Effect of task-specific reinforcement learning on performance-model construction

Determine whether task-specific reinforcement learning that improves mapping-reasoning accuracy also improves the construction of complete analytical performance models, including reliable predictions of memory traffic and buffer capacity.

Background

The paper evaluates LLMs on two related tasks: answering qualitative questions about mappings and generating executable analytical performance-model code. Verifier-guided reinforcement learning improves the Qwen3-4B model’s mapping-reasoning accuracy from 54.3% to 70.0%, but the reported experiment does not establish whether this improvement transfers to quantitative model synthesis.

Reliable construction requires more than recognizing which mapping is preferable: the generated program must correctly compute both off-chip memory accesses and on-chip buffer capacity for unseen workload, architecture, and mapping specifications. Establishing whether reinforcement learning can bridge this reasoning-to-construction gap is therefore left unresolved.

References

Accuracy rises from $54.3\%$ to $70.0\%$, showing that additional domain training can improve the reasoning prerequisite, although its effect on full model construction remains open.

— PerfReasoning: How Well Do LLMs Reason on Hardware Performance?  (2609.04476 - Zhao et al., 3 Sep 2026) in Section 3, paragraph “Task-specific RL improves mapping reasoning”