arXiv subjects

Zhi Qiu

Publications and source records attributed to Zhi Qiu.

15 recordsLinked to original sources

arXiv2026

MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents

Long horizon Large Language Model (LLM) agents rely on external memory systems to preserve user preferences and task knowledge across extended interactions. Conventional retrieval mechanisms optimize semantic compatibility rather than downstream utility, frequently introducing outdated, misleading, or conflicting evidence into the active context. We present MeClear, a task conditioned memory clearance framework that identifies memories featuring negative downstream utility through cooperative attribution and selectively suppresses them from agent execution. MeClear combines Leave One Out screening with sampled cooperative Shapley attribution to distribute utility across interacting evidence, effectively resolving redundant conflict masking where single removal evaluations fail. Utilizing attribution rankings, MeClear executes a query scoped minimal clearance strategy over a nested filtration, verifying task recovery on the cleared context without permanently altering the persistent memory bank. Comprehensive experimental evaluations across ten long dialogue memory pools demonstrate that MeClear achieves a target recall of 85.9% and an overall task recovery rate of 82.3%, representing a 25.5 percentage point improvement over Leave One Out (LOO) baselines.

cs.AI↗

arXiv2026

CI4A: Semantic Component Interfaces for Agents Empowering Web Automation

While Large Language Models demonstrate remarkable proficiency in high-level semantic planning, they remain limited in handling fine-grained, low-level web component manipulations. To address this limitation, extensive research has focused on enhancing model grounding capabilities through techniques such as Reinforcement Learning. However, rather than compelling agents to adapt to human-centric interfaces, we propose constructing interaction interfaces specifically optimized for agents. This paper introduces Component Interface for Agent (CI4A), a semantic encapsulation mechanism that abstracts the complex interaction logic of UI components into a set of unified tool primitives accessible to agents. We implemented CI4A within Ant Design, an industrial-grade front-end framework, covering 23 categories of commonly used UI components. Furthermore, we developed a hybrid agent featuring an action space that dynamically updates according to the page state, enabling flexible invocation of available CI4A tools. Leveraging the CI4A-integrated Ant Design, we refactored and upgraded the WebArena benchmark to evaluate existing SoTA methods. Experimental results demonstrate that the CI4A-based agent significantly outperforms existing approaches, achieving a new SoTA task success rate of 86.3%, alongside substantial improvements in execution efficiency.

cs.AI↗

arXiv2015

Collision Geometry and Flow in Uranium+Uranium Collisions

Using event-by-event viscous fluid dynamics to evolve fluctuating initial density profiles from the Monte-Carlo Glauber model for U+U collisions, we report a "knee"-like structure in the elliptic flow as a function of collision centrality, located around the 0.5% most central collisions as measured by the final charged multiplicity. This knee is due to the preferential selection of tip-on-tip collision geometries by a high-multiplicity trigger. Such a knee structure is not seen in the STAR data. This rules out the two-component MC-Glauber model for initial energy and entropy production. Hence an enrichment of tip-tip configurations by triggering solely on high-multiplicity in the U+U collisions does not work. On the other hand, by using the Zero Degree Calorimeters (ZDCs) coupled with event-shape engineering such a selection is possible. We identify the selection purity of body-body and tip-tip events in full-overlap U+U collisions. By additionally constraining the asymmetry of the ZDC signals we can further increase the probability of selecting tip-tip events in U+U collisions.

nucl-th↗

arXiv2015

Shape and flow fluctuations in ultra-central Pb+Pb collisions at the LHC

In ultra-central heavy-ion collisions, anisotropic hydrodynamic flow is generated by density fluctuations in the initial state rather than by geometric overlap effects. For a given centrality class, the initial fluctuation spectrum is sensitive to the method chosen for binning the events into centrality classes. We show that sorting events by total initial entropy or by total final multiplicity yields event classes with equivalent statistical fluctuation properties, in spite of viscous entropy production during the fireball evolution. With this initial entropy-based centrality definition we generate several classes of ultra-central Pb+Pb collisions at LHC energies and evolve the events using viscous hydrodynamics with non-zero shear but vanishing bulk viscosity. Comparing the predicted anisotropic flow coefficients for charged hadrons with CMS data we find that both the Monte Carlo Glauber (MC-Glb) and Monte Carlo Kharzeev-Levin-Nardi (MC-KLN) models produce initial fluctuation spectra that are incompatible with the measured final anisotropic flow power spectrum, for any choice of the specific shear viscosity. In spite of this failure, we show that the hydrodynamic model can qualitatively explain, in terms of event-by-event fluctuations of the anisotropic flow coefficients and flow angles, the breaking of flow factorization for elliptic, triangular and quadrangular flow measured by the CMS experiment. For elliptic flow, this factorization breaking is large in ultra-central collisions. We conclude that the bulk of the experimentally observed flow factorization breaking effects are qualitatively explained by hydrodynamic evolution of initial-state fluctuations, but that their quantitative description requires a better understanding of the initial fluctuation spectrum.

