SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2604.13073 · 2026

OmniTrace: A Unified Framework for Generation-Time Attribution in Omni-Modal LLMs

Yang Zhao, Xin Wang, Hang Yin, Shan Jiang, Qianqi Yan, Ching-Chen Kuo, Yichen Guo

arXiv · PDF · Open in the Atlas

Code that ran

18 of 24 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; 5 honoured the contract we drafted, 3 violated it and 0 have no contract verdict; 10 failures of our instrument; 6 no recorded run), measured 2026-11-22; 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/2604.13073.json · badge: /api/ran/2604.13073.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
openai/whisper canonical (classifier label)
classifier basis: census
9 ran · 4 no recorded run
modelscope/FunASR canonical (classifier label)
classifier basis: census
9 ran · 2 no recorded run
FunctionStatusWhere it lives
add_file_root_path Ran
ran; our contract draft was wrong, not the code
modelscope/FunASR/funasr/download/download_model_from_hub.py
code served (permissive licence) · get_code("e736efd2088f5330")
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 os
import tempfile
import shutil

# Create a temporary directory structure for testing
test_dir = tempfile.mkdtemp()
subdir = os.path.join(test_dir, "models")
os.makedirs(subdir, exist_ok=True)

# Create some dummy files
file1_path = os.path.join(subdir, "file1.txt")
file2_path = os.path.join(subdir, "file2.txt")
nested_file_path = os.path.join(subdir, "nested_file.txt")

with open(file1_path, "w") as f:
    f.write("dummy")
with open(file2_path, "w") as f:
    f.write("dummy")
with open(nested_file_path, "w") as f:
    f.write("dummy")

# Construct file_path_metas with various structures
file_path_metas = {
    "simple_file": "file1.txt",
    "nested_config": {
        "inner_file": "file2.txt",
    },
    "file_list": ["file1.txt", "file2.txt"],
}

# Construct cfg as a dict
cfg = {}

ARGS = {
    "model_or_path": subdir,
    "file_path_metas": file_path_metas,
    "cfg": cfg,
}

# Cleanup will happen after function execution
try:
    pass
finally:
    shutil.rmtree(test_dir, ignore_errors=True)
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 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
check_vllm_applicable Ran
ran, violated the contract we drafted
modelscope/FunASR/funasr/auto/auto_model_vllm.py
code served (permissive licence) · get_code("d94882bbd36f686d")
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 bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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

# Test case 1: LLM-based model (should return True)
ARGS = {
    "model_name": "FunASRNano"
}

# Additional test cases can be verified by calling the function with different model names:
# - "LLMASR" -> True
# - "Paraformer" -> raises ValueError
# - "SenseVoice" -> raises ValueError
# - "CTTransformer" -> raises ValueError
# - "Conformer" -> raises ValueError
# - "Qwen3ASR" -> raises ValueError
# - "UnknownModel" -> 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 bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
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
cr_ctc_loss Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
modelscope/FunASR/funasr/losses/cr_ctc.py
code served (permissive licence) · get_code("6c9da623e97fa4fe")
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 Tensor, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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

torch.manual_seed(0)

# Small batch size: 2
# Max sequence length: 5
# Vocabulary size: 4
batch_size = 2
max_len = 5
vocab_size = 4

# Create log probabilities for augmented input (B, T, V)
# Values should be log probabilities, so in range roughly [-inf, 0]
log_probs_aug = torch.randn(batch_size, max_len, vocab_size) - 2.0
log_probs_aug = torch.log_softmax(log_probs_aug, dim=-1)

# Create log probabilities for clean input (B, T, V)
log_probs_clean = torch.randn(batch_size, max_len, vocab_size) - 2.0
log_probs_clean = torch.log_softmax(log_probs_clean, dim=-1)

# Create input lengths (B,) - some sequences may be shorter than max_len
input_lengths = torch.tensor([3, 4], dtype=torch.long)

ARGS = {
    "log_probs_aug": log_probs_aug,
    "log_probs_clean": log_probs_clean,
    "input_lengths": input_lengths,
}
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 Tensor, ndim 0, dtype family float, shape () — shape () as a later run recorded it
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
exact_div Ran
ran, honoured the contract we drafted
openai/whisper/whisper/utils.py
pointer only (licence: MIT); the stored copy get_code serves, from qwenlm/qwen-audio, is pointer only (licence: NOASSERTION) · get_code("6e0808fc4828513a")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from qwenlm/qwen-audio, and GitHub could not classify the licence of that repository (NOASSERTION). The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

get_human_readable_count Ran
ran; our contract draft was wrong, not the code
modelscope/FunASR/funasr/train_utils/model_summary.py
code served (permissive licence) · get_code("e40f7b6dc5c3977a")
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 numpy as np

