arXiv subjects

Shihan Deng

Publications and source records attributed to Shihan Deng.

4 recordsLinked to original sources

arXiv2026

When Search Agents Should Ask: DiscoBench for Clarification-Aware Deep Search

Search agents powered by large language models (LLMs) are increasingly used to solve complex information-seeking tasks, requiring multi-step retrieval and reasoning to fulfill user goals. However, existing benchmarks often assume that user queries are complete and explicit, overlooking the fact that real-world search requests are frequently vague, underspecified, or even factually incorrect. In deep search scenarios, such ambiguity can propagate along multi-step reasoning chains and lead agents toward incorrect search trajectories. To address this gap, we introduce DiscoBench, a benchmark for clarification-aware deep search, designed to evaluate whether search agents can proactively identify ambiguity, ask effective clarification questions, and recover correct reasoning paths through user interaction. DiscoBench contains 211 samples and 463 ambiguity instances across 11 real-world domains, covering four ambiguity types. We further design a user simulator for multi-turn interaction and evaluate model performance from four perspectives: task utility, ambiguity detection, interaction strategy, and cost efficiency. Experiments on representative LLMs show that ambiguity detection and effective clarification are distinct capabilities, and that repeatedly searching instead of asking for clarification often performs worse than direct guessing, highlighting a critical gap between retrieval ability and interactive problem-solving in current search agents.

cs.CL↗

arXiv2026

Unified definition of ferroelectricity

Recent theoretical and experimental advances in quantum ferroelectrics suggest that ferroelectricity can also emerge in non-polar space group, highlighting the limitations of conventional polar space group criteria in identifying ferroelectric materials. Here, we introduce a unified definition based on switchable polarization differences between energetically equivalent states, which naturally encompasses conventional and quantum ferroelectrics. Guided by this principle, we implement a high-throughput screening strategy that systematically identifies both conventional and quantum ferroelectrics among experimentally synthesized materials. In particular, we identify a new type of quantum ferroelectric in which the quantized polarization arises from arbitrary ionic displacements, in contrast to previous quantum ferroelectrics (including both fractional and integer quantum ferroelectrics) where quantized polarization results from fractional or integer ionic displacements. Notably, we find that materials such as Ba3I6 and Cs2PdC2 exhibit low switching barriers and robust insulating behavior, highlighting their experimental viability. Our results reconcile conventional and quantum ferroelectrics, expand the accessible materials landscape, and provide a practical roadmap for discovering next-generation ferroelectrics with advanced switchable functionalities.

cond-mat.mtrl-sci↗

arXiv2024

Deterministic and Efficient Switching of Sliding Ferroelectrics

Recent studies highlight the scientific importance and broad application prospects of two-dimensional (2D) sliding ferroelectrics, which prevalently exhibit vertical polarization with suitable stackings. It is crucial to understand the mechanisms of sliding ferroelectricity and to deterministically and efficiently switch the polarization with optimized electric fields. Here, applying our newly developed DREAM-Allegro multi-task equivariant neural network, which simultaneously predicts interatomic potentials and Born effective charges, we construct a comprehensive potential for boron nitride ($\mathrm{BN}$) bilayer. The molecular dynamics simulations reveal a remarkably high Curie temperature of up to 1500K, facilitated by robust intralayer chemical bonds and delicate interlayer van der Waals(vdW) interactions. More importantly, it is found that, compared to the out-of-plane electric field, the inclined field not only leads to deterministic switching of electric polarization, but also largely lower the critical strength of field, due to the presence of the in-plane polarization in the transition state. This strategy of an inclined field is demonstrated to be universal for other sliding ferroelectric systems with monolayer structures belonging to the symmetry group $p \bar{6} m 2$, such as transition metal dichalcogenides (TMDs).

cond-mat.mtrl-sci↗

arXiv2024

Mobile-Bench: An Evaluation Benchmark for LLM-based Mobile Agents

With the remarkable advancements of large language models (LLMs), LLM-based agents have become a research hotspot in human-computer interaction. However, there is a scarcity of benchmarks available for LLM-based mobile agents. Benchmarking these agents generally faces three main challenges: (1) The inefficiency of UI-only operations imposes limitations to task evaluation. (2) Specific instructions within a singular application lack adequacy for assessing the multi-dimensional reasoning and decision-making capacities of LLM mobile agents. (3) Current evaluation metrics are insufficient to accurately assess the process of sequential actions. To this end, we propose Mobile-Bench, a novel benchmark for evaluating the capabilities of LLM-based mobile agents. First, we expand conventional UI operations by incorporating 103 collected APIs to accelerate the efficiency of task completion. Subsequently, we collect evaluation data by combining real user queries with augmentation from LLMs. To better evaluate different levels of planning capabilities for mobile agents, our data is categorized into three distinct groups: SAST, SAMT, and MAMT, reflecting varying levels of task complexity. Mobile-Bench comprises 832 data entries, with more than 200 tasks specifically designed to evaluate multi-APP collaboration scenarios. Furthermore, we introduce a more accurate evaluation metric, named CheckPoint, to assess whether LLM-based mobile agents reach essential points during their planning and reasoning steps.

cs.AI↗