance-grounded alignment, which casts creative tool use as a preference learning problem. Using Direct Preference Optimization, we encourage models to prefer attribute-affordance reasoning grounded in visual evidence over hallucinated alternatives. In addition, we incorporate supervision derived from an affordance knowledge base to guide broader entity exploration and multi-turn planning. Our results show consistent gains in selecting the correct entities and parts, while substantially reducing hallucination and grounding-related errors.

Zhongyang Li, Yaqian Li, Faming Fang et al.

Multimodal large language models suffer from severe computational and memory bottlenecks, as the number of visual tokens far exceeds that of textual tokens. While recent methods employ projector modules to align and compress visual tokens into text-aligned features, they typically depend on fixed heuristics that limit adaptability across diverse scenarios. In this paper, we first propose Query Guided Mixture-of-Projector (QMoP), a novel and flexible framework that adaptively compresses visual tokens via three collaborative branches: (1) a pooling-based branch for coarse-grained global semantics, (2) a resampler branch for extracting high-level semantic representations, and (3) a pruning-based branch for fine-grained token selection to preserve critical visual detail. To adaptively coordinate these branches, we introduce the Query Guided Router (QGR), which dynamically selects and weights the outputs from different branches based on both visual input and textual queries. A Mixture-of-Experts-style fusion mechanism is designed to aggregate the outputs, harnessing the strengths of each strategy while suppressing noise. To systematically evaluate the effects of Visual Token Compression, we also develop VTCBench, a dedicated benchmark for evaluating the information loss induced by visual token compression. Extensive experiments demonstrate that despite relying on fundamental compression modules, QMoP outperforms strong baselines and delivers significant savings in memory, computation, and inference time.

Cheng Qian, Julen Urain, Kevin Zakka et al.

In this work, we introduce PianoMime, a framework for training a piano-playing agent using internet demonstrations. The internet is a promising source of large-scale demonstrations for training our robot agents. In particular, for the case of piano-playing, Youtube is full of videos of professional pianists playing a wide myriad of songs. In our work, we leverage these demonstrations to learn a generalist piano-playing agent capable of playing any arbitrary song. Our framework is divided into three parts: a data preparation phase to extract the informative features from the Youtube videos, a policy learning phase to train song-specific expert policies from the demonstrations and a policy distillation phase to distil the policies into a single generalist agent. We explore different policy designs to represent the agent and evaluate the influence of the amount of training data on the generalization capability of the agent to novel songs not available in the dataset. We show that we are able to learn a policy with up to 56\% F1 score on unseen songs.

Rui Yang, Hanyang Chen, Junyu Zhang et al.

Leveraging Multi-modal Large Language Models (MLLMs) to create embodied agents offers a promising avenue for tackling real-world tasks. While language-centric embodied agents have garnered substantial attention, MLLM-based embodied agents remain underexplored due to the lack of comprehensive evaluation frameworks. To bridge this gap, we introduce EmbodiedBench, an extensive benchmark designed to evaluate vision-driven embodied agents. EmbodiedBench features: (1) a diverse set of 1,128 testing tasks across four environments, ranging from high-level semantic tasks (e.g., household) to low-level tasks involving atomic actions (e.g., navigation and manipulation); and (2) six meticulously curated subsets evaluating essential agent capabilities like commonsense reasoning, complex instruction understanding, spatial awareness, visual perception, and long-term planning. Through extensive experiments, we evaluated 24 leading proprietary and open-source MLLMs within EmbodiedBench. Our findings reveal that: MLLMs excel at high-level tasks but struggle with low-level manipulation, with the best model, GPT-4o, scoring only 28.9\% on average. EmbodiedBench provides a multifaceted standardized evaluation platform that not only highlights existing challenges but also offers valuable insights to advance MLLM-based embodied agents. Our code and dataset are available at https://embodiedbench.github.io.

Xiusi Chen, Gaotang Li, Ziqi Wang et al.

