Researchers extended the autoresearch loop by inserting a literature‑search stage before code edits. Targeting the open‑source llama.cpp project, they deployed four cloud VMs and let the agent scan arXiv papers and competing forks such as ik_llama.cpp. Within three hours the system produced five optimizations, including a fused AVX2 FMA loop in real‑world scenarios that accelerated flash‑attention generation by 15% on x86 and 5% on ARM.

Pure‑code agents often chase SIMD tweaks that yield marginal gains because they lack the broader performance picture. Early experiments on llama.cpp’s CPU path tried prefetching and loop unrolling, delivering sub‑1% improvements before regressing. Adding a research phase revealed that the workload is memory‑bandwidth bound, a fact only visible from papers and fork analyses, saving countless fruitless trials for production.

The entire run cost roughly $29—$20 for CPU VMs and $9 for API calls—and required no GPUs. By fusing three passes over the flash‑attention QK tile and merging gate‑up weights, the agent cut inference latency while preserving correctness across 974 unit tests. This experiment proves that feeding agents scholarly context can unlock optimizations that pure code introspection misses.