Study Examines How Inference-Time Expert Deactivation Affects Routing and Weights in Sparse Models

A new arXiv paper analyzes expert deactivation at inference time, showing its performance impact depends on how changes to token routing and expert weights are implemented.

Read the full story at arXiv cs.CL ↗

Timeline · 1 report

2026-11-05 05:00 · arXiv cs.CL
Matching the Training-Time Weight Rule in Inference-Time Expert Deactivation

More stories

OpenAI says autonomous pipeline found and patched 948 Linux privilege escalation flaws — OpenAI News
OpenAI launches Images 2.6 with animated GIF creation and faster image generation — OpenAI News
OpenAI to acquire developer tooling startup Pierre Computer Company — OpenAI News
Google turns off AI features by default for school accounts after backlash — Tom's Guide AI
Anthropic reports progress on provable inference for large language models — Anthropic News
Developers Expand Claude Code Beyond Interactive Use With CI, Cloud, and Multi-Agent Setups — Runway Research
Researchers say LLMs can be good at natively reading and writing neural network weights; GPT-6 Astra edited Qwen3.5-0.8B's weights to make it better at math (Rikiya Takehi) — Techmeme
Nothing teases screen-free AI-native device launching in December — The Verge AI

Get the daily brief of stories like this at 6:30 every morning →