Yikai Zhang

Qianyu He, Jie Zeng, Wenhao Huang et al.

Large language models (LLMs) can understand human instructions, showing their potential for pragmatic applications beyond traditional NLP tasks. However, they still struggle with complex instructions, which can be either complex task descriptions that require multiple tasks and constraints, or complex input that contains long context, noise, heterogeneous information and multi-turn format. Due to these features, LLMs often ignore semantic constraints from task descriptions, generate incorrect formats, violate length or sample count constraints, and be unfaithful to the input text. Existing benchmarks are insufficient to assess LLMs' ability to understand complex instructions, as they are close-ended and simple. To bridge this gap, we propose CELLO, a benchmark for evaluating LLMs' ability to follow complex instructions systematically. We design eight features for complex instructions and construct a comprehensive evaluation dataset from real-world scenarios. We also establish four criteria and develop corresponding metrics, as current ones are inadequate, biased or too strict and coarse-grained. We compare the performance of representative Chinese-oriented and English-oriented models in following complex instructions through extensive experiments. Resources of CELLO are publicly available at https://github.com/Abbey4799/CELLO.

Saumya Gupta, Yikai Zhang, Xiaoling Hu et al.

Segmentation of curvilinear structures such as vasculature and road networks is challenging due to relatively weak signals and complex geometry/topology. To facilitate and accelerate large scale annotation, one has to adopt semi-automatic approaches such as proofreading by experts. In this work, we focus on uncertainty estimation for such tasks, so that highly uncertain, and thus error-prone structures can be identified for human annotators to verify. Unlike most existing works, which provide pixel-wise uncertainty maps, we stipulate it is crucial to estimate uncertainty in the units of topological structures, e.g., small pieces of connections and branches. To achieve this, we leverage tools from topological data analysis, specifically discrete Morse theory (DMT), to first capture the structures, and then reason about their uncertainties. To model the uncertainty, we (1) propose a joint prediction model that estimates the uncertainty of a structure while taking the neighboring structures into consideration (inter-structural uncertainty); (2) propose a novel Probabilistic DMT to model the inherent uncertainty within each structure (intra-structural uncertainty) by sampling its representations via a perturb-and-walk scheme. On various 2D and 3D datasets, our method produces better structure-wise uncertainty maps compared to existing works. Code available at https://github.com/Saumya-Gupta-26/struct-uncertainty

Minwei Kong, Chonghe Jiang, Ao Qu et al.

Large language models (LLMs) are increasingly used for optimization modeling and solver-code generation, yet practical operations research and optimization problems often require a harder capability: designing scalable algorithms that exploit problem structure and outperform direct formulation-and-solve baselines. Existing benchmarks are limited to small or simplified examples far below real-world scale and complexity. We introduce FrontierOR, among the first benchmarks to systematically evaluate LLM-based efficient algorithm design for realistic large-scale optimization problems. FrontierOR includes 180 tasks derived from methodologically diverse papers published in top-tier operations research venues, each with standardized instances and a hidden, expert-verified evaluation suite. We evaluate seven LLMs spanning frontier, cost-effective, and open-source models both in one-shot and test-time evolution settings. The results reveal that frontier models still struggle to move from executable formulations to efficient optimization algorithms: the strongest one-shot model outperforms Gurobi in only 31% of cases in both solution quality and computational efficiency, and even strong coding agents with test-time evolution achieve only 50% on selected hard tasks. FrontierOR establishes a practical evaluation platform for LLM-based optimization algorithm design, which enables future LLMs and agents to be systematically tested on whether they can move beyond correct formulation toward a feasible, high-quality, and efficient algorithm. Our FrontierOR Benchmark is available at https://anonymous.4open.science/r/efficient-opt-bench-F03D.

Wenjia Zhang, Yikai Zhang, Xiaoling Hu et al.

The adversarial risk of a machine learning model has been widely studied. Most previous works assume that the data lies in the whole ambient space. We propose to take a new angle and take the manifold assumption into consideration. Assuming data lies in a manifold, we investigate two new types of adversarial risk, the normal adversarial risk due to perturbation along normal direction, and the in-manifold adversarial risk due to perturbation within the manifold. We prove that the classic adversarial risk can be bounded from both sides using the normal and in-manifold adversarial risks. We also show with a surprisingly pessimistic case that the standard adversarial risk can be nonzero even when both normal and in-manifold risks are zero. We finalize the paper with empirical studies supporting our theoretical results. Our results suggest the possibility of improving the robustness of a classifier by only focusing on the normal adversarial risk.

Jiachen Yao, Yikai Zhang, Songzhu Zheng et al.

Noisy labels can significantly affect the performance of deep neural networks (DNNs). In medical image segmentation tasks, annotations are error-prone due to the high demand in annotation time and in the annotators' expertise. Existing methods mostly assume noisy labels in different pixels are \textit{i.i.d}. However, segmentation label noise usually has strong spatial correlation and has prominent bias in distribution. In this paper, we propose a novel Markov model for segmentation noisy annotations that encodes both spatial correlation and bias. Further, to mitigate such label noise, we propose a label correction method to recover true label progressively. We provide theoretical guarantees of the correctness of the proposed method. Experiments show that our approach outperforms current state-of-the-art methods on both synthetic and real-world noisy annotations.

Yikai Zhang, Jiachen Yao, Yusu Wang et al.

Topological loss based on persistent homology has shown promise in various applications. A topological loss enforces the model to achieve certain desired topological property. Despite its empirical success, less is known about the optimization behavior of the loss. In fact, the topological loss involves combinatorial configurations that may oscillate during optimization. In this paper, we introduce a general purpose regularized topology-aware loss. We propose a novel regularization term and also modify existing topological loss. These contributions lead to a new loss function that not only enforces the model to have desired topological behavior, but also achieves satisfying convergence behavior. Our main theoretical result guarantees that the loss can be optimized efficiently, under mild assumptions.