ARGS = {
    "number": 1234567
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — 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
install_requirements Ran
ran, violated the contract we drafted
modelscope/FunASR/funasr/utils/install_model_requirements.py
code served (permissive licence) · get_code("d6d4ea8b2806c520")
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 bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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 tempfile
import os

# Create a temporary requirements file for testing
with tempfile.NamedTemporaryFile(mode='w', suffix='.txt', delete=False) as f:
    # Write some valid pip package requirements
    f.write('numpy==1.21.0\n')
    f.write('requests==2.26.0\n')
    requirements_path = f.name

ARGS = {
    'requirements_path': requirements_path
}
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 bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
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
median_filter Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
openai/whisper/whisper/timing.py
code served (permissive licence) · get_code("6e70207ad77f0ae6")
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 Tensor, ndim 3, dtype family float, shape (2, 3, 8) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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.functional as F

torch.manual_seed(0)

# Create a small tensor for testing
# The function expects a tensor and applies median filtering along the last dimension
# We'll create a 3D tensor to test the general case
x = torch.randn(2, 3, 8)

# filter_width must be odd and positive
filter_width = 3

ARGS = {
    "x": x,
    "filter_width": filter_width,
}
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 Tensor, ndim 3, dtype family float, shape (2, 3, 8) — shape (2, 3, 8) as a later run recorded it
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
mel_filters Ran
ran, honoured the contract we drafted
openai/whisper/whisper/audio.py
code served (permissive licence) · get_code("12c6c409a1e359c3")
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 Tensor, 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.

import os
import numpy as np
import torch
from unittest.mock import patch, MagicMock

torch.manual_seed(0)

# Create a mock mel_filters.npz file in memory
# The function expects to load mel_80 and mel_128 from an npz file
mel_80_data = np.random.randn(80, 201).astype(np.float32)
mel_128_data = np.random.randn(128, 201).astype(np.float32)

mock_npz_data = {
    'mel_80': mel_80_data,
    'mel_128': mel_128_data
}

# Patch np.load to return our mock data
original_load = np.load

def mock_np_load(path, allow_pickle=False):
    if 'mel_filters.npz' in path:
        class MockNpz:
            def __init__(self, data):
                self.data = data
            
            def __enter__(self):
                return self
            
            def __exit__(self, *args):
                pass
            
            def __getitem__(self, key):
                return self.data[key]
        
        return MockNpz(mock_npz_data)
    return original_load(path, allow_pickle=allow_pickle)

# Apply the patch
import builtins
_original_np_load = np.load
np.load = mock_np_load

ARGS = {
    'device': 'cpu',
    'n_mels': 80
}
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 Tensor, ndim 2, dtype family float — no output shape is stated: observe_signature.py could not observe it (call:NameError)
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
model_summary Ran
ran; our contract draft was wrong, not the code
modelscope/FunASR/funasr/train_utils/model_summary.py
code served (permissive licence) · get_code("26200c7d0299db96")
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 torch
import torch.nn as np

torch.manual_seed(0)

class SimpleModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear1 = torch.nn.Linear(4, 3)
        self.linear2 = torch.nn.Linear(3, 2)
    
    def forward(self, x):
        x = self.linear1(x)
        x = self.linear2(x)
        return x

model = SimpleModel()

ARGS = {
    'model': model
}
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 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
optional_int Ran
ran, honoured the contract we drafted
openai/whisper/whisper/utils.py
code served (permissive licence) · get_code("e56996b20b9c19cb")
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 int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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

def optional_int(string):
    return None if string == 'None' else int(string)

# The function takes a single parameter 'string' which is a string
# It returns None if the string equals 'None', otherwise converts to int

ARGS = {
    'string': '42'
}
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 int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
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
pad_or_trim Ran
ran; our contract draft was wrong, not the code
openai/whisper/whisper/audio.py
code served (permissive licence) · get_code("9f13de1c23e4003d")
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 Tensor, ndim 2, dtype family float, shape (2, 480000) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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
import torch.nn.functional as F

torch.manual_seed(0)

