SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2601.00423 · 2026

E-GRPO: High Entropy Steps Drive Effective Reinforcement Learning for Flow Models

Zhang Zhang, Yueqi Duan, Shengjun Zhang, Chensheng Dai

arXiv · PDF · Open in the Atlas

Code that ran

10 of 10 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 2 honoured the contract we drafted, 2 violated it and 0 have no contract verdict; 6 failures of our instrument; 0 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/2601.00423.json · badge: /api/ran/2601.00423.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
shengjun-zhang/VisualGRPO canonical (classifier label)
classifier basis: census
10 ran · 0 no recorded run
FunctionStatusWhere it lives
balance_pos_neg Ran
ran; our contract draft was wrong, not the code
shengjun-zhang/VisualGRPO/fastvideo/reward_model/utils.py
code served (permissive licence) · get_code("d92441b3beb9fce7")
How it ran

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

import torch
import random

# Set seed for reproducibility
torch.manual_seed(0)
random.seed(0)

# Create sample data
# Each sample is a dict with 'advantages' key containing a tensor
samples = [
    {'advantages': torch.tensor(0.5)},   # positive
    {'advantages': torch.tensor(-0.3)},  # negative
    {'advantages': torch.tensor(0.8)},   # positive
    {'advantages': torch.tensor(-0.2)},  # negative
    {'advantages': torch.tensor(-0.1)},  # negative
]

ARGS = {
    'samples': samples,
    'use_random': 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 list, ndim 1, dtype family float — a list output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_reward Ran
ran; our contract draft was wrong, not the code
shengjun-zhang/VisualGRPO/fastvideo/reward_model/utils.py
code served (permissive licence) · get_code("a6ffa44bdd5d788c")
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 torch.nn as nn

torch.manual_seed(0)

# Define minimal reward model classes that satisfy the function's type checks
class HPSClipRewardModel(nn.Module):
    def __init__(self):
        super().__init__()
    
    def forward(self, images, input_prompts):
        return [0.5 + i * 0.1 for i in range(len(input_prompts))]


class CLIPScoreRewardModel(nn.Module):
    def __init__(self):
        super().__init__()
    
    def forward(self, input_prompts, images):
        return [0.6 + i * 0.1 for i in range(len(input_prompts))]


class ImageRewardModel(nn.Module):
    def __init__(self):
        super().__init__()
    
    def forward(self, images, input_prompts):
        return [0.7 + i * 0.1 for i in range(len(input_prompts))]


class UnifiedRewardModel(nn.Module):
    def __init__(self):
        super().__init__()
    
    def forward(self, images, input_prompts):
        rewards = [0.8 + i * 0.1 for i in range(len(input_prompts))]
        successes = [True] * len(input_prompts)
        return (rewards, successes)


class PickScoreRewardModel(nn.Module):
    def __init__(self):
        super().__init__()
    
    def forward(self, images, input_prompts):
        return [0.4 + i * 0.1 for i in range(len(input_prompts))]


# Create sample data
batch_size = 2
image_height = 4
image_width = 3
num_channels = 3

images = [torch.randn(num_channels, image_height, image_width) for _ in range(batch_size)]
input_prompts = ["a cat", "a dog"]

# Create reward models
reward_models = [
    HPSClipRewardModel(),
    CLIPScoreRewardModel(),
    ImageRewardModel(),
]

# Define reward weights
reward_weights = {
    "HPSClipRewardModel": 0.5,
    "CLIPScoreRewardModel": 0.3,
    "ImageRewardModel": 0.2,
}

ARGS = {
    "images": images,
    "input_prompts": input_prompts,
    "reward_models": reward_models,
    "reward_weights": reward_weights,
}
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 — 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
contains_chinese Ran
ran, violated the contract we drafted
shengjun-zhang/VisualGRPO/fastvideo/data_preprocess/preprocess_flux_embedding.py
code served (permissive licence) · get_code("33e3c69c78826a6b")
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 re

def contains_chinese(text):
    return bool(re.search('[\\u4e00-\\u9fff]', text))

ARGS = {
    "text": "Hello world with Chinese characters: 你好世界"
}
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
flux_step Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
shengjun-zhang/VisualGRPO/other_algorithms/dancegrpo_train_grpo_flux.py
code served (permissive licence) · get_code("fe8a641669476f3e")
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 math

torch.manual_seed(0)

# Create small tensors with distinct dimensions for clarity
batch_size = 2
latent_dim = 4
num_timesteps = 5

# model_output: gradient or prediction tensor, shape (batch, latent_dim)
model_output = torch.randn(batch_size, latent_dim)

