AINewsnow

Random logit bias rivals weight perturbation for model gains

Researchers reproduce gains attributed to Neural Thickets, but show they do not require dense expert neighborhoods in weight space. Perturbing one decoder block recovers much improvement, and adding a random bias to frozen-model logits matches or exceeds full-weight perturbation on most reasoning t…

Read the full story at DEV Community — Machine Learning ↗

Timeline · 1 report

  1. 2026-10-31 07:02 · DEV Community — Machine Learning
    Random logit bias rivals weight perturbation for model gains

More stories

  1. Mistral releases weights for trillion-parameter 'Le Chonk' model — Mistral AI News
  2. Anthropic files for blockbuster IPO as Blackstone flags likely junk rating — Fast Company AI
  3. Malicious Claude Code extension on Cursor marketplace exposed up to 5,000 developers — Microsoft Research Blog
  4. Sam Altman says it's 'quite terrible' for personal PCs to hold RAM amid shortage — Tom's Guide AI
  5. S&P Revises Meta's Credit Outlook to Negative on AI Spending Concerns — Bloomberg AI
  6. OpenAI cuts funding target by $10 billion to $20 billion at $1.25 trillion valuation — Financial Times AI
  7. Google unveils Gemini 4 Carbon, a cybersecurity-focused frontier model — Google DeepMind Blog
  8. OpenAI, Google, Microsoft and Anthropic to honor shared robots.txt opt-out for AI training — OpenAI News

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