Qianyu He, Yikai Zhang, Jiaqing Liang et al.

Similes play an imperative role in creative writing such as story and dialogue generation. Proper evaluation metrics are like a beacon guiding the research of simile generation (SG). However, it remains under-explored as to what criteria should be considered, how to quantify each criterion into metrics, and whether the metrics are effective for comprehensive, efficient, and reliable SG evaluation. To address the issues, we establish HAUSER, a holistic and automatic evaluation system for the SG task, which consists of five criteria from three perspectives and automatic metrics for each criterion. Through extensive experiments, we verify that our metrics are significantly more correlated with human ratings from each perspective compared with prior automatic metrics.

Sirui Xia, Yikai Zhang, Aili Chen et al.

Discovering improved policy optimization algorithms for language models remains a costly manual process requiring repeated mechanism-level modification and validation. Unlike simple combinatorial code search, this problem requires searching over algorithmic mechanisms tightly coupled with training dynamics while reusing empirical evidence across iterations. We propose POISE, a closed-loop framework for automated discovery of policy optimization algorithms for language models. POISE maintains a structured, genealogically linked archive linking proposals, executable implementations, standardized evaluations, and natural-language reflections to support evidence-driven iteration. In mathematical reasoning experiments starting from GRPO, POISE evaluates 64 candidate algorithms and discovers improved mechanisms, including analytic-variance scaling and validity masking. The best variant improves weighted Overall from 47.8 to 52.5 (+4.6) and increases AIME25 pass@32 from 26.7% to 43.3%, demonstrating the feasibility of automated policy optimization discovery while supporting interpretable design principles.

Siye Wu, Jian Xie, Yikai Zhang et al.

The emergence of large reasoning models demonstrates that scaling inference-time compute significantly enhances performance on complex tasks. However, it often falls into another trap: overthinking simple problems, where repetitive rationales yield minimal accuracy gains at a disproportionately high cost. This motivates adaptive reasoning: dynamically aligning reasoning depth with instance difficulty. In this paper, we study adaptive reasoning from an optimality perspective, formalizing it as a utility maximization problem where tokens are allocated until the marginal accuracy gain falls below the incremental cost. Based on this, we propose CODA (Compute Allocation by Difficulty Awareness), a method that operationalizes this principle by allocating tokens via a policy-internal difficulty signal. Specifically, CODA estimates difficulty via group-based rollouts and maps it to two non-negative gates that modulate a length-dependent shaping term on top of the binary base reward. The easy-side gate penalizes verbosity on simple instances, whereas the hard-side gate encourages more deliberative rollouts on challenging ones. Across model scales and benchmarks, CODA achieves adaptive reasoning without external annotations or user-provided budgets: on easy tasks, CODA reduces token costs by over 60% while maintaining strong accuracy, whereas on hard tasks it incentivizes more deliberative rollouts to maximize performance.

Yikai Zhang, Junlong Li, Pengfei Liu

Large Language Models (LLMs) are known to have limited extrapolation ability beyond their pre-trained context window, constraining their application in downstream tasks with lengthy inputs. Recent studies have sought to extend LLMs' context window by modifying rotary position embedding (RoPE), a popular position encoding method adopted by well-known LLMs such as LLaMA, PaLM, and GPT-NeoX. However, prior works like Position Interpolation (PI) and YaRN are resource-intensive and lack comparative experiments to assess their applicability. In this work, we identify the inherent need for LLMs' attention entropy (i.e. the information entropy of attention scores) to maintain stability and introduce a novel extension to RoPE which combines adjusting RoPE's base frequency and scaling the attention logits to help LLMs efficiently adapt to a larger context window. We validate the superiority of our method in both fine-tuning performance and robustness across different context window sizes on various context-demanding tasks. Notably, our method extends the context window of LLaMA-2-7B-Chat to 16,384 with only 100 samples and 6 training steps, showcasing extraordinary efficiency. Finally, we also explore how data compositions and training curricula affect context window extension for specific downstream tasks, suggesting fine-tuning LLMs with lengthy conversations as a good starting point. We release our code and SFT data at https://github.com/GAIR-NLP/Entropy-ABF.

Yikai Zhang, Jiaxin Pei, Kenan Li et al.

Large language model agents have achieved remarkable progress on software engineering tasks, yet current approaches suffer from a fundamental context coupling problem: the standard code editing interface conflates code inspection, modification planning, and edit execution within a single context window, forcing agents to interleave exploratory viewing with strictly formatted edit generation. This causes irrelevant information to accumulate and degrades agent performance. To address this, we propose SWE-Edit, which decomposes code editing into two specialized subagents: a Viewer that extracts task-relevant code on demand, and an Editor that executes modifications from high-level plans--allowing the main agent to focus on reasoning while delegating context-intensive operations to clean context windows. We further investigate what makes an effective editing model: observing that the prevalent find-and-replace format is error-prone, we train Qwen3-8B with GRPO to adaptively select editing modes, yielding improved editing efficiency over single-format baselines. On SWE-bench Verified, SWE-Edit improves resolved rate by 2.1% while reducing inference cost by 17.9%. We additionally propose a code editing benchmark that reliably predicts downstream agentic performance, providing practical guidance for editing model selection. Our code is publicly available at https://github.com/microsoft/SWE-Edit.

Junlong Li, Fan Zhou, Shichao Sun et al.