# Define constants as in the function
SAMPLE_RATE = 16000
CHUNK_LENGTH = 30
N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE

# Test case 1: torch tensor that needs padding
array_torch = torch.randn(2, 100)  # Short audio, needs padding

# Test case 2: numpy array that needs trimming
array_numpy = np.random.randn(3, N_SAMPLES + 1000)  # Long audio, needs trimming

# We'll use the torch tensor case for the main ARGS
ARGS = {
    "array": array_torch,
    "length": N_SAMPLES,
    "axis": -1
}
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 Tensor, ndim 2, dtype family float, shape (2, 480000) — shape (2, 480000) as a later run recorded it
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
pip_install_r Ran
ran; our contract draft was wrong, not the code
modelscope/FunASR/funasr/utils/install_model_requirements.py
code served (permissive licence) · get_code("229df00957f1445e")
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 CompletedProcess. 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 tempfile
import os

# Create a temporary requirements file for testing
with tempfile.NamedTemporaryFile(mode='w', suffix='.txt', delete=False) as f:
    f.write("numpy\n")
    f.write("scipy\n")
    requirements_file = f.name

try:
    ARGS = {
        'requirements_path': requirements_file
    }
finally:
    # Clean up will happen after execution
    pass
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 CompletedProcess — a CompletedProcess output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a CompletedProcess 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
prepare_vllm_weights Ran
ran; our contract draft was wrong, not the code
modelscope/FunASR/funasr/auto/auto_model_vllm.py
code served (permissive licence) · get_code("fc2c2ae72cabe336")
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 os
import tempfile
import torch
import json
from omegaconf import OmegaConf

torch.manual_seed(0)

# Create a temporary directory structure for the FunASR model
temp_model_dir = tempfile.mkdtemp()

# Create config.yaml
config_dict = {
    'llm_conf': {
        'init_param_path': 'qwen_model'
    }
}
config_path = os.path.join(temp_model_dir, 'config.yaml')
with open(config_path, 'w') as f:
    OmegaConf.save(config_dict, f)

# Create the llm model subdirectory with config files
llm_model_dir = os.path.join(temp_model_dir, 'qwen_model')
os.makedirs(llm_model_dir, exist_ok=True)

# Add some dummy config files to the llm directory
config_json = os.path.join(llm_model_dir, 'config.json')
with open(config_json, 'w') as f:
    json.dump({'model_type': 'qwen', 'hidden_size': 128}, f)

tokenizer_config = os.path.join(llm_model_dir, 'tokenizer_config.json')
with open(tokenizer_config, 'w') as f:
    json.dump({'type': 'qwen_tokenizer'}, f)

# Create model.pt with LLM weights
model_state_dict = {
    'llm.layers.0.self_attn.q_proj.weight': torch.randn(64, 128),
    'llm.layers.0.self_attn.k_proj.weight': torch.randn(64, 128),
    'llm.layers.0.mlp.up_proj.weight': torch.randn(256, 128),
    'llm.embed_tokens.weight': torch.randn(32000, 128),
    'other_module.weight': torch.randn(10, 10),  # non-llm weights to be filtered
}

model_pt_path = os.path.join(temp_model_dir, 'model.pt')
torch.save({'state_dict': model_state_dict}, model_pt_path)

# Create output directory path
temp_output_dir = tempfile.mkdtemp()

ARGS = {
    'model_dir': temp_model_dir,
    'output_dir': temp_output_dir,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
omegaconf, safetensors, torch — 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
remove_symbols Ran
ran; our contract draft was wrong, not the code
openai/whisper/whisper/normalizers/basic.py
code served (permissive licence) · get_code("2206d84b6900ab3e")
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 unicodedata

def remove_symbols(s: str):
    """
    Replace any other markers, symbols, punctuations with a space, keeping diacritics
    """
    return ''.join((' ' if unicodedata.category(c)[0] in 'MSP' else c for c in unicodedata.normalize('NFKC', s)))

ARGS = {
    's': 'Hello, World! © 2024 — café'
}
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
remove_symbols_and_diacritics Ran
ran; our contract draft was wrong, not the code
openai/whisper/whisper/normalizers/basic.py
code served (permissive licence) · get_code("cef27f679a755971")
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 unicodedata