nucl-th↗

arXiv2015

Collision Geometry and Flow in Uranium+Uranium Collisions

Using event-by-event viscous fluid dynamics to evolve fluctuating initial density profiles from the Monte-Carlo Glauber model for U+U collisions, we report a "knee"-like structure in the elliptic flow as a function of collision centrality, located near 0.5% centrality as measured by the final charged multiplicity. This knee is due to the preferential selection of tip-on-tip collision geometries by a high-multiplicity trigger. Such a knee structure is not seen in the STAR data. This rules out the two-component MC-Glauber model for initial energy and entropy production. An enrichment of tip-tip configurations by triggering solely on high-multiplicity in the U+U collisions thus does not work. On the other hand, using the Zero Degree Calorimeters (ZDCs) coupled with event-shape engineering, we identify the selection purity of body-body and tip-tip events in the full-overlap U+U collisions. With additional constraints on the asymmetry of the ZDC signals one can further increases the probability of selecting tip-tip events in U+U collisions.

nucl-th↗

arXiv2014

The iEBE-VISHNU code package for relativistic heavy-ion collisions

The iEBE-VISHNU code package performs event-by-event simulations for relativistic heavy-ion collisions using a hybrid approach based on (2+1)-dimensional viscous hydrodynamics coupled to a hadronic cascade model. We present the detailed model implementation, accompanied by some numerical code tests for the package. iEBE-VISHNU forms the core of a general theoretical framework for model-data comparisons through large scale Monte-Carlo simulations. A numerical interface between the hydrodynamically evolving medium and thermal photon radiation is also discussed. This interface is more generally designed for calculations of all kinds of rare probes that are coupled to the temperature and flow velocity evolution of the bulk medium, such as jet energy loss and heavy quark diffusion.

nucl-th↗

arXiv2013

Event-by-event Hydrodynamic Simulations for Relativistic Heavy-ion Collisions

In this thesis, I show my Ph.D. work on event-by-event hydrodynamic simulations for relativistic heavy-ion collision. I show that event-by-event hydrodynamic simulations have become an indispensable tool for studying relativistic heavy-ion collisions and how it can be used to explain many phenomena. Different chapters focus on different topics; it mainly includes: Chap 2: comparison between single-shot hydrodynamics event-by-event hydrodynamic simulations. Chap 3: using the elliptic and triangular flow data measured by the ALICE collaboration at the LHC to constrain initial condition models. Chap 4: study on correlations between event-plane angles. Chap 5: how resonance decay calculation can be speed up by a factor of 10. Chap 6: study on fluctuations of event planes angle {\Psi}n(pT) and their theoretical and experimental consequences. Chap 7: sampling particles according to the Cooper-Frye formula.

nucl-th↗

arXiv2013

Fluctuating flow angles and anisotropic flow measurements

Event-by-event fluctuations in the initial density distributions of the fireballs created in relativistic heavy-ion collisions lead to event-by-event fluctuations of the final anisotropic flow angles, and density inhomogeneities in the initial state cause these flow angles to vary with the transverse momentum of the emitted particles. It is shown that these effects lead to characteristically different transverse momentum dependencies for anisotropic flow coefficients extracted from different experimental methods. These differences can be used to experimentally constrain flow angle fluctuations in the final state of heavy-ion collisions which, in turn, are sensitive to the initial state density fluctuations and the shear viscosity of the expanding fireball medium.

nucl-th↗

arXiv2012

Resonance Decay Contributions to Higher-Order Anisotropic Flow Coefficients

We show that in hydrodynamic simulations for relativistic heavy-ion collisions, strong resonance decay calculations can be performed with fewer species of particle resonances while preserving good accuracy in single particle spectra and flow anisotropies. Such partial resonance calculations boost computation efficiency by a factor of 10 which is essential for large scale event-by-event simulations.

