Selective Left-Shift: Turning Test-Time Compute and Difficulty-based Curation into Training Data for Low-Resource Code Generation

• URL: http://arxiv.org/abs/2607.07748v1
• Date: Wed, 08 Jul 2026 09:18:56 GMT
• Title: Selective Left-Shift: Turning Test-Time Compute and Difficulty-based Curation into Training Data for Low-Resource Code Generation
• Abstract summary: Large Language Models achieve strong code generation for high resource languages like Python and Java but suffer sharp performance drops on Low-Resource Programming Languages(LRPLs) such as Julia.<n>We propose a three-phase pipeline that resolves this trilemma by decoupling syntax acquisition from algorithmic reasoning.<n>Applying to Qwen3-8B, our pipeline improves pass@1 by up to +7.6 points on MultiPL-E and +14.2 points on the Agnostics LiveCodeBench for Julia compared to SOTA results.
• Score: 0.8602553195689513
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Large Language Models achieve strong code generation for high resource languages like Python and Java but suffer sharp performance drops on Low-Resource Programming Languages~(LRPLs) such as Julia. Improving Small Language Models~(SLMs) for these languages faces a trilemma: Supervised Fine-Tuning~(SFT) is bottlenecked by data scarcity, inference-time scaling is too expensive for deployment, and Reinforcement Learning from scratch yields near zero advantages. We propose a three-phase pipeline that resolves this trilemma by decoupling syntax acquisition from algorithmic reasoning. First, we \emph{left-shift} inference-time compute to an offline data synthesis engine that uses iterative compiler and test feedback to generate verified training examples. Second, we fine-tune an SLM on this synthetic, verified data to embed strong syntactic priors. Third, we apply Reinforcement Learning with Verifiable Reward~(RLVR) grounded by language-agnostic Input/Output tests, where the SFT prior constrains exploration away from syntax errors. Applied to Qwen3-8B, our pipeline improves pass@1 by up to +7.6 points on MultiPL-E and +14.2 points on the Agnostics LiveCodeBench for Julia compared to SOTA results. Furthermore, the pipeline only used $\frac{1}{3}$ data and $\frac{1}{6}$ cost over the previous state-of-the-art. We further demonstrate that the pipeline generalizes to Ballerina achieving 49.7\% MultiPL-E Pass@1, a language with near-zero pretraining representation. Ablations confirm that both the SFT phase and execution-grounded rewards are necessary for stable training.

Related papers

• Translate-R1: Cost-Aware Translation Tool Use via Reinforcement Learning [25.323183334678195]
  We learn a policy that decides when to translate from reward alone, developing language- and domain-adaptive introspection.<n>We analyze how tool use emerges over training, per language and per domain.
  arXiv  Detail & Related papers  (2026-06-05T02:21:41Z)
• Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective [1.786053901581251]
  GRPO yields consistent improvements on all 13 languages, up to $+$5.03 chrF++ for Traditional Chinese, and, without any target-language data, competes with 3-epoch supervised fine-tuning on morphologically complex languages.<n>We identify a consistent empirical pattern in which gains are largest where baseline performance is weakest and reward discriminability is highest.
  arXiv  Detail & Related papers  (2026-05-15T14:11:23Z)
• Scaling Latent Reasoning via Looped Language Models [109.64154923739842]
  We present and open-source Ouro, a family of pre-trained Looped Language Models (LoopLM)<n>Ouro builds reasoning into the pre-training phase through (i) iterative computation in latent space, (ii) an entropy-regularized objective for learned depth allocation, and (iii) scaling to 7.7T tokens.
  arXiv  Detail & Related papers  (2025-10-29T17:45:42Z)
• Agnostics: Learning to Code in Any Programming Language via Reinforcement with a Universal Learning Environment [44.47221244109926]
  Agnostics is a language-agnostic post-training pipeline that eliminates per-language engineering.<n>Key idea is to judge code solely by its externally observable behavior, so a single verifier can test solutions written in any language.<n>We will release the language-agnostic training datasets (Ag-MBPP-X, Ag-Codeforces-X, Ag-LiveCodeBench-X), training code, and ready-to-use configurations, making RL post-training in any programming language as simple as editing a short YAML file.
  arXiv  Detail & Related papers  (2025-08-06T20:30:55Z)
• $\textbf{PLUM}$: Improving Code LMs with Execution-Guided On-Policy Preference Learning Driven By Synthetic Test Cases [28.791570350483816]
  We propose PLUM, an on-policy $textbfP$reference $textbfL$earning framework A$textbfu$gmented with test cases for code L$textbfM$ s. PLUM levitates the need to train reward models, allowing for large scale on-policy and online preference data collation. We show PLUM's benefits are consistent across various widely-used code LMs even they have been well-trained with SFT.
  arXiv  Detail & Related papers  (2024-06-11T02:07:18Z)
• LINC: A Neurosymbolic Approach for Logical Reasoning by Combining Language Models with First-Order Logic Provers [60.009969929857704]
  Logical reasoning is an important task for artificial intelligence with potential impacts on science, mathematics, and society. In this work, we reformulating such tasks as modular neurosymbolic programming, which we call LINC. We observe significant performance gains on FOLIO and a balanced subset of ProofWriter for three different models in nearly all experimental conditions we evaluate.
  arXiv  Detail & Related papers  (2023-10-23T17:58:40Z)
• GECTurk: Grammatical Error Correction and Detection Dataset for Turkish [1.804922416527064]
  Grammatical Error Detection and Correction (GEC) tools have proven useful for native speakers and second language learners. Synthetic data generation is a common practice to overcome the scarcity of such data. We present a flexible and synthetic data generation pipeline for Turkish covering more than 20 expert-curated grammar and spelling rules.
  arXiv  Detail & Related papers  (2023-09-20T14:25:44Z)
• LeTI: Learning to Generate from Textual Interactions [60.425769582343506]
  We explore LMs' potential to learn from textual interactions (LETI) that not only check their correctness with binary labels but also pinpoint and explain errors in their outputs through textual feedback. Our focus is the code generation task, where the model produces code based on natural language instructions. LETI iteratively fine-tunes the model, using the objective LM, on a concatenation of natural language instructions, LM-generated programs, and textual feedback.
  arXiv  Detail & Related papers  (2023-05-17T15:53:31Z)
• Improving Code Generation by Training with Natural Language Feedback [69.52985513422381]
  We formalize an algorithm for learning from natural language feedback at training time instead, which we call learning from Language Feedback (ILF) ILF requires only a small amount of human-written feedback during training and does not require the same feedback at test time, making it both user-friendly and sample-efficient. We use ILF to improve a Codegen-Mono 6.1B model's pass@1 rate by 38% relative (and 10% absolute) on the Mostly Basic Python Problems (MBPP) benchmark.
  arXiv  Detail & Related papers  (2023-03-28T16:15:31Z)
• Rejuvenating Low-Frequency Words: Making the Most of Parallel Data in Non-Autoregressive Translation [98.11249019844281]
  Knowledge distillation (KD) is commonly used to construct synthetic data for training non-autoregressive translation (NAT) models. We propose reverse KD to rejuvenate more alignments for low-frequency target words. Results demonstrate that the proposed approach can significantly and universally improve translation quality.
  arXiv  Detail & Related papers  (2021-06-02T02:41:40Z)
• Emergent Communication Pretraining for Few-Shot Machine Translation [66.48990742411033]
  We pretrain neural networks via emergent communication from referential games. Our key assumption is that grounding communication on images---as a crude approximation of real-world environments---inductively biases the model towards learning natural languages.
  arXiv  Detail & Related papers  (2020-11-02T10:57:53Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.