Scholarlink

Large language models can draft plausible answers to textbook-style science questions, yet online platforms still lack a fast way to score and compare those answers on non-GPU hardware. We tackle this gap by framing solution verification as a lightweight binary classification task built from problem-solution pairs in STEM disciplines derived from the SCP-116K corpus. Using this corpus we design and evaluate two CPU-friendly baselines: (i) a sparse Gradient Boosted Trees classifier trained on TF–IDF features, and (ii) a single-layer Text CNN operating on Byte-Pair-Encoded subwords. In our setup, both models deliver F1 \(~\approx \)  0.78 and can score approximately 1,000 pairs/s on a single CPU, suggesting that near-real-time scoring is feasible for this baseline task without GPUs. We detail a reproducible methodology and provide efficiency metrics to guide practitioners and propose future dataset refinements to better align the task with real-world scenarios.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CPU-Efficient Verification of Science Problem–Solution Pairs: Design Rationale and Baselines

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nicolas Douard,
  • Ahmed Samet,
  • George Giakos,
  • Denis Cavallucci

摘要

Large language models can draft plausible answers to textbook-style science questions, yet online platforms still lack a fast way to score and compare those answers on non-GPU hardware. We tackle this gap by framing solution verification as a lightweight binary classification task built from problem-solution pairs in STEM disciplines derived from the SCP-116K corpus. Using this corpus we design and evaluate two CPU-friendly baselines: (i) a sparse Gradient Boosted Trees classifier trained on TF–IDF features, and (ii) a single-layer Text CNN operating on Byte-Pair-Encoded subwords. In our setup, both models deliver F1 \(~\approx \)  0.78 and can score approximately 1,000 pairs/s on a single CPU, suggesting that near-real-time scoring is feasible for this baseline task without GPUs. We detail a reproducible methodology and provide efficiency metrics to guide practitioners and propose future dataset refinements to better align the task with real-world scenarios.