CompressionBench: How can your coding agents compress LLMs?
arXiv cs.AI · 5 Nov 2026, 04:14 AEST
The study explores how coding agents can autonomously compress large language models (LLMs) while retaining their performance. Results show that 29 out of 36 submissions passed verification, with agents achieving notable improvements in perplexity and downstream accuracy.
Key facts
- Recursive self-improvement requires AI systems to improve the models and infrastructure used by subsequent systems.
- We study whether coding agents can carry out this step autonomously.
- Across six agents and six model--size settings, 29 of 36 submissions pass verification.
- Selecting the lowest-perplexity agent design per task yields an average 3.4% perplexity increase over BF16 and 77.33% downstream accuracy, compared with 79.56% for BF16.
- The results identify representation design, experimental judgment, and capability validation as central to agent-driven model compression.
Read aloud
The study explores how coding agents can autonomously compress large language models L L M s while retaining their performance. Results show that twenty nine out of thirty six submissions passed verification, with agents achieving notable improvements in perplexity and downstream accuracy. Recursive self-improvement requires A I systems to improve the models and infrastructure used by subsequent systems. We study whether coding agents can carry out this step autonomously. Across six agents and six model--size settings, twenty nine of thirty six submissions pass verification. Selecting the lowest-perplexity agent design per task yields an average three point four percent perplexity increase over BF16 and seventy seven point three three percent downstream accuracy, compared with seventy nine point five six percent for BF16. The results identify representation design, experimental judgment, and capability validation as central to agent-driven model compression.
Sources in this cluster
- CompressionBench: How can your coding agents compress LLMs?
arXiv cs.AI · 5 Nov 2026, 04:14 AEST