As a relative quality comparison of model responses, human and Large Language Model (LLM) preferences serve as common alignment goals in model fine-tuning and criteria in evaluation. Yet, these preferences merely reflect broad tendencies, resulting in less explainable and controllable models with potential safety risks. In this work, we dissect the preferences of human and 32 different LLMs to understand their quantitative composition, using annotations from real-world user-model conversations for a fine-grained, scenario-wise analysis. We find that humans are less sensitive to errors, favor responses that support their stances, and show clear dislike when models admit their limits. On the contrary, advanced LLMs like GPT-4-Turbo emphasize correctness, clarity, and harmlessness more. Additionally, LLMs of similar sizes tend to exhibit similar preferences, regardless of their training methods, and fine-tuning for alignment does not significantly alter the preferences of pretrained-only LLMs. Finally, we show that preference-based evaluation can be intentionally manipulated. In both training-free and training-based settings, aligning a model with the preferences of judges boosts scores, while injecting the least preferred properties lowers them. This results in notable score shifts: up to 0.59 on MT-Bench (1-10 scale) and 31.94 on AlpacaEval 2.0 (0-100 scale), highlighting the significant impact of this strategic adaptation. Interactive Demo: https://huggingface.co/spaces/GAIR/Preference-Dissection-Visualization Dataset: https://huggingface.co/datasets/GAIR/preference-dissection Code: https://github.com/GAIR-NLP/Preference-Dissection

Yikai Zhang, Songzhu Zheng, Mina Dalirrooyfard et al.

Learning and decision-making in domains with naturally high noise-to-signal ratio, such as Finance or Healthcare, is often challenging, while the stakes are very high. In this paper, we study the problem of learning and acting under a general noisy generative process. In this problem, the data distribution has a significant proportion of uninformative samples with high noise in the label, while part of the data contains useful information represented by low label noise. This dichotomy is present during both training and inference, which requires the proper handling of uninformative data during both training and testing. We propose a novel approach to learning under these conditions via a loss inspired by the selective learning theory. By minimizing this loss, the model is guaranteed to make a near-optimal decision by distinguishing informative data from uninformative data and making predictions. We build upon the strength of our theoretical guarantees by describing an iterative algorithm, which jointly optimizes both a predictor and a selector, and evaluates its empirical performance in a variety of settings.

Yang Liu, Enxi Wang, Yufei Gao et al.

Despite the success of reinforcement learning for large language models, a common failure mode is reduced sampling diversity, where the policy repeatedly generates similar erroneous behaviors. Classical entropy regularization encourages randomness under the current policy, but does not explicitly discourage recurrent failure patterns across rollouts. We propose MEDS, a Memory-Enhanced Dynamic reward Shaping framework that incorporates historical behavioral signals into reward design. By storing and leveraging intermediate model representations, we capture features of past rollouts and use density-based clustering to identify frequently recurring error patterns. Rollouts assigned to more prevalent error clusters are penalized more heavily, encouraging broader exploration while reducing repeated mistakes. Across five datasets and three base models, MEDS consistently improves average performance over existing baselines, achieving gains of up to 4.13 pass@1 points and 4.37 pass@128 points. Additional analyses using both LLM-based annotations and quantitative diversity metrics show that MEDS increases behavioral diversity during sampling.

Jiangjie Chen, Xintao Wang, Rui Xu et al.

Recent advancements in large language models (LLMs) have significantly boosted the rise of Role-Playing Language Agents (RPLAs), i.e., specialized AI systems designed to simulate assigned personas. By harnessing multiple advanced abilities of LLMs, including in-context learning, instruction following, and social intelligence, RPLAs achieve a remarkable sense of human likeness and vivid role-playing performance. RPLAs can mimic a wide range of personas, ranging from historical figures and fictional characters to real-life individuals. Consequently, they have catalyzed numerous AI applications, such as emotional companions, interactive video games, personalized assistants and copilots, and digital clones. In this paper, we conduct a comprehensive survey of this field, illustrating the evolution and recent progress in RPLAs integrating with cutting-edge LLM technologies. We categorize personas into three types: 1) Demographic Persona, which leverages statistical stereotypes; 2) Character Persona, focused on well-established figures; and 3) Individualized Persona, customized through ongoing user interactions for personalized services. We begin by presenting a comprehensive overview of current methodologies for RPLAs, followed by the details for each persona type, covering corresponding data sourcing, agent construction, and evaluation. Afterward, we discuss the fundamental risks, existing limitations, and future prospects of RPLAs. Additionally, we provide a brief review of RPLAs in AI applications, which reflects practical user demands that shape and drive RPLA research. Through this work, we aim to establish a clear taxonomy of RPLA research and applications, and facilitate future research in this critical and ever-evolving field, and pave the way for a future where humans and RPLAs coexist in harmony.

Yiran Song, Yikai Zhang, Shuang Zhou et al.

Multiple instance learning (MIL) has emerged as the dominant paradigm for whole slide image (WSI) analysis in computational pathology, achieving strong diagnostic performance through patch-level feature aggregation. However, existing MIL methods face critical limitations: (1) they rely on attention mechanisms that lack causal interpretability, and (2) they fail to integrate patient demographics (age, gender, race), leading to fairness concerns across diverse populations. These shortcomings hinder clinical translation, where algorithmic bias can exacerbate health disparities. We introduce \textbf{MeCaMIL}, a causality-aware MIL framework that explicitly models demographic confounders through structured causal graphs. Unlike prior approaches treating demographics as auxiliary features, MeCaMIL employs principled causal inference -- leveraging do-calculus and collider structures -- to disentangle disease-relevant signals from spurious demographic correlations. Extensive evaluation on three benchmarks demonstrates state-of-the-art performance across CAMELYON16 (ACC/AUC/F1: 0.939/0.983/0.946), TCGA-Lung (0.935/0.979/0.931), and TCGA-Multi (0.977/0.993/0.970, five cancer types). Critically, MeCaMIL achieves superior fairness -- demographic disparity variance drops by over 65% relative reduction on average across attributes, with notable improvements for underserved populations. The framework generalizes to survival prediction (mean C-index: 0.653, +0.017 over best baseline across five cancer types). Ablation studies confirm causal graph structure is essential -- alternative designs yield 0.048 lower accuracy and 4.2x times worse fairness. These results establish MeCaMIL as a principled framework for fair, interpretable, and clinically actionable AI in digital pathology. Code will be released upon acceptance.

Zhen Huang, Zengzhi Wang, Shijie Xia et al.

The evolution of Artificial Intelligence (AI) has been significantly accelerated by advancements in Large Language Models (LLMs) and Large Multimodal Models (LMMs), gradually showcasing potential cognitive reasoning abilities in problem-solving and scientific discovery (i.e., AI4Science) once exclusive to human intellect. To comprehensively evaluate current models' performance in cognitive reasoning abilities, we introduce OlympicArena, which includes 11,163 bilingual problems across both text-only and interleaved text-image modalities. These challenges encompass a wide range of disciplines spanning seven fields and 62 international Olympic competitions, rigorously examined for data leakage. We argue that the challenges in Olympic competition problems are ideal for evaluating AI's cognitive reasoning due to their complexity and interdisciplinary nature, which are essential for tackling complex scientific challenges and facilitating discoveries. Beyond evaluating performance across various disciplines using answer-only criteria, we conduct detailed experiments and analyses from multiple perspectives. We delve into the models' cognitive reasoning abilities, their performance across different modalities, and their outcomes in process-level evaluations, which are vital for tasks requiring complex reasoning with lengthy solutions. Our extensive evaluations reveal that even advanced models like GPT-4o only achieve a 39.97% overall accuracy, illustrating current AI limitations in complex reasoning and multimodal integration. Through the OlympicArena, we aim to advance AI towards superintelligence, equipping it to address more complex challenges in science and beyond. We also provide a comprehensive set of resources to support AI research, including a benchmark dataset, an open-source annotation platform, a detailed evaluation tool, and a leaderboard with automatic submission features.

Yikai Zhang, Siyu Yuan, Caiyu Hu et al.

Despite remarkable advancements in emulating human-like behavior through Large Language Models (LLMs), current textual simulations do not adequately address the notion of time. To this end, we introduce TimeArena, a novel textual simulated environment that incorporates complex temporal dynamics and constraints that better reflect real-life planning scenarios. In TimeArena, agents are asked to complete multiple tasks as soon as possible, allowing for parallel processing to save time. We implement the dependency between actions, the time duration for each action, and the occupancy of the agent and the objects in the environment. TimeArena grounds to 30 real-world tasks in cooking, household activities, and laboratory work. We conduct extensive experiments with various state-of-the-art LLMs using TimeArena. Our findings reveal that even the most powerful models, e.g., GPT-4, still lag behind humans in effective multitasking, underscoring the need for enhanced temporal awareness in the development of language agents.

Jian Xie, Kexun Zhang, Jiangjie Chen et al.

Autonomous planning has been an ongoing pursuit since the inception of artificial intelligence. Based on curated problem solvers, early planning agents could deliver precise solutions for specific tasks but lacked generalization. The emergence of large language models (LLMs) and their powerful reasoning capabilities has reignited interest in autonomous planning by automatically generating reasonable solutions for given tasks. However, prior research and our experiments show that current language agents still lack human-level planning abilities. Even the state-of-the-art reasoning model, OpenAI o1, achieves only 15.6% on one of the complex real-world planning benchmarks. This highlights a critical question: What hinders language agents from achieving human-level planning? Although existing studies have highlighted weak performance in agent planning, the deeper underlying issues and the mechanisms and limitations of the strategies proposed to address them remain insufficiently understood. In this work, we apply the feature attribution study and identify two key factors that hinder agent planning: the limited role of constraints and the diminishing influence of questions. We also find that although current strategies help mitigate these challenges, they do not fully resolve them, indicating that agents still have a long way to go before reaching human-level intelligence.

Ruihan Yang, Fanghua Ye, Jian Li et al.

Large language models (LLMs) have recently transformed from text-based assistants to autonomous agents capable of planning, reasoning, and iteratively improving their actions. While numerical reward signals and verifiers can effectively rank candidate actions, they often provide limited contextual guidance. In contrast, natural language feedback better aligns with the generative capabilities of LLMs, providing richer and more actionable suggestions. However, parsing and implementing this feedback effectively can be challenging for LLM-based agents. In this work, we introduce Critique-Guided Improvement (CGI), a novel two-player framework, comprising an actor model that explores an environment and a critic model that generates detailed nature language feedback. By training the critic to produce fine-grained assessments and actionable revisions, and the actor to utilize these critiques, our approach promotes more robust exploration of alternative strategies while avoiding local optima. Experiments in three interactive environments show that CGI outperforms existing baselines by a substantial margin. Notably, even a small critic model surpasses GPT-4 in feedback quality. The resulting actor achieves state-of-the-art performance, demonstrating the power of explicit iterative guidance to enhance decision-making in LLM-based agents.

Zefan Cai, Wen Xiao, Hanshi Sun et al. · microsoft-research

Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 16% of the KV cache. This KV-cache reduction also leads to a 90% memory saving and a 6.6X throughput over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.

Rong Ye, Yongxin Zhang, Yikai Zhang et al. · bytedance

Achieving Artificial General Intelligence (AGI) requires AI agents that can not only make stratigic decisions but also engage in flexible and meaningful communication. Inspired by Wittgenstein's language game theory in Philosophical Investigations, we propose that language agents can learn through in-context interaction rather than traditional multi-stage frameworks that separate decision-making from language expression. Using Werewolf, a social deduction game that tests language understanding, strategic interaction, and adaptability, we develop the Multi-agent Kahneman & Tversky's Optimization (MaKTO). MaKTO engages diverse models in extensive gameplay to generate unpaired desirable and unacceptable responses, then employs KTO to refine the model's decision-making process. In 9-player Werewolf games, MaKTO achieves a 61% average win rate across various models, outperforming GPT-4o and two-stage RL agents by relative improvements of 23.0% and 10.9%, respectively. Notably, MaKTO also demonstrates human-like performance, winning 60% against expert players and showing only 49% detectability in Turing-style blind tests.

