SabaPivot's picture
Publish canonical reproduction with fresh CPU audit
af62971 verified
|
Raw
History Blame Contribute Delete
1.9 kB

Claim 4: Optimization scale


Claim. Agents perform better at function-level optimizations than repository-level ones, with OpenHands + Claude 4.0 Sonnet showing strongest module-level performance that degrades at function-level, while OpenHands + GPT-5 shows the inverse pattern.

Verdict: falsified as a blanket statement. The specific OpenHands model patterns are reproduced, but the released stratified aggregate contradicts "agents perform better at function-level optimizations." In f2_stratified.json from https://github.com/formula-code/formula-code.github.io/tree/a79001bfdf46cd6c6e11e742f8b98893294cbc7a, mean function-level Advantage is -0.032143, while mean module-level Advantage is -0.003323. Module-level is therefore less negative overall.

Slice Recomputed Advantage
Mean function-level -0.032143
Mean class-level -0.028500
Mean module-level -0.003323
OpenHands + Claude function -0.003835
OpenHands + Claude module 0.295726
OpenHands + GPT-5 function 0.048876
OpenHands + GPT-5 module -0.009541

So the nuanced model interaction is credible: OpenHands + Claude is strongest at module scale and OpenHands + GPT-5 is better at function than module. The broader statement is not supported by the released aggregate. The recomputation is in outputs/formulacode_audit/summary.json and was repeated by HF Job https://huggingface.co/jobs/Srishti280992/6a6d1d066b79c09949c1dc23.

Related artifacts: dataset https://huggingface.co/datasets/formulacode/formulacode-all/tree/897e48cab8a27d32ba20ddb970b1fc397d96ee95, eval repo https://github.com/formula-code/fc-eval/tree/c08f665e7bf3b4de225b72dc02ce9b15b7aaba2b, paper https://arxiv.org/abs/2603.16011.