arXiv subjects

Zhao Cao

Publications and source records attributed to Zhao Cao.

At least 19 recordsLinked to original sources

arXiv2026

Evidence Blindness in Direct Corpus Interaction: Persistent Navigation with AtlasNav

Large language model agents are moving beyond conventional retrieval-augmented generation toward direct interaction with external corpora. Direct Corpus Interaction (DCI) keeps the full corpus accessible, yet reachable evidence can remain unusable under finite interaction budgets. Required evidence may fail to surface, a surfaced supporting document may remain unopened, or an opened document may fail to expose its decisive fragment. We call this progressive silent loss Evidence Blindness and quantify it through stage-wise evidence realization. Within the DCI paradigm, raw interaction adds little reusable corpus organization, while dynamic-workspace methods reconstruct a query-conditioned interaction space from each query and trajectory. In both cases, useful structure is recovered largely online. We instead formulate large-scale agentic search as finite-budget navigation over reusable corpus structure. We introduce AtlasNav, a persistent multi-view corpus-navigation framework that retains direct corpus interaction but organizes the corpus once into a Corpus Atlas, allowing each query to navigate adaptively rather than reconstruct shared structure. On BrowseComp-Plus, AtlasNav achieves 92.05% strict accuracy while reducing recorded online inference cost by 30.21% relative to the prior dynamic-workspace state of the art. Under matched budgets, it realizes the complete required evidence earlier and approaches the same model's evidence-supplied empirical reference more rapidly. The same representation principle remains effective under PhantomWiki's distinct corpus organization and controlled 10K-1M scaling, and transfers competitively to heterogeneous enterprise knowledge. These results show that agentic search depends not only on accessible evidence, but also on how the corpus is represented so that limited interaction becomes effective navigation.

cs.AI↗

arXiv2026

AgentOmnia: Scaling Agentic Models for Full-Scenario Applications

Large language model agents have advanced rapidly, yet progress remains fragmented across domains, capabilities, task difficulty, and interaction settings. We frame this as full-scenario agentic scaling and present AgentOmnia, a framework coordinating task-space definition, data synthesis, post-training, evaluation, and improvement across To-Consumer (ToC), To-Business (ToB), and To-Employee (ToE) applications. An extensible Domain x Capability x Atomic Difficulty taxonomy aligns these stages and enables fine-grained diagnosis with OmniaBench. AgentOmnia combines bidirectional environment-task synthesis with tool-dependency, program-structured, and solver-based pipelines, constructing 5,018 stateful environments with 255,375 tools and 52,361 tasks. Programs, solvers, and verifiers provide correctness signals, while supervised fine-tuning, online agentic reinforcement learning, and a rollback curriculum support post-training. Evaluation failures translate into Product Requirement Documents (PRDs) for targeted self-evolution. Starting from Qwen3-30B-A3B-Thinking-2507, AgentOmnia raises the pass rate on the OmniaBench challenging subset from 9.16% to 37.11% and the macro-average across OmniaBench, $\tau^2$-Bench, DeepPlanning, and VitaBench from 22.86% to 41.69%. Under a unified protocol,it leads the evaluated agentic post-trained baselines on OmniaBench and retains the highest four-benchmark macro-average. It also surpasses Qwen3-235B-A22B-Thinking-2507 on all four benchmarks and exceeds Qwen3.5-35B-A3B on the macro-average. Gains span three application splits, ten capability dimensions, eight atomic-difficulty factors, and 76 of 90 level-1 domains, indicating broad rather than category-specific improvement. A one-round study provides initial evidence for PRD-guided self-evolution, motivating validation at larger scales and in industrial settings.