Reward modeling is essential for aligning large language models with human preferences through reinforcement learning from human feedback. To provide accurate reward signals, a reward model (RM) should stimulate deep thinking and conduct interpretable reasoning before assigning a score or a judgment. Inspired by recent advances of long chain-of-thought on reasoning-intensive tasks, we hypothesize and validate that integrating reasoning capabilities into reward modeling significantly enhances RMs interpretability and performance. To this end, we introduce a new class of generative reward models - Reasoning Reward Models (ReasRMs) - which formulate reward modeling as a reasoning task. We propose a reasoning-oriented training pipeline and train a family of ReasRMs, RM-R1. RM-R1 features a chain-of-rubrics (CoR) mechanism - self-generating sample-level chat rubrics or math/code solutions, and evaluating candidate responses against them. The training of RM-R1 consists of two key stages: (1) distillation of high-quality reasoning chains and (2) reinforcement learning with verifiable rewards. Empirically, our models achieve state-of-the-art performance across three reward model benchmarks on average, outperforming much larger open-weight models (e.g., INF-ORM-Llama3.1-70B) and proprietary ones (e.g., GPT-4o) by up to 4.9%. Beyond final performance, we perform thorough empirical analyses to understand the key ingredients of successful ReasRM training. To facilitate future research, we release six REASRM models along with code and data at https://github.com/RM-R1-UIUC/RM-R1.

Kunlun Zhu, Hongyi Du, Zhaochen Hong et al.

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, gpt-4o-mini reaches the average highest task score, graph structure performs the best among coordination protocols in the research scenario, and cognitive planning improves milestone achievement rates by 3%. Code and datasets are public available at https://github.com/MultiagentBench/MARBLE.

Cheng Qian, Haode Qi, Gengyu Wang et al.

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.

Yaxi Lu, Shenzhi Yang, Cheng Qian et al. · tsinghua

Agents powered by large language models have shown remarkable abilities in solving complex tasks. However, most agent systems remain reactive, limiting their effectiveness in scenarios requiring foresight and autonomous decision-making. In this paper, we tackle the challenge of developing proactive agents capable of anticipating and initiating tasks without explicit human instructions. We propose a novel data-driven approach for this problem. Firstly, we collect real-world human activities to generate proactive task predictions. These predictions are then labeled by human annotators as either accepted or rejected. The labeled data is used to train a reward model that simulates human judgment and serves as an automatic evaluator of the proactiveness of LLM agents. Building on this, we develop a comprehensive data generation pipeline to create a diverse dataset, ProactiveBench, containing 6,790 events. Finally, we demonstrate that fine-tuning models with the proposed ProactiveBench can significantly elicit the proactiveness of LLM agents. Experimental results show that our fine-tuned model achieves an F1-Score of 66.47% in proactively offering assistance, outperforming all open-source and close-source models. These results highlight the potential of our method in creating more proactive and effective agent systems, paving the way for future advancements in human-agent collaboration.

Cheng Qian, Emre Can Acikgoz, Bingxuan Li et al.

Agents built on vision-language models increasingly face tasks that demand anticipating future states rather than relying on short-horizon reasoning. Generative world models offer a promising remedy: agents could use them as external simulators to foresee outcomes before acting. This paper empirically examines whether current agents can leverage such world models as tools to enhance their cognition. Across diverse agentic and visual question answering tasks, we observe that some agents rarely invoke simulation (fewer than 1%), frequently misuse predicted rollouts (approximately 15%), and often exhibit inconsistent or even degraded performance (up to 5%) when simulation is available or enforced. Attribution analysis further indicates that the primary bottleneck lies in the agents' capacity to decide when to simulate, how to interpret predicted outcomes, and how to integrate foresight into downstream reasoning. These findings underscore the need for mechanisms that foster calibrated, strategic interaction with world models, paving the way toward more reliable anticipatory cognition in future agent systems.