# latents: current latent state, shape (batch, latent_dim)
latents = torch.randn(batch_size, latent_dim)

# eta: scalar noise scaling parameter
eta = 0.5

# sigmas: noise schedule, shape (num_timesteps,)
sigmas = torch.linspace(1.0, 0.0, num_timesteps)

# index: current step index (must be valid for indexing sigmas and sigmas[index+1])
index = 2

# prev_sample: previous sample or None, shape (batch, latent_dim)
prev_sample = None

# grpo: boolean flag for GRPO mode
grpo = True

# sde_solver: boolean flag for SDE solver mode
sde_solver = False

ARGS = {
    "model_output": model_output,
    "latents": latents,
    "eta": eta,
    "sigmas": sigmas,
    "index": index,
    "prev_sample": prev_sample,
    "grpo": grpo,
    "sde_solver": sde_solver,
}
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 — 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
generate_experiment_name Ran
ran; our contract draft was wrong, not the code
shengjun-zhang/VisualGRPO/other_algorithms/branchgrpo_train_grpo_flux.py
code served (permissive licence) · get_code("dc4b90fb67d85374")
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.

from datetime import datetime

class Args:
    def __init__(self):
        self.tree_split_points = "0.3,0.5,0.7"
        self.tree_split_rounds = 5
        self.tree_split_noise_scale = 0.1
        self.learning_rate = 0.001
        self.clip_range = 0.2
        self.depth_pruning = "2,4,6"
        self.depth_pruning_slide = True
        self.width_pruning_mode = 1
        self.width_pruning_ratio = 0.5
        self.tree_prob_weighted = True
        self.mix_ode_sde_tree = True
        self.mix_sde_window_size = 4

ARGS = {
    "args": Args()
}
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
pack_latents Ran
ran; our contract draft was wrong, not the code
shengjun-zhang/VisualGRPO/other_algorithms/train_grpo_flux.py
code served (permissive licence) · get_code("c6e6dc7696a9dd53")
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, 12, 16) 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)

# The function expects:
# - latents: a tensor that will be reshaped and permuted
# - batch_size: int
# - num_channels_latents: int
# - height: int
# - width: int

# Working backwards from the operations:
# latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
# This means the input tensor must have total elements equal to:
# batch_size * num_channels_latents * (height // 2) * 2 * (width // 2) * 2
# = batch_size * num_channels_latents * height * width

batch_size = 2
num_channels_latents = 4
height = 8
width = 6

# Create a latents tensor with the right number of elements
# Total elements needed: 2 * 4 * 8 * 6 = 384
latents = torch.randn(batch_size, num_channels_latents, height, width)

ARGS = {
    "latents": latents,
    "batch_size": batch_size,
    "num_channels_latents": num_channels_latents,
    "height": height,
    "width": 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, 12, 16) — shape (2, 12, 16) 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
parse_split_points Ran
ran, honoured the contract we drafted
shengjun-zhang/VisualGRPO/other_algorithms/branchgrpo_train_grpo_flux.py
code served (permissive licence) · get_code("4f6bf719daba09bd")
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
from dataclasses import dataclass

torch.manual_seed(0)

@dataclass
class MockArgs:
    tree_split_points: str = None
    tree_split_rounds: int = 0

# Test case 1: args with tree_split_points
args_with_points = MockArgs(tree_split_points="10, 20, 30")
total_steps_1 = 100

# Test case 2: args with tree_split_rounds
args_with_rounds = MockArgs(tree_split_rounds=4)
total_steps_2 = 100

# Test case 3: args with neither (empty case)
args_empty = MockArgs()
total_steps_3 = 50

# Choose one of the test cases for ARGS
ARGS = {
    "args": args_with_points,
    "total_steps": total_steps_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 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
prepare_latent_image_ids Ran
ran, honoured the contract we drafted
shengjun-zhang/VisualGRPO/other_algorithms/dancegrpo_train_grpo_flux.py
code served (permissive licence) · get_code("4080cb55d923dfba")
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 (12, 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 torch

torch.manual_seed(0)

ARGS = {
    "batch_size": 2,
    "height": 4,
    "width": 3,
    "device": "cpu",
    "dtype": torch.float32,
}
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 2, dtype family float, shape (12, 3) — shape (12, 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
sd3_time_shift Ran
ran, violated the contract we drafted
shengjun-zhang/VisualGRPO/other_algorithms/branchgrpo_train_grpo_flux.py
code served (permissive licence) · get_code("a34c4529d41a7731")
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 1, dtype family float, shape (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
import numpy as np

torch.manual_seed(0)

# The function sd3_time_shift takes two parameters: shift and t
# It performs a mathematical operation: shift * t / (1 + (shift - 1) * t)
# Both parameters should be numeric values (scalars or tensors)

shift = 2.0  # scalar shift value
t = torch.tensor([0.1, 0.5, 0.9, 0.3], dtype=torch.float32)  # tensor of time values

