The Archive

User critique of Anthropic's customer support responsiveness following Fable suspension due to US government directive.

COCOLogic-V2 benchmark dataset for visual logical reasoning on real-world images with fine-grained hard negatives for concept bottleneck model evaluation.

Time-series and ML models predict Sri Lankan remittances using exchange rates and oil prices; domain-specific economics work, not AI frontier.

Uses LLM reasoning traces to predict item difficulty in educational assessment via interpretable process evidence extraction.

LLawCo framework enables embodied multi-agent cooperation by learning coordination laws through reflection on failures.

CPAgents uses agentic iteration to auto-generate cardiac phenotypes for disease association studies via composite feature discovery.

Tandem RL training pairs weak and strong agents to maintain verifiable reasoning quality while improving compatibility and readability.

EchoSonar-R is a vision-language model for multi-view echocardiography that generates diagnostic reports with anatomical reasoning.

Physics-informed neural networks with wavelet encoding for inverse conductivity problems; applied math, not AI systems frontier.

Mechanistic analysis reveals jailbreak attacks suppress specific attention heads (ACHs) while safety-aligned heads (SAHs) remain robust.

KL-Coupled Policy Regularization framework for reward-punishment RL with coupled soft-optimal policies via KL-regularized Bellman operators.

Concept-Guided In-Context Segmentation improves robustness of few-shot segmentation via high-level semantic concept extraction.

Dimensionality reduction via autoencoders and PCA for compressing wearable runner telemetry into scalar performance scores.

Low-rank cross-channel mixing module for test-time adaptation under distribution shift in normalization layers.

ADC-GNN: diffusion-guided graph neural network for fraud detection with sparse, imbalanced labels on transaction graphs.

PhysisForcing: physics constraints in video diffusion models to correct implausible robot manipulations and object deformations.

Conceptual analysis: LLMs as degenerate world models; latent-space architectures as strict generalization of autoregressive token prediction.

Automation-in-design framework using deep learning and generative AI for high-tech system design synthesis.

Theoretical result: non-affine gradient aggregation cannot preserve monotonicity in convex learning, blocking last-iterate convergence.

Higher-order Fourier Neural Operator: mode mixing for nonlinear PDE surrogate models beyond linear constant-coefficient regimes.

Physics-constrained neural networks projecting onto Stiefel manifold for energy-conserving surrogate modeling of optical structures.

Low-rank redundancy in task-specific LoRA adapters; single adapters can represent multiple continual learning tasks effectively.

Mechanism-driven internal monitors detect LLM training instability before loss/gradient metrics surface failures, reducing costly compute waste.

BiDeMem uses bidirectional degradation memory with explainable queries for interpretable image restoration beyond PSNR metrics.

It's been two weeks since Anthropic took its Mythos-class models offline after a Friday evening ultimatum from the Trump administration. The company sprang into action immediately, sending a barrage of executives to Washington, DC. But updates have been suspiciously lacking, with no resolution in sight. Anthropic declined to comment multiple times this week about the state of the talks, saying there was no news to share. But the lack of news is the story here. After 14 days of high-intensity negotiations, nobody knows when or if Anthropic's most powerful AI models will come back, let alone wh...

Quantitative theory of Random Language Model in scaling limit reveals condensation phase transition via large-deviation principles.

CME-AQA framework uses cross-view multimodal vision for action quality assessment in Traditional Chinese Medicine rehabilitation.

Nvidia has dominated the AI chip market for years, but the era of total dependence might be ending. OpenAI just shared its plans to spice things up with Jalapeño, its custom inference chip built with Broadcom, joining Google, Apple, and SpaceX in a growing list of companies building their way out of single-supplier risk. The goal is less of a […]

Post-hoc controllable fair classification algorithm balances fairness-accuracy trade-offs without retraining.

OSOR one-step diffusion inpainting removes objects with non-local effects (shadows, reflections) efficiently on edge devices.