Chronological feed of everything captured from Ted Cruz.
paper / tedcruz / Jun 13
Pre-trained language models struggle with keyphrase extraction in long documents due to context window limitations that prevent capturing scattered evidence. The paper proposes an attention expansion mechanism that augments PLM embeddings with information from out-of-context chunks using pre-trained word embeddings, expanding effective context without requiring expensive full-document attention or LLM-based inference. Evaluated across five PLM backbones and five benchmark corpora, attention expansion consistently improves F1 scores and outperforms state-of-the-art baselines, suggesting the mechanism provides complementary information beyond simple context extension.
keyphrase-extractionlong-contextplmattention-mechanismnlparxiv-paper
“Salient keyphrase evidence in long documents is often scattered across distant sections beyond the limited context window of standard PLMs”
paper / tedcruz / Jun 13
Direct self-distillation for long-horizon tool-use agents amplifies both useful skills and harmful shortcuts indiscriminately because it mimics teacher behavior without distinguishing which actions the reward signal actually endorses. Sibling-Guided Credit Distillation (SGCD) decouples credit assignment from policy imitation by using contrastive pairs of successful/failed rollouts to generate dense stepwise credit signals, then reshaping policy advantages through bounded detached weights rather than competing actor losses. This approach maintains policy gradient control while enabling fine-grained credit reassignment, yielding measurable improvements on tool-use benchmarks without exposing the deployed agent to external LLM or oracle dependencies.
reinforcement-learningcredit-assignmenttool-use-agentsself-distillationlong-horizon-reasoningllm-training
“Direct token-level self-distillation can amplify both useful skills and harmful shortcuts together when applied to long-horizon tool-use RL”
paper / tedcruz / Jun 13
Belle and Belle II collaborations report the first observation of B-meson decays to pairs of charmed baryon-antibaryon states: B⁺ → Σc(2455)⁺⁺ Ξ̄c'⁻ and B⁰ → Σc(2455)⁰ Ξ̄c'⁰, with statistical significances of 6.4σ and 5.3σ respectively. The measured branching fractions are (1.68 ± 0.31 ± 0.12⁺¹·⁴⁹₋₀.₅₄) × 10⁻³ and (1.28 ± 0.32 ± 0.10⁺⁰·³⁰₋₀.₂₁) × 10⁻³ respectively, based on 771.6 million Υ(4S) decays from Belle and 520.6 million from Belle II. This marks the first instance of B-meson decays into baryon-antibaryon pairs from the same SU(3) flavor sextet.
particle-physicsb-meson-decayscharmed-baryonsbelle-experimenthigh-energy-physicsexperimental-measurement
“Belle and Belle II observed the decays B⁺ → Σc(2455)⁺⁺ Ξ̄c'⁻ and B⁰ → Σc(2455)⁰ Ξ̄c'⁰ with statistical significances of 6.4σ and 5.3σ respectively”
paper / tedcruz / Jun 10
This paper analyzes the circumcentered-reflection method (CRM) for projecting onto the intersection of two subspaces, $U$ and $V$. The authors demonstrate that when CRM is initialized in $V$, it achieves a sharp and optimal convergence rate of $\rho_V = (\sin^2\theta_p - \sin^2\theta_F) / (\sin^2\theta_p + \sin^2\theta_F)$, which is strictly better than the previously known rate $c_F$ unless the Friedrichs angle $\theta_F$ equals $\pi/2$. This rate is shown to be equivalent to existing linesearch methods and can be improved only by using memory-intensive methods like Chebyshev semi-iteration.
optimization algorithmslinear convergencesubspace methodsnumerical analysismathematical optimization
“The circumcentered-reflection method (CRM) converges at a rate of $\rho_V=(\sin^2\theta_p-\sin^2\theta_F)/(\sin^2\theta_p+\sin^2\theta_F)$ when initialized in subspace V, which is strictly smaller than the known rate $c_F$ (cosine of the Friedrichs angle) for CRM initialized in $U$, provided $\theta_F < \pi/2$.”
paper / tedcruz / Jun 10
This paper provides a complete spectral analysis of the Lévy area matrix — a skew-symmetric random matrix from d-dimensional Brownian motion. The authors derive exact density formulas for its singular values, characterize the spectrum as a determinantal point process, and establish asymptotic results: the empirical spectral distribution converges to an absolute Cauchy distribution, the largest singular values grow as O(d) with Gaussian fluctuations, and the smallest decay as O(1/d) with sine-kernel universality in the bulk.
stochastic-analysisprobabilityrandom-matrix-theorymathematical-physicsbrownian-motion
“An explicit formula for the density of singular values of the Lévy area matrix is derived.”
paper / tedcruz / Jun 1
Free electrons can excite waveguided optical modes with high spatial precision, but coupling efficiency has historically been constrained by short interaction times along fixed straight-line trajectories. This theoretical work demonstrates that applying an electrostatic bias to a silicon waveguide introduces a tunable turning point in grazing electron trajectories, dramatically extending the interaction and enabling unity-order coupling. The configuration simultaneously suppresses lossy high-energy channels (above the silicon band gap) via aloof interaction geometry, while preserving strong excitation of targeted modes. Using 100 keV electrons, the authors predict yields exceeding ten photons per electron with voltage-tunable control.
free-electron-photon-couplingwaveguide-opticsquantum-opticsnanophotonicsfock-stateselectrostatic-steeringcondensed-matter-physics
“Electrostatic steering of grazing electrons via a biased silicon waveguide enables unity-order coupling to waveguided photonic modes.”
paper / tedcruz / Jun 1
This paper introduces the Dimming Space-Time Code (DSTC), a structured coding scheme for color-shift keying (CSK) visible light communication (VLC) systems that decouples optical power control from chromaticity, addressing a critical gap in VLC dimming research. The DSTC enforces physical feasibility, color preservation, and identifiability constraints simultaneously via a temporal dimming matrix. Two receiver architectures are proposed: a pilot-assisted zero-forcing (ZF) receiver and a tensor-based semi-blind PARAFAC receiver that jointly estimates channel and symbols using a single training slot. The semi-blind approach substantially reduces training overhead and delivers diversity gains with improved BER, especially in large-scale MIMO configurations.
visible-light-communicationspace-time-codingmimosignal-processingwireless-communicationtensor-methodsoptical-wireless
“The proposed DSTC achieves diversity gains and substantial BER reductions compared to conventional CSK in VLC systems.”
paper / tedcruz / May 18 / failed
paper / tedcruz / May 18 / failed
paper / tedcruz / May 18 / failed
paper / tedcruz / May 18 / failed