ADDITIONAL_DIACRITICS = {'œ': 'oe', 'Œ': 'OE', 'ø': 'o', 'Ø': 'O', 'æ': 'ae', 'Æ': 'AE', 'ß': 'ss', 'ẞ': 'SS', 'đ': 'd', 'Đ': 'D', 'ð': 'd', 'Ð': 'D', 'þ': 'th', 'Þ': 'th', 'ł': 'l', 'Ł': 'L'}

def remove_symbols_and_diacritics(s: str, keep=''):
    """
    Replace any other markers, symbols, and punctuations with a space,
    and drop any diacritics (category 'Mn' and some manual mappings)
    """
    return ''.join((c if c in keep else ADDITIONAL_DIACRITICS[c] if c in ADDITIONAL_DIACRITICS else '' if unicodedata.category(c) == 'Mn' else ' ' if unicodedata.category(c)[0] in 'MSP' else c for c in unicodedata.normalize('NFKD', s)))

# Test with various inputs
test_string = "Héllo Wørld! Café œuvre"
keep_chars = ""

ARGS = {
    "s": test_string,
    "keep": keep_chars
}
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
sinusoids Ran
ran, honoured the contract we drafted
openai/whisper/whisper/model.py
code served (permissive licence) · get_code("e529c9641c178fe0")
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 Tensor, ndim 2, dtype family float, shape (8, 4) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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 numpy as np
import torch

torch.manual_seed(0)

ARGS = {
    "length": 8,
    "channels": 4,
    "max_timescale": 10000,
}
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 Tensor, ndim 2, dtype family float, shape (8, 4) — shape (8, 4) as a later run recorded it
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
str2bool Ran
ran, violated the contract we drafted
openai/whisper/whisper/utils.py
code served (permissive licence) · get_code("176b6ba6c2c2940c")
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 bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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
import numpy as np

torch.manual_seed(0)

# The function str2bool takes a single parameter: string
# It expects a string that is either 'True' or 'False'
# and returns the corresponding boolean value

ARGS = {
    'string': 'True'
}
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 bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
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
to_bytes Ran
ran, honoured the contract we drafted
modelscope/FunASR/funasr/train_utils/model_summary.py
code served (permissive licence) · get_code("54138299a3e55e00")
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 int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape 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)

# The function takes a dtype and converts it to bytes
# It does: int(str(dtype)[-2:]) // 8
# This expects a dtype that when converted to string has last 2 chars as digits
# torch.float32 -> "torch.float32" -> "32" -> 32 // 8 = 4
# torch.float64 -> "torch.float64" -> "64" -> 64 // 8 = 8
# torch.float16 -> "torch.float16" -> "16" -> 16 // 8 = 2

dtype = torch.float32

ARGS = {
    "dtype": dtype
}
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 int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
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_or_download_model_dir Not yet run
unverified (harvested, no recorded run)
modelscope/FunASR/funasr/download/download_model_from_hub.py
code served (permissive licence) · get_code("5288c864dbd0538d")
How it ran

no invocation recorded.

get_or_download_model_dir_hf Not yet run
unverified (harvested, no recorded run)
modelscope/FunASR/funasr/download/download_model_from_hub.py
code served (permissive licence) · get_code("1c80497412c69732")
How it ran

no invocation recorded.

load_audio Not yet run
unverified (harvested, no recorded run)
openai/whisper/whisper/audio.py
code served (permissive licence) · get_code("80abc0b281d75575")
How it ran

no invocation recorded.

median_filter_cuda Not yet run
unverified (harvested, no recorded run)
openai/whisper/whisper/triton_ops.py
code served (permissive licence) · get_code("9c0b804792b26762")
How it ran

no invocation recorded.

median_kernel Not yet run
unverified (harvested, no recorded run)
openai/whisper/whisper/triton_ops.py
code served (permissive licence) · get_code("61d0c7cb5ccfdd5b")
How it ran

no invocation recorded.

transcribe Not yet run
unverified (harvested, no recorded run)
openai/whisper/whisper/transcribe.py
code served (permissive licence) · get_code("700202ff899b0c2d")
How it ran

no invocation recorded.

Also present in other papers' code

Of 24 samples lifted from this paper's linked repositories, 24 carry a function identity and 0 do not. Of the 24 distinct functions with an identity, 22 are also present in other papers whose code we hold (22 with at least one paper dated earlier, 0 the same month, 0 later) and 2 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.

