Coding agents shrink LLMs on their own

🔥 What happened Researchers from Rice and Carnegie Mellon let six coding agents autonomously compress LLMs, given only a storage cap, time budget, and perplexity feedback. 29 of 36 submissions passed independent verification. 💡 Why it matters The best agent design per task hits just 3.4% higher perplexity than BF16 and 77.33% downstream accuracy versus 79.56% for BF16. Agents cut PPL in 21 of 24 revisions at nearly fixed storage — real, measurable gains. ⚡ Our take Recursive self-improvement just got a concrete benchmark. Anyone assuming model compression stays a human craft should note that full-model feedback flipped 17 of 79 prior local rankings.