The Archive

In a Saturday morning post, Microsoft's CEO wrote that it’s time “to step back and assess the trust architecture” of AI.

Is Apple hoping to get into the AI-generated podcast business?

Artists are taking to social media to complain that DistroKid has unceremoniously removed their work without notice. Now DistroKid has confirmed to The Verge that the takedowns are a direct response to claims made by UMG. The label filed a lawsuit in September claiming that DistroKid has created an "AI-slop pipeline." Artists are saying that their non-AI works are being caught up in the purge, however. Musician and self-described "Video Game boy" McGwire reported that six of his songs had been removed from streaming. As has been the case with others, McGwire says that there was no communicati...

Structural Chain-of-Thought trains LLMs to generate verifiable executable computational graphs instead of free-form reasoning, rewarding structural validity and answer correctness.

RELAGRAM applies latent recursion and compact transformers to combinatorial graph optimization, specifically the capacitated vehicle routing problem using edge token encoding.

RegExtract benchmark evaluates LLM performance on extracting structured regulatory data from legally dense PFAS documents with source traceability requirements.

SEED-MAX algorithm optimizes random feature selection for multi-class classification by recursively identifying effective features and reducing redundancy.

AnchorDLM improves controllability of diffusion language models by generating anchor fragments positioned to guide constrained text generation from arbitrary sequence positions.

SwaT (Swarm Telepathy) enables cross-family KV-cache transfer between LLMs without prefilling, increasing bandwidth and capabilities for multi-agent text-based systems.

TurnCost analyzes loop engineering for coding agents, showing median observations are small and identifying practical optimization opportunities beyond compression.

VLCQ post-training quantization method uses variable-length coding instead of fixed-point integers to achieve higher compression rates on resource-constrained devices.

NARS combines retrieval-augmented LLM reasoning with gradual argumentation semantics to detect persuasiveness in online debate discussions via neuro-symbolic structures.

Marvo self-evolves visual agents by using a challenger-solver pair to learn tool invocation from unlabeled images with margin reward feedback.

GAMW embeds watermarks during TTS generation rather than post-hoc to improve robustness of audio provenance tracking and authenticity verification.

Depth-recurrent model generalizes across Sudoku-like discrete worlds with novel rule combinations, testing transfer learning at four levels of compositional novelty.

Theoretical analysis shows unconstrained machine-learned interatomic potentials face exponentially ill-conditioned prediction-sampling trade-offs in high dimensions.

5.5M-parameter 3D U-Net reconstructs magnetic field and density from sparse plasma diagnostics, proof-of-concept toward multimodal foundation model for FLARE.

Inference on Demand (IoD) exposes split AI inference as a 5G network service via CAMARA-like API, enabling context-aware edge inference.

Comparative benchmark of PSPNet, DeepLabv3, UNet, UNet++ for dental radiograph segmentation under standardized experimental conditions.

Tilted temperature decoding rule derived as closed-form solution to pass@k maximization, explaining why higher temperatures benefit larger sample budgets in LLMs.

Oracle-guided fine-tuning using N-best ASR hypotheses enables post-editing without retraining recognizer, improving WER via curated correction and regularization pools.

Trace Evolution Bench protocol isolates genuine agent learning from frozen rescoring effects, scoring self-improving agents on coding tasks with hidden unit tests.

Belief Embedding Tree Search approximates continuous belief states in POMDPs using learned embeddings within Monte Carlo planning trees for tractable inference.

Depression detection on social media via personality-driven transfer learning and post-level supervision to reduce false positives.

Online entropy-based temperature sampling for LLM decoding reduces repetitive generation using information-theoretic optimization.

Sparse Mixture-of-Experts models achieve compute savings but not sample efficiency; oracle inequalities bound adaptive routing cost.

Generative search engines (ChatGPT, Gemini, Perplexity) cite web pages based on title/paragraph matching, not structured data or speed.

InCLAD benchmark for continual visual anomaly detection across industrial components addresses deployment-time task increments.

Decision foundation model (Clef-Flash, 9B) fine-tuned with QLoRA for 24-class disease classification balances speed and context.

UltraBERT modernizes encoder pre-training; saturates beyond few trillion tokens and matches LLM innovations for classification and retrieval.