Pith. sign in

REVIEW 8 cited by

CodeGen2: Lessons for Training LLMs on Programming and Natural Languages

Not yet reviewed by Pith; the record is open.

Re-run · record.json Download PDF Read on arXiv ↗

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2305.02309 v2 pith:TTCQHR3N submitted 2023-05-03 cs.LG

Erik Nijkamp , Hiroaki Hayashi , Caiming Xiong , Silvio Savarese , Yingbo Zhou This is my paper
classification cs.LG
keywords modeltraininglearningllmsdatamodelsattemptcodegen2
verification ladder T0 review T1 audit T2 compute T3 formal
0 comments
read the original abstract

Large language models (LLMs) have demonstrated remarkable abilities in representation learning for program synthesis and understanding tasks. The quality of the learned representations appears to be dictated by the neural scaling laws as a function of the number of model parameters and observations, while imposing upper bounds on the model performance by the amount of available data and compute, which is costly. In this study, we attempt to render the training of LLMs for program synthesis more efficient by unifying four key components: (1) model architectures, (2) learning methods, (3) infill sampling, and, (4) data distributions. Specifically, for the model architecture, we attempt to unify encoder and decoder-based models into a single prefix-LM. For learning methods, (i) causal language modeling, (ii) span corruption, (iii) infilling are unified into a simple learning algorithm. For infill sampling, we explore the claim of a "free lunch" hypothesis. For data distributions, the effect of a mixture distribution and multi-epoch training of programming and natural languages on model performance is explored. We conduct a comprehensive series of empirical experiments on 1B LLMs, for which failures and successes of this exploration are distilled into five lessons. We will provide a final recipe for training and release CodeGen2 models in size 1B, 3.7B, 7B, and, 16B parameters, along with the training framework as open-source: https://github.com/salesforce/CodeGen.

Discussion (0). Sign in to comment.

Forward citations

Cited by 8 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. AutoVeriFix: Automatically Correcting Errors and Enhancing Functional Correctness in LLM-Generated Verilog Code

    cs.AR 2025-09 conditional novelty 6.0 of 10

    AutoVeriFix improves LLM-generated Verilog functional correctness by generating a high-level Python reference model, deriving a high-coverage testbench, and iteratively fixing Verilog simulation mismatches.

  2. Dr. Boot: Bootstrapping Program Synthesis Language Models to Perform Repairing

    cs.SE 2025-07 conditional novelty 6.0 of 10

    A bootstrapping algorithm that trains program synthesis models on their own successful outputs and on repaired failures modestly improves pass@k over regular fine-tuning on MBPP, with mixed results on APPS.

  3. Failure by Interference: Language Models Make Balanced Parentheses Errors When Faulty Mechanisms Overshadow Sound Ones

    cs.CL 2025-06 conditional novelty 6.0 of 10

    Language models fail at balanced parentheses because unreliable internal components that promote wrong tokens can outvote reliable ones, and amplifying reliable components fixes the errors.

  4. EFIM: Efficient Serving of LLMs for Infilling Tasks with Improved KV Cache Reuse

    cs.CL 2025-05 conditional novelty 6.0 of 10

    EFIM rearranges the fill-in-the-middle prompt so prefix and suffix stay fixed across user edits, enabling more KV cache reuse, while a fragment tokenization training method fixes the resulting partial-word generation problem.

  5. GRACE: Graph-Guided Repository-Aware Code Completion through Hierarchical Code Fusion

    cs.SE 2025-09 conditional novelty 5.0 of 10

    GRACE combines a multi-level code graph, hybrid text-structure retrieval, and graph fusion to improve repository-level code completion over vanilla and graph-based RAG baselines.

  6. Semantic Source Code Segmentation using Small and Large Language Models

    cs.SE 2025-07 conditional novelty 5.0 of 10

    Fine-tuned encoder-only models such as CodeBERT outperform zero-shot and few-shot LLMs at semantic line-level segmentation of R code, and a new annotated R dataset, StatCodeSeg, is introduced.

  7. LaMAGIC2: Advanced Circuit Formulations for Language Model-Based Analog Topology Generation

    cs.LG 2025-06 conditional novelty 5.0 of 10

    LaMAGIC2's SFCI representation cuts output token length from quadratic to linear and improves precision-sensitive generation of analog power converter topologies.

  8. Boosting Open-Source LLMs for Program Repair via Reasoning Transfer and LLM-Guided Reinforcement Learning

    cs.SE 2025-06 conditional novelty 4.0 of 10

    Repairity lifts Qwen2.5-Coder-32B from 38.2% to 62.7% on SWE-bench Verified by distilling Claude's reasoning traces and then fine-tuning with a reward model trained on Claude preferences.

Pith tools