cs.AI↗

arXiv2026

OmniaBench: Benchmarking General AI Agents Across Diverse Scenarios

Large language models are increasingly evolving from text generators into general agents capable of understanding user requests, invoking external tools, and completing complex tasks through interaction. However, existing agent benchmarks often focus on limited scenarios, tool ecosystems, or interaction formats, making it difficult to systematically characterize model capabilities across heterogeneous application settings. We introduce OmniaBench, a benchmark for evaluating general agents across diverse scenarios with explicit state spaces. We derive application-oriented scenario knowledge from app stores, product documents, industry resources, Web retrieval, and human refinement, forming a hierarchical taxonomy that spans ToC, ToB and ToE with 90 level-1 and 354 level-2 domains. Based on this taxonomy, we construct executable environments and synthesize single-turn and multi-turn tasks through four complementary routes: DAG, DAG-S, Solver, and Program. OmniaBench further introduces a ten-dimensional capability taxonomy and eight compositional atomic difficulty factors to support fine-grained evaluation and analysis. The resulting dataset contains 1,431 tasks, together with a challenging subset of 644 tasks designed to reduce evaluation cost and mitigate potential contamination of the full set after public release. The bench presents substantial challenges to current frontier models, with even Claude-Sonnet-5 and GPT-5.6-Sol achieving Overall Pass@1 scores of only 58.54 and 57.14, respectively. Further analyses reveal clear differences across domains and capabilities, as well as persistent limitations in planning, constraint maintenance, and adaptive correction. OmniaBench provides a broad and diagnostic benchmark for characterizing the capability boundaries of general agents.

cs.CL↗

arXiv2026

LLM Evolution as an Industry-Scale Ecosystem: A Lifecycle Perspective on Continual Learning

Continual learning capability is critical for Industrial LLMs, as deployed models must be continuously updated to meet evolving requirements and environments, rather than repeatedly retrained from scratch. However, most existing research focuses on improvements on static benchmarks, failing to capture real industrial needs. In this survey, we reformulate Industrial Continual Learning (ICL) for LLMs as a closed-loop update-and-release problem in a versioned ecosystem, where updates propagate hierarchically to industrial, application-specific models and LLM-powered applications, with capability inheritance and transfer across versions and model families. From this ecosystem perspective, we identify three core challenges: repeated adaptation erodes model plasticity, foundation-model upgrades break capability inheritance, and long-term sustainability is constrained by deployment requirements. We then organize the technical landscape of ICL around five lifecycle design principles: preserving plasticity headroom, treating upgrades as capability transfer, enabling trustworthy continual reinforcement learning, making training recipes self-optimizing, and building accountability as a base layer for long-term iteration. For each principle, we synthesize representative technical directions. Finally, we evaluate the maturity of each principle and its technical components via an evidence-based lens, identify key gaps hindering real-world deployment, and outline a practical ICL deployment blueprint and a pathway for feeding industrial realities back into academic research.

cs.LG↗

arXiv2026

When Planning Fails Despite Correct Execution: On Epistemic Calibration for LLM-Based Multi-Agent Systems

LLM-based multi-agent systems can fail even when planned actions are executed correctly because agents may misjudge their knowledge when evaluating plan feasibility, a phenomenon we term epistemic miscalibration in planning. Unlike execution errors, epistemic miscalibration is latent during planning, as generated plans can remain self-consistent and executable without observable errors; the miscalibration is also dynamic, as new information can alter feasibility assessments, potentially obscuring past miscalibration signals and causing them to recur over time. To address this, we propose the Epistemic Planning Calibration Agentic Workflow (EPC-AW), which assesses whether plans remain supported under varying information conditions rather than directly verifying feasibility. EPC-AW employs Information-consistency-based Plan Selection, selecting plans whose evaluations are stable across agents, together with Consistency-guided Epistemic State Refinement to adapt calibration over time by leveraging past discrepancies to guide future planning. Experiments show that EPC-AW improves system-level success by an average of 9.75%.

