Prior Knowledge or Search? A Study of LLM Agents in Hardware-Aware Code Optimization
Dmitry RedkoApplied AI Institute [email protected], Albert FazlyevAI Talent Hub, ITMO University [email protected], Konstantin SozykinApplied AI Institute [email protected] Maria IvanovaYSDA, Applied AI Institute [email protected] Evgeny BurnaevApplied AI Institute [email protected] Egor ShvetsovApplied AI Institute [email protected] · 2026
Read Prior Knowledge or Search? A Study of LLM Agents in Hardware-Aware Code Optimization by Dmitry RedkoApplied AI Institute [email protected], Albert FazlyevAI Talent Hub, ITMO University [email protected], Konstantin SozykinApplied AI Institute [email protected] Maria IvanovaYSDA, Applied AI Institute [email protected] Evgeny BurnaevApplied AI Institute [email protected] Egor ShvetsovApplied AI Institute [email protected] (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-19782.txt.
Description
Redko, Fazlyev, Sozykin, Ivanova, Burnaev and Shvetsov test whether language-model agents optimise code from feedback or merely lean on prior knowledge. In black-box optimisation the models behave greedily; in zero-shot kernel generation, input-size hints change nothing and unusual sizes degrade performance in any language. With iterative feedback, CUDA kernels improve steadily while TVM intermediate code worsens, showing optimisation collapses in low-density languages. The verdict: priors, not feedback or agent scaffolding, drive most gains.
Catalog record
| CREATOR | Dmitry RedkoApplied AI Institute [email protected], Albert FazlyevAI Talent Hub, ITMO University [email protected], Konstantin SozykinApplied AI Institute [email protected] Maria IvanovaYSDA, Applied AI Institute [email protected] Evgeny BurnaevApplied AI Institute [email protected] Egor ShvetsovApplied AI Institute [email protected] |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 66848 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.19782