The Archive

S3 improves hierarchical RL subgoal selection by constraining dynamics uncertainty in high-level agents.

Preregistered replication study on label agreement and monotonicity in NLI datasets; methodological validation.

Investigates cost-quality tradeoffs in RL post-training for neural machine translation with reasoning verification.

AdaFlash improves speculative decoding via adaptive diffusion drafters with on-policy distillation for LLM inference acceleration.

RLAES framework uses RL with rubric-based rewards to jointly optimize essay scoring and feedback generation, introducing RFE for measurable feedback evaluation.

Report on drone computing infrastructure vision addressing software-hardware capability gaps for large-scale logistics, disaster response, and infrastructure inspection.

Methods for assessing student team performance in tabletop crisis-response exercises using recorded actions and communication data.

Treasury Secretary Scott Bessent said the U.S. could sanction Chinese open AI models over alleged IP theft, expanding the Trump administration's campaign to slow China's AI advances.

MIRA-Ev: multilingual clinical NLP benchmark with span-level evidence detection and argumentation graphs on Spanish MIR exam cases.

Offline RL method using adaptive regularization and conservative query selection for preference-based policy improvement without environment interaction.

ATLAS: diffusion-based sampler for generating Boltzmann-distributed amorphous material structures, addressing rare-event sampling in molecular systems.

Large deviations theory applied to derive free energy functionals for dense associative memory systems with polynomial interactions.

ABot-World-0: action-conditioned video world model enabling long-horizon closed-loop interaction on single GPU, trained on games/simulation/web video.

Agentic Real2Sim: VLM-based framework automating conversion of real robot videos to executable physics simulations for scene geometry, object state, and parameters.

Google DeepMind releases Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber models for inference.

Comparison of reasoning-capable LLMs, SFT, and RL for legal machine translation, showing structured reasoning improves precision on domain-specific terminology.

Automated extraction of 3.2M data points from 76K energy studies using NLP/ML to improve meta-analysis transparency.

Neural Kolmogorov Equations reformulate neural SDEs for parallelizable stochastic dynamics learning under general noise.

Physics-informed neural networks with boundary enforcement for elliptic PDEs and mean escape time computation.

Agentic AI shifts more of the critical execution path onto the CPU. Agents operate in sandboxes to execute code, invoke tools, retrieve context, interact with... Agentic AI shifts more of the critical execution path onto the CPU. Agents operate in sandboxes to execute code, invoke tools, retrieve context, interact with databases, and analyze results before returning information to the model. As these loops run concurrently across an AI factory, CPU performance increasingly shapes both per-agent responsiveness and overall factory throughput. Source

Google is launching Gemini 3.6 Flash alongside a new security model dedicated to quickly finding and patching security vulnerabilities. In a blog post on Tuesday, Google describes Gemini 3.5 Flash Cyber as a "cost-efficient and highly capable alternative" to larger, more expensive AI systems, such as the one offered by Anthropic's Mythos. The cybersecurity model is built upon Gemini 3.5 Flash and will be available first to governments and trusted partners via CodeMender, Google's security-focused coding agent. As noted by Google, CodeMender can call upon 3.5 Flash Cyber "multiple times at hig...

What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale.... What began as discrete AI model training and human-facing chat interfaces has evolved into always-on AI factories dedicated to producing intelligence at scale. These factories are now tasked with powering agentic workflows that reason, plan, use tools, verify intermediate results, and execute complex multistep tasks across vast contexts. Agentic workloads are not defined by a single prompt… Source

Frontier model pre-training has converged on mixture of experts (MoE), which is fundamentally changing what limits large-scale AI training. As compute per token... Frontier model pre-training has converged on mixture of experts (MoE), which is fundamentally changing what limits large-scale AI training. As compute per token falls, communication increasingly determines how efficiently models scale across thousands of GPUs. NVIDIA GB300 NVL72 set a world record for pre-training DeepSeek-V3 671B at 1,648 TFLOPs per GPU, showing how advances across the entire AI… Source

Multi-generator adversarial learning addresses class imbalance and non-homogeneous failure modes in predictive maintenance.

Generative state-space model learns ocean dynamics from sparse, noisy observations without complete reanalysis data.

MIRAGE residual U-Net infers breast MRI contrast enhancement with lesion-aware supervision combining reconstruction and perceptual losses.

Nativ wraps MLX in a macOS app for local inference, offering chat UI and localhost API similar to LM Studio.

Vision-language models as unified backbone for attributed graphs with heterogeneous modalities (text, visual, mixed).

Parallel Noising algorithm strengthens Neural Markov Logic Networks for generative modeling on larger relational structures.

Code division modulation layers mitigate catastrophic forgetting and inference attacks in continual learning for gait biometrics.