SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2608.07548 · ICML · 2026

SC 2 -WM: A Self-Correcting World Model with Closed-Loop Feedback for Vision-and-Language Navigation in Continuous Environments

Junyu Gao, Xuan Yao, Changsheng Xu, Yuze Zhu, Zongmeng Wang

arXiv · PDF · Open in the Atlas

Code that ran

5 of 7 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; 1 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 4 failures of our instrument; 2 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/2608.07548.json · badge: /api/ran/2608.07548.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
sunrise-ikun/SC2_WM canonical (classifier label)
classifier basis: census
5 ran · 2 no recorded run
FunctionStatusWhere it lives
convert_decoder_to_target Ran
ran; our contract draft was wrong, not the code
sunrise-ikun/SC2_WM/utils_p/convert.py
code served (permissive licence) · get_code("d1b19a3f10dca79d")
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 DecoderNet. 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 CustomNorm(nn.Module):
    """Custom normalization layer that can be used as target norm"""
    def __init__(self, num_features, warm_n):
        super().__init__()
        self.num_features = num_features
        self.warm_n = warm_n
        self.weight = nn.Parameter(torch.ones(num_features))
        self.bias = nn.Parameter(torch.zeros(num_features))
    
    def load_old_dict(self, old_norm):
        """Custom loading method"""
        pass
    
    def forward(self, x):
        return x

class SimpleLayer(nn.Module):
    """Simplified layer with batch norm"""
    def __init__(self, num_features):
        super().__init__()
        self.bn = nn.BatchNorm2d(num_features)
        self.num_features = num_features
    
    def forward(self, x):
        return self.bn(x)

class DecoderNet(nn.Module):
    """Minimal decoder network with 5 stages"""
    def __init__(self):
        super().__init__()
        self.layer0 = SimpleLayer(64)
        self.layer1 = SimpleLayer(128)
        self.layer2 = SimpleLayer(256)
        self.layer3 = SimpleLayer(512)
        self.layer4 = nn.BatchNorm2d(1024)
        self.layer4.num_features = 1024
    
    def __getitem__(self, idx):
        return [self.layer0, self.layer1, self.layer2, self.layer3, self.layer4][idx]
    
    def __setitem__(self, idx, value):
        layers = [self.layer0, self.layer1, self.layer2, self.layer3, self.layer4]
        layers[idx] = value

net = DecoderNet()
norm = CustomNorm

ARGS = {
    'net': net,
    'norm': norm,
    'start': 0,
    'end': 5,
    'verbose': True,
    'input_size': 512,
    'warm_n': 5
}
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 DecoderNet — a DecoderNet output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DecoderNet output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
convert_encoder_to_target Ran
ran; our contract draft was wrong, not the code
sunrise-ikun/SC2_WM/utils_p/convert.py
code served (permissive licence) · get_code("ca50b22abf292757")
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 MinimalNet. 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)

# Create a custom norm layer that the function expects to convert to
class CustomNorm(nn.Module):
    def __init__(self, num_features, warm_n):
        super().__init__()
        self.num_features = num_features
        self.warm_n = warm_n
        self.weight = nn.Parameter(torch.ones(num_features))
        self.bias = nn.Parameter(torch.zeros(num_features))
    
    def load_old_dict(self, old_norm):
        pass
    
    def forward(self, x):
        return x

# Create a minimal ResNet-like network structure
class MinimalBlock(nn.Module):
    def __init__(self, in_channels, out_channels, bottleneck=False):
        super().__init__()
        self.bn1 = nn.BatchNorm2d(out_channels)
        self.bn2 = nn.BatchNorm2d(out_channels)
        if bottleneck:
            self.bn3 = nn.BatchNorm2d(out_channels)
        self.downsample = nn.Sequential(
            nn.Conv2d(in_channels, out_channels, kernel_size=1),
            nn.BatchNorm2d(out_channels)
        ) if in_channels != out_channels else None

class MinimalNet(nn.Module):
    def __init__(self, bottleneck=False):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)
        self.bn1 = nn.BatchNorm2d(64)
        
        # Create layers with blocks
        self.layer1 = nn.ModuleList([MinimalBlock(64, 64, bottleneck=bottleneck)])
        self.layer2 = nn.ModuleList([MinimalBlock(64, 128, bottleneck=bottleneck)])
        self.layer3 = nn.ModuleList([MinimalBlock(128, 256, bottleneck=bottleneck)])
        self.layer4 = nn.ModuleList([MinimalBlock(256, 512, bottleneck=bottleneck)])

