SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2303.16133 · 2023

Exposing and Addressing Cross-Task Inconsistency in Unified Vision-Language Models

arXiv · PDF · Open in the Atlas

Code that ran

9 of 9 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; 0 honoured the contract we drafted, 0 violated it and 8 have no contract verdict; 1 failure of our instrument; 0 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2303.16133.json · badge: /api/ran/2303.16133.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
adymaharana/cococon canonical (classifier label)
classifier basis not recorded on the link
9 ran · 0 no recorded run
FunctionStatusWhere it lives
cross_entropy_with_logits Ran
ran on a synthesized input
adymaharana/cococon/unified-io/uio/model.py
code served (permissive licence) · get_code("69563a7420a49fa8")
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 ArrayImpl, ndim 1, dtype family float, shape (2,) 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 jax
import jax.numpy as jnp
import numpy as np

# Set random seed for reproducibility
np.random.seed(0)

# Create logits: shape (batch_size=2, num_classes=4)
logits = jnp.array([
    [1.0, 2.0, 0.5, 1.5],
    [0.2, 0.8, 1.2, 0.1]
], dtype=jnp.float32)

# Create targets (one-hot encoded): shape (batch_size=2, num_classes=4)
targets = jnp.array([
    [0.0, 1.0, 0.0, 0.0],
    [1.0, 0.0, 0.0, 0.0]
], dtype=jnp.float32)

ARGS = {
    "logits": logits,
    "targets": targets,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 1, dtype family float, shape (2,) — shape (2,) 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
dot_product_attention Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
adymaharana/cococon/unified-io/uio/t5x_layers.py
code served (permissive licence) · get_code("dac0288079ac279f")
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 ArrayImpl, ndim 4, dtype family float, shape (2, 3, 2, 6) 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 jax
import jax.numpy as jnp
from jax import random

# Set seed for reproducibility
key = random.PRNGKey(0)

# Dimensions
batch = 2
q_length = 3
kv_length = 4
num_heads = 2
qk_depth_per_head = 5
v_depth_per_head = 6

# Create query, key, value tensors with distinct sizes
query = jnp.ones((batch, q_length, num_heads, qk_depth_per_head), dtype=jnp.float32)
key = jnp.ones((batch, kv_length, num_heads, qk_depth_per_head), dtype=jnp.float32)
value = jnp.ones((batch, kv_length, num_heads, v_depth_per_head), dtype=jnp.float32)

# Create bias (optional) - broadcastable to [batch, num_heads, q_length, kv_length]
bias = jnp.zeros((batch, num_heads, q_length, kv_length), dtype=jnp.float32)

# Generate dropout RNG key
dropout_rng = random.PRNGKey(1)

# Define arguments
ARGS = {
    "query": query,
    "key": key,
    "value": value,
    "bias": bias,
    "dropout_rng": dropout_rng,
    "dropout_rate": 0.1,
    "deterministic": False,
    "dtype": jnp.float32,
    "float32_logits": False,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 4, dtype family float, shape (2, 3, 2, 6) — shape (2, 3, 2, 6) 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
evaluate_contrast_sets Ran
ran on a synthesized input
adymaharana/cococon/unified-io/evaluate_contrast_all_tasks.py
code served (permissive licence) · get_code("8acbb3b977150c88")
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. 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
from scipy.stats import spearmanr

torch.manual_seed(0)

# The function expects two list-like objects (task_1_logprobs, task_2_logprobs)
# that support:
# - len() to get the number of contrast sets
# - indexing with [idx]
# - passing to spearmanr()

# Create simple lists of log probabilities
task_1_logprobs = [-0.5, -0.3, -0.8, -0.1, -0.6]
task_2_logprobs = [-0.4, -0.7, -0.2, -0.9, -0.55]

ARGS = {
    "task_1_logprobs": task_1_logprobs,
    "task_2_logprobs": task_2_logprobs,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is 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
evaluate_contrast_sets Ran
ran on a synthesized input
adymaharana/cococon/unified-io/evaluate_task.py
code served (permissive licence) · get_code("032edece79949c43")
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 int; 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)

# The function expects two lists (or list-like objects) of log probabilities
# task_1_logprobs and task_2_logprobs should have the same length
# The function iterates from index 1 to len(task_2_logprobs)-1
# It compares logprob values, so they should be numeric (float)

task_1_logprobs = [-2.5, -1.8, -3.2, -1.5]
task_2_logprobs = [-1.9, -2.1, -1.3, -2.8]

ARGS = {
    "task_1_logprobs": task_1_logprobs,
    "task_2_logprobs": task_2_logprobs,
}
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 list, ndim 1, dtype family int — 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_open_img_annotations Ran
ran on a synthesized input
adymaharana/cococon/unified-io/evaluate_contrast_all_tasks.py
code served (permissive licence) · get_code("50fa314af669c957")
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 tempfile
import os

# Create a temporary JSON file with the expected structure
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.json')
data = {
    'images': [
        {'id': 1, 'file_name': 'image1.jpg', 'width': 640, 'height': 480},
        {'id': 2, 'file_name': 'image2.jpg', 'width': 640, 'height': 480},
        {'id': 3, 'file_name': 'image3.jpg', 'width': 640, 'height': 480}
    ],
    'annotations': [
        {'id': 1, 'image_id': 1, 'caption': 'A cat on the mat'},
        {'id': 2, 'image_id': 1, 'caption': 'A sleeping cat'},
        {'id': 3, 'image_id': 2, 'caption': 'A dog playing fetch'},
        {'id': 4, 'image_id': 3, 'caption': 'A bird in flight'}
    ]
}
json.dump(data, temp_file)
temp_file.close()

