AI researcher team autonomously discovers linear attention beating GDN2 at 1.3B
UCSantaBarbara · hf · 2026-10-22
UCSB, Penn State and LinkedIn built a team of persistent AI researchers with distinct disciplinary backgrounds, private research histories, and adaptive experiment allocation to balance exploration/exploitation under limited budgets. Deployed to search linear attention architectures, the team discovered models outperforming GDN2 at 1.3B scale on language modeling, commonsense reasoning, and retrieval tasks. Ablations show removing any organizational component markedly weakens discovery: no backgrounds → early plateau; no private histories → less progress despite shared evidence; uniform selection loses to adaptive allocation.
More from coding & agent
- AI-agent slideshow farm pulls 10M+ monthly TikTok views with zero creators or ad spend — PrajwalTomar_ · 2026-10-22
- Agent Standup turns your Claude Code sessions into a live video call — stickygoose · 2026-10-22
- Microsoft's ScholarEvolve lets agents improve by reading papers, lifting Qwen's AppWorld score from 69% to 81.4% — bendee983 · 2026-10-22
- Crypto KOL says project was a clone scam with fake Vitalik ties, recommends Claude Code for vetting — econoar · 2026-10-22
- claude-howto: Visual Claude Code guide hits 42.6k GitHub stars — luongnv89 · 2026-10-22
- Dev builds a browser-based 3D world with Hyperfy, vibe-coding a realistic pool table in 15 minutes — remorej · 2026-10-22