ARGS = {
    "shift": shift,
    "t": t
}
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 1, dtype family float, shape (4,) — shape (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
unpack_latents Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
shengjun-zhang/VisualGRPO/other_algorithms/train_grpo_flux.py
code served (permissive licence) · get_code("41cac86bce0769df")
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 4, dtype family float, shape (2, 4, 16, 20) 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)

# Construct valid arguments for unpack_latents
# The function expects:
# - latents: a tensor of shape (batch_size, num_patches, channels)
# - height: an integer
# - width: an integer
# - vae_scale_factor: an integer

# Based on the function logic:
# batch_size = 2
# num_patches should equal (height // 2) * (width // 2) after the transformation
# channels should be divisible by 4 (for the view operation) and by 4 again (for reshape)
# So channels should be divisible by 16

batch_size = 2
height = 64
width = 80
vae_scale_factor = 4

# Calculate derived dimensions
# height = 2 * (int(height) // (vae_scale_factor * 2)) = 2 * (64 // 8) = 2 * 8 = 16
# width = 2 * (int(width) // (vae_scale_factor * 2)) = 2 * (80 // 8) = 2 * 10 = 20
calculated_height = 2 * (int(height) // (vae_scale_factor * 2))
calculated_width = 2 * (int(width) // (vae_scale_factor * 2))

# num_patches = (height // 2) * (width // 2) = 8 * 10 = 80
num_patches = (calculated_height // 2) * (calculated_width // 2)

# channels must be divisible by 16 for the operations
channels = 16

# Create latents tensor with shape (batch_size, num_patches, channels)
latents = torch.randn(batch_size, num_patches, channels)

ARGS = {
    "latents": latents,
    "height": height,
    "width": width,
    "vae_scale_factor": vae_scale_factor,
}
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 4, dtype family float, shape (2, 4, 16, 20) — shape (2, 4, 16, 20) 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 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 6 are also present in other papers whose code we hold (6 with at least one paper dated earlier, 0 the same month, 5 later) and 4 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
prepare_latent_image_ids 3 other papers
2 dated earlier, 1 later · 1 only as an assembly brick
2505.07818 (2025-05)
DanceGRPO: Unleashing GRPO on Visual Generation
2509.25774 (2025-09)
PCPO: Proportionate Credit Policy Optimization for Aligning Image Generation Models
2605.08354 (2026-05)
Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria
sd3_time_shift 3 other papers
2 dated earlier, 1 later · 1 only as an assembly brick
2505.07818 (2025-05)
DanceGRPO: Unleashing GRPO on Visual Generation
2509.25774 (2025-09)
PCPO: Proportionate Credit Policy Optimization for Aligning Image Generation Models
2605.08354 (2026-05)
Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria
flux_step 2 other papers
1 dated earlier, 1 later
2505.07818 (2025-05)
DanceGRPO: Unleashing GRPO on Visual Generation
2605.08354 (2026-05)
Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria
pack_latents 2 other papers
1 dated earlier, 1 later · 1 only as an assembly brick
2509.25774 (2025-09)
PCPO: Proportionate Credit Policy Optimization for Aligning Image Generation Models
2605.08354 (2026-05)
Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria
unpack_latents 2 other papers
1 dated earlier, 1 later · 1 only as an assembly brick
2509.25774 (2025-09)
PCPO: Proportionate Credit Policy Optimization for Aligning Image Generation Models
2605.08354 (2026-05)
Auto-Rubric as Reward: From Implicit Preferences to Explicit Multimodal Generative Criteria
contains_chinese 1 other paper
1 dated earlier
2502.11190 (2025-02)
ReLearn: Unlearning via Learning for Large Language Models

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

Recent reinforcement learning has enhanced the flow matching models on human preference alignment. While stochastic sampling enables the exploration of denoising directions, existing methods which optimize over multiple denoising steps suffer from sparse and ambiguous reward signals. We observe that the high entropy steps enable more efficient and effective exploration while the low entropy steps result in undistinguished roll-outs. To this end, we propose E-GRPO, an entropy aware Group Relative Policy Optimization to increase the entropy of SDE sampling steps. Since the integration of stochastic differential equations suffer from ambiguous reward signals due to stochasticity from multiple steps, we specifically merge consecutive low entropy steps to formulate one high entropy step for SDE sampling, while applying ODE sampling on other steps. Building upon this, we introduce multi-step group normalized advantage, which computes group-relative advantages within samples sharing the same consolidated SDE denoising step. Experimental results on different reward settings have demonstrated the effectiveness of our methods. Our code is available at https: //github.com/shengjun-zhang/VisualGRPO.

For agents

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

get_harvested_code_for_paper("2601.00423")
get_code_for_paper("2601.00423")
have("2601.00423")

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

curl https://syntology.ai/api/ran/2601.00423.json

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

[![Syntology run record](https://syntology.ai/api/ran/2601.00423.svg)](https://syntology.ai/paper/2601.00423)

Connect an agent — have() is free.