Engineering
HyQuant: Hybrid-Precision Quantization for LLM Attention
HyQuant is a hybrid-precision quantization framework that addresses accuracy loss in low-bit quantization of LLM attention by selectively keeping high…
Engineering
Auto-RecSys: Harnessing Autonomous Research Agents for Industry-Scale Recommender Systems
The paper presents Auto-RecSys, an autonomous research-agent system for running large-scale recommender-system experimentation, using distributed asynchronous execution to run multiple…
Engineering
SearchAtlas: Analyzing Agentic Search Strategies via Evidential Query Graphs
SearchAtlas is a framework that converts raw search trajectories produced by LLM agents into structured evidence graphs showing how evidence…
Engineering
RAG-Safety-Bench: Reliable Evaluation of Retrieval-Augmented LLM Safety
The paper introduces RAG-Safety-Bench, a benchmark isolating how retrieval-augmented generation affects LLM safety by testing four controlled conditions: no-RAG baseline…
Engineering
LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation
The paper introduces LOCUS, a post-training method that curbs verbosity in large language models by applying task-aware low-rank adaptations that…
Engineering
Structured Transforms for Low-Overhead Quantization of Language Models
Researchers present an enhanced quantization method for large language models that improves on Kashin-decomposition-based approaches by replacing dense random orthogonal…
Engineering
vLLM x AgentX: Optimizing for Real-World Agentic Serving
The post describes a three-layer stack vLLM uses to optimize serving for agentic workloads: a data plane with hybrid KV…
Engineering
Tiered KV Cache Offloading in vLLM
vLLM’s tiered KV cache offloading system preserves evicted key-value cache data across host memory, storage, and remote peers instead of…
Engineering
An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics
The paper details training Nemotron 3 Ultra with supervised fine-tuning and reinforcement learning to produce natural-language mathematical proofs, built around…
Engineering
A Fragility Spectrum for Recursive Language-Model Training
The paper studies how 13 different language model checkpoints degrade under recursive training on model-generated text, using a shared corpus…
Engineering
Helion x Hugging Face Kernels: Building and Shipping Out-of-the-box Performant Kernels
This PyTorch blog post details how the Hugging Face Kernels project added support for Helion, a tile-based DSL from Meta…
Engineering
TurboQuant Quantization Explained
TurboQuant is a vector-embedding compression technique that applies a random orthogonal rotation to embedding coordinates, then maps the rotated values…