SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2603.02635 · 2026

SaFeR-ToolKit: Structured Reasoning via Virtual Tool Calling for Multimodal Safety

Yang Liu, Xi Chen, Kun Wang, Kang Li, Zixuan Xu, Zhigang Zeng, Huahui Yi, Qiankun Li, Gongli Xi, Tiancheng He

arXiv · PDF · Open in the Atlas

Code that ran

9 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 0 have no contract verdict; 9 failures of our instrument; 6 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2603.02635.json · badge: /api/ran/2603.02635.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
hiyouga/EasyR1 canonical (classifier label)
classifier basis: census
9 ran · 6 no recorded run
FunctionStatusWhere it lives
collate_fn Ran
ran; our contract draft was wrong, not the code
hiyouga/EasyR1/verl/utils/dataset.py
code served (permissive licence) · get_code("6dabc7804e03b1d0")
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 torch
import numpy as np
from collections import defaultdict

torch.manual_seed(0)

# Create sample features as a list of dicts
# Each dict contains both tensor and non-tensor values
features = [
    {
        "tensor_field_1": torch.randn(4, 3),
        "tensor_field_2": torch.randn(2, 5),
        "non_tensor_field_1": "sample_string_1",
        "non_tensor_field_2": 42,
    },
    {
        "tensor_field_1": torch.randn(4, 3),
        "tensor_field_2": torch.randn(2, 5),
        "non_tensor_field_1": "sample_string_2",
        "non_tensor_field_2": 99,
    },
]

ARGS = {
    "features": features,
}
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 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
get_kl_controller Ran
ran; our contract draft was wrong, not the code
hiyouga/EasyR1/verl/trainer/core_algos.py
code served (permissive licence) · get_code("7633fd09bcf8972a")
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 AdaptiveKLController. 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

class AlgorithmConfig:
    def __init__(self, kl_type, kl_coef, kl_target=None, kl_horizon=None):
        self.kl_type = kl_type
        self.kl_coef = kl_coef
        self.kl_target = kl_target
        self.kl_horizon = kl_horizon

ARGS = {
    'algorithm_config': AlgorithmConfig(
        kl_type='adaptive',
        kl_coef=0.1,
        kl_target=0.01,
        kl_horizon=10000.0
    )
}
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 AdaptiveKLController — a AdaptiveKLController output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a AdaptiveKLController 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_rope_index Ran
ran; our contract draft was wrong, not the code
hiyouga/EasyR1/verl/models/transformers/qwen3_vl.py
code served (permissive licence) · get_code("0758af83dd12ea52")
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 int, shape (3, 8) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

class MockImageProcessor:
    def __init__(self):
        self.merge_size = 2

class MockQwen3VLProcessor:
    def __init__(self):
        self.image_processor = MockImageProcessor()
        self.image_token_id = 100
        self.video_token_id = 101
        self.vision_start_token_id = 102

processor = MockQwen3VLProcessor()

# Create input_ids: 1D tensor representing a single example
# Include vision_start_token, image_token, and regular text tokens
input_ids = torch.tensor([1, 2, 3, processor.vision_start_token_id, processor.image_token_id, 4, 5, 6], dtype=torch.long)

# Create image_grid_thw: tensor with shape (num_images, 3) for T, H, W
image_grid_thw = torch.tensor([[1, 4, 4]], dtype=torch.long)

# Create video_grid_thw: tensor with shape (num_videos, 3) for T, H, W
video_grid_thw = None

# Create attention_mask: 1D tensor same shape as input_ids
attention_mask = torch.ones_like(input_ids)

ARGS = {
    "processor": processor,
    "input_ids": input_ids,
    "image_grid_thw": image_grid_thw,
    "video_grid_thw": video_grid_thw,
    "attention_mask": attention_mask,
}
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 int, shape (3, 8) — shape (3, 8) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
prepare_fa2_from_position_ids Ran
ran; our contract draft was wrong, not the code
hiyouga/EasyR1/verl/models/transformers/flash_attention_utils.py
code served (permissive licence) · get_code("72f505c72eeef3e8")
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

torch.manual_seed(0)

# Create small tensors for query, key, value
# Shape: (batch_size, seq_len, hidden_dim)
batch_size = 2
seq_len = 3
hidden_dim = 4

query = torch.randn(batch_size, seq_len, hidden_dim)
key = torch.randn(batch_size, seq_len, hidden_dim)
value = torch.randn(batch_size, seq_len, hidden_dim)

# Create position_ids tensor
# Must be 2D according to the assertion
# Position IDs typically contain indices including 0 at sequence starts
position_ids = torch.tensor([
    [0, 1, 2],
    [0, 1, 2]
], dtype=torch.int64)

ARGS = {
    'query': query,
    'key': key,
    'value': value,
    'position_ids': position_ids
}
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
qwen2_vl_model_forward Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
hiyouga/EasyR1/verl/models/transformers/qwen2_vl.py
code served (permissive licence) · get_code("c2c77f3bf52d2305")
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 Qwen2VLCausalLMOutputWithPast. 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
from transformers.models.qwen2_vl.modeling_qwen2_vl import Qwen2VLCausalLMOutputWithPast, Qwen2VLForConditionalGeneration, Qwen2VLModel, Qwen2VLModelOutputWithPast

torch.manual_seed(0)

# Build a minimal Qwen2VLModel that returns the expected output structure
class MinimalQwen2VLModel(nn.Module):
    def __init__(self, hidden_size=4):
        super().__init__()
        self.hidden_size = hidden_size
        self.embedding = nn.Embedding(100, hidden_size)
    
    def forward(self, input_ids, **kwargs):
        hidden_states = self.embedding(input_ids)
        return Qwen2VLModelOutputWithPast(last_hidden_state=hidden_states)

# Build a minimal Qwen2VLForConditionalGeneration
class MinimalQwen2VLForConditionalGeneration(nn.Module):
    def __init__(self):
        super().__init__()
        self.hidden_size = 4
        self.vocab_size = 100
        self.model = MinimalQwen2VLModel(hidden_size=self.hidden_size)
        self.lm_head = nn.Linear(self.hidden_size, self.vocab_size)
    
    def forward(self, input_ids, labels=None, **kwargs):
        outputs = self.model(input_ids=input_ids, **kwargs)
        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)
        return Qwen2VLCausalLMOutputWithPast(logits=logits)

# Create instances
model = MinimalQwen2VLForConditionalGeