Fengbin Zhu
Publications
Fetch-then-Explore: Decoupling Selection from Extraction over a Persistent Workspace for Search Agents
Search agents now answer questions that take dozens of searches to settle, yet how such an agent reads a page has drawn far less attention than how it finds one. Nearly all of them use one of two document interfaces, and both tie a page to the moment it is opened. \emph{Visit-and-read} injects a reading of the page into the message history at fetch time, fixing that reading before the agent knows which fact it will need. Stateful \emph{browsing} instead extracts on demand from the page in hand, but holds one page at a time and releases it as soon as the agent opens another. Either way, a page that turns out to matter many turns later has to be fetched and rendered into context all over again. We propose \textbf{Fetch-then-Explore}, which separates page selection from evidence extraction and keeps what it selects: pages are recorded in a per-question workspace on the filesystem rather than the context window or a transient session, and evidence is pulled from them on demand later. Selection becomes almost free, extraction can wait until the agent knows what to look for and be repeated as its hypothesis sharpens, and pages are not released when the agent moves on, so evidence accumulates across the trajectory. In a unified ReAct harness with fixed search, we compare Fetch-then-Explore against snippet-only, visit-and-read, and browsing baselines on two open-web benchmarks, BrowseComp and WideSearch, across three agent backbones. It leads BrowseComp accuracy at every backbone and generally matches or exceeds the baselines on WideSearch, and a behavioral analysis traces the gains to the workspace's defining move: returning to a page after leaving it, which it does far more than any transient interface, so evidence missed on a first pass can still be recovered later.
Diagnosing Search Behavior and Failure Modes in Long-Horizon Search Agents
Deep search agents answer difficult information-seeking questions by iteratively issuing search queries to gather supporting evidence, but it remains unclear whether and how greater search effort leads to better answers. We study these questions through a trajectory-level diagnosis of long-horizon search agents. Using human-annotated document-level relevance judgments, we evaluate the evidence retrieved at each search step and separate two stages of agent behavior: what evidence an agent retrieves and how effectively it uses that evidence. This distinction further allows us to decompose failures into retrieval gaps, where the necessary evidence is never found, and utilization gaps, where relevant evidence is retrieved but not used correctly. With the retrieval model and evaluation harness held fixed, we compare six agents on BrowseComp-Plus and further validate our findings on BrowseComp with an open-web search API. Across settings, we find that search effort and answer quality are only weakly aligned. Answer accuracy is better correlated with the quality of retrieved evidence, especially cumulative retrieval recall, than with the number of searches or the amount of context consumed. Useful evidence often appears early in the trajectory, yet agents tend to continue searching, producing a long tail of low-yield retrieval steps. At the query level, exploratory reformulations remain useful, but the best-performing agents issue far fewer redundant queries. Overall, by systematically characterizing the search behavior and failure modes of long-horizon search agents, this work points to practical directions for building better deep research systems, including stronger query formulation, more effective evidence selection and context management, and stopping criteria based on whether sufficient supporting evidence has been retrieved.
Plant, Persist, Trigger: Sleeper Attack on Large Language Model Agents
Large Language Model (LLM) agents remain vulnerable to safety threats from the external environment, where attackers inject adversarial content into external observations such as tool-returned data, webpages, or MCP context, causing harmful agentic behaviors such as unsafe actions or incorrect outputs. Existing studies typically focus on single-interaction attacks, where the agent observes adversarial content and immediately exhibits harmful behavior within one user request. However, we show that adversarial content can also persist across interactions served by the same agent, making such threats harder to detect and mitigate. Specifically, adversarial content may persist in the agent state, remain dormant across interactions, and later be activated by a benign user query. We formalize this type of safety threat as Sleeper Attack. To evaluate it, we construct a benchmark with 1,896 instances covering six real-world harmful outcomes, three attack strategies, and three agent state targets: session context, memory, and reusable skills. Experiments on seven strong open-source and closed-source LLMs show that state-of-the-art LLM agents remain vulnerable to Sleeper Attack, even when they achieve low attack success rates under a single-interaction baseline. Our code and data are available at https://anonymous.4open.science/r/skdvnfu23ihr9wdscnksf1asdffsaef.
Herculean: An Agentic Benchmark for Financial Intelligence
As AI agents improve, the central question is no longer whether they can solve isolated well-defined financial tasks, but whether they can reliably carry out financial professional work. Existing financial benchmarks offer only a partial view of this ability, as they primarily evaluate static competencies such as question answering, retrieval, summarization, and classification. We introduce Herculean, the first skilled benchmark for agentic financial intelligence spanning four representative workflows, including Trading, Hedging, Market Insights, and Auditing. Each workflow is instantiated as a standardized MCP-based skill environment with its own tools, interaction dynamics, constraints, and success criteria, enabling consistent end-to-end assessment of heterogeneous agent systems. Across frontier agents, we find agents perform relatively well on Trading and Market Insights, but struggle substantially on Hedging and Auditing, where long-horizon coordination, state consistency, and structured verification are critical. Overall, our results point to a key gap in current agents in turning financial reasoning into dependable workflow execution in high-stakes financial workflows.
Scale over Preference: The Impact of AI-Generated Content on Online Content Ecology
The rapid proliferation of Artificial Intelligence-Generated Content (AIGC) is fundamentally restructuring online content ecologies, necessitating a rigorous examination of its behavioral and distributional implications. Leveraging a comprehensive longitudinal dataset comprising tens of millions of users from a leading Chinese video-sharing platform, this study elucidated the distinct creation and consumption behaviors characterizing AIGC versus Human-Generated Content (HGC). We identified a prevalent scale-over-preference dynamic, wherein AIGC creators achieve aggregate engagement comparable to HGC creators through high-volume production, despite a marked consumer preference for HGC. Deeper analysis uncovered the ability of the algorithmic content distribution mechanism in moderating these competing interests regarding AIGC. These findings advocated for the implementation of AIGC-sensitive distribution algorithms and precise governance frameworks to ensure the long-term health of the online content platforms.
CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization
Subcellular localization is a crucial biological task for drug target identification and function annotation. Although it has been biologically realized that subcellular localization is closely associated with protein structure, no existing dataset offers comprehensive 3D structural information with detailed subcellular localization annotations, thus severely hindering the application of promising structure-based models on this task. To address this gap, we introduce a new benchmark called $\mathbf{CAPSUL}$, a $\mathbf{C}$omprehensive hum$\mathbf{A}$n $\mathbf{P}$rotein benchmark for $\mathbf{SU}$bcellular $\mathbf{L}$ocalization. It features a dataset that integrates diverse 3D structural representations with fine-grained subcellular localization annotations carefully curated by domain experts. We evaluate this benchmark using a variety of state-of-the-art sequence-based and structure-based models, showcasing the importance of involving structural features in this task. Furthermore, we explore reweighting and single-label classification strategies to facilitate future investigation on structure-based methods for this task. Lastly, we showcase the powerful interpretability of structure-based methods through a case study on the Golgi apparatus, where we discover a decisive localization pattern $α$-helix from attention mechanisms, demonstrating the potential for bridging the gap with intuitive biological interpretability and paving the way for data-driven discoveries in cell biology.
MURE: Hierarchical Multi-Resolution Encoding via Vision-Language Models for Visual Document Retrieval
Visual Document Retrieval (VDR) requires representations that capture both fine-grained visual details and global document structure to ensure retrieval efficacy while maintaining computational efficiency. Existing VDR models struggle to balance effectiveness and efficiency when processing high-resolution documents: they often either lose fine-grained information or generate an excessive number of visual tokens, resulting in significant indexing overhead and high retrieval latency. In this work, we rethink the visual encoding mechanism and propose a new X-VisEmb paradigm that progresses from multi-resolution sampling and encoding, through cross-granularity feature fusion, to adaptive representation distillation. A preliminary study validates its feasibility and effectiveness in capturing complementary visual cues at varying scales. Building on the insights, we develop MURE, a novel framework that employs VLMs as a hierarchical multi-resolution encoder, integrates resolution-level Matryoshka representation learning (RMRL) for effective feature fusion, and applies a semantic-aware hierarchical clustering mechanism for visual token compression. Experiments on two widely used VDR benchmarks show that our MURE framework consistently beats strong baselines. Furthermore, it significantly outperforms ColPali with only 50% of its visual token budget.
One Adapts to Any: Meta Reward Modeling for Personalized LLM Alignment
Alignment of Large Language Models (LLMs) aims to align outputs with human preferences, and personalized alignment further adapts models to individual users. This relies on personalized reward models that capture user-specific preferences and automatically provide individualized feedback. However, developing these models faces two critical challenges: the scarcity of feedback from individual users and the need for efficient adaptation to unseen users. We argue that addressing these constraints requires a paradigm shift from fitting data to learn user preferences to learn the process of preference adaptation. To realize this, we propose Meta Reward Modeling (MRM), which reformulates personalized reward modeling as a meta-learning problem. Specifically, we represent each user's reward model as a weighted combination of base reward functions, and optimize the initialization of these weights using a Model-Agnostic Meta-Learning (MAML)-style framework to support fast adaptation under limited feedback. To ensure robustness, we introduce the Robust Personalization Objective (RPO), which places greater emphasis on hard-to-learn users during meta optimization. Extensive experiments on personalized preference datasets validate that MRM enhances few-shot personalization, improves user robustness, and consistently outperforms baselines.