Qian Liu

A single attention module dynamically switches between linear prefilling and full decoding, achieving $21.9\times$ faster prefilling and Transformer-level long-context recall without the quadratic bottleneck.

Doubling photocathode coverage with cost-effective MCP-PMTs cuts underground muon direction reconstruction error to just 6° while boosting detection efficiency by 59%, successfully isolating CJPL's first up-going neutrino-induced event.

This work proposes an integrated view on the use of LLMs for EDA and establishes an LLM-enabled behavior driven hardware development workflow, introducing and defining Formal Verification Gherkin Scenarios (FV Gherkin Scenarios), unlocking CNL specifications as the foundation for formally verified hardware designs via Formal Property Verification (FPV).

Decoupling structured sensor artifacts with a Deep Image Prior allows modular spectral unmixing to slash reconstruction error by up to 69.5% without corrupting physical abundance estimates under flawed endmember priors.

Escalating to a larger LLM is counterproductive when it corrupts correct answers, meaning optimal cascading requires routing on net rescue-versus-harm rather than model uncertainty.

Video-grounded diagnostics reveal a staggering array of failure types in one-shot web application generation, pushing the boundaries of evaluation beyond mere artifact scoring.

Achieving accurate material decomposition in sparse-view DECT could revolutionize medical imaging by enabling safer, lower-radiation scans without sacrificing detail.