Identifying Latent Declarative Representations of Code for Assisting Repository Migration

Shraddha Surana; Ashwin Srinivasan; Michael Bain

cs.SEAug 22, 2026

Identifying Latent Declarative Representations of Code for Assisting Repository Migration

Authors: Shraddha Surana, Ashwin Srinivasan, Michael Bain

Organizations: Department of Computer Science and Information Systems, BITS Pilani, K. K. Birla Goa Campus, NH 17B, Zuarinagar, 403726, Goa, India · School of Computer Science and Engineering, UNSW Sydney, Sydney, NSW, 2052, Australia

Abstract

Legacy software repositories embed decades of domain knowledge in undocumented code, making understanding and modernization difficult. We treat a program as the implementation of an unobserved, declarative description of its computation and investigate whether making this latent declarative representation explicit improves repository-scale porting. ADFD-Migrate approximates the latent representation with an annotated data-flow diagram (ADFD) of processes, data stores, external entities, flows, and behavioral contracts. An LLM infers the source ADFD from bounded repository context, guided by static-analysis coverage checks. Dependency-aware chunking orders bounded process groups for target-language generation. Differences between the source ADFD and a statically recovered target ADFD then guide regeneration. We evaluate ADFD-Migrate on f2x50, a new benchmark of 50~Fortran repositories spanning 1.5k--1.6M lines of code and three complexity tiers, and assess the resulting ports along two dimensions: porting soundness, measured by source-oracle behavioral agreement, and porting completeness, measured by a composite migration outcome index. Against 382 curated Fortran-oracle probes, the generated Python passes 327 (85.6%), with 40 repositories passing every attempted probe. ADFD-Migrate exposes all 382 planned behaviors as runnable targets, compared with 99 and 98 for direct and repository-context translation and 69 and 30 for the static-profile and dependency-chunking ablations. It also achieves a 93.1% mean migration outcome index and a 17--59 percentage-point outcome-index advantage over direct translation on 47 repositories. These results suggest that an inspectable semantic bottleneck can improve the coverage and integration of repository-scale migration while enabling lower-cost generation for many repositories.

Figures & tables

Figure 1: ADFD-mediated porting workflow. The

\lambda

and gear symbols denote LLM and deterministic agents. The first diagnostic loop refines a source ADFD; after optional human ratification, the second refines target code. Experiments automate both loops; human exchanges indicate where experts can optionally inspect and correct the intermediate ADFD and target code.

Property	f2py	fable	LFortran	Ours
Target language	Py (bind)	C++	Py (AST)	Python
Readable output	✗	Partial	✗	✓
Idiomatic target	✗	✗	✗	✓
Preserves structure	N/A	✗	✗	✓
Inspectable IR	✗	✗	LLVM IR	ADFD
Human checkpoint	✗	✗	✗	✓

Table 1: Comparison of Fortran migration approaches.

Tier	$n$	LoC Range	Files	Examples
Low	20	1.5k–41k	4–112	ABAQUS, fftpack, tsunami
Medium	20	4.8k–192k	9–405	minpack, stdlib, xtb
Complex	10	29k–1.6M	158–3.6k	hdf5, lapack, pymc2

Table 2: f2x50 : 50 Fortran repository benchmark.

Setting	Value
Generator (low)	Qwen-3-Coder 30B (local)
Generator (medium)	Qwen-3-Coder 30B (4), 480B (16)
Generator (complex)	Claude Sonnet 4.5 (API)
Judge model	Claude Opus 4.6 (evaluation only)
Temperature	0.3 (generation), 0.2 (judge)
ADFD iterations $k$	3–5

Table 3: Experimental settings.

Method	Planned	Runnable	Pass	Failures		Behavioral agreement
Method	Planned	Runnable	Pass	Wrong output	No runnable target	Over all 382	Over runnable only
ADFD-Migrate	382	382	327	55	0	85.6%	85.6%
Direct translation	382	99	83	16	283	21.7%	83.8%
Repo-context direct	382	98	83	15	284	21.7%	84.7%
Static profile (no ADFD)	382	69	42	27	313	11.0%	60.9%
Dependency chunking (no ADFD)	382	30	14	16	352	3.7%	46.7%

Table 4: Cross-method behavioral comparison on the same 382 planned Fortran-vs-Python probes. Runnable : probes for which the generated Python target could be located, imported, invoked, and compared with the Fortran oracle. A probe fails either because the generated code returned a wrong output or because no runnable target existed for it at all ( missing functionality : missing generated target, non-generated repository, import/syntax defect, or incompatible interface). The last two columns give the same pass counts as a rate over all 382 planned probes, which is the reported agreement rate, and over only the probes that ran.

Tier	$n$	Sem.	Syn.	Imp.	ADFD outcome	Direct $n$	Direct outcome	$\Delta$
Low	20	95.2%	99.9%	77.2%	93.0%	20	76.1%	+16.9
Medium	20	98.3%	98.1%	72.5%	91.9%	20	65.3%	+26.6
Complex	10	99.0%	100.0%	79.8%	95.7%	7	37.2%	+58.5
All	50	97.2%	99.2%	75.9%	93.1%

Table 5: Cross-tier migration outcome results. Sem. = ADFD semantic-alignment score; Syn. = syntax validity; Imp. = import success. Direct comparison uses 20 low-, 20 medium-, and 7 complex-tier repositories.

\Delta

is the ADFD-minus-direct difference in percentage points.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A.1 : Full ADFD-Migrate system architecture. Static-analysis coverage refines the inferred source ADFD; dependency-aware chunks guide generation; source–target ADFD differences guide regeneration. The LLM judge supplies supporting representation-level evidence. Behavioral agreement is the primary evaluation.

