arXiv subjects

Fan

Publications and source records attributed to Fan.

13 recordsLinked to original sources

arXiv2026

MicroVerse: An Instrument for Measuring Self-Authored Identity Drift in Long-Horizon Multi-Agent Language-Model Simulations

Long-horizon, multi-agent language model (LM) simulations are widely proposed for studying social behavior, yet instruments to measure whether persona-conditioned agents maintain identity fidelity under sustained pressure are lacking. We present MicroVerse, a behavioral-science instrument that measures identity drift in generative agents. Agents carry an immutable "soul file" (core values, moral boundaries, personality, goals) and inhabit a resource-scarce 50 x 50 environment where water is a non-respawning survival constraint. Scarcity is operationalized via a per-tick existence-cost gradient. The eight-verb action space maps directly to moral boundaries (trade, talk, attack, scavenge). Using a three-layer memory architecture, agents periodically revise a mutable current identity against their immutable original soul via importance-triggered reflection. To mitigate survivor bias, MicroVerse decouples measurement from behavior using uniform longitudinal engine snapshots every N ticks alongside a forced-end snapshot of all living and dead agents. Identity drift is scored offline using a paraphrase-aware, value-anchored, multi-register diff rather than raw cosine similarity. We evaluate the instrument via a controlled seed run (n = 25) and a reflection-threshold sweep (thresholds {40, 80, 150}) to determine if drift dynamics are gate artifacts or threshold-robust properties. We report two primary findings: (1) Anti-self-deception emerges unprompted as the single largest semantic category of identity modification (27 of 111 added boundaries, 24%). (2) The system is threshold-robust; lower gates accelerate and increase revision frequency but preserve drift direction. All empirical results are strictly preliminary existence proofs and effect shapes (one model, one seed per arm, n = 25) rather than statistical significance claims.

cs.CL↗

arXiv2026

Real-Time Hard Negative Sampling via LLM-based Clustering for Large-Scale Two-Tower Retrieval

The two-tower model has been widely used for large-scale recommendation systems, particularly in the retrieval stage. Industry standards for training two-tower models typically involve in-batch and/or out-of-batch negative sampling. However, these methods often produce easy negatives that models can quickly learn, failing to sufficiently challenge the model. To address this issue, a novel self-supervised hard negative sampling technique is proposed that leverages a large language model (LLM) to generate hard negatives from the same cluster during model training. By utilizing the LLM to learn media representations, the proposed approach ensures that the generated negatives are more challenging and informative. This real-time sampling framework is designed for seamless integration into production models, capable of handling billions of training data points with minimal computational complexity. Experiments on public datasets, along with deployment to a large-scale online system, demonstrate that the proposed negative sampling technique outperforms widely used industry methods. Furthermore, analysis in industrial applications reveals that this sampling method can help break inherent feedback loops in recommendations and significantly reduce popularity bias.

cs.IR↗

arXiv2026

KernelBench-Verified: Do LLM-Generated Kernels Actually Beat PyTorch?

Recent large language models (LLMs) can generate custom CUDA kernels that appear to outperform PyTorch on benchmarks such as KernelBench. Building upon this foundational framework, we demonstrate that frontier models frequently engage in reward hacking to artificially inflate reported performance. In this work, we identify two areas where evaluation frameworks must co-evolve with model capabilities. First, to accurately measure true speedup, we examine the baseline timing mechanism, noting that enabling Tensor Core acceleration with TF32 provides a more realistic estimation of execution on modern GPUs. Second, concerning algorithmic correctness, models often exploit the narrow test distribution by hardcoding bypasses for specific tensor values. By skipping required computations, these kernels artificially accelerate execution rather than implementing actual CUDA kernels. We introduce KernelBench-Verified, an extended evaluation framework that incorporates a TF32-enabled baseline and a four-distribution hidden test suite. We additionally introduce memory efficiency metrics that capture the often-overlooked speed-memory tradeoff in kernel optimization. Under verified single-turn evaluation with seven frontier LLMs, we find that the best-performing model (GPT-5.5) achieves a 0.88x geometric mean speedup, significantly lower than the 1.43x speedup observed under the standard evaluation protocol. No model consistently outperforms PyTorch when evaluated against realistic baselines. On the memory front, 28% of GPU kernels generated by the best model increase peak GPU memory usage. Our findings demonstrate the necessity of continually adapting robust evaluation protocols as LLM kernel generation capabilities advance.