Cheng Qian, Di Xiu, Minghao Tian

In this technical report, we present the 2nd place solution of 2023 Waymo Open Sim Agents Challenge (WOSAC)[4]. We propose a simple yet effective autoregressive method for simulating multi-agent behaviors, which is built upon a well-known multimodal motion forecasting framework called Motion Transformer (MTR)[5] with postprocessing algorithms applied. Our submission named MTR+++ achieves 0.4697 on the Realism Meta metric in 2023 WOSAC. Besides, a modified model based on MTR named MTR_E is proposed after the challenge, which has a better score 0.4911 and is ranked the 3rd on the leaderboard of WOSAC as of June 25, 2023.

Bingxuan Li, Jeonghwan Kim, Cheng Qian et al.

Overlapping calendar invitations force busy professionals to repeatedly decide which meetings to attend, reschedule, or decline. We refer to this preference-driven decision process as calendar conflict resolution. Automating this decision process is crucial yet challenging. Scheduling logistics can drain hours, and human delegation often fails at scale, which motivates us to ask: Can we trust large language models (LLMs) or language agents to manage time? To enable a systematic study of this question, we introduce CalConflictBench, a benchmark for long-horizon calendar conflict resolution. In CalConflictBench, conflicts are presented to agents round-by-round over a calendar year, requiring them to infer and adapt to user preferences progressively. Our experiments show that current LLM agents perform poorly with high error rates, e.g., Qwen-3-30B-Think has an average error rate of 35%. To address this gap, we propose PEARL, a reinforcement-learning framework that (i) augments the language agent with an external preference memory that stores and updates inferred strategies (e.g., attendee priorities, topic importance, time/location preferences), and (ii) optimizes the agent with round-wise rewards that directly supervise decision correctness, ranking quality, and memory usage across rounds. Experiments on CalConflictBench show that PEARL achieves an error reduction rate of 0.76 and a 55% improvement in average error rate compared to the strongest baseline.

Xuying Ning, Katherine Tieu, Dongqi Fu et al.

Recent large language models (LLMs) have demonstrated strong capabilities in understanding and generating code, from competitive programming to repository-level software engineering. In emerging agentic systems, code is no longer only a target output. It increasingly serves as an operational substrate for agent reasoning, acting, environment modeling, and execution-based verification. We frame this shift through the lens of agent harnesses and introduce code as agent harness: a unified view that centers code as the basis for agent infrastructure. To systematically study this perspective, we organize the survey around three connected layers. First, we study the harness interface, where code connects agents to reasoning, action, and environment modeling. Second, we examine harness mechanisms: planning, memory, and tool use for long-horizon execution, together with feedback-driven control and optimization that make harness reliable and adaptive. Third, we discuss scaling the harness from single-agent systems to multi-agent settings, where shared code artifacts support multi-agent coordination, review, and verification. Across these layers, we summarize representative methods and practical applications of code as agent harness, spanning coding assistants, GUI/OS automation, embodied agents, scientific discovery, personalization and recommendation, DevOps, and enterprise workflows. We further outline open challenges for harness engineering, including evaluation beyond final task success, verification under incomplete feedback, regression-free harness improvement, consistent shared state across multiple agents, human oversight for safety-critical actions, and extensions to multimodal environments. By centering code as the harness of agentic AI, this survey provides a unified roadmap toward executable, verifiable, and stateful AI agent systems.

Yixia Li, Hongru Wang, Jiahao Qiu et al.

Agentic reinforcement learning increasingly relies on experience-driven scaling, yet real-world environments remain non-adaptive, limited in coverage, and difficult to scale. World models offer a potential way to improve learning efficiency through simulated experience, but it remains unclear whether large language models can reliably serve this role and under what conditions they meaningfully benefit agents. We study these questions in text-based environments, which provide a controlled setting to reinterpret language modeling as next-state prediction under interaction. We introduce a three-level framework for evaluating LLM-based world models: (i) fidelity and consistency, (ii) scalability and robustness, and (iii) agent utility. Across five representative environments, we find that sufficiently trained world models maintain coherent latent state, scale predictably with data and model size, and improve agent performance via action verification, synthetic trajectory generation, and warm-starting reinforcement learning. Meanwhile, these gains depend critically on behavioral coverage and environment complexity, delineating clear boundry on when world modeling effectively supports agent learning.

