Recursive self‑improvement requires AI systems to enhance the models and infrastructure used by later generations. LLM compression offers a concrete efficiency goal: retain a pretrained model’s behavior within a smaller storage budget. We investigate whether coding agents can perform this step autonomously.
We formulate compression as a terminal task. An agent receives a source checkpoint, a storage limit, an inference environment, and a time budget, then produces compressed weights and executable inference code. The agent controls its own algorithm and experiments; perplexity feedback guides the search, and an independent verification step checks the final model.
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. Recorded experiments reveal:
- Agents lower perplexity in 21 of 24 revisions while keeping storage nearly constant.
- Full‑model feedback reverses 17 of 79 prior local rankings.
- Generation and inference tests expose failures that likelihood alone misses.
The results identify representation design, experimental judgment, and capability validation as central to agent‑driven model compression, a measurable component of efficiency‑oriented self‑improvement.
Review: The paper presents a systematic evaluation framework, demonstrates the promise and limits of coding agents in LLM compression, and points to key research directions for future self‑improving systems.