SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2503.09573 · ICLR · 2025

BLOCK DIFFUSION: INTERPOLATING BETWEEN AU-TOREGRESSIVE AND DIFFUSION LANGUAGE MODELS

Justin Chiu, Volodymyr Kuleshov, Jiaqi Han, Marianne Arriola, Aaron Kerem Gokaslan, Zhihan Yang, Zhixuan Qi, ¶ Subham, Sekhar Sahoo

arXiv · PDF · Open in the Atlas

Code that ran

14 of 15 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 7 have no contract verdict; 7 failures of our instrument; 1 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/2503.09573.json · badge: /api/ran/2503.09573.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
kuleshov-group/bd3lms canonical (classifier label)
classifier basis not recorded on the link
10 ran · 1 no recorded run
MindSpore-scientific/code-12 community (classifier label)
classifier basis not recorded on the link
4 ran · 0 no recorded run
FunctionStatusWhere it lives
bias_dropout_add_scale Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
kuleshov-group/bd3lms/models/autoregressive.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from hxixixh/gumbel-distill, is pointer only (licence: NONE) · get_code("fbbc6722b74bda27")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from hxixixh/gumbel-distill, and GitHub detected no licence in that repository when it was checked. 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.

bias_dropout_add_scale_fused_train Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
kuleshov-group/bd3lms/models/autoregressive.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from hxixixh/gumbel-distill, is pointer only (licence: NONE) · get_code("ee894c881614c3ec")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from hxixixh/gumbel-distill, and GitHub detected no licence in that repository when it was checked. 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.

block_diff_mask Ran
ran; our contract draft was wrong, not the code
kuleshov-group/bd3lms/models/dit.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from hxixixh/gumbel-distill, is pointer only (licence: NONE) · get_code("14959ccf1e594504")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from hxixixh/gumbel-distill, and GitHub detected no licence in that repository when it was checked. 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.

decode Ran
ran on a synthesized input
MindSpore-scientific/code-12/AECRNet/data.py
code served (permissive licence) · get_code("72814d7de33e5aef")
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 Image. 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
from PIL import Image

# Create a small valid image array (Height=4, Width=3, Channels=3 for RGB)
# PIL.Image.fromarray expects a numpy array with shape (H, W) for grayscale
# or (H, W, 3) for RGB or (H, W, 4) for RGBA
img_array = np.random.randint(0, 256, size=(4, 3, 3), dtype=np.uint8)

ARGS = {
    "img": img_array
}
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 Image — a Image output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Image output the recorded ndim 3 and dtype family int 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_bias_dropout_add_scale Ran
ran; our contract draft was wrong, not the code
kuleshov-group/bd3lms/models/autoregressive.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from jdeschena/blockgen, is pointer only (licence: NONE) · get_code("06d2d8e50f91481f")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from jdeschena/blockgen, and GitHub detected no licence in that repository when it was checked. 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_noise Ran
ran on a synthesized input
kuleshov-group/bd3lms/noise_schedule.py
code served (permissive licence) · get_code("e1fdbc4d0bebba92")
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 LogLinearNoise. 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 nn

torch.manual_seed(0)

class Config:
    def __init__(self):
        self.noise = NoiseConfig()

class NoiseConfig:
    def __init__(self):
        self.type = 'loglinear'

config = Config()
noise_type = 'loglinear'

ARGS = {
    'config': config,
    'noise_type': noise_type
}
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 LogLinearNoise — a LogLinearNoise output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a LogLinearNoise 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_time_variables Ran
ran on a synthesized input
kuleshov-group/bd3lms/ssd-lm/ssd_model_decode.py
code served (permissive licence) · get_code("0e5797e82608fcdc")
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 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 torch
import math

torch.manual_seed(0)

# Construct arguments for get_time_variables
t = torch.tensor([0, 5, 10, 15], dtype=torch.float32)  # time steps
total_t = torch.tensor(100.0, dtype=torch.float32)  # total time
device = 'cpu'

ARGS = {
    't': t,
    'total_t': total_t,
    'device': device
}
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 2, 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
lm1b_detokenizer Ran
ran; our contract draft was wrong, not the code
kuleshov-group/bd3lms/dataloader.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from jdeschena/blockgen, is pointer only (licence: NONE) · get_code("576e1b9b9bcae8ff")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from jdeschena/blockgen, and GitHub detected no licence in that repository when it was checked. 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.