# Build the network
net = MinimalNet(bottleneck=False)

# Define arguments
ARGS = {
    'net': net,
    'norm': CustomNorm,
    'start': 0,
    'end': 5,
    'verbose': True,
    'bottleneck': False,
    'input_size': 512,
    'warm_n': 5
}
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 MinimalNet — a MinimalNet output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a MinimalNet 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
jaccard Ran
ran, honoured the contract we drafted
sunrise-ikun/SC2_WM/utils_p/metrics.py
code served (permissive licence) · get_code("12b5ba241c136bb3")
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 float, 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 numpy as np

# Create test and reference numpy arrays with matching shapes
test = np.array([
    [[1, 0, 1, 0],
     [0, 1, 0, 1]],
    [[1, 1, 0, 0],
     [0, 0, 1, 1]]
], dtype=np.int32)

reference = np.array([
    [[1, 0, 0, 0],
     [0, 1, 0, 0]],
    [[1, 0, 0, 0],
     [0, 0, 1, 0]]
], dtype=np.int32)

# Create a ConfusionMatrix instance
class ConfusionMatrix:
    def __init__(self, test=None, reference=None):
        self.tp = None
        self.fp = None
        self.tn = None
        self.fn = None
        self.size = None
        self.reference_empty = None
        self.reference_full = None
        self.test_empty = None
        self.test_full = None
        self.set_reference(reference)
        self.set_test(test)

    def set_test(self, test):
        self.test = test
        self.reset()

    def set_reference(self, reference):
        self.reference = reference
        self.reset()

    def reset(self):
        self.tp = None
        self.fp = None
        self.tn = None
        self.fn = None
        self.size = None
        self.test_empty = None
        self.test_full = None
        self.reference_empty = None
        self.reference_full = None

    def compute(self):
        if self.test is None or self.reference is None:
            raise ValueError("'test' and 'reference' must both be set to compute confusion matrix.")
        self.tp = int(((self.test != 0) * (self.reference != 0)).sum())
        self.fp = int(((self.test != 0) * (self.reference == 0)).sum())
        self.tn = int(((self.test == 0) * (self.reference == 0)).sum())
        self.fn = int(((self.test == 0) * (self.reference != 0)).sum())
        self.size = int(np.prod(self.reference.shape, dtype=np.int64))
        self.test_empty = not np.any(self.test)
        self.test_full = np.all(self.test)
        self.reference_empty = not np.any(self.reference)
        self.reference_full = np.all(self.reference)

    def get_matrix(self):
        for entry in (self.tp, self.fp, self.tn, self.fn):
            if entry is None:
                self.compute()
                break
        return (self.tp, self.fp, self.tn, self.fn)

    def get_size(self):
        if self.size is None:
            self.compute()
        return self.size

    def get_existence(self):
        for case in (self.test_empty, self.test_full, self.reference_empty, self.reference_full):
            if case is None:
                self.compute()
                break
        return (self.test_empty, self.test_full, self.reference_empty, self.reference_full)

confusion_matrix = ConfusionMatrix(test, reference)

ARGS = {
    'test': None,
    'reference': None,
    'confusion_matrix': confusion_matrix,
    'nan_for_nonexisting': True
}
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 float, 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
precision Ran
ran; our contract draft was wrong, not the code
sunrise-ikun/SC2_WM/utils_p/metrics.py
code served (permissive licence) · get_code("035dc40fdafe0275")
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 float, 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 numpy as np

# Create test and reference numpy arrays
# They need to be the same shape for the assertion to pass
test_array = np.array([[1, 0, 1, 0],
                       [1, 1, 0, 0]])

reference_array = np.array([[1, 0, 1, 1],
                            [1, 0, 0, 0]])

# Create a ConfusionMatrix instance
confusion_matrix_obj = None  # Will be None initially, letting the function create it

ARGS = {
    'test': test_array,
    'reference': reference_array,
    'confusion_matrix': confusion_matrix_obj,
    'nan_for_nonexisting': True
}
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 float, 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
static_to_grid Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
sunrise-ikun/SC2_WM/habitat_extensions/maps.py
code served (permissive licence) · get_code("e3de2e386766be00")
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 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)

# Analyzing the function:
# - realworld_x: float - a real-world x coordinate
# - realworld_y: float - a real-world y coordinate
# - grid_resolution: Tuple[int, int] - grid dimensions (width, height)
# - bounds: Dict[str, Tuple[float, float]] - dictionary with 'upper' and 'lower' keys
#   Each containing tuples with at least 3 elements (indices 0 and 2 are accessed)

