The Archive

JustFit enables 200K-token LLM inference on 24GB laptop via just-in-time state management and KV compression.

Coupled Calibration and Learning mitigates teacher bias in LLM distillation without target-domain reward feedback.

Google DeepMind releases Gemini 3.8 Live and 3.8 Live Extended Thinking variants with real-time interaction and enhanced reasoning capabilities.

Meta One bundles expanded access to the company’s AI tools with premium features across Facebook, Instagram and WhatsApp.

Multi-agent decomposition trees sacrifice output yield to maintain integrity; discovery rate scales as N^(1-δ).

How can a 30B-parameter model activate only 3B parameters per token, and still use the capacity of the larger model? Nemotron 3.5 Lightning illustrates the... How can a 30B-parameter model activate only 3B parameters per token, and still use the capacity of the larger model? Nemotron 3.5 Lightning illustrates the answer: It uses a Mixture-of-Experts (MoE) architecture that selects only a subset of its parameters for each token. There are two dominant model architectures: Dense model and MoE. How a model organizes its parameters matters as much as… Source

DELTA and TARQA improve table understanding via structured text over vision-language models; supports multilingual documents.

Power is a defining constraint for AI factories. As AI workloads demand a full compute platform to serve them, each component of that platform must maximize... Power is a defining constraint for AI factories. As AI workloads demand a full compute platform to serve them, each component of that platform must maximize output within the factory’s limited power budget. This makes performance per watt—rather than raw, unnormalized throughput—the ultimate measure of an AI platform’s value. The NVIDIA Vera Rubin platform is designed to enable power… Source

For operators of large-scale AI factories, maximizing continuous output is essential for productivity. In massive-scale AI training, every GPU in the cluster... For operators of large-scale AI factories, maximizing continuous output is essential for productivity. In massive-scale AI training, every GPU in the cluster must synchronize gradients across thousands of collective operations per second. Similarly, during inference, unplanned downtime directly reduces the total volume of requests served, strictly limiting revenue generation. Source

Reduced-space multi-fidelity Bayesian optimization combines GSA and Gaussian processes for expensive industrial simulations.

Quantum-Harbor benchmark evaluates AI agent autonomy and reliability in quantum engineering tasks.

MéTRON-FR: 125M GPT-2 model pretrained on French scores 85.97% on QFrBLiMP; reveals tokenizer sensitivity in cross-lingual evaluation.

CareMirror: HCI system for dementia caregiver wellbeing tracking via longitudinal reflection and clinician-facing interfaces.

LP-BTS: learning-guided planning architecture for large dynamic action spaces applied to mobile sensor charging optimization.

Occlusion-robust multi-object tracking framework combining YOLOv11n, Kalman filtering, and appearance models for defense/surveillance.

Plain Language Adaptation for medical texts using GPT-4o-mini, Gemini-1.5-pro, LLaMA via zero/few-shot learning to improve accessibility.

SWE-bench leaderboard audit: top coding agents converge (0.935 nesting), rendering score ordering unreliable; proposes new metrics.

FlashVector: AI agent optimizing multi-layer model serving stacks (GPU kernels, frameworks, servers) for recommender system cost reduction.

CoTS: post-hoc temperature scaling calibration for test-time prompt tuning preserving accuracy while reducing confidence gap.

OPEN-1B: fully auditable 1B-parameter LLM training run with hardware-agnostic reproducibility verification against undisclosed data/backdoors.

LLMs form linear belief-state geometry for HMM hidden states during in-context learning; probed across 6 models on 40 HMMs.

Google highlights Gemma model applications for societal impact across expert and community use cases.

Google outlines Gemma's role in accelerating scientific research and improving real-world outcomes.

Google extends Gemma multilingual capabilities beyond text translation to preserve nuance across world languages.

ECHO: Chinese full-duplex dialogue benchmark with matched-contrast pairs testing context-sensitive floor-yield vs. backchannel decisions.

Hybrid variational quantum circuits with classical post-processing for multivariate regression; theoretical approximation guarantees on synthetic benchmarks.

OpenAI confirms weeks of AI safety talks with Anthropic and Google DeepMind, as Trump's team dismisses safety concerns and pushes to keep pace with China.

Empirical comparison of context windows, fine-tuning, and latent representations for knowledge adaptation in LLMs across five benchmarks.

Non-contrastive representation learning for detecting semantically equivalent code clones; addresses Type-IV clone detection challenges.