ARGS = {
    'data_file': temp_file.name
}
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
rank Ran
ran on a synthesized input
adymaharana/cococon/unified-io/evaluate_contrast_all_tasks.py
code served (permissive licence) · get_code("08a5061ff824fcc4")
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 int; 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)

# The function rank() takes a reference parameter and an optional reverse boolean
# It sorts the reference by value while tracking original indices
# reference should be iterable with comparable elements

reference = [3.2, 1.5, 4.1, 2.0]

ARGS = {
    "reference": reference,
    "reverse": 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 list, ndim 1, dtype family int — 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
read_image Ran
ran on a synthesized input
adymaharana/cococon/evaluators/evaluators.py
code served (permissive licence) · get_code("9e2a5c7d00f0cba5")
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 ndarray, ndim 3, dtype family int, shape (3, 4, 3) 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
from PIL import Image
import tempfile
import os

# Create a temporary image file for testing
with tempfile.TemporaryDirectory() as tmpdir:
    # Create a simple test image
    test_image = Image.new('RGB', (4, 3), color=(255, 0, 0))
    temp_filepath = os.path.join(tmpdir, 'test_image.png')
    test_image.save(temp_filepath)
    
    # Since we can't persist the temp file after the context, we need to create it differently
    # Create a persistent test image in memory and save it
    test_image = Image.new('RGB', (4, 3), color=(255, 0, 0))
    temp_filepath = '/tmp/test_read_image.png'
    test_image.save(temp_filepath)

ARGS = {
    'filename': temp_filepath
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 3, dtype family int, shape (3, 4, 3) — shape (3, 4, 3) 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
reverse_space_to_depth Ran
ran on a synthesized input
adymaharana/cococon/unified-io/uio/t5x_layers.py
code served (permissive licence) · get_code("c04f426b52feac4e")
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 ArrayImpl, ndim 4, dtype family float, shape (2, 8, 6, 1) 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 jax.numpy as jnp

# For 4D case: (batch, height, width, channels)
# spatial_block_size=2 means the channel dimension should be divisible as (dh * dw * c)
# So if we want spatial_block_size=2, we need channels to encode dh=2, dw=2, c=1
# Shape: (batch=2, height=4, width=3, channels=4) where 4 = 2*2*1
frames_4d = jnp.ones((2, 4, 3, 4), dtype=jnp.float32)

# For 5D case: (batch, time, height, width, channels)
# temporal_block_size=2, spatial_block_size=2
# channels should encode dt=2, dh=2, dw=2, c=1 = 8
frames_5d = jnp.ones((2, 2, 4, 3, 8), dtype=jnp.float32)

ARGS = {
    "frames": frames_4d,
    "temporal_block_size": 1,
    "spatial_block_size": 2,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 4, dtype family float, shape (2, 8, 6, 1) — shape (2, 8, 6, 1) 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
space_to_depth Ran
ran on a synthesized input
adymaharana/cococon/unified-io/uio/t5x_layers.py
code served (permissive licence) · get_code("ca0c4d33f9443b8d")
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 ArrayImpl, ndim 3, dtype family float, shape (2, 6, 12) 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 jax.numpy as jnp

# Create a 4D tensor (batch, height, width, channels)
# Shape: (batch=2, height=4, width=6, channels=3)
# Height and width are divisible by spatial_block_size for proper rearrangement
frames_4d = jnp.ones((2, 4, 6, 3))

# Create a 5D tensor (batch, time, height, width, channels)
# Shape: (batch=2, time=2, height=4, width=6, channels=3)
# Time, height, and width are divisible by their respective block sizes
frames_5d = jnp.ones((2, 2, 4, 6, 3))

ARGS = {
    "frames": frames_4d,
    "temporal_block_size": 1,
    "spatial_block_size": 2,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
einops, jax — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ArrayImpl, ndim 3, dtype family float, shape (2, 6, 12) — shape (2, 6, 12) 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

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 1 later) and 8 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
dot_product_attention 5 other papers
4 dated earlier, 1 later
2106.03598 (2021-06)
SciFive: a text-to-text transformer model for biomedical literature
2112.00791 (2021-12)
Controlling Conditional Language Models without Catastrophic Forgetting
2206.05408 (2022-06)
Multi-instrument Music Synthesis with Spectrogram Diffusion
and 2 more not named 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

As general purpose vision models get increasingly effective at a wide set of tasks, it is imperative that they be consistent across the tasks they support. Inconsistent AI models are considered brittle and untrustworthy by human users and are more challenging to incorporate into larger systems that take dependencies on their outputs. Measuring consistency between very heterogeneous tasks that might include outputs in different modalities is challenging since it is difficult to determine if the predictions are consistent with one another. As a solution, we introduce a benchmark dataset, CocoCon, where we create contrast sets by modifying test instances for multiple tasks in small but semantically meaningful ways to change the gold label and outline metrics for measuring if a model is consistent by ranking the original and perturbed instances across tasks. We find that state-of-the-art vision-language models suffer from a surprisingly high degree of inconsistent behavior across tasks, especially for more heterogeneous tasks. To alleviate this issue, we propose a rank correlation-based auxiliary training objective, computed over large automatically created cross-task contrast sets, that improves the multi-task consistency of large unified models while retaining their original accuracy on downstream tasks.

For agents

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

get_harvested_code_for_paper("2303.16133")
get_code_for_paper("2303.16133")
have("2303.16133")

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

curl https://syntology.ai/api/ran/2303.16133.json

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

[![Syntology run record](https://syntology.ai/api/ran/2303.16133.svg)](https://syntology.ai/paper/2303.16133)

Connect an agent — have() is free.