Yan Xie, Changkui Mao, Changsong Wu et al.

As deep learning-based AI technologies gain momentum, the demand for general-purpose AI computing architectures continues to grow. While GPGPU-based architectures offer versatility for diverse AI workloads, they often fall short in efficiency and cost-effectiveness. Various Domain-Specific Architectures (DSAs) excel at particular AI tasks but struggle to extend across broader applications or adapt to the rapidly evolving AI landscape. M100 is Li Auto's response: a performant, cost-effective architecture for AI inference in Autonomous Driving (AD), Large Language Models (LLMs), and intelligent human interactions, domains crucial to today's most competitive automobile platforms. M100 employs a dataflow parallel architecture, where compiler-architecture co-design orchestrates not only computation but, more critically, data movement across time and space. Leveraging dataflow computing efficiency, our hardware-software co-design improves system performance while reducing hardware complexity and cost. M100 largely eliminates caching: tensor computations are driven by compiler- and runtime-managed data streams flowing between computing elements and on/off-chip memories, yielding greater efficiency and scalability than cache-based systems. Another key principle was selecting the right operational granularity for scheduling, issuing, and execution across compiler, firmware, and hardware. Recognizing commonalities in AI workloads, we chose the tensor as the fundamental data element. M100 demonstrates general AI computing capability across diverse inference applications, including UniAD (for AD) and LLaMA (for LLMs). Benchmarks show M100 outperforms GPGPU architectures in AD applications with higher utilization, representing a promising direction for future general AI computing.

Cheng Qian, Hainan Zhang, Lei Sha et al.

With the growing deployment of LLMs in daily applications like chatbots and content generation, efforts to ensure outputs align with human values and avoid harmful content have intensified. However, increasingly sophisticated jailbreak attacks threaten this alignment, aiming to induce unsafe outputs. Current defense efforts either focus on prompt rewriting or detection, which are limited in effectiveness due to the various design of jailbreak prompts, or on output control and detection, which are computationally expensive as they require LLM inference. Therefore, designing a pre-inference defense method that resists diverse jailbreak prompts is crucial for preventing LLM jailbreak attacks. We observe that jailbreak attacks, safe queries, and harmful queries exhibit different clustering patterns within the LLM's hidden state representation space. This suggests that by leveraging the LLM's hidden state representational capabilities, we can analyze the LLM's forthcoming behavior and proactively intervene for defense. In this paper, we propose a jailbreak attack defense strategy based on a Hidden State Filter (HSF), a lossless architectural defense mechanism that enables the model to preemptively identify and reject adversarial inputs before the inference process begins. We activate its defensive potential through an additional plugin module, effectively framing the defense task as a classification problem. Experimental results on two benchmark datasets, utilizing three different LLMs, show that HSF significantly enhances resilience against six cutting-edge jailbreak attacks. It significantly reduces the success rate of jailbreak attacks while minimally impacting responses to benign user queries, with negligible inference overhead, and outperforming defense baselines.Our code and data are available at https://anonymous.4open.science/r/Hidden-State-Filtering-8652/

Haode Qi, Cheng Qian, Jian Ni et al.

In an enterprise Virtual Assistant (VA) system, intent classification is the crucial component that determines how a user input is handled based on what the user wants. The VA system is expected to be a cost-efficient SaaS service with low training and inference time while achieving high accuracy even with a small number of training samples. We pretrain a transformer-based sentence embedding model with a contrastive learning objective and leverage the embedding of the model as features when training intent classification models. Our approach achieves the state-of-the-art results for few-shot scenarios and performs better than other commercial solutions on popular intent classification benchmarks. However, generating features via a transformer-based model increases the inference time, especially for longer user inputs, due to the quadratic runtime of the transformer's attention mechanism. On top of model distillation, we introduce a practical multi-task adaptation approach that configures dynamic token pruning without the need for task-specific training for intent classification. We demonstrate that this approach improves the inference speed of popular sentence transformer models without affecting model performance.

Peixuan Han, Cheng Qian, Xiusi Chen et al.

