SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.03411 · 2026

SWE-Master: Unleashing the Potential of Software Engineering Agents via Post-Training

Yang Song, Ji-Rong Wen, Jinhao Jiang, Wayne Zhao, Daixuan Cheng, Tao Zhang, Guoxin Chen, Shuang Sun, Zongchao Chen, Lisheng Huang, Ran Le, Yiwen Hu, and 2 more

arXiv · PDF · Open in the Atlas

Code that ran

9 of 10 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 2 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 7 failures of our instrument; 1 no recorded run), measured 2026-11-21; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2602.03411.json · badge: /api/ran/2602.03411.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
RUCAIBox/SWE-Master canonical (classifier label)
classifier basis: census
9 ran · 1 no recorded run
FunctionStatusWhere it lives
collapse_subdicts Ran
ran; our contract draft was wrong, not the code
RUCAIBox/SWE-Master/R2E-Gym/SWE-smith/swesmith/bug_gen/combine/same_module.py
code served (permissive licence) · get_code("412ae192e1acff1f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

# Call torch.manual_seed for reproducibility if needed
torch.manual_seed(0)

# Construct test arguments for collapse_subdicts
# The function takes: current (dict), depth (int), prefix (str, default='')

# Test case 1: nested dict with depth=2
current_test = {
    "foo": {
        "bar": {
            "baz": 1
        }
    }
}

depth_test = 2
prefix_test = ''

ARGS = {
    'current': current_test,
    'depth': depth_test,
    'prefix': prefix_test
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convert_to_path Ran
ran; our contract draft was wrong, not the code
RUCAIBox/SWE-Master/R2E-Gym/SWE-smith/swesmith/bug_gen/combine/same_module.py
code served (permissive licence) · get_code("afb84aa1334897a8")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re

def convert_to_path(folder: str):
    DUNDER_PATTERN = '____[a-zA-z\\d]+__\\.py$'
    path = '/'.join(folder.split('__')[2:])
    if re.search(DUNDER_PATTERN, folder):
        path = path.replace('//', '/__').replace('/.py', '__.py')
    return path

ARGS = {
    "folder": "__root__module__path__to__file____abc123__.py"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
generate_hash Ran
ran; our contract draft was wrong, not the code
RUCAIBox/SWE-Master/R2E-Gym/SWE-smith/swesmith/constants.py
code served (permissive licence) · get_code("48cd515a4d64b121")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import hashlib
import random
import string

def generate_hash(s):
    rng = random.Random(int(hashlib.sha256(s.encode()).hexdigest(), 16))
    return ''.join((rng.choice(string.ascii_lowercase + string.digits) for _ in range(8)))

ARGS = {
    's': 'hello_world'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_combos Ran
ran; our contract draft was wrong, not the code
RUCAIBox/SWE-Master/R2E-Gym/SWE-smith/swesmith/bug_gen/utils.py
code served (permissive licence) · get_code("90549129b147cef1")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 2, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

from itertools import combinations

def get_combos(items, r, max_combos) -> list[tuple]:
    """Get `max_combos` combinations of items of length r or greater."""
    all_combos = []
    for new_combo in combinations(items, r):
        all_combos.append(new_combo)
        if max_combos != -1 and len(all_combos) >= max_combos:
            break
    return sorted(all_combos, key=len)


ARGS = {
    "items": ["a", "b", "c", "d"],
    "r": 2,
    "max_combos": 5
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_patches_from_folder Ran
ran, honoured the contract we drafted
RUCAIBox/SWE-Master/R2E-Gym/SWE-smith/swesmith/bug_gen/combine/same_module.py
code served (permissive licence) · get_code("6601304bdf38974f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned list, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import os
import tempfile

# Create a temporary directory structure with some .diff files
with tempfile.TemporaryDirectory() as tmpdir:
    # Create some patch files
    patch1_path = os.path.join(tmpdir, "patch1.diff")
    patch2_path = os.path.join(tmpdir, "patch2.diff")
    patch3_path = os.path.join(tmpdir, "other.txt")
    
    with open(patch1_path, 'w') as f:
        f.write("dummy patch content")
    with open(patch2_path, 'w') as f:
        f.write("dummy patch content")
    with open(patch3_path, 'w') as f:
        f.write("not a patch")
    
    # Create a subdirectory with another patch
    subdir = os.path.join(tmpdir, "subdir")
    os.makedirs(subdir, exist_ok=True)
    patch3_path = os.path.join(subdir, "patch3.diff")
    with open(patch3_path, 'w') as f:
        f.write("dummy patch content")
    
    ARGS = {
        "folder_path": tmpdir,
        "include_patches": []
    }
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
main Ran
ran, honoured the contract we drafted
RUCAIBox/SWE-Master/R2E-Gym/SWE-smith/swesmith/bug_gen/get_cost.py
code served (permissive licence) · get_code("241ab2afb5b092fe")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import json
import os
import tempfile
import shutil

# Create a temporary directory structure with test metadata files
temp_dir = tempfile.mkdtemp()

# Create subdirectories
subdir1 = os.path.join(temp_dir, 'subdir1')
subdir2 = os.path.join(temp_dir, 'subdir2')
os.makedirs(subdir1, exist_ok=True)
os.makedirs(subdir2, exist_ok=True)

# Create metadata files with different bug types
metadata_all_1 = {
    'cost': 10.5,
    'bug_id': 'bug001'
}
with open(os.path.join(subdir1, 'metadata__all_1.json'), 'w') as f:
    json.dump(metadata_all_1, f)

metadata_all_2 = {
    'cost': 20.3,
    'bug_id': 'bug002'
}
with open(os.path.join(subdir2, 'metadata__all_2.json'), 'w') as f:
    json.dump(metadata_all_2, f)