# The function accesses:
# bounds['upper'][2] - z-coordinate of upper bound
# bounds['lower'][2] - z-coordinate of lower bound
# bounds['upper'][0] - x-coordinate of upper bound
# bounds['lower'][0] - x-coordinate of lower bound

ARGS = {
    'realworld_x': 5.5,
    'realworld_y': 3.2,
    'grid_resolution': (10, 8),
    'bounds': {
        'upper': (10.0, 0.0, 20.0),  # (x_max, y, z_max)
        'lower': (0.0, 0.0, 0.0)      # (x_min, y, z_min)
    }
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family int — 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
colorize_topdown_map Not yet run
unverified (harvested, no recorded run)
sunrise-ikun/SC2_WM/habitat_extensions/maps.py
code served (permissive licence) · get_code("3559a5421dd17ce5")
How it ran

no invocation recorded.

dice Not yet run
unverified (harvested, no recorded run)
sunrise-ikun/SC2_WM/utils_p/metrics.py
code served (permissive licence) · get_code("a85d8587c247d950")
How it ran

no invocation recorded.

Also present in other papers' code

Of 7 samples lifted from this paper's linked repositories, 7 carry a function identity and 0 do not. Of the 7 distinct functions with an identity, 7 are also present in other papers whose code we hold (7 with at least one paper dated earlier, 0 the same month, 0 later) and 0 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
colorize_topdown_map 7 other papers
7 dated earlier
2110.02207 (2021-10)
Waypoint Models for Instruction-guided Navigation in Continuous Environments
2203.02764 (2022-03)
Bridging the Gap Between Learning in Discrete and Continuous Environments for Vision-and-Language Navigation
2210.16822 (2022-10)
Towards Versatile Embodied Navigation
and 4 more not named here
static_to_grid 6 other papers
6 dated earlier
2110.02207 (2021-10)
Waypoint Models for Instruction-guided Navigation in Continuous Environments
2210.16822 (2022-10)
Towards Versatile Embodied Navigation
2304.03047 (2023-04)
ETPNav: Evolving Topological Planning for Vision-Language Navigation in Continuous Environments
and 3 more not named here
dice 4 other papers
4 dated earlier
1809.10486 (2018-09)
nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation
2212.04497 (2022-12)
UNETR++: Delving into Efficient and Accurate 3D Medical Image Segmentation
2410.09299 (2024-10)
Hierarchical uncertainty estimation for learning-based registration in neuroimaging
and 1 more not named here
jaccard 4 other papers
4 dated earlier
1809.10486 (2018-09)
nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation
2212.04497 (2022-12)
UNETR++: Delving into Efficient and Accurate 3D Medical Image Segmentation
2410.09299 (2024-10)
Hierarchical uncertainty estimation for learning-based registration in neuroimaging
and 1 more not named here
precision 4 other papers
4 dated earlier
1809.10486 (2018-09)
nnU-Net: Self-adapting Framework for U-Net-Based Medical Image Segmentation
2212.04497 (2022-12)
UNETR++: Delving into Efficient and Accurate 3D Medical Image Segmentation
2410.09299 (2024-10)
Hierarchical uncertainty estimation for learning-based registration in neuroimaging
and 1 more not named here
convert_decoder_to_target 1 other paper
1 dated earlier
2506.23468 (2025-06)
NavM rph: A Self-Evolving World Model for Vision-and-Language Navigation in Continuous Environments
convert_encoder_to_target 1 other paper
1 dated earlier
2506.23468 (2025-06)
NavM rph: A Self-Evolving World Model for Vision-and-Language Navigation in Continuous Environments

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

Vision-and-Language Navigation in Continuous Environments (VLN-CE) requires agents to make fine-grained navigation decisions under partial observability. However, most existing methods rely on open-loop execution, lacking mechanisms to detect and correct internal state drift during inference. We propose SC 2 -WM, a self-correcting world model framework that introduces internal feedback for closed-loop decision making in VLN-CE. Our method derives feedback from world-model foresight to perform state-level plan refinement before action execution. To handle challenging scenarios, we further introduce conditional world-aware adaptation, which enables model-level correction by selectively updating the world model at test time when feedback indicates model capacity insufficiency. Experiments on standard VLN-CE benchmarks demonstrate improved navigation robustness and generalization. Our code is available at https: //github.com/sunrise-ikun/SC2_WM.

For agents

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

get_harvested_code_for_paper("2608.07548")
get_code_for_paper("2608.07548")
have("2608.07548")

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

curl https://syntology.ai/api/ran/2608.07548.json

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

[![Syntology run record](https://syntology.ai/api/ran/2608.07548.svg)](https://syntology.ai/paper/2608.07548)

Connect an agent — have() is free.