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. Codes for this work are available at https://github.com/Hanpx20/SafeSwitch.

Runchu Tian, Yanghao Li, Yuepeng Fu et al. · tsinghua

Positional bias in large language models (LLMs) 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. Thorough experiments are conducted with five commercial and six open-source models. These experiments reveal that while most current models are robust against the "lost in the middle" issue, there exist significant biases related to the spacing of relevant information pieces. These findings highlight the importance of evaluating and reducing positional biases to advance LLM's capabilities.

Kun Wang, Cheng Qian, Miao Yu et al.

Multimodal Large Language Models (MLLMs) have achieved remarkable success in cross-modal understanding and generation, yet their deployment is threatened by critical safety vulnerabilities. While prior works have demonstrated the feasibility of backdoors in MLLMs via fine-tuning data poisoning to manipulate inference, the underlying mechanisms of backdoor attacks remain opaque, complicating the understanding and mitigation. To bridge this gap, we propose ProjLens, an interpretability framework designed to demystify MLLMs backdoors. We first establish that normal downstream task alignment--even when restricted to projector fine--tuning--introduces vulnerability to backdoor injection, whose activation mechanism is different from that observed in text-only LLMs. Through extensive experiments across four backdoor variants, we uncover:(1) Low-Rank Structure: Backdoor injection updates appear overall full-rank and lack dedicated ``trigger neurons'', but the backdoor-critical parameters are encoded within a low-rank subspace of the projector;(2) Activation Mechanism: Both clean and poisoned embedding undergoes a semantic shift toward a shared direction aligned with the backdoor target, but the shifting magnitude scales linearly with the input norm, resulting in the distinct backdoor activation on poisoned samples. Our code is available at: https://anonymous.4open.science/r/ProjLens-8FD7

Emre Can Acikgoz, Cheng Qian, Hongru Wang et al.

Recent advances in Large Language Models (LLMs) have propelled conversational AI from traditional dialogue systems into sophisticated agents capable of autonomous actions, contextual awareness, and multi-turn interactions with users. Yet, fundamental questions about their capabilities, limitations, and paths forward remain open. This survey paper presents a desideratum for next-generation Conversational Agents - what has been achieved, what challenges persist, and what must be done for more scalable systems that approach human-level intelligence. To that end, we systematically analyze LLM-driven Conversational Agents by organizing their capabilities into three primary dimensions: (i) Reasoning - logical, systematic thinking inspired by human intelligence for decision making, (ii) Monitor - encompassing self-awareness and user interaction monitoring, and (iii) Control - focusing on tool utilization and policy following. Building upon this, we introduce a novel taxonomy by classifying recent work on Conversational Agents around our proposed desideratum. We identify critical research gaps and outline key directions, including realistic evaluations, long-term multi-turn reasoning skills, self-evolution capabilities, collaborative and multi-agent task completion, personalization, and proactivity. This work aims to provide a structured foundation, highlight existing limitations, and offer insights into potential future research directions for Conversational Agents, ultimately advancing progress toward Artificial General Intelligence (AGI). We maintain a curated repository of papers at: https://github.com/emrecanacikgoz/awesome-conversational-agents.

Emre Can Acikgoz, Cheng Qian, Jonas Hübotter et al.

Large language models (LLMs) are becoming the foundation for autonomous agents that can use tools to solve complex tasks. Reinforcement learning (RL) has emerged as a common approach for injecting such agentic capabilities, but typically under tightly controlled training setups. It often depends on carefully constructed task-solution pairs and substantial human supervision, which creates a fundamental obstacle to open-ended self-evolution toward superintelligent systems. In this paper, we propose Tool-R0 framework for training general purpose tool-calling agents from scratch with self-play RL, under a zero-data assumption. Initialized from the same base LLM, Tool-R0 co-evolves a Generator and a Solver with complementary rewards: one proposes targeted challenging tasks at the other's competence frontier and the other learns to solve them with real-world tool calls. This creates a self-evolving cycle that requires no pre-existing tasks or datasets. Evaluation on different tool-use benchmarks show that Tool-R0 yields 92.5 relative improvement over the base model and surpasses fully supervised tool-calling baselines under the same setting. Our work further provides empirical insights into self-play LLM agents by analyzing co-evolution, curriculum dynamics, and scaling behavior.

