STRIDE formalizes training data attribution as a sparse recovery problem in activation space, achieving an order of magnitude faster results than gradient-based methods.
Streaming-based multi-agent reasoning reduces latency through pipelining while simultaneously improving accuracy because early, more reliable reasoning steps protect against erroneous later steps.
KVarN reduces error accumulation when quantizing KV-caches to 2-bit precision through improved token-scale normalization and achieves state-of-the-art results on MATH500, AIME24, and HumanEval.
Gemma 4 12B runs on standard laptops with 16 GB RAM and enables local API endpoints via the LiteRT-LM CLI for agent-driven workflows without cloud dependency.
NVIDIA automates workflows in Physical AI research through new Agent Skills that make scene reconstruction, data generation, and policy training for autonomous vehicles, robotics, and Vision AI scalable.
Context Engineering is the discipline of systematically and at runtime filling the context window of language models with the right information in optimal form—far more comprehensive than prompt engineering.
A CPU-based RL controller optimizes adaptive sampling during test-time scaling, reducing computational overhead and latency compared to heuristic methods.
VaSE achieves higher accuracy than existing sparse-attention methods at 4x KV-cache compression, thereby reducing the memory bottleneck of reasoning models.
Successful domain specialization of LLMs requires careful tuning of learning rate, data-mixing ratios, and checkpoint selection to avoid catastrophic forgetting.
PaW trains environment models during policy training using the same RL rollouts, consistently improving agent performance without requiring additional simulators or inference costs.