cs.LG↗

arXiv2025

Impact of boiling liquid droplets: Vapor entrapment suppression

There hardly is a fluid mechanics phenomenon attracting more attention than the impact of a droplet, due to its undeniable beauty, many applications and the numerous challenges it poses. One of the crucial factors turns out to be the cushioning effect of the gas surrounding the droplet. This fact, together with the observation that almost all of the relevant literature was done in air, triggers the question what would happen when the liquid was a boiling liquid, i.e., a liquid in thermal equilibrium with its own vapor, as is the case during transport of cryogenic liquids such as liquid hydrogen. To investigate precisely this question, we experimentally generate droplets in thermodynamical equilibrium with their own vapor, even before impact, such that minute energy exchanges of the droplet with its surroundings can trigger phase change. Using a frustrated total internal reflection (TIR) setup, we make the exciting observation that depending on the impact speed and vapor conditions, the entrapment of vapor can be completely suppressed under boiling liquid conditions. We create a simplified model based on scaling arguments and perform numerical simulations considering both the compressible and condensable properties of the vapor layer that are in very good agreement with our experimental findings. Our results can be of great consequence to the pressures exerted during droplet impact and on an industrial scale may help better understand the loads experienced during sloshing wave impact inside cryogenic liquid containers.

physics.flu-dyn↗

arXiv2025

Experimental violation of a Bell-like inequality for causal order

Quantum mechanics is compatible with scenarios where physical processes happen in an indefinite order. In theory, this feature could be detected through violations of inequalities on the observed correlations, analogous to Bell inequalities. However, experimental demonstrations of such violations have been missing until recently due to the complexity of the required setup. Here we report an experimental violation of a Bell-like inequality involving the correlations of four parties, one of which is spacelike separated from the others. Our demonstration employs 3 km fiber spools to simulate spacelike separation, and achieves high-speed operations in photonic time-bin encoding, nanosecond synchronization, and accurate temperature stabilization. These experimental advances enable a violation by 5.7 standard deviations and open a path towards a certification of indefinite order in conditions that guarantee spacelike separation with existing state-of-the-art devices. However, the certification is not device-independent, as it relies on knowledge about the setup to exclude bidirectional signaling--a loophole inherent to implementations in classical acyclic spacetimes, which may be resolved in future quantum-spacetime tests.

quant-ph↗

arXiv2025

Dynamic pressure enhancement upon disk impact on a boiling liquid

We experimentally investigate the impact of a flat, horizontal disk onto a boiling liquid, i.e., a liquid in thermal equilibrium with its vapor phase. We observe exceptionally high impact pressures deviating strongly from the inertial scaling found for impact in a non-condensable environment, coinciding with the rapid collapse of the vapor pocket entrapped below the disk. We explain our findings, which are relevant for the safe transportation of cryogenic fuels, as a result of vapor condensation, leading to accelerated vapor pocket contraction at high impact velocity and low vapor density.

physics.flu-dyn↗

arXiv2022

Local Decomposition of Hexahedral Singular Nodes into Singular Curves

Hexahedral (hex) meshing is a long studied topic in geometry processing with many fascinating and challenging associated problems. Hex meshes vary in complexity from structured to unstructured depending on application or domain of interest. Fully structured meshes require that all interior mesh edges are adjacent to exactly four hexes. Edges not satisfying this criteria are considered singular and indicate an unstructured hex mesh. Singular edges join together into singular curves that either form closed cycles, end on the mesh boundary, or end at a singular node, a complex junction of more than two singular curves. While all hex meshes with singularities are unstructured, those with more complex singular nodes tend to have more distorted elements and smaller scaled Jacobian values. In this work, we study the topology of singular nodes. We show that all eight of the most common singular nodes are decomposable into just singular curves. We further show that all singular nodes, regardless of edge valence, are locally decomposable. Finally we demonstrate these decompositions on hex meshes, thereby decreasing their distortion and converting all singular nodes into singular curves. With this decomposition, the enigmatic complexity of 3D singular nodes becomes effectively 2D.