Kenan Li, Qirui Jin, Liao Zhu et al.

Recent advances in language model (LM) agents have significantly improved automated software engineering (SWE). Prior work has proposed various agentic workflows and training strategies as well as analyzed failure modes of agentic systems on SWE tasks, focusing on several contextual information signals: Reproduction Test, Regression Test, Edit Location, Execution Context, and API Usage. However, the individual contribution of each signal to overall success remains underexplored, particularly their ideal contribution when intermediate information is perfectly obtained. To address this gap, we introduce Oracle-SWE, a unified method to isolate and extract oracle information signals from SWE benchmarks and quantify the impact of each signal on agent performance. To further validate the pattern, we evaluate the performance gain of signals extracted by strong LMs when provided to a base agent, approximating real-world task-resolution settings. These evaluations aim to guide research prioritization for autonomous coding systems.

Ruihan Yang, Yikai Zhang, Aili Chen et al.

Large language models (LLMs) have enabled agents to perform complex reasoning and decision-making through free-form language interactions. However, in open-ended language action environments (e.g., negotiation or question-asking games), the action space can be formulated as a joint distribution over tokens, resulting in an exponentially large action space. Sampling actions in such a space can lead to extreme reward sparsity, which brings large reward variance, hindering effective reinforcement learning (RL). To address this, we propose ARIA, a method that Aggregates Rewards in Intention space to enable efficient and effective language Agents training. ARIA aims to project natural language actions from the high-dimensional joint token distribution space into a low-dimensional intention space, where semantically similar actions are clustered and assigned shared rewards. This intention-aware reward aggregation reduces reward variance by densifying reward signals, fostering better policy optimization. Extensive experiments demonstrate that ARIA not only significantly reduces policy gradient variance, but also delivers substantial performance gains of an average of 9.95% across four downstream tasks, consistently outperforming offline and online RL baselines.

Sirui Xia, Aili Chen, Xintao Wang et al.

Recent advances in Large Language Models (LLMs) have demonstrated strong capabilities in tasks such as code and mathematics. However, their potential to internalize structured spatial knowledge remains underexplored. This study investigates whether LLMs, grounded in locally relative human observations, can construct coherent global spatial cognition by integrating fragmented relational descriptions. We focus on two core aspects of spatial cognition: spatial perception, where models infer consistent global layouts from local positional relationships, and spatial navigation, where models learn road connectivity from trajectory data and plan optimal paths between unconnected locations. Experiments conducted in a simulated urban environment demonstrate that LLMs not only generalize to unseen spatial relationships between points of interest (POIs) but also exhibit latent representations aligned with real-world spatial distributions. Furthermore, LLMs can learn road connectivity from trajectory descriptions, enabling accurate path planning and dynamic spatial awareness during navigation.

Aili Chen, Chengyu Du, Jiangjie Chen et al.

To advance personalized applications such as recommendation systems and user behavior prediction, recent research increasingly adopts large language models (LLMs) for human -readable persona modeling. In dynamic real -world scenarios, effective persona modeling necessitates leveraging streaming behavior data to continually optimize user personas. However, existing methods -whether regenerating personas or incrementally extending them with new behaviors -often fail to achieve sustained improvements in persona quality or future behavior prediction accuracy. To address this, we propose DEEPER, a novel approach for dynamic persona modeling that enables continual persona optimization. Specifically, we enhance the model's direction -search capability through an iterative reinforcement learning framework, allowing it to automatically identify effective update directions and optimize personas using discrepancies between user behaviors and model predictions. Extensive experiments on dynamic persona modeling involving 4800 users across 10 domains highlight the superior persona optimization capabilities of DEEPER, delivering an impressive 32.2% average reduction in user behavior prediction error over four update rounds -outperforming the best baseline by a remarkable 22.92%.

Lida Chen, Dong Xu, Chenxin An et al.

Large Language Models (LLMs) face efficiency bottlenecks due to the quadratic complexity of the attention mechanism when processing long contexts. Sparse attention methods offer a promising solution, but existing approaches often suffer from incomplete effective context and/or require complex implementation of pipeline. We present a comprehensive analysis of sparse attention for autoregressive LLMs from the respective of receptive field, recognize the suboptimal nature of existing methods for expanding the receptive field, and introduce PowerAttention, a novel sparse attention design that facilitates effective and complete context extension through the theoretical analysis. PowerAttention achieves exponential receptive field growth in $d$-layer LLMs, allowing each output token to attend to $2^d$ tokens, ensuring completeness and continuity of the receptive field. Experiments demonstrate that PowerAttention outperforms existing static sparse attention methods by $5\sim 40\%$, especially on tasks demanding long-range dependencies like Passkey Retrieval and RULER, while maintaining a comparable time complexity to sliding window attention. Efficiency evaluations further highlight PowerAttention's superior speedup in both prefilling and decoding phases compared with dynamic sparse attentions and full attention ($3.0\times$ faster on 128K context), making it a highly effective and user-friendly solution for processing long sequences in LLMs.

Yikai Zhang, Jiahe Lin, Fengpei Li et al.

In this work, we study the weighted empirical risk minimization (weighted ERM) schema, in which an additional data-dependent weight function is incorporated when the empirical risk function is being minimized. We show that under a general ``balanceable" Bernstein condition, one can design a weighted ERM estimator to achieve superior performance in certain sub-regions over the one obtained from standard ERM, and the superiority manifests itself through a data-dependent constant term in the error bound. These sub-regions correspond to large-margin ones in classification settings and low-variance ones in heteroscedastic regression settings, respectively. Our findings are supported by evidence from synthetic data experiments.

Kenan Li, Rongzhi Li, Linghao Zhang et al.

