Research map / Large language models / Agents, code and math
Regression Language Models for Code
Yash Akhauri, Xingyou Song, Arissa Wongpanich, Bryan Lewandowski, Mohamed S. Abdelfattah
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use decoder numeric head instead of normalized MSE regression head (ablation) | RLM · NASBench-101 | Spearman rho: 0.717 -> 0.800 | Improved |
| normalize y-values for MSE regression head (ablation) | Regression Head · NAS spaces | Spearman rho: 0.478 -> 0.717 | Improved |
| use larger T5Gemma encoder (300M to 600M) (ablation) | RLM · CodeNet | Spearman rho: 0.744 -> 0.782 | Improved |
| use learned 8K ONNX tokenizer instead of T5 32K (ablation) | RLM · NASBench-101 | Spearman rho: 0.533 -> 0.723 | Improved |
| increase encoder context from 1K to 4K (ablation) | RLM · NASBench-101 | Spearman rho: 0.819 -> 0.838 | Improved |
| use explicit digit tokenization instead of T5Gemma tokenizer (ablation) | RLM · CodeNet (random decoder init) | Spearman rho: 0.654 -> 0.747 | Improved |
| initialize from language-pretrained T5Gemma encoder (ablation) | RLM · Triton Kernel | Spearman rho: 0.504 -> 0.532 | Improved |
| pretrain on synthetic FLOPS before accuracy training (ablation) | RLM · NASBench-101 | Spearman rho: 0.83 -> 0.85 | Improved |
| add NAS graph data to code training (ablation) | RLM · coding benchmarks avg | Spearman rho: 0.701 -> 0.706 | No clear effect |
| fine-tune on 1024 language-specific samples | RLM · CodeNet C++ | Spearman rho: 0.730 -> 0.595 | Hurt |
| fine-tune on 1K target NAS samples | RLM · NASBench-201 | Spearman rho: 0.681 -> 0.738 | Improved |
| fine-tune on 1K target NAS samples | RLM · Amoeba | Spearman rho: 0.045 -> 0.501 | Improved |
Closest papers by meaning
- Unsupervised Evaluation of Code LLMs with Round-Trip Correctness (ICML 2024 · similarity 0.86)
- EffiCoder: Enhancing Code Generation in Large Language Models through Efficiency-Aware Fine-tuning (ICML 2025 · similarity 0.85)
- CODE REPRESENTATION LEARNING AT SCALE (ICLR 2024 · similarity 0.85)
- DyCodeEval: Dynamic Benchmarking of Reasoning Capabilities in Code Large Language Models Under Data Contamination (ICML 2025 · similarity 0.85)
- rStar-Coder: Scaling Competitive Code Reasoning with a Large-Scale Verified Dataset (NeurIPS 2025 · similarity 0.84)
- LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code (ICLR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