Figure B.1 : Worked MINPACK ADFD excerpt using Gane–Sarson notation. Rounded rectangles are processes, the open-ended rectangle is a data store, square-cornered rectangles are external entities, and labeled arrows are flows. P1–P5 identify corresponding computational roles in the source and generated-Python ADFDs.

Repository	LoC	Files	$\#P$	Chunks	Syntax	Import	Semantic	Beh.	Outcome	Time	Model
fortran2018-ex.	3.1k	75	19	7	100%	100%	100%	4/4	100.0%	25 min	30B
roots-fortran	5.2k	4	1	1	100%	100%	95%	0/7	97.0%	13 min	30B
forpy	13.8k	9	3	1	100%	100%	95%	4/4	97.0%	2 min	30B
tsunami	3.1k	33	31	11	100%	100%	95%	7/7	97.0%	214 min	30B
IO-Fortran-Lib.	40.9k	34	6	2	100%	70%	100%	4/4	96.4%	21 min	30B
functional-fortran	6.6k	27	10	4	100%	67%	100%	7/7	96.0%	13 min	30B

Appendix

Table E.1 : Repository-level results for the low-complexity tier.

Repository	LoC	Files	$\#P$	Chunks	Syntax	Import	Semantic	Beh.	Outcome	Time	Model
stdlib	35k	405	16	6	100%	100%	100%	11/11	98.9%	61 min	480B
FOODIE	19k	44	14	5	100%	97%	100%	4/4	96.8%	31 min	480B
arpack-ng	148k	334	16	6	100%	69%	100%	7/7	96.3%	71 min	480B
atomsk	86k	166	12	4	100%	72%	100%	4/4	96.1%	47 min	480B
WPS	32k	123	16	6	100%	73%	100%	4/4	96.1%	53 min	480B
packmol	13k	40	132	42	99%	74%	100%	8/8	96.0%	177 min	480B

Appendix

Table E.2 : Repository-level results for the medium-complexity tier.

Repository	LoC	Files	$\#P$	Chunks	Syntax	Import	Semantic	Beh.	Outcome	Time	Model
lapack	1.56M	3,587	52	18	100%	90%	100%	16/16	98.3%	105 min	S4.5
pymc2	132k	410	22	7	100%	80%	100%	10/10	96.7%	98 min	S4.5
ccpp-physics	325k	240	45	15	100%	78%	99.9%	7/11	96.6%	135 min	S4.5
Nek5000	281k	327	33	12	100%	90%	98%	13/13	96.5%	232 min	S4.5
petsc	29k	200	25	9	100%	80%	100%	8/8	96.2%	115 min	S4.5
hdf5	94k	158	85	29	100%	77%	100%	10/10	95.7%	107 min	S4.5

Appendix

Table E.3 : Repository-level results for the complex tier.

Code Translation

Code translation converts programs between languages, frameworks, and representations while aiming to preserve behavior and sometimes improve safety or performance.

See all 42 papers

CodeMeepo: A Multi-Agent Workflow for Repository-Level Code Migration and Optimization
Oct 25, 2026Farzad Habibi, Changrong Wu, Jinzhou Jiang +1Code Translation Multi-Agent Coding
University of California, Irvine · Microsoft · OpenAI
Entropy-based Code Adversarial Translation for Real-world Repository Migration
Aug 10, 2026Yushun Tang, Yisen Cao, Zhicheng Chen +4Code Translation Software Engineering Agents
Huawei Technologies Co., Ltd. · The Chinese University of Hong Kong
SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration?
Aug 24, 2026Deyao Hong, Yizhe Chi, Wenyi Li +7Automated Software Engineering AI Coding Agents
Navers Lab, Einsia.AI · Tsinghua University

Oct 25, 2026J/K move · Enter open · S save

CodeMeepo: A Multi-Agent Workflow for Repository-Level Code Migration and Optimization

Farzad Habibi, Changrong Wu, Jinzhou Jiang, Guohan Lu

University of California, Irvine · Microsoft · OpenAI

Code Translation Multi-Agent Coding LLM Agent Workflow Optimization

Open paper AskPDF arXiv

Repository-level code migration rewrites an entire project in a new programming language while preserving its observable behavior. It is challenging because behavior spans interdependent modules, dependencies and build infrastructure must be adapted to the target ecosystem, and the source code is often the only specification. Large language model agents can now attempt these migrations end to end, but three challenges remain. First, an agent that writes both the translation and its acceptance tests may overlook the same requirements in both, leaving missing or incorrect behavior undetected. Second, large repositories can exceed a model's context window, and dividing the work into smaller tasks does not ensure that all required behavior is migrated. Third, even behaviorally correct translations may be slower than the original implementation. We present CodeMeepo, a multi-agent pipeline for complete and efficient migration of large repositories. Each milestone is evaluated against a fixed external oracle that the agents neither author nor modify. A separate parity assessment compares source and target components in scope and adds identified gaps back into the plan as required work. Ordered milestones limit the scope of each task, while cumulative regression checks guard against breaking already-migrated behavior. A benchmark-guided optimization phase addresses performance, followed by full-suite conformance testing and repair. Eight specialized agent roles communicate through file artifacts, and a deterministic state machine governs repair, deferral, scope changes, and termination. On nine subjects in three language pairs, CodeMeepo achieves the best pass rate on every subject and passes 18% more tests than a state-of-the-art multi-agent baseline using the same model. On a production daemon, its optimization phase improves half of the benchmarks, by 25% on average.

Abstract

Code Translation

Explore similar work

CodeMeepo: A Multi-Agent Workflow for Repository-Level Code Migration and Optimization

Entropy-based Code Adversarial Translation for Real-world Repository Migration

SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration?