Parallax boosts LLM efficiency AgingBench shows AI decline

Parallax adds a scalable, parameterized Local Linear Attention mechanism that replaces the softmax constant estimate with a learned linear probe, improving the bias‑variance trade‑off in large language models. The new attention kernel matches or exceeds FlashAttention performance across batch sizes and context lengths, yielding consistent perplexity reductions at 0.6B‑1.7B scales and downstream gains.

Researchers introduce AgingBench, a longitudinal benchmark that measures how AI coding agents deteriorate after deployment. The suite isolates four aging mechanisms, compression, interference, revision, and maintenance, and shows that agents can retain surface performance while factual accuracy drops, requiring stage‑specific repairs rather than stronger initial models.

Agent Judge is a new evaluation harness that tackles the shortcomings of simple LLM judges for long‑horizon, stateful agents. By adding searchable trajectories, verification of tool actions against ground‑truth systems, and adaptive rubrics, it lets teams automatically assess complex workflows without manual review.

The open‑weights Pixal3D model from Tencent ARC, originally built for CUDA GPUs, has been fully ported to Apple Silicon. The blog details the conversion process and links to a GitHub repo with a working inference pipeline, letting Mac users generate textured 3D meshes locally.