Building software repositories typically requires significant manual effort. Recent advances in large language model (LLM) agents have accelerated automation in software engineering (SWE). We introduce RepoLaunch, the first agent capable of automatically resolving dependencies, compiling source code, and extracting test results for repositories across arbitrary programming languages and operating systems. To demonstrate its utility, we further propose a fully automated pipeline for SWE dataset creation, where task design is the only human intervention. RepoLaunch automates the remaining steps, enabling scalable benchmarking and training of coding agents and LLMs. Notably, several works on agentic benchmarking and training have recently adopted RepoLaunch for automated task generation.

Yikai Zhang, Ye Rong, Siyu Yuan et al.

Existing language agents often encounter difficulties in dynamic adversarial games due to poor strategic reasoning. To mitigate this limitation, a promising approach is to allow agents to learn from game interactions automatically, without relying on costly expert-labeled data. Unlike static environments where agents receive fixed feedback or rewards, selecting appropriate opponents in dynamic adversarial games can significantly impact learning performance. However, the discussion of opponents in adversarial environments remains an area under exploration. In this paper, we propose a Step-level poliCy Optimization method through Play-And-Learn, SCO-PAL. Leveraging SCO-PAL, we conduct a detailed analysis of opponent selection by setting opponents at different levels and find that self-play is the most effective way to improve strategic reasoning in such adversarial environments. Utilizing SCO-PAL with self-play, we increase the average win rate against four opponents by approximately 30% compared to baselines and achieve a 54.76% win rate against GPT-4 in six adversarial games.

Hankun Dai, Maoquan Wang, Mengnan Qi et al.

Large language models (LLMs) are increasingly being applied to programming tasks, ranging from single-turn code completion to autonomous agents. Current code agent designs frequently depend on complex, hand-crafted workflows and tool sets. However, this reliance on elaborate scaffolding presents several challenges: agent performance becomes overly dependent on prompt tuning and custom design choices, heavy human intervention obscures a model's true underlying capabilities, and intricate pipelines are costly to build and maintain. Furthermore, optimizing complex task prompts increases the risk of data leakage. Currently, when introducing new models, LLM providers like OpenAI and Anthropic often publish benchmark scores to demonstrate their models' coding proficiency, but keep their proprietary evaluation frameworks confidential. To address these limitations, we introduce Lita (Lite Agent), which operationalizes liteness, a principle of minimizing manual design while retaining the essential elements of a fully autonomous agent. Lita enables a more faithful and unified evaluation without elaborate scaffolding. Experiments on the Aider Polyglot and SWE-Bench with frontier models demonstrate that Lita achieves competitive or superior performance compared to workflow-based and agentic baselines. Crucially, Lita also consumes fewer tokens and requires significantly less design effort. Our results suggest that Lita is sufficient to reveal the underlying coding competence of modern LLMs. Finally, we propose the Agent Complexity Law: the performance gap between agents of varying complexity, from simple to sophisticated designs, will shrink as the core model improves, ultimately converging to a negligible difference.

Yiran Song, Yikai Zhang, Silvia Orengo-Nania et al.

Glaucoma is one of the leading causes of irreversible blindness worldwide. Glaucoma prognosis is essential for identifying at-risk patients and enabling timely intervention to prevent blindness. Many existing approaches rely on historical sequential data but are constrained by fixed-length inputs, limiting their flexibility. Additionally, traditional glaucoma prognosis methods often employ end-to-end models, which struggle with the limited size of glaucoma datasets. To address these challenges, we propose a Two-Stage Decoupling Framework (TSDF) for variable-length glaucoma prognosis. In the first stage, we employ a feature representation module that leverages self-supervised learning to aggregate multiple glaucoma datasets for training, disregarding differences in their supervisory information. This approach enables datasets of varying sizes to learn better feature representations. In the second stage, we introduce a temporal aggregation module that incorporates an attention-based mechanism to process sequential inputs of varying lengths, ensuring flexible and efficient utilization of all available data. This design significantly enhances model performance while maintaining a compact parameter size. Extensive experiments on two benchmark glaucoma datasets:the Ocular Hypertension Treatment Study (OHTS) and the Glaucoma Real-world Appraisal Progression Ensemble (GRAPE),which differ significantly in scale and clinical settings,demonstrate the effectiveness and robustness of our approach.

Qifeng Tan, Shusen Yang, Xuebin Ren et al.

Layer-wise Gaussian mechanisms (LGM) enhance flexibility in differentially private deep learning by injecting noise into partitioned gradient vectors. However, existing methods often rely on heuristic noise allocation strategies, lacking a rigorous understanding of their theoretical grounding in connecting noise allocation to formal privacy-utility tradeoffs. In this paper, we present a unified analytical framework that systematically connects layer-wise noise injection strategies with their implicit optimization objectives and associated privacy budget allocations. Our analysis reveals that several existing approaches optimize ill-posed objectives -- either ignoring inter-layer signal-to-noise ratio (SNR) consistency or leading to inefficient use of the privacy budget. In response, we propose a SNR-Consistent noise allocation strategy that unifies both aspects, yielding a noise allocation scheme that achieves better signal preservation and more efficient privacy budget utilization. Extensive experiments in both centralized and federated learning settings demonstrate that our method consistently outperforms existing allocation strategies, achieving better privacy-utility tradeoffs. Our framework not only offers diagnostic insights into prior methods but also provides theoretical guidance for designing adaptive and effective noise injection schemes in deep models.

Jiahe Lin, Yikai Zhang, George Michailidis

Graphical models are widely used in diverse application domains to model the conditional dependencies amongst a collection of random variables. In this paper, we consider settings where the graph structure is covariate-dependent, and investigate a deep neural network-based approach to estimate it. The method allows for flexible functional dependency on the covariate, and fits the data reasonably well in the absence of a Gaussianity assumption. Theoretical results with PAC guarantees are established for the method, under assumptions commonly used in an Empirical Risk Minimization framework. The performance of the proposed method is evaluated on several synthetic data settings and benchmarked against existing approaches. The method is further illustrated on real datasets involving data from neuroscience and finance, respectively, and produces interpretable results.