load_variavle Ran
ran on a synthesized input
MindSpore-scientific/code-12/Alexnet-ABeffect/feature_predict_abm.py
code served (permissive licence) · get_code("ea40ca66b8b3349e")
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 pickle
import tempfile
import os

# Create a temporary file and pickle some data to it
temp_dir = tempfile.gettempdir()
temp_file = os.path.join(temp_dir, "test_pickle_data.pkl")

# Create test data to pickle
test_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "dict"}
}

# Write the test data to a pickle file
with open(temp_file, 'wb') as f:
    pickle.dump(test_data, f)

ARGS = {
    "filename": temp_file
}
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
logits_projection Ran
ran on a synthesized input
kuleshov-group/bd3lms/ssd-lm/ssd_model_decode.py
code served (permissive licence) · get_code("5f25d11e3f7d8a07")
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, 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 torch

torch.manual_seed(0)

# Create a 3D logits tensor: (batch_size=2, sequence_length=3, vocab_size=4)
logits = torch.randn(2, 3, 4)

# top_p parameter for nucleus sampling
top_p = 0.9

# one_hot_value parameter
one_hot_value = 1.0

ARGS = {
    "logits": logits,
    "top_p": top_p,
    "one_hot_value": one_hot_value,
}
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, 4) — shape (2, 3, 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
ptb_detokenizer Ran
ran; our contract draft was wrong, not the code
kuleshov-group/bd3lms/dataloader.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from jdeschena/blockgen, is pointer only (licence: NONE) · get_code("9f7fbdbe7ca280e0")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from jdeschena/blockgen, and GitHub detected no licence in that repository when it was checked. 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.

save_variable Ran
ran on a synthesized input
MindSpore-scientific/code-12/Alexnet-ABeffect/feature_predict_abm.py
code served (permissive licence) · get_code("f34618312c4fa9c9")
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 pickle
import tempfile
import os

# Create a temporary file path for the output
temp_dir = tempfile.gettempdir()
temp_filename = os.path.join(temp_dir, "test_variable.pkl")

# Define a variable to save - can be any picklable object
v = {
    "key1": "value1",
    "key2": [1, 2, 3, 4],
    "key3": {"nested": "dict"}
}

ARGS = {
    "v": v,
    "filename": temp_filename
}
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
set_random_seed Ran
ran on a synthesized input
MindSpore-scientific/code-12/AECRNet/data.py
code served (permissive licence) · get_code("14c181fabe865cbf")
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 random
import torch

torch.manual_seed(0)

ARGS = {
    "img_name": "test_image.jpg",
    "seed": 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 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
wt_detokenizer Ran
ran; our contract draft was wrong, not the code
kuleshov-group/bd3lms/dataloader.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from jdeschena/blockgen, is pointer only (licence: NONE) · get_code("46c10627206a7795")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from jdeschena/blockgen, and GitHub detected no licence in that repository when it was checked. 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.

