The Archive

Information bottleneck analysis shows multi-agent LLM systems gain advantage only under bandwidth constraints, simulating single-agent with more context.

ToolSciVer framework uses visual tools and RL to improve multimodal scientific claim verification from figures, tables, and charts.

Lightweight framework for auditable trustworthiness assessments in AI lifecycle governance with formal representation and monitoring.

CRAFT converts rubric-based evaluations into capability diagnoses and generates targeted fine-tuning data addressing model weaknesses.

Stratechery weekly digest covering mainframe obsolescence, OpenAI developments, and Netflix competitive position.

Rate-utility frontier analysis compares tokens, bytes, and pixels as language encodings across 13 languages under controlled capacity.

Methodology for harmonized AI safety thresholds across misuse, malfunction, and systemic risks to prevent race-to-the-bottom in standards.

Honest Quorum Problem introduces epistemic faults for Byzantine fault tolerance in agentic validators, extending BFT guarantees to reasoning errors.

Empirical study of how pretraining choices affect RL post-training efficiency and what RL changes in model internals using controlled experiments.

DADiff uses diffusion models for online domain adaptation in RL with limited target interactions and source domain pretraining.

Interpretability study reveals AlphaFold2's learned protein conformational landscapes via Evoformer weight analysis and Gaussian smoothing.

Technique to mitigate shortcut learning in transformer-based spoken English proficiency assessment systems.

Pick-to-Learn methodology applied to hyperparameter calibration for Model Predictive Control in aircraft flight planning.

Official estimates Google and Apple likely made millions in nudify app fees.

Physics-informed U-Net for 10–90 min precipitation nowcasting using multi-variable radar data in high-resolution settings.

HCIG framework uses hierarchical cross-modal graph networks to detect sarcasm and cyberbullying via text-visual incongruity.

JoyNexus: multi-tenant post-training service for Vision-Language-Action models with efficient GPU resource pooling.

Frontier LLMs fail at exact string copying due to positional encoding bias; 2D-RoPE organizes text as 2D grid to fix.

Tutorial and survey on agentic AI for 5G/6G networks covering reasoning, planning, multi-agent coordination, and standardization.

Spatial normalization technique for cross-domain retinal OCT layer segmentation in clinical neurodegenerative disease analysis.

Model merging (TIES, RAM+) matches jointly trained multi-task RL on Qwen3-8B AppWorld agent benchmark; geometry analysis explains parity.

New benchmark evaluates frontier LLMs on real analytical knowledge work—synthesizing information, judgment under uncertainty, strategic thinking—beyond factual recall and coding.

Methods for gene regulatory network inference from genome-wide data using deep probabilistic models with uncertainty quantification.

Loopie: MoE Transformer models (20B and 6B parameters) that outperform vanilla scaling by efficiently looping under fixed compute budgets.

DELUGE: multimodal deep learning system for continental-scale daily pluvial flood damage prediction at 1km resolution using foundation model embeddings.

Patreon is strengthening its defenses against AI scraping by working with Cloudflare to block bots that train AI models on creators’ content without permission. The move marks a shift away from relying on websites using robots.txt alone to actively block unauthorized AI training.

SciForge: AI-native multimodal workbench for scientific discovery with agent-accessible services for code, datasets, workflow execution, and paper management.

Tabular foundation model for pre-fault dynamic security assessment in power systems using in-context learning to reduce labeling and improve contingency generalization.

Quantum elastic weight consolidation: QFI-based regularization to mitigate catastrophic forgetting in variational quantum classifiers on sequential tasks.