cs.AI↗

arXiv2026

Reasoning emerges from constrained inference manifolds in large language models

Reasoning in large language models is predominantly evaluated through labeled benchmarks, conflating task performance with the quality of internal inference. Here we study reasoning as an intrinsic dynamical process by examining the evolution of internal representations during inference. We find that inference-time dynamics consistently self-organize into low-dimensional manifolds embedded within high-dimensional representation spaces. we find that such geometric compression, although pervasive, is not sufficient for stable or reliable reasoning. Instead, effective reasoning dynamics emerge within a constrained structural regime characterized by three conditions: adequate representational expressivity, spontaneous manifold compression, and preservation of non-degenerate information volume within the compressed subspace. Models outside this regime exhibit characteristic pathological inference dynamics. Based on these insights, we introduce a unified, label-free diagnostic computed solely from internal dynamics. These findings suggest that reasoning in LLMs is fundamentally governed by geometric and informational constraints, offering a complementary framework to benchmark-centric assessment.

cs.LG↗

arXiv2026

PaperScope: A Multi-Modal Multi-Document Benchmark for Agentic Deep Research Across Massive Scientific Papers

Leveraging Multi-modal Large Language Models (MLLMs) to accelerate frontier scientific research is promising, yet how to rigorously evaluate such systems remains unclear. Existing benchmarks mainly focus on single-document understanding, whereas real scientific workflows require integrating evidence from multiple papers, including their text, tables, and figures. As a result, multi-modal, multi-document scientific reasoning remains underexplored and lacks systematic evaluation. To address this gap, we introduce PaperScope, a multi-modal multi-document benchmark designed for agentic deep research. PaperScope presents three advantages: (1) Structured scientific grounding. It is built on a knowledge graph of over 2,000 AI papers spanning three years, providing a structured foundation for research-oriented queries. (2) Semantically dense evidence construction. It integrates semantically related key information nodes and employs optimized random-walk article selector to sample thematically coherent paper sets, thereby ensuring adequate semantic density and task complexity. (3) Multi-task evaluation of scientific reasoning. It contains over 2,000 QA pairs across reasoning, retrieval, summarization, and problem solving, enabling evaluation of multi-step scientific reasoning. Experimental results show that even advanced systems such as OpenAI Deep Research and Tongyi Deep Research achieve limited scores on PaperScope, highlighting the difficulty of long-context retrieval and deep multi-source reasoning. PaperScope thus provides a rigorous benchmark alongside a scalable pipeline for constructing large-scale multi-modal, multi-source deep research datasets.

cs.AI↗

arXiv2026

MemoBrain: Executive Memory as an Agentic Brain for Reasoning

Complex reasoning in tool-augmented agent frameworks is inherently long-horizon, causing reasoning traces and transient tool artifacts to accumulate and strain the bounded working context of large language models. Without explicit memory mechanisms, such accumulation disrupts logical continuity and undermines task alignment. This positions memory not as an auxiliary efficiency concern, but as a core component for sustaining coherent, goal-directed reasoning over long horizons. We propose MemoBrain, an executive memory model for tool-augmented agents that constructs a dependency-aware memory over reasoning steps, capturing salient intermediate states and their logical relations. Operating as a co-pilot alongside the reasoning agent, MemoBrain organizes reasoning progress without blocking execution and actively manages the working context. Specifically, it prunes invalid steps, folds completed sub-trajectories, and preserves a compact, high-salience reasoning backbone under a fixed context budget. Together, these mechanisms enable explicit cognitive control over reasoning trajectories rather than passive context accumulation. We evaluate MemoBrain on challenging long-horizon benchmarks, including GAIA, WebWalker, and BrowseComp-Plus, demonstrating consistent improvements over strong baselines.

cs.AI↗

arXiv2025

MomentSeeker: A Task-Oriented Benchmark For Long-Video Moment Retrieval

Accurately locating key moments within long videos is crucial for solving long video understanding (LVU) tasks. However, existing benchmarks are either severely limited in terms of video length and task diversity, or they focus solely on the end-to-end LVU performance, making them inappropriate for evaluating whether key moments can be accurately accessed. To address this challenge, we propose MomentSeeker, a novel benchmark for long-video moment retrieval (LMVR), distinguished by the following features. First, it is created based on long and diverse videos, averaging over 1200 seconds in duration and collected from various domains, e.g., movie, anomaly, egocentric, and sports. Second, it covers a variety of real-world scenarios in three levels: global-level, event-level, object-level, covering common tasks like action recognition, object localization, and causal reasoning, etc. Third, it incorporates rich forms of queries, including text-only queries, image-conditioned queries, and video-conditioned queries. On top of MomentSeeker, we conduct comprehensive experiments for both generation-based approaches (directly using MLLMs) and retrieval-based approaches (leveraging video retrievers). Our results reveal the significant challenges in long-video moment retrieval in terms of accuracy and efficiency, despite improvements from the latest long-video MLLMs and task-specific fine-tuning. We have publicly released MomentSeeker(https://yhy-2000.github.io/MomentSeeker/) to facilitate future research in this area.

cs.CV↗

arXiv2024

ICLEval: Evaluating In-Context Learning Ability of Large Language Models

In-Context Learning (ICL) is a critical capability of Large Language Models (LLMs) as it empowers them to comprehend and reason across interconnected inputs. Evaluating the ICL ability of LLMs can enhance their utilization and deepen our understanding of how this ability is acquired at the training stage. However, existing evaluation frameworks primarily focus on language abilities and knowledge, often overlooking the assessment of ICL ability. In this work, we introduce the ICLEval benchmark to evaluate the ICL abilities of LLMs, which encompasses two key sub-abilities: exact copying and rule learning. Through the ICLEval benchmark, we demonstrate that ICL ability is universally present in different LLMs, and model size is not the sole determinant of ICL efficacy. Surprisingly, we observe that ICL abilities, particularly copying, develop early in the pretraining process and stabilize afterward. Our source codes and benchmark are released at https://github.com/yiye3/ICLEval.

cs.CL↗

arXiv2024

An Analysis on Matching Mechanisms and Token Pruning for Late-interaction Models

With the development of pre-trained language models, the dense retrieval models have become promising alternatives to the traditional retrieval models that rely on exact match and sparse bag-of-words representations. Different from most dense retrieval models using a bi-encoder to encode each query or document into a dense vector, the recently proposed late-interaction multi-vector models (i.e., ColBERT and COIL) achieve state-of-the-art retrieval effectiveness by using all token embeddings to represent documents and queries and modeling their relevance with a sum-of-max operation. However, these fine-grained representations may cause unacceptable storage overhead for practical search systems. In this study, we systematically analyze the matching mechanism of these late-interaction models and show that the sum-of-max operation heavily relies on the co-occurrence signals and some important words in the document. Based on these findings, we then propose several simple document pruning methods to reduce the storage overhead and compare the effectiveness of different pruning methods on different late-interaction models. We also leverage query pruning methods to further reduce the retrieval latency. We conduct extensive experiments on both in-domain and out-domain datasets and show that some of the used pruning methods can significantly improve the efficiency of these late-interaction models without substantially hurting their retrieval effectiveness.

cs.IR↗

arXiv2023

Argue with Me Tersely: Towards Sentence-Level Counter-Argument Generation

Counter-argument generation -- a captivating area in computational linguistics -- seeks to craft statements that offer opposing views. While most research has ventured into paragraph-level generation, sentence-level counter-argument generation beckons with its unique constraints and brevity-focused challenges. Furthermore, the diverse nature of counter-arguments poses challenges for evaluating model performance solely based on n-gram-based metrics. In this paper, we present the ArgTersely benchmark for sentence-level counter-argument generation, drawing from a manually annotated dataset from the ChangeMyView debate forum. We also propose Arg-LlaMA for generating high-quality counter-argument. For better evaluation, we trained a BERT-based evaluator Arg-Judge with human preference data. We conducted comparative experiments involving various baselines such as LlaMA, Alpaca, GPT-3, and others. The results show the competitiveness of our proposed framework and evaluator in counter-argument generation tasks. Code and data are available at https://github.com/amazingljy1206/ArgTersely.

cs.CL↗

arXiv2023

Hi-ArG: Exploring the Integration of Hierarchical Argumentation Graphs in Language Pretraining

The knowledge graph is a structure to store and represent knowledge, and recent studies have discussed its capability to assist language models for various applications. Some variations of knowledge graphs aim to record arguments and their relations for computational argumentation tasks. However, many must simplify semantic types to fit specific schemas, thus losing flexibility and expression ability. In this paper, we propose the Hierarchical Argumentation Graph (Hi-ArG), a new structure to organize arguments. We also introduce two approaches to exploit Hi-ArG, including a text-graph multi-modal model GreaseArG and a new pre-training framework augmented with graph information. Experiments on two argumentation tasks have shown that after further pre-training and fine-tuning, GreaseArG supersedes same-scale language models on these tasks, while incorporating graph information during further pre-training can also improve the performance of vanilla language models. Code for this paper is available at https://github.com/ljcleo/Hi-ArG .

cs.CL↗

arXiv2023

IAG: Induction-Augmented Generation Framework for Answering Reasoning Questions

Retrieval-Augmented Generation (RAG), by incorporating external knowledge with parametric memory of language models, has become the state-of-the-art architecture for open-domain QA tasks. However, common knowledge bases are inherently constrained by limited coverage and noisy information, making retrieval-based approaches inadequate to answer implicit reasoning questions. In this paper, we propose an Induction-Augmented Generation (IAG) framework that utilizes inductive knowledge along with the retrieved documents for implicit reasoning. We leverage large language models (LLMs) for deriving such knowledge via a novel prompting method based on inductive reasoning patterns. On top of this, we implement two versions of IAG named IAG-GPT and IAG-Student, respectively. IAG-GPT directly utilizes the knowledge generated by GPT-3 for answer prediction, while IAG-Student gets rid of dependencies on GPT service at inference time by incorporating a student inductor model. The inductor is firstly trained via knowledge distillation and further optimized by back-propagating the generator feedback via differentiable beam scores. Experimental results show that IAG outperforms RAG baselines as well as ChatGPT on two Open-Domain QA tasks. Notably, our best models have won the first place in the official leaderboards of CSQA2.0 (since Nov 1, 2022) and StrategyQA (since Jan 8, 2023).

cs.CL↗

arXiv2023

UFIN: Universal Feature Interaction Network for Multi-Domain Click-Through Rate Prediction

Click-Through Rate (CTR) prediction, which aims to estimate the probability of a user clicking on an item, is a key task in online advertising. Numerous existing CTR models concentrate on modeling the feature interactions within a solitary domain, thereby rendering them inadequate for fulfilling the requisites of multi-domain recommendations in real industrial scenarios. Some recent approaches propose intricate architectures to enhance knowledge sharing and augment model training across multiple domains. However, these approaches encounter difficulties when being transferred to new recommendation domains, owing to their reliance on the modeling of ID features (e.g., item id). To address the above issue, we propose the Universal Feature Interaction Network (UFIN) approach for CTR prediction. UFIN exploits textual data to learn universal feature interactions that can be effectively transferred across diverse domains. For learning universal feature representations, we regard the text and feature as two different modalities and propose an encoder-decoder network founded on a Large Language Model (LLM) to enforce the transfer of data from the text modality to the feature modality. Building upon the above foundation, we further develop a mixtureof-experts (MoE) enhanced adaptive feature interaction model to learn transferable collaborative patterns across multiple domains. Furthermore, we propose a multi-domain knowledge distillation framework to enhance feature interaction learning. Based on the above methods, UFIN can effectively bridge the semantic gap to learn common knowledge across various domains, surpassing the constraints of ID-based models. Extensive experiments conducted on eight datasets show the effectiveness of UFIN, in both multidomain and cross-platform settings. Our code is available at https://github.com/RUCAIBox/UFIN.

