Engineering
vLLM Reaches 25K Total TPS/GPU on Qwen3.5
vLLM reached 25,000 tokens per second per GPU serving Qwen3.5 on Blackwell GPUs via a disaggregated prefill-decode architecture. Three optimizations…
Engineering
LEMUR and Mean Centering for Late-Interaction Retrieval in txtai
txtai adds LEMUR and mean-centering techniques for late-interaction retrieval. LEMUR learns a fixed-dimensional encoding of multi-vector representations so late-interaction models…
Engineering
AX-Ray: Finding Causal-Leakage Defects in Two General-Purpose Public Models
VIDRAFT’s AX-Ray diagnostic framework evaluates AI models for deployment safety beyond standard capability benchmarks, using 117 structured diagnostic items across…
Engineering
Stealing Reasoning Traces from Proprietary LLM APIs
Researchers discovered a security vulnerability in proprietary LLM APIs that allowed encrypted chain-of-thought reasoning traces to be extracted and read.…
Engineering
Don’t classify. Hallucinate!
Simon Willison describes Doug Turnbull’s technique for tagging blog content without constraining a model to a fixed vocabulary. Rather than…
Opinions
EFF: The Youth AI Privacy Act Creates a Privacy Paradox for Minors
The Electronic Frontier Foundation argues that the Youth AI Privacy Act’s age-verification and “safe design” mandates would require platforms to…
Engineering
LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers
Researchers from UIUC, the University of Maryland, Nanyang Technological University, Purdue, and the University of Illinois Chicago introduce LLMRouter, a…
Engineering
DarwinX: Evolving Agent Harnesses Through Natural Selection
DarwinX proposes evolving the harness around a frozen LLM — its prompts, tools, memory, and control flow — using a…
Engineering
AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
AutoDesign targets systems that turn multimodal content into structured design outputs, arguing that most such pipelines stay static rather than…
Engineering
Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence
This paper proposes the Spatial Memory Agent (SMA), a framework that improves spatial reasoning in frozen vision-language models without any…
Engineering
Massive Activations in Hybrid Linear Attention Large Language Models
This research examines internal activation patterns in hybrid linear-attention language models and identifies two distinct phenomena: sharp pre-attention spikes that…
Engineering
Full-Bandwidth Transformer
The paper identifies a bottleneck in autoregressive transformers where only sampled output tokens are fed back into the model, discarding…