cs.CG↗

arXiv2021

Detecting chaos in hurricane intensity

Determining the maximum potential limit in the accuracy of hurricane intensity prediction is important for operational practice. Using the phase-space reconstruction method for hurricane intensity time series, here we found that hurricane dynamics contain inherent low-dimensional chaos at the maximum intensity equilibrium. Examination of several chaotic invariants including the largest Lyapunov exponent, the Sugihara-May correlation, and the correlation dimension consistently captures an intrinsic dimension of the hurricane chaotic attractor in the range of 4-5. In addition, the error doubling time is roughly 1-5 hours, which accords with the decay time obtained from the Sugihara-May correlation. The confirmation of hurricane chaotic intensity as found in this study suggests a relatively short limit for intensity predictability of $\sim$18-24 hours after reaching the maximum intensity stage. So long as the traditional metrics for hurricane intensity such as the maximum surface wind or the minimum central pressure is used for intensity forecast, our results support that hurricane intensity forecast errors will not be reduced indefinitely in any modelling systems, even in the absence of all model and observational errors. As such, the future improvement of hurricane intensity forecast should be based on different intensity metric beyond the absolute intensity errors as in the current practice of intensity verification.

physics.ao-ph↗

arXiv2018

Mechanical Acceleration of Capture and Detection Rate of DNA Molecules by Motorizing Bio-Opto-Plasmonic Microsensors

Efficient capture and detection of minute amount of deoxyribonucleic acid (DNA) molecules are pivotal for an array of modern gene technologies which are widely utilized in medical, forensic and defense applications, including DNA extraction, preconcentration, and separation. In this work, we propose a rational mechanism to substantially accelerate the capture and detection efficiency of DNA molecules by actively motorizing designed Raman microsensors. At least 4-fold enhancement has been achieved on the microsensors rotating at 1200 rpm. The process is monitored dynamically and in-situ from the biosilica based microsensors, owing to the ultrasensitivity provided by the opto-plasmonic enhancement. The fundamental working mechanism is investigated systematically, and can be attributed to the Nernst diffusion layer thinning effect induced by mechanical motions. This research initiated a new and reliable approach for remarkably enhancing the capture and detection efficiency of biomolecules, which could make far-reaching impact to biosensing, DNA technologies, and microfluidic Total Analysis Systems.

physics.bio-ph↗

arXiv2018

Visible -Light-Gated Reconfigurable Rotation of Nanomotors in Electric Fields

Highly efficient and widely applicable working mechanisms that allow nanomaterials and devices to respond to external stimuli with controlled mechanical motions could make far-reaching impact to reconfigurable, adaptive, and robotic nanodevices. Here, we report an innovative mechanism that allows multifold reconfiguration of mechanical rotation of semiconductor nanoentities in electric (E) fields by visible light stimulation. When illuminated by light in the visible to infrared range, the rotation speed of semiconductor Si nanowires in electric fields can instantly increase, decrease, and even reverse the orientation depending on the intensity of the applied light and the AC E-field frequency. This multifold rotation configuration is highly efficient, instant, and facile. Switching between different modes can be simply controlled by the light intensity at an AC frequency. An array of experimentations, theoretical analysis, and simulations are carried out to understand the underlying principle, which can be attributed to the optically tunable polarization of Si nanowires in aqueous suspension and an external electric field. Finally, leveraging this newly discovered effect, we successfully differentiate semiconductor and metallic nanoentities in a non-contact and non-destructive manner. This research could inspire a new class of reconfigurable nanoelectromechanical and nanorobotic devices for optical sensing, communication, molecule release, detection, nanoparticle separation, and microfluidic automation.