Cheng Qian, Lucas Glass, Nikos Sidiropoulos

Forecasting project expenses is a crucial step for businesses to avoid budget overruns and project failures. Traditionally, this has been done by financial analysts or data science techniques such as time-series analysis. However, these approaches can be uncertain and produce results that differ from the planned budget, especially at the start of a project with limited data points. This paper proposes a constrained non-negative matrix completion model that predicts expenses by learning the likelihood of the project correlating with certain expense patterns in the latent space. The model is constrained on three probability simplexes, two of which are on the factor matrices and the third on the missing entries. Additionally, the predicted expense values are guaranteed to meet the budget constraint without the need of post-processing. An inexact alternating optimization algorithm is developed to solve the associated optimization problem and is proven to converge to a stationary point. Results from two real datasets demonstrate the effectiveness of the proposed method in comparison to state-of-the-art algorithms.

Cheng Qian, Zuxin Liu, Akshara Prabhakar et al. · princeton

Reinforcement learning (RL) has shown promise in training agentic models that move beyond static benchmarks to engage in dynamic, multi-turn interactions. Yet, the ultimate value of such agents lies in their ability to assist users, a setting where diversity and dynamics of user interaction pose challenges. In this work, we propose UserRL, a unified framework for training and evaluating user-centric abilities through standardized gym environments paired with simulated users. We systematically vary turn-level reward assignment and trajectory-level score calculation to analyze how different formulations affect learning under the GRPO algorithm. Our experiments across Qwen3 models reveal three key findings: (i) SFT cold start is critical for unlocking initial interaction ability and enabling sustained RL improvements; (ii) deliberate trajectory scoring yields more efficient and effective multi-turn interactions; and (iii) while stronger simulated users (e.g., GPT-4o) facilitates training, open-source simulators (e.g., Qwen3-32B) remain a cost-effective and transferable option. Together, these results highlight that careful design of reward shaping and user simulation choice is as crucial as model scale, and establish UserRL as a practical pathway for developing robust user-centric agentic models. All codes and data are public for future research.

Qi He, Cheng Qian, Xiusi Chen et al.

Claim verification with large language models (LLMs) has recently attracted growing attention, due to their strong reasoning capabilities and transparent verification processes compared to traditional answer-only judgments. However, existing approaches to online claim verification, which requires iterative evidence retrieval and reasoning, still mainly rely on prompt engineering or pre-designed reasoning workflows, without unified training to improve necessary skills. Therefore, we introduce Veri-R1, an online reinforcement learning (RL) framework that enables an LLM to interact with a search engine and to receive reward signals that explicitly shape its planning, retrieval, and reasoning behaviors. This dynamic interaction of LLM with retrieval systems more accurately reflects real-world verification scenarios and fosters comprehensive verification skills. Empirical results show that Veri-R1 improves joint accuracy by up to 30% and doubles the evidence score, often surpassing its larger-scale model counterparts. Ablation studies further reveal the impact of reward components, and the link between output logits and label accuracy. Our results highlight the effectiveness of online RL for precise and faithful claim verification, providing an important foundation for future research. We release our code to support community progress in LLM empowered claim verification.

Jielin Qiu, Zuxin Liu, Zhiwei Liu et al.

