CodeForge-MA: Execution-Verified Multi-Agent Learning with Language-Conditioned LoRA for Multilingual Code Generation

Zhizhou Gu; Xianting Wu; Siyu Gu; Tian Zhang; Kejian Tong

cs.AIOct 4, 2026

CodeForge-MA: Execution-Verified Multi-Agent Learning with Language-Conditioned LoRA for Multilingual Code Generation

Authors: Zhizhou Gu, Xianting Wu, Siyu Gu, Tian Zhang, Kejian Tong

Organizations: Northeastern University Bellevue, USA · Columbia University New York, USA · University of Pennsylvania Philadelphia, USA · University of Southern California San Jose, USA · Independent Researcher Mukilteo, USA

Abstract

Large language models for code generation often fail on execution, multilingual coverage, and contamination control, especially under frozen backbone constraints. We present CodeForge-MA, a unified framework that improves code synthesis through a multi-agent data forge, execution verified reinforced instruction tuning, and a language conditioned mixture of LoRA adapters. Four specialized agents, Composer, Reviewer, Executor, and Curator, iteratively refine instruction code pairs, validate them with tests, and filter duplicates and benchmark leakage. During training, we combine masked supervised fine tuning with a test driven reinforcement objective to align generations with executable correctness. For the larger model, we use sparse expert routing over low rank adapters to improve cross language transfer while keeping the base model unchanged at inference. Experiments show that joint data, objective, and adapter design yields robust gains across programming languages.

Figures & tables

Fig. 1: Overview of the CodeForge-MA framework. Four sequential stages (Multi-Agent Data Forge, EV-RIT Training, LC-MoLoRA Adaptation, Inference) operate above a frozen Qwen-1.8B/72B backbone, faithfully respecting the competition’s no-modification rule.

Fig. 2: The Multi-Agent Data Forge. Four specialized agents $\{A_{\mathrm{Cmp}},A_{\mathrm{Rev}},A_{\mathrm{Exe}},A_{\mathrm{Cur}}\}$ cooperate in a closed Reflexion-style loop; a sample is accepted only when $\pi_{k}=1$ , $q_{k}\geq\tau_{q}$ , and $k\leq K_{\max}$ .

Fig. 3: Execution-Verified Reinforced Instruction Tuning (EV-RIT). The upper lane implements masked cross-entropy on the assistant span; the lower lane implements a PPO surrogate whose reward is the unit-test pass ratio. The two streams are combined by the warm-up coefficient $\lambda_{t}$ .

Symbol	Value	Description
$\tau_{q}$	$0.75$	Forge acceptance threshold
$K_{\max}$	$3$	Max forge revisions per sample
$M$	$5$	Reviewer self-consistency samples
$\alpha$	$0.3$	Language sampling temperature
$\tau_{J}$	$0.8$	MinHash deduplication threshold
$\beta_{\mathrm{KL}}$	$0.02$	KL penalty in PPO surrogate

TABLE I: Key hyperparameters of CodeForge-MA.

Fig. 4: Chain-of-Repair prompting schema. The model emits an ordered triple $(\Phi_{\mathrm{loc}},\Phi_{\mathrm{rat}},\Phi_{\mathrm{fix}})$ ; only $\Phi_{\mathrm{fix}}$ is forwarded to the hidden unit tests at evaluation time, while $\mathcal{L}_{\mathrm{CoR}}$ ties the three fields together during training.

1.8B-class backbones
Model / Configuration	pass@1 / Score	CodeBLEU	CR	BMPS
Qwen-1.8B (base)	13.6	22.1	48.3	10.9
CodeGen-2B-mono	16.6	25.8	55.7	13.1
StarCoderBase-3B	21.0	29.6	62.4	17.4
Qwen-1.8B + Evol-Instruct	21.9	30.2	63.8	17.9
DeepSeek-Coder-1.3B-Instruct	27.1	34.7	71.2	22.8

TABLE II: Consolidated results across four metrics. Top: 1.8B-class backbones, pass@1 averaged over HumanEvalSynthesize / HumanEvalFixTests / MBPP. Middle: 70B-class backbones, where the first column is

\mathrm{Score}_{\mathrm{final}}

from ( 32 ). Bottom: ablation of CodeForge-MA on Qwen-1.8B. All numbers in %.

Explore similar work

SolidCoder: Bridging the Mental-Reality Gap in LLM Code Generation through Concrete Execution
Apr 20, 2026Woojin Lee, Jin-Xia HuangLlm-Driven Code Synthesis Code Optimization
Electronics and Telecommunications Research Institute, Republic of Korea
Selective Left-Shift: Turning Test-Time Compute and Difficulty-based Curation into Training Data for Low-Resource Code Generation
Jul 8, 2026Didula Samaraweera, Anjana Supun, Srinath PereraCode Generation Small Large Language Models
WSO2 Santa Clara, CA, USA
Agnostics: Learning to Code in Any Programming Language via Reinforcement with a Universal Learning Environment
Aug 6, 2025Aleksander Boruch-Gruszecki, Yangtian Zi, Zixuan Wu +4Raw Judge Outputs Low-Resource Languages
1Northeastern University · University of Texas · 3Wellesley College

Apr 20, 2026, cs.SEJ/K move · Enter open · S save

SolidCoder: Bridging the Mental-Reality Gap in LLM Code Generation through Concrete Execution

Woojin Lee, Jin-Xia Huang

Electronics and Telecommunications Research Institute, Republic of Korea

Llm-Driven Code Synthesis Code Optimization Self-Verification

Open paper AskPDF arXiv

State-of-the-art code generation frameworks rely on mental simulation, where LLMs internally trace execution to verify correctness. We expose a fundamental limitation: the Mental-Reality Gap -- where models hallucinate execution traces and confidently validate buggy code. This gap manifests along two orthogonal dimensions: the Specification Gap (overlooking edge cases during planning) and the Verification Gap (hallucinating correct behavior for flawed code). We propose SolidCoder with a simple principle: don't imagine -- execute. The S.O.L.I.D. architecture addresses both dimensions by forcing edge-case awareness before algorithm design and replacing imagined traces with sandboxed execution using property-based oracles. With GPT-4o, SolidCoder achieves state-of-the-art pass@1 performance: 95.7% on HumanEval (+0.6%p), 77.0% on CodeContests (+4.3%p), and 26.7% on APPS (+3.4%p). Ablation reveals that edge-case awareness provides the largest individual gain, while execution grounding catches categorically different errors that specification improvements cannot address. These gains generalize to RL post-trained models, validating that bridging both gap dimensions is essential for robust code synthesis. We release our code and framework to facilitate future research.

Abstract

Explore similar work

SolidCoder: Bridging the Mental-Reality Gap in LLM Code Generation through Concrete Execution

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

Agnostics: Learning to Code in Any Programming Language via Reinforcement with a Universal Learning Environment