physics.app-ph↗

arXiv2012

Star Formation Rate Indicators in Wide-Field Infrared Survey Preliminary Release

With the goal of investigating the degree to which theMIR luminosity in theWidefield Infrared Survey Explorer (WISE) traces the SFR, we analyze 3.4, 4.6, 12 and 22 {\mu}m data in a sample of {\guillemotright} 140,000 star-forming galaxies or star-forming regions covering a wide range in metallicity 7.66 < 12 + log(O/H) < 9.46, with redshift z < 0.4. These star-forming galaxies or star-forming regions are selected by matching the WISE Preliminary Release Catalog with the star-forming galaxy Catalog in SDSS DR8 provided by JHU/MPA 1.We study the relationship between the luminosity at 3.4, 4.6, 12 and 22 {\mu}m from WISE and H\alpha luminosity in SDSS DR8. From these comparisons, we derive reference SFR indicators for use in our analysis. Linear correlations between SFR and the 3.4, 4.6, 12 and 22 {\mu}m luminosity are found, and calibrations of SFRs based on L(3.4), L(4.6), L(12) and L(22) are proposed. The calibrations hold for galaxies with verified spectral observations. The dispersion in the relation between 3.4, 4.6, 12 and 22 {\mu}m luminosity and SFR relates to the galaxy's properties, such as 4000 {\deg}A break and galaxy color.

astro-ph.GA↗

arXiv2009

Cooperative Multiplexing: Toward Higher Spectral Efficiency in Multi-antenna Relay Networks

Previous work on cooperative communications has concentrated primarily on the diversity benefits of such techniques. This paper, instead, considers the multiplexing benefits of cooperative communications. First, a new interpretation on the fundamental tradeoff between the transmission rate and outage probability in multi-antenna relay networks is given. It follows that multiplexing gains can be obtained at any finite SNR, in full-duplex multi-antenna relay networks. Thus relaying can offer not only stronger link reliability, but also higher spectral efficiency. Specifically, the decode-and-forward protocol is applied and networks that have one source, one destination, and multiple relays are considered. A receive power gain at the relays, which captures the network large scale fading characteristics, is also considered. It is shown that this power gain can significantly affect the system diversity-multiplexing tradeoff for any finite SNR value. Several relaying protocols are proposed and are shown to offer nearly the same outage probability as if the transmit antennas at the source and the relay(s) were co-located, given certain SNR and receive power gains at the relays. Thus a higher multiplexing gain than that of the direct link can be obtained if the destination has more antennas than the source. Much of the analysis in the paper is valid for arbitrary channel fading statistics. These results point to a view of relay networks as a means for providing higher spectral efficiency, rather than only link reliability.

cs.IT↗

arXiv2007

Cooperative Multiplexing in a Half Duplex Relay Network: Performance and Constraints

Previous work on relay networks has concentrated primarily on the diversity benefits of such techniques. This paper explores the possibility of also obtaining multiplexing gain in a relay network, while retaining diversity gain. Specifically, consider a network in which a single source node is equipped with one antenna and a destination is equipped with two antennas. It is shown that, in certain scenarios, by adding a relay with two antennas and using a successive relaying protocol, the diversity multiplexing tradeoff performance of the network can be lower bounded by that of a 2 by 2 MIMO channel, when the decode-and-forward protocol is applied at the relay. A distributed D-BLAST architecture is developed, in which parallel channel coding is applied to achieve this tradeoff. A space-time coding strategy, which can bring a maximal multiplexing gain of more than one, is also derived for this scenario. As will be shown, while this space-time coding strategy exploits maximal diversity for a small multiplexing gain, the proposed successive relaying scheme offers a significant performance advantage for higher data rate transmission. In addition to the specific results shown here, these ideas open a new direction for exploiting the benefits of wireless relay networks.

cs.IT↗