The emergence of long-context language models with context windows extending to millions of tokens has created new opportunities for sophisticated code understanding and software development evaluation. We propose LoCoBench, a comprehensive benchmark specifically designed to evaluate long-context LLMs in realistic, complex software development scenarios. Unlike existing code evaluation benchmarks that focus on single-function completion or short-context tasks, LoCoBench addresses the critical evaluation gap for long-context capabilities that require understanding entire codebases, reasoning across multiple files, and maintaining architectural consistency across large-scale software systems. Our benchmark provides 8,000 evaluation scenarios systematically generated across 10 programming languages, with context lengths spanning 10K to 1M tokens, a 100x variation that enables precise assessment of long-context performance degradation in realistic software development settings. LoCoBench introduces 8 task categories that capture essential long-context capabilities: architectural understanding, cross-file refactoring, multi-session development, bug investigation, feature implementation, code comprehension, integration testing, and security analysis. Through a 5-phase pipeline, we create diverse, high-quality scenarios that challenge LLMs to reason about complex codebases at unprecedented scale. We introduce a comprehensive evaluation framework with 17 metrics across 4 dimensions, including 8 new evaluation metrics, combined in a LoCoBench Score (LCBS). Our evaluation of state-of-the-art long-context models reveals substantial performance gaps, demonstrating that long-context understanding in complex software development represents a significant unsolved challenge that demands more attention. LoCoBench is released at: https://github.com/SalesforceAIResearch/LoCoBench.

Xiusi Chen, Shanyong Wang, Cheng Qian et al.

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.

Guangwei Zhang, Jianing Zhu, Cheng Qian et al.

We present Copyright Detective, the first interactive forensic system for detecting, analyzing, and visualizing potential copyright risks in LLM outputs. The system treats copyright infringement versus compliance as an evidence discovery process rather than a static classification task due to the complex nature of copyright law. It integrates multiple detection paradigms, including content recall testing, paraphrase-level similarity analysis, persuasive jailbreak probing, and unlearning verification, within a unified and extensible framework. Through interactive prompting, response collection, and iterative workflows, our system enables systematic auditing of verbatim memorization and paraphrase-level leakage, supporting responsible deployment and transparent evaluation of LLM copyright risks even with black-box access.

Cheng Qian, Hyeonjeong Ha, Jiayu Liu et al.

Recent advances in large language models have led to strong performance on reasoning and environment-interaction tasks, yet their ability for creative problem-solving remains underexplored. We study this capability through the lens of creative tool use, where a model repurposes available objects by reasoning about their affordances and attributes rather than relying on canonical usage. As a first step, we introduce CreativityBench, a benchmark for evaluating affordance-based creativity in LLMs. To this end, we build a large-scale affordance knowledge base (KB) with 4K entities and 150K+ affordance annotations, explicitly linking objects, parts, attributes, and actionable uses. Building on this KB, we generate 14K grounded tasks that require identifying non-obvious yet physically plausible solutions under constraints. Evaluations across 10 state-of-the-art LLMs, including closed and open-source models, show that models can often select a plausible object, but fail to identify the correct parts, their affordances, and the underlying physical mechanism needed to solve the task, leading to a significant drop in performance. Furthermore, improvements from model scaling quickly saturate, strong general reasoning does not reliably translate to creative affordance discovery, and common inference-time strategies such as Chain-of-Thought yield limited gains. These results suggest that creative tool use remains a major challenge for current models, and that CreativityBench provides a useful testbed for studying this missing dimension of intelligence, with potential implications for planning and reasoning modules in future agents.

Guangwei Zhang, Qisheng Su, Jiateng Liu et al.

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. The implementation is available on GitHub.\footnote{https://github.com/changhu73/Internal_states_leakage}

Revanth Gangi Reddy, Tanay Dixit, Jiaxin Qin et al.

Wikipedia, a vast and continuously consulted knowledge base, faces significant challenges in maintaining up-to-date content due to its reliance on manual human editors. Inspired by the vision of continuous knowledge acquisition in NELL and fueled by advances in LLM-based agents, this paper introduces WiNELL, an agentic framework for continuously updating Wikipedia articles. Our approach employs a multi-agent framework to aggregate online information, select new and important knowledge for a target entity in Wikipedia, and then generate precise edit suggestions for human review. Our fine-grained editing models, trained on Wikipedia's extensive history of human edits, enable incorporating updates in a manner consistent with human editing behavior. Our editor models outperform both open-source instruction-following baselines and closed-source LLMs (e.g., GPT-4o) in key information coverage and editing efficiency. End-to-end evaluation on high-activity Wikipedia pages demonstrates WiNELL's ability to identify and suggest timely factual updates. This opens up a promising research direction in LLM agents for automatically updating knowledge bases in a never-ending fashion.