fused_flex_attention Not yet run
unverified (harvested, no recorded run)
kuleshov-group/bd3lms/models/dit.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from hxixixh/gumbel-distill, is pointer only (licence: NONE) · get_code("b4f2749f771988c2")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from hxixixh/gumbel-distill, and GitHub detected no licence in that repository when it was checked. 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. No invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 distinct functions with an identity, 12 are also present in other papers whose code we hold (10 with at least one paper dated earlier, 0 the same month, 8 later) and 3 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
lm1b_detokenizer 11 other papers
3 dated earlier, 8 later
2310.16834 (2023-10)
Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution
2406.03736 (2024-06)
Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data
2412.10193 (2024-12)
Simple Guidance Mechanisms for Discrete Diffusion Models
and 8 more not named here
ptb_detokenizer 10 other papers
2 dated earlier, 8 later
2310.16834 (2023-10)
Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution
2406.03736 (2024-06)
Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data
2505.18883 (2025-05)
Partition Generative Modeling: Masked Modeling Without Masks
and 7 more not named here
wt_detokenizer 10 other papers
2 dated earlier, 8 later
2310.16834 (2023-10)
Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution
2406.03736 (2024-06)
Your Absorbing Discrete Diffusion Secretly Models the Conditional Distributions of Clean Data
2505.18883 (2025-05)
Partition Generative Modeling: Masked Modeling Without Masks
and 7 more not named here
get_bias_dropout_add_scale 8 other papers
4 dated earlier, 4 later
2310.16834 (2023-10)
Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution
2405.20971 (2024-05)
Amortizing intractable inference in diffusion models for vision, language, and control
2410.21035 (2024-10)
Beyond Autoregression: Fast LLMs via Self-Distillation Through Time
and 5 more not named here
bias_dropout_add_scale
brick of this paper's assembly: helper
7 other papers
1 dated earlier, 6 later
2410.21035 (2024-10)
Beyond Autoregression: Fast LLMs via Self-Distillation Through Time
2505.18883 (2025-05)
Partition Generative Modeling: Masked Modeling Without Masks
2509.15188 (2025-09)
Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning
and 4 more not named here
bias_dropout_add_scale_fused_train
brick of this paper's assembly: helper
5 other papers
1 dated earlier, 4 later
2410.21035 (2024-10)
Beyond Autoregression: Fast LLMs via Self-Distillation Through Time
2505.18883 (2025-05)
Partition Generative Modeling: Masked Modeling Without Masks
2509.15188 (2025-09)
Fast and Fluent Diffusion Language Models via Convolutional Decoding and Rejective Fine-tuning
and 2 more not named here
block_diff_mask 3 other papers
3 later
2603.22216 (2026-03)
Gumbel Distillation for Parallel Text Generation
2603.25702 (2026-03)
S2D2: Fast Decoding for Diffusion LLMs via Training-Free Self-Speculation
2606.24982 (2026-06)
Latent Block-Diffusion Temporal Point Processes: A Semi-Autoregressive Framework for Asynchronous Event Sequence Generation
decode 2 other papers
2 dated earlier
2311.18815 (2023-11)
IMMA: Immunizing text-to-image Models against Malicious Adaptation
2408.14512 (2024-08)
LLMs as Zero-shot Graph Learners: Alignment of GNN Representations with LLM Token Embeddings
set_random_seed 2 other papers
2 dated earlier
2311.18815 (2023-11)
IMMA: Immunizing text-to-image Models against Malicious Adaptation
2408.14512 (2024-08)
LLMs as Zero-shot Graph Learners: Alignment of GNN Representations with LLM Token Embeddings
fused_flex_attention
brick of this paper's assembly: model
1 other paper
1 later
2603.22216 (2026-03)
Gumbel Distillation for Parallel Text Generation
load_variavle 1 other paper
1 dated earlier
2311.18815 (2023-11)
IMMA: Immunizing text-to-image Models against Malicious Adaptation
save_variable 1 other paper
1 dated earlier
2311.18815 (2023-11)
IMMA: Immunizing text-to-image Models against Malicious Adaptation

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

Diffusion language models offer unique benefits over autoregressive models due to their potential for parallelized generation and controllability, yet they lag in likelihood modeling and are limited to fixed-length generation. In this work, we introduce a class of block diffusion language models that interpolate between discrete denoising diffusion and autoregressive models. Block diffusion overcomes key limitations of both approaches by supporting flexible-length generation and improving inference efficiency with KV caching and parallel token sampling. We propose a recipe for building effective block diffusion models that includes an efficient training algorithm, estimators of gradient variance, and data-driven noise schedules to minimize the variance. Block diffusion sets a new state-of-the-art performance among diffusion models on language modeling benchmarks and enables generation of arbitrary-length sequences. We provide the code 1 , along with the model weights and blog post on the project page: https://m-arriola.com/bd3lms

For agents

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

get_harvested_code_for_paper("2503.09573")
get_code_for_paper("2503.09573")
have("2503.09573")

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

curl https://syntology.ai/api/ran/2503.09573.json

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

[![Syntology run record](https://syntology.ai/api/ran/2503.09573.svg)](https://syntology.ai/paper/2503.09573)

Connect an agent — have() is free.