FunctionAlso present inEarliest-datedOthers named
sinusoids 9 other papers
9 dated earlier
2212.04356 (2022-12)
Robust Speech Recognition via Large-Scale Weak Supervision
2306.05284 (2023-06)
Simple and Controllable Music Generation
2307.03183 (2023-07)
Whisper-AT: Noise-Robust Automatic Speech Recognizers are Also Strong General Audio Event Taggers
and 6 more not named here
exact_div 6 other papers
6 dated earlier
2212.04356 (2022-12)
Robust Speech Recognition via Large-Scale Weak Supervision
2306.01428 (2023-06)
Improved DeepFake Detection Using Whisper Features
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
and 3 more not named here
pad_or_trim 5 other papers
5 dated earlier
2212.04356 (2022-12)
Robust Speech Recognition via Large-Scale Weak Supervision
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
and 2 more not named here
optional_int 4 other papers
4 dated earlier
2212.04356 (2022-12)
Robust Speech Recognition via Large-Scale Weak Supervision
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
2309.15701 (2023-09)
HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models
and 1 more not named here
str2bool 4 other papers
4 dated earlier
2212.04356 (2022-12)
Robust Speech Recognition via Large-Scale Weak Supervision
2308.06112 (2023-08)
Lip2Vec: Efficient and Robust Visual Speech Recognition via Latent-to-Latent Visual to Audio Representation Mapping
2309.15701 (2023-09)
HyPoradise: An Open Baseline for Generative Speech Recognition with Large Language Models
and 1 more not named here
get_human_readable_count 3 other papers
3 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
2407.05361 (2024-07)
Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation
2505.00598 (2025-05)
Fast and Low-Cost Genomic Foundation Models via Outlier Removal
to_bytes 3 other papers
3 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
2407.05361 (2024-07)
Emilia: An Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech Generation
2505.00598 (2025-05)
Fast and Low-Cost Genomic Foundation Models via Outlier Removal
load_audio 2 other papers
2 dated earlier
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
2510.26096 (2025-10)
ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio-Language Models
mel_filters 2 other papers
2 dated earlier
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
2510.26096 (2025-10)
ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio-Language Models
add_file_root_path 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
check_vllm_applicable 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
cr_ctc_loss 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
get_or_download_model_dir 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
get_or_download_model_dir_hf 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
install_requirements 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
median_filter 1 other paper
1 dated earlier
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
median_filter_cuda 1 other paper
1 dated earlier
2406.10082 (2024-06)
Whisper-Flamingo: Integrating Visual Features into Whisper for Audio-Visual Speech Recognition and Translation
model_summary 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
pip_install_r 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation
prepare_vllm_weights 1 other paper
1 dated earlier
2312.15185 (2023-12)
emotion2vec: Self-Supervised Pre-Training for Speech Emotion Representation

Shown: the 20 functions present in the most other papers. 2 more shared functions are not listed here.

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 2026-11-22.

Repositories linked to this paper

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

Abstract

Modern multimodal large language models (MLLMs) generate fluent responses from interleaved text, image, audio, and video inputs. However, identifying which input sources support each generated statement remains an open challenge. Existing attribution methods are primarily designed for classification settings, fixed prediction targets, or single-modality architectures, and do not naturally extend to autoregressive, decoderonly models performing open-ended multimodal generation. We introduce OmniTrace, a lightweight and model-agnostic framework that formalizes attribution as a generation-time tracing problem over the causal decoding process. OmniTrace provides a unified protocol that converts arbitrary token-level signals such as attention weights or gradient-based scores into coherent span-level, cross-modal explanations during decoding. It traces each generated token to multimodal inputs, aggregates signals into semantically meaningful spans, and selects concise supporting sources through confidence-weighted and temporally coherent aggregation, without retraining or supervision. Evaluations on Qwen2.5-Omni and MiniCPM-o-4.5 across visual, audio, and video tasks demonstrate that generation-aware span-level attribution produces more stable and interpretable explanations than naive self-attribution and embedding-based baselines, while remaining robust across multiple underlying attribution signals. Our results suggest that treating attribution as a structured generation-time tracing problem provides a scalable foundation for transparency in omni-modal language models.

For agents

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

get_harvested_code_for_paper("2604.13073")
get_code_for_paper("2604.13073")
have("2604.13073")

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

curl https://syntology.ai/api/ran/2604.13073.json

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

[![Syntology run record](https://syntology.ai/api/ran/2604.13073.svg)](https://syntology.ai/paper/2604.13073)

Connect an agent — have() is free.