# Create type-specific metadata files
metadata_syntax = {
    'cost': 5.2,
    'bug_id': 'bug_syntax_001'
}
with open(os.path.join(subdir1, 'metadata__syntax_1.json'), 'w') as f:
    json.dump(metadata_syntax, f)

metadata_logic = {
    'cost': 15.7,
    'bug_id': 'bug_logic_001'
}
with open(os.path.join(subdir2, 'metadata__logic_1.json'), 'w') as f:
    json.dump(metadata_logic, f)

# Create a non-matching file (should be ignored)
with open(os.path.join(subdir1, 'other_file.json'), 'w') as f:
    json.dump({'cost': 100}, f)

ARGS = {
    'repo_path': temp_dir,
    'bug_type': 'all'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
preprocess_data Ran
ran; our contract draft was wrong, not the code
RUCAIBox/SWE-Master/OpenRLHF_SFT/scripts_swe_master/sft_data_pre_tokenize.py
code served (permissive licence) · get_code("a48d5458f99a6b87")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create a mock chat template function
def mock_chat_template(messages, tokenize=False, add_generation_prompt=False):
    """Mock function that simulates a chat template formatter"""
    if isinstance(messages, list):
        result = "".join(str(m) for m in messages)
    else:
        result = str(messages)
    
    if add_generation_prompt:
        result += " [GENERATION_PROMPT]"
    
    return result

# Test case 1: with apply_chat_template
data_with_chat = {
    'input': ['message1', 'message2', 'message3']
}

# Test case 2: with input_template and output_key
data_with_template = {
    'input': 'user question',
    'output': 'model answer'
}

# Test case 3: with only input_key
data_simple = {
    'input': 'some input text'
}

# Build ARGS for a basic execution
ARGS = {
    'data': data_with_template,
    'input_template': 'Question: {}',
    'input_key': 'input',
    'output_key': 'output',
    'apply_chat_template': None,
    'multiturn': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
preprocess_data Ran
ran; our contract draft was wrong, not the code
RUCAIBox/SWE-Master/OpenRLHF_SFT/scripts_swe_master/sft_data_pre_tokenize_toolcall.py
code served (permissive licence) · get_code("31b049c55ece2f82")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Stand-in class for apply_chat_template callable
class ChatTemplateCallable:
    def __call__(self, messages, tools=None, tokenize=False, add_generation_prompt=False):
        # Simulates a chat template that processes messages
        if isinstance(messages, list):
            result = "Processed: " + " ".join(str(m) for m in messages)
        else:
            result = "Processed: " + str(messages)
        if add_generation_prompt:
            result += " [GENERATION_PROMPT]"
        return result

# Stand-in class for input_template callable
class InputTemplate:
    def format(self, text):
        return f"Template: {text}"

# Create sample data dict
data = {
    'input': ["What is machine learning?"],
    'output': "Machine learning is a subset of artificial intelligence."
}

# Test case 1: with apply_chat_template and output_key
apply_chat_template_fn = ChatTemplateCallable()

# Test case 2: with input_template
input_template_obj = InputTemplate()

ARGS = {
    'data': data,
    'input_template': input_template_obj,
    'input_key': 'input',
    'output_key': 'output',
    'apply_chat_template': None,
    'multiturn': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
preprocess_data Ran
ran; our contract draft was wrong, not the code
RUCAIBox/SWE-Master/OpenRLHF_SFT/scripts_swe_master/sft_dataset.py
code served (permissive licence) · get_code("3e9101bb92d47eb7")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Stand-in class for apply_chat_template function
class ChatTemplate:
    def __call__(self, messages, tokenize=False, add_generation_prompt=False):
        # Simulates a chat template that processes messages
        if isinstance(messages, list):
            if add_generation_prompt:
                return "prompt_text_" + str(len(messages))
            else:
                return "full_text_" + str(len(messages))
        return "text"

# Build sample data
sample_data = {
    'input': [
        {'role': 'user', 'content': 'Hello'},
        {'role': 'assistant', 'content': 'Hi there'}
    ]
}

# Create a mock apply_chat_template callable
apply_chat_template_func = ChatTemplate()

ARGS = {
    'data': sample_data,
    'input_template': None,
    'input_key': 'input',
    'output_key': None,
    'apply_chat_template': apply_chat_template_func,
    'multiturn': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
parse_oai_response Not yet run
unverified (harvested, no recorded run)
RUCAIBox/SWE-Master/DeepSWE_RL/rllm/rllm/agents/swe_agent.py
code served (permissive licence) · get_code("c7e743994eca38db")
How it ran

no invocation recorded.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 10 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-21.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

In this technical report, we present SWE-Master, an open-source and fully reproducible post-training framework for building effective software engineering agents. SWE-Master systematically explores the complete agent development pipeline, including teacher-trajectory synthesis and data curation, long-horizon SFT, RL with real execution feedback, and inference framework design. Starting from an open-source base model with limited initial SWE capability, SWE-Master demonstrates how systematical optimization method can elicit strong long-horizon SWE task solving abilities. We evaluate SWE-Master on SWE-bench Verified, a standard benchmark for realistic software engineering tasks. Under identical experimental settings, our approach achieves a resolve rate of 61.4% with Qwen2.5-Coder-32B, substantially outperforming existing open-source baselines. By further incorporating test-time scaling (TTS) with LLM-based environment feedback, SWE-Master reaches 70.8% at TTS@8, demonstrating a strong performance potential. SWE-Master provides a practical and transparent foundation for advancing reproducible research on software engineering agents. The code is available at https://github.com/RUCAIBox/SWE-Master.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2602.03411")
get_code_for_paper("2602.03411")
have("2602.03411")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2602.03411.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2602.03411.svg)](https://syntology.ai/paper/2602.03411)

Connect an agent — have() is free.