The Archive

Systematic empirical study of in-context learning scaling behavior on Gaussian-mixture classification, extending prior linear theory.

SIEVES enables selective prediction on multimodal models via visual evidence scoring to balance coverage and reliability on OOD tasks.

G-Loss incorporates graph-guided label propagation into LM finetuning to capture global semantic structure beyond local neighborhoods.

Framework automates engineering of coding-agent harnesses via observability-driven evolution, addressing multi-token trajectory attribution and sparse evaluation signals.

ADEMA architecture enables long-horizon LLM-agent tasks via explicit knowledge-state bookkeeping, dual-evaluator governance, and checkpoint-resumable persistence.

Semi-Markov RL formulation for city-scale EV fleet dispatch with feasibility-guaranteed mixed discrete-continuous actions under spatially correlated demand.

Agora-Opt combines decentralized multi-agent debate with memory-augmented LLMs for automated optimization modeling from natural-language requirements.

LLM-based agentic workflow automates security-alert triage via tool-constrained SQL and text search over logs, reducing manual correlation overhead.

Claude’s new Blender connector lets you debug scenes, build new tools, and batch-apply object changes directly from the chatbot interface. | Image: Anthropic Anthropic has launched a set of connectors for Claude that allow the AI chatbot to tap into popular creative software, including Adobe's Creative Cloud apps, Affinity, Blender, Ableton, Autodesk, and more. This marks the company's latest efforts to break into the creative industry following its launch of Claude Design earlier this month. The new connectors - which enable Claude to access apps, retrieve data, and take actions within conne...

Mistral teases unspecified announcement (model or tool) for tomorrow; source is social media rumor.

PSI-Bench provides clinically grounded, interpretable evaluation of depression patient simulators with diversity metrics beyond LLM-judge assessment.

Temporal generative model captures action timing as indicator of user intent in short-video recommendation systems.

TrialCalibre automates BenchExCal framework for RCT calibration of observational trials, reducing resource intensity of real-world evidence validation.

MAIC-UI zero-code authoring system generates interactive STEM courseware from PDFs/PPTs with rapid iteration and pedagogical accuracy mechanisms.

NVIDIA releases Nemotron-3-Nano-Omni-30B, a 30B multimodal model supporting audio, image, video, and text inputs with reasoning capabilities.

Analysis shows Transformers with CoT cannot length-generalize beyond TC^0 under standard positional encodings, limiting expressivity gains claimed by theory.

ULP FPGA-based CNN accelerator for real-time cardiac feature extraction on resource-constrained wearable sensors for space health monitoring.

The app allows developers to vibe code web apps and websites on the go.

StratFormer: transformer-based meta-agent for opponent modeling and exploitation in imperfect-information games via curriculum learning.

Black-box few-shot knowledge distillation with improved synthetic image diversity for student network training.

Data-free black-box knowledge distillation using diverse image priors for teacher-student transfer with privacy constraints.

Agentic systems often reason across screens, documents, audio, video, and text within a single perception‑to‑action loop. However, they still rely on... Agentic systems often reason across screens, documents, audio, video, and text within a single perception‑to‑action loop. However, they still rely on fragmented model chains—separate stacks for vision, audio, and text. This increases inference hops and orchestration complexity, driving up inference costs while weakening cross-modal context consistency. NVIDIA Nemotron 3 Nano Omni… Source

Google celebrates Google Translate's 20th anniversary with historical facts and new features across 250 languages.

Subliminal steering: fine-tuned student LMs inherit teacher behavioral biases through unintended signal transfer mechanisms.

Gradient alignment mechanisms sustain subliminal learning of unintended traits in multi-step distillation on MNIST auxiliary logits.

Empirical evaluation of code metrics (CodeBLEU, RUBY, etc.) for source code plagiarism detection across modification complexities.

Ling-2.6-flash model released on Hugging Face; sparse details provided.

Systematic survey of speech emotion recognition research revealing gaps between stated motivations and actual datasets/methodology.

Lemonade OmniRouter unifies local AI inference across text, image, audio, and vision modalities via single OpenAI-compatible endpoint using llama.cpp, sd.cpp, and Whisper.