Caiyu Hu, Yikai Zhang, Tinghui Zhu et al.

Multimodal Large Language Models (MLLMs) have advanced in integrating diverse modalities but frequently suffer from hallucination. A promising solution to mitigate this issue is to generate text with citations, providing a transparent chain for verification. However, existing work primarily focuses on generating citations for text-only content, leaving the challenges of multimodal scenarios largely unexplored. In this paper, we introduce MCiteBench, the first benchmark designed to assess the ability of MLLMs to generate text with citations in multimodal contexts. Our benchmark comprises data derived from academic papers and review-rebuttal interactions, featuring diverse information sources and multimodal content. Experimental results reveal that MLLMs struggle to ground their outputs reliably when handling multimodal input. Further analysis uncovers a systematic modality bias and reveals how models internally rely on different sources when generating citations, offering insights into model behavior and guiding future directions for multimodal citation tasks.

Zhouhong Gu, Lin Zhang, Xiaoxuan Zhu et al.

Detecting evidence within the context is a key step in the process of reasoning task. Evaluating and enhancing the capabilities of LLMs in evidence detection will strengthen context-based reasoning performance. This paper proposes a benchmark called DetectBench for verifying the ability to detect and piece together implicit evidence within a long context. DetectBench contains 3,928 multiple-choice questions, with an average of 994 tokens per question. Each question contains an average of 4.55 pieces of implicit evidence, and solving the problem typically requires 7.62 logical jumps to find the correct answer. To enhance the performance of LLMs in evidence detection, this paper proposes Detective Reasoning Prompt and Finetune. Experiments demonstrate that the existing LLMs' abilities to detect evidence in long contexts are far inferior to humans. However, the Detective Reasoning Prompt effectively enhances the capability of powerful LLMs in evidence detection, while the Finetuning method shows significant effects in enhancing the performance of weaker LLMs. Moreover, when the abilities of LLMs in evidence detection are improved, their final reasoning performance is also enhanced accordingly.

Yikai Zhang, Qianyu He, Xintao Wang et al.

Multi-Modal Knowledge Graphs (MMKGs) have proven valuable for various downstream tasks. However, scaling them up is challenging because building large-scale MMKGs often introduces mismatched images (i.e., noise). Most entities in KGs belong to the long tail, meaning there are few images of them available online. This scarcity makes it difficult to determine whether a found image matches the entity. To address this, we draw on the Triangle of Reference Theory and suggest enhancing vision-language models with concept guidance. Specifically, we introduce COG, a two-stage framework with COncept-Guided vision-language models. The framework comprises a Concept Integration module, which effectively identifies image-text pairs of long-tailed entities, and an Evidence Fusion module, which offers explainability and enables human verification. To demonstrate the effectiveness of COG, we create a dataset of 25k image-text pairs of long-tailed entities. Our comprehensive experiments show that COG not only improves the accuracy of recognizing long-tailed image-text pairs compared to baselines but also offers flexibility and explainability.

Ruihan Yang, Jiangjie Chen, Yikai Zhang et al.

Language agents powered by large language models (LLMs) are increasingly valuable as decision-making tools in domains such as gaming and programming. However, these agents often face challenges in achieving high-level goals without detailed instructions and in adapting to environments where feedback is delayed. In this paper, we present SelfGoal, a novel automatic approach designed to enhance agents' capabilities to achieve high-level goals with limited human prior and environmental feedback. The core concept of SelfGoal involves adaptively breaking down a high-level goal into a tree structure of more practical subgoals during the interaction with environments while identifying the most useful subgoals and progressively updating this structure. Experimental results demonstrate that SelfGoal significantly enhances the performance of language agents across various tasks, including competitive, cooperative, and deferred feedback environments. Project page: https://selfgoal-agent.github.io.

Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu et al.

New NLP benchmarks are urgently needed to align with the rapid development of large language models (LLMs). We present C-Eval, the first comprehensive Chinese evaluation suite designed to assess advanced knowledge and reasoning abilities of foundation models in a Chinese context. C-Eval comprises multiple-choice questions across four difficulty levels: middle school, high school, college, and professional. The questions span 52 diverse disciplines, ranging from humanities to science and engineering. C-Eval is accompanied by C-Eval Hard, a subset of very challenging subjects in C-Eval that requires advanced reasoning abilities to solve. We conduct a comprehensive evaluation of the most advanced LLMs on C-Eval, including both English- and Chinese-oriented models. Results indicate that only GPT-4 could achieve an average accuracy of over 60%, suggesting that there is still significant room for improvement for current LLMs. We anticipate C-Eval will help analyze important strengths and shortcomings of foundation models, and foster their development and growth for Chinese users.

Yu Chen, Wei Deng, Shikai Fang et al.

The Schrödinger bridge problem (SBP) is gaining increasing attention in generative modeling and showing promising potential even in comparison with the score-based generative models (SGMs). SBP can be interpreted as an entropy-regularized optimal transport problem, which conducts projections onto every other marginal alternatingly. However, in practice, only approximated projections are accessible and their convergence is not well understood. To fill this gap, we present a first convergence analysis of the Schrödinger bridge algorithm based on approximated projections. As for its practical applications, we apply SBP to probabilistic time series imputation by generating missing values conditioned on observed data. We show that optimizing the transport cost improves the performance and the proposed algorithm achieves the state-of-the-art result in healthcare and environmental data while exhibiting the advantage of exploring both temporal and feature patterns in probabilistic time series imputation.

Songzhu Zheng, Yikai Zhang, Hubert Wagner et al.

