arXiv 2608.07401 — Lets space-group symmetry co-evolve alongside geometry and elements via diffusion. Uses shared Wyckoff codebook to ensure valid group-subgroup transitions. Achieves best-in-class stability, novelty, and low structural displacement on relaxation.

arXiv 2608.07427 — Encode time-series numerical data as 2D plots; feed to vision-language models instead of text. Cuts tokens 3.6–10x, inference energy down to 2.5x, while boosting anomaly detection precision 220% over text and 144% over LSTM/ARIMA baselines.

arXiv 2608.07428 — Variational quantum-circuit receiver for multi-antenna communication. Classical detectors have lower average error, but VQC stays stable under extreme channel complexity, offering a predictable complement to classical methods.

arXiv 2608.07429 — TEPA solves memory pollution via explicit validity revocation. When world state reverses, standard append-only scores 0.2; TEPA reaches 0.95. Preserves revoked history for auditability.

arXiv 2608.07430 — Diffusion language models inherit safety footprints from autoregressive parents. Pruning safety neurons raises attack success from 2.6% to 73.8% on some models; SN-Guided Diffusion achieves 74%+ transfer against Gemini.

arXiv 2608.07431 — GALA embeds AOD motion constraints directly into quantum error correction via group-action lifts, achieving rate-1/2 codes on just 132 qubits with projected logical error <10^-8.

arXiv 2608.07432 — By Baichen Tan, Audrey Yang, Cynthia Rush. Proves that the full singular-value spectrum, not just aspect ratio or average power, governs convex-estimator performance limits under high dimensions.

arXiv 2608.07433 — Wasserstein Policy Gradient reduces entropy-regularized LQ control to ODEs, achieving exponential convergence without singularities even as temperature approaches zero.

arXiv 2608.07434 — COVER wraps any video temporal grounder in statistical guarantees via conformal prediction, converting point estimates into certified intervals that don't degrade with changes in world state.

arXiv 2608.07435 — SABRE: a dynamic benchmark that stress-tests vision-language models on unusual contexts and counterfactual scenarios, revealing how often they rely on shortcuts rather than true visual understanding. Authors: Jian, Okhonko, Turian.