Cheng Qian, Emre Can Acikgoz, Qi He et al.

Current Large Language Models (LLMs) often undergo supervised fine-tuning (SFT) to acquire tool use capabilities. However, SFT struggles to generalize to unfamiliar or complex tool use scenarios. Recent advancements in reinforcement learning (RL), particularly with R1-like models, have demonstrated promising reasoning and generalization abilities. Yet, reward design for tool use presents unique challenges: multiple tools may be invoked with diverse parameters, and coarse-grained reward signals, such as answer matching, fail to offer the finegrained feedback required for effective learning. In this work, we present the first comprehensive study on reward design for tool selection and application tasks within the RL paradigm. We systematically explore a wide range of reward strategies, analyzing their types, scales, granularity, and temporal dynamics. Building on these insights, we propose a principled reward design tailored for tool use tasks and apply it to train LLMs using Group Relative Policy Optimization (GRPO). Empirical evaluations across diverse benchmarks demonstrate that our approach yields robust, scalable, and stable training, achieving a 17% improvement over base models and a 15% gain over SFT models. These results highlight the critical role of thoughtful reward design in enhancing the tool use capabilities and generalization performance of LLMs. All the codes are released to facilitate future research.

Duo Zhou, Yuji Zhang, Tianxin Wei et al.

Large language models (LLMs) can internalize private or harmful content, motivating unlearning that removes a forget set while preserving retaining knowledge. However, forgetting updates often cause collateral degradation on retaining knowledge, creating a persistent trade-off. Existing LLM unlearning methods are often heuristic, and other theoretical approaches rely on offline feature constructions that do not capture update-time forget-retain interaction in LLMs. To address this limitation, we aim to develop an LLM unlearning method that reduces the forget-retain trade-off with theoretical guarantees. We take a first-principles view by formalizing "no side effects" as local retain invariance under small parameter updates, and prove an equivalence under optimizer-induced geometry: the retain loss is locally invariant if and only if the update direction is orthogonal to the subspace spanned by retain gradients. Based on the insight, we propose Geometric-disentanglement Unlearning (GU), a lightweight and theoretically grounded projection that can be plug-and-play to existing gradient-based unlearning methods to mitigate forget-retain side effects. Experiments on TOFU, MUSE, and WMDP-cyber show that GU strengthens forgetting while reducing retain drift. When added to SimNPO, it achieves up to 62\% improved forgetting Extraction Strength (ES) and 31\% higher retain ES. We open-sourced our code in https://github.com/Lemutisme/Geometric-Unlearning.

Mingxuan Cui, Yilan Jiang, Duo Zhou et al.

Drug shortages pose critical risks to patient care and healthcare systems worldwide, yet the effectiveness of regulatory interventions remains poorly understood due to fundamental information asymmetries in pharmaceutical supply chains. We present \textbf{ShortageSim}, the first Large Language Model (LLM)-based multi-agent simulation framework that captures the complex, strategic interactions between drug manufacturers, institutional buyers, and regulatory agencies in response to shortage alerts. Unlike traditional game-theoretic models that assume perfect rationality and complete information, \textbf{ShortageSim} leverages LLMs to simulate bounded-rational decision-making under uncertainty. Through a sequential production game spanning multiple quarters, we model how FDA announcements, both reactive alerts about existing shortages and proactive warnings about potential disruptions, propagate through the supply chain and influence capacity investment and procurement decisions. Our experiments on historical shortage events reveal that \textbf{ShortageSim} reduces the resolution-lag percentage for discontinued-disclosed cases by 83\%, bringing simulated durations more aligned to ground truth than the zero-shot baseline. We open-source \textbf{ShortageSim} and a dataset of 2,925 FDA shortage events at https://github.com/Lemutisme/Sortage_Management, providing a novel computational framework for designing and testing interventions in complex, information-scarce supply chains.