cs.IR↗

arXiv2023

EulerNet: Adaptive Feature Interaction Learning via Euler's Formula for CTR Prediction

Learning effective high-order feature interactions is very crucial in the CTR prediction task. However, it is very time-consuming to calculate high-order feature interactions with massive features in online e-commerce platforms. Most existing methods manually design a maximal order and further filter out the useless interactions from them. Although they reduce the high computational costs caused by the exponential growth of high-order feature combinations, they still suffer from the degradation of model capability due to the suboptimal learning of the restricted feature orders. The solution to maintain the model capability and meanwhile keep it efficient is a technical challenge, which has not been adequately addressed. To address this issue, we propose an adaptive feature interaction learning model, named as EulerNet, in which the feature interactions are learned in a complex vector space by conducting space mapping according to Euler's formula. EulerNet converts the exponential powers of feature interactions into simple linear combinations of the modulus and phase of the complex features, making it possible to adaptively learn the high-order feature interactions in an efficient way. Furthermore, EulerNet incorporates the implicit and explicit feature interactions into a unified architecture, which achieves the mutual enhancement and largely boosts the model capabilities. Such a network can be fully learned from data, with no need of pre-designed form or order for feature interactions. Extensive experiments conducted on three public datasets have demonstrated the effectiveness and efficiency of our approach. Our code is available at: https://github.com/RUCAIBox/EulerNet.

cs.IR↗

arXiv2023

Optimizing Factual Accuracy in Text Generation through Dynamic Knowledge Selection

Language models (LMs) have revolutionized the way we interact with information, but they often generate nonfactual text, raising concerns about their reliability. Previous methods use external knowledge as references for text generation to enhance factuality but often struggle with the knowledge mix-up(e.g., entity mismatch) of irrelevant references. Besides,as the length of the output text grows, the randomness of sampling can escalate, detrimentally impacting the factual accuracy of the generated text. In this paper, we present DKGen, which divide the text generation process into an iterative process. In each iteration, DKGen takes the input query, the previously generated text and a subset of the reference passages as input to generate short text. During the process, the subset is dynamically selected from the full passage set based on their relevance to the previously generated text and the query, largely eliminating the irrelevant references from input. To further enhance DKGen's ability to correctly use these external knowledge, DKGen distills the relevance order of reference passages to the cross-attention distribution of decoder. We train and evaluate DKGen on a large-scale benchmark dataset. Experiment results show that DKGen outperforms all baseline models.

cs.CL↗

arXiv2023

Information Retrieval Meets Large Language Models: A Strategic Report from Chinese IR Community

The research field of Information Retrieval (IR) has evolved significantly, expanding beyond traditional search to meet diverse user information needs. Recently, Large Language Models (LLMs) have demonstrated exceptional capabilities in text understanding, generation, and knowledge inference, opening up exciting avenues for IR research. LLMs not only facilitate generative retrieval but also offer improved solutions for user understanding, model evaluation, and user-system interactions. More importantly, the synergistic relationship among IR models, LLMs, and humans forms a new technical paradigm that is more powerful for information seeking. IR models provide real-time and relevant information, LLMs contribute internal knowledge, and humans play a central role of demanders and evaluators to the reliability of information services. Nevertheless, significant challenges exist, including computational costs, credibility concerns, domain-specific limitations, and ethical considerations. To thoroughly discuss the transformative impact of LLMs on IR research, the Chinese IR community conducted a strategic workshop in April 2023, yielding valuable insights. This paper provides a summary of the workshop's outcomes, including the rethinking of IR's core values, the mutual enhancement of LLMs and IR, the proposal of a novel IR technical paradigm, and open challenges.

cs.IR↗