Deep neural networks are known to have security issues. One particular threat is the Trojan attack. It occurs when the attackers stealthily manipulate the model's behavior through Trojaned training samples, which can later be exploited. Guided by basic neuroscientific principles we discover subtle -- yet critical -- structural deviation characterizing Trojaned models. In our analysis we use topological tools. They allow us to model high-order dependencies in the networks, robustly compare different networks, and localize structural abnormalities. One interesting observation is that Trojaned models develop short-cuts from input to output layers. Inspired by these observations, we devise a strategy for robust detection of Trojaned models. Compared to standard baselines it displays better performance on multiple benchmarks.

Yikai Zhang, Songzhu Zheng, Pengxiang Wu et al.

Label noise is frequently observed in real-world large-scale datasets. The noise is introduced due to a variety of reasons; it is heterogeneous and feature-dependent. Most existing approaches to handling noisy labels fall into two categories: they either assume an ideal feature-independent noise, or remain heuristic without theoretical guarantees. In this paper, we propose to target a new family of feature-dependent label noise, which is much more general than commonly used i.i.d. label noise and encompasses a broad spectrum of noise patterns. Focusing on this general noise family, we propose a progressive label correction algorithm that iteratively corrects labels and refines the model. We provide theoretical guarantees showing that for a wide variety of (unknown) noise patterns, a classifier trained with this strategy converges to be consistent with the Bayes classifier. In experiments, our method outperforms SOTA baselines and is robust to various noise types and levels.

Yikai Zhang, Wenjia Zhang, Sammy Bald et al.

Stochastic Gradient Descent (SGD) based methods have been widely used for training large-scale machine learning models that also generalize well in practice. Several explanations have been offered for this generalization performance, a prominent one being algorithmic stability [18]. However, there are no known examples of smooth loss functions for which the analysis can be shown to be tight. Furthermore, apart from the properties of the loss function, data distribution has also been shown to be an important factor in generalization performance. This raises the question: is the stability analysis of [18] tight for smooth functions, and if not, for what kind of loss functions and data distributions can the stability analysis be improved? In this paper we first settle open questions regarding tightness of bounds in the data-independent setting: we show that for general datasets, the existing analysis for convex and strongly-convex loss functions is tight, but it can be improved for non-convex loss functions. Next, we give a novel and improved data-dependent bounds: we show stability upper bounds for a large class of convex regularized loss functions, with negligible regularization parameters, and improve existing data-dependent bounds in the non-convex setting. We hope that our results will initiate further efforts to better understand the data-dependent setting under non-convex loss functions, leading to an improved understanding of the generalization abilities of deep networks.

Yikai Zhang, Hui Qu, Qi Chang et al.

Recently, Generative Adversarial Networks (GANs) have demonstrated their potential in federated learning, i.e., learning a centralized model from data privately hosted by multiple sites. A federatedGAN jointly trains a centralized generator and multiple private discriminators hosted at different sites. A major theoretical challenge for the federated GAN is the heterogeneity of the local data distributions. Traditional approaches cannot guarantee to learn the target distribution, which isa mixture of the highly different local distributions. This paper tackles this theoretical challenge, and for the first time, provides a provably correct framework for federated GAN. We propose a new approach called Universal Aggregation, which simulates a centralized discriminator via carefully aggregating the mixture of all private discriminators. We prove that a generator trained with this simulated centralized discriminator can learn the desired target distribution. Through synthetic and real datasets, we show that our method can learn the mixture of largely different distributions where existing federated GAN methods fail.

Qi Chang, Zhennan Yan, Lohendran Baskaran et al.

As deep learning technologies advance, increasingly more data is necessary to generate general and robust models for various tasks. In the medical domain, however, large-scale and multi-parties data training and analyses are infeasible due to the privacy and data security concerns. In this paper, we propose an extendable and elastic learning framework to preserve privacy and security while enabling collaborative learning with efficient communication. The proposed framework is named distributed Asynchronized Discriminator Generative Adversarial Networks (AsynDGAN), which consists of a centralized generator and multiple distributed discriminators. The advantages of our proposed framework are five-fold: 1) the central generator could learn the real data distribution from multiple datasets implicitly without sharing the image data; 2) the framework is applicable for single-modality or multi-modality data; 3) the learned generator can be used to synthesize samples for down-stream learning tasks to achieve close-to-real performance as using actual samples collected from multiple data centers; 4) the synthetic samples can also be used to augment data or complete missing modalities for one single data center; 5) the learning process is more efficient and requires lower bandwidth than other distributed deep learning methods.

Hui Qu, Yikai Zhang, Qi Chang et al.

In this work, we propose a method for training distributed GAN with sequential temporary discriminators. Our proposed method tackles the challenge of training GAN in the federated learning manner: How to update the generator with a flow of temporary discriminators? We apply our proposed method to learn a self-adaptive generator with a series of local discriminators from multiple data centers. We show our design of loss function indeed learns the correct distribution with provable guarantees. The empirical experiments show that our approach is capable of generating synthetic data which is practical for real-world applications such as training a segmentation model.

Qi Chang, Hui Qu, Yikai Zhang et al.

In this paper, we propose a data privacy-preserving and communication efficient distributed GAN learning framework named Distributed Asynchronized Discriminator GAN (AsynDGAN). Our proposed framework aims to train a central generator learns from distributed discriminator, and use the generated synthetic image solely to train the segmentation model.We validate the proposed framework on the application of health entities learning problem which is known to be privacy sensitive. Our experiments show that our approach: 1) could learn the real image's distribution from multiple datasets without sharing the patient's raw data. 2) is more efficient and requires lower bandwidth than other distributed deep learning methods. 3) achieves higher performance compared to the model trained by one real dataset, and almost the same performance compared to the model trained by all real datasets. 4) has provable guarantees that the generator could learn the distributed distribution in an all important fashion thus is unbiased.