nucl-th↗

arXiv2012

Imprinting Quantum Fluctuations on Hydrodynamic Initial Conditions

Starting from the two-point covariance function derived in [1], we have developed a toy model to investigate the effect of gluonic fluctuations on the transverse energy profile predicted by MC-KLN Color-Glass Condensate initial conditions. We find that the effect of these fluctuations on the eccentricity harmonics varies strongly with the correlation length and consequently the value of the fixed saturation momentum Qs used in [1]. Using approximate values for the minimum and maximum saturation momenta probed at RHIC in Au-Au collisions, we estimate the increase in the eccentricity coefficients resulting from gluonic fluctuations is as large as 10 to 20 percent in central collisions.

nucl-th↗

arXiv2012

Hydrodynamic event-plane correlations in Pb+Pb collisions at sqrt(s)=2.76ATeV

The recently measured correlations between the flow angles associated with higher harmonics in the anisotropic flow generated in relativistic heavy-ion collisions are shown to be of hydrodynamic origin. The correlation strength is found to be sensitive to both the initial conditions and the shear viscosity of the expanding fireball medium.

nucl-th↗

arXiv2011

Hydrodynamic elliptic and triangular flow in Pb-Pb collisions at sqrt(s)=2.76ATeV

It is shown that a simultaneous comparison of both elliptic and triangular flow from (2+1)-dimensional viscous fluid dynamics with recent measurements in Pb+Pb collisions at the Large Hadron Collider (LHC) favors a small specific shear viscosity (eta/s)_QGP approximately equal to 1/(4pi) for the quark-gluon plasma. Using this viscosity value, the relative magnitude of the elliptic and triangular flow is well described with Monte-Carlo Glauber (MC-Glauber) initial conditions while Monte-Carlo Kharzeev-Levin-Nardi (MC-KLN) initial conditions require twice as large viscosity to reproduce the elliptic flow and then underpredict triangular flow by about 30%.

nucl-th↗

arXiv2011

Event-by-event hydrodynamics for heavy-ion collisions

We compare v2/{\epsilon}2 and v3/{\epsilon}3 from single-shot and event-by-event (2+1)-dimensional hydrodynamic calculations and discuss the validity of using single-shot calculations as substitutes for event-by-event calculations. Further we present a proof- of-concept calculation demonstrating that v2 and v3 together can be used to strongly reduce initial condition ambiguities.

nucl-th↗

arXiv2011

The QGP shear viscosity -- elusive goal or just around the corner?

With the new viscous hydrodynamic + hadron cascade hybrid code VISHNU, a rather precise (O(25%)) extraction of the QGP shear viscosity (eta/s)_QGP from heavy-ion elliptic flow data is possible if the initial eccentricity of the collision fireball is known with <5% accuracy. At this point, eccentricities from initial state models differ by up to 20%, leading to an O(100%) uncertainty for (eta/s)_QGP. It is shown that a simultaneous comparison of elliptic and triangular flow, v_2 and v_3, puts strong constraints on initial state models and can largely eliminate the present uncertainty in (eta/s)_QGP. The variation of the differential elliptic flow v_2(p_T) for identified hadrons between RHIC and LHC energies provides additional tests of the evolution model.

nucl-th↗

arXiv2011

Event-by-event shape and flow fluctuations of relativistic heavy-ion collision fireballs

Heavy-ion collisions create deformed quark-gluon plasma (QGP) fireballs which explode anisotropically. The viscosity of the fireball matter determines its ability to convert the initial spatial deformation into momentum anisotropies that can be measured in the final hadron spectra. A quantitatively precise empirical extraction of the QGP viscosity thus requires a good understanding of the initial fireball deformation. This deformation fluctuates from event to event, and so does the finally observed momentum anisotropy. We present a harmonic decomposition of the initial fluctuations in shape and orientation of the fireball and perform event-by-event ideal fluid dynamical simulations to extract the resulting fluctuations in the magnitude and direction of the corresponding harmonic components of the final anisotropic flow at midrapidity. The final harmonic flow coefficients are found to depend non-linearly on the initial harmonic eccentricity coefficients. We show that, on average, initial density fluctuations suppress the buildup of elliptic flow relative to what one obtains from a smooth initial profile of the same eccentricity, and discuss implications for the phenomenological extraction of the QGP shear viscosity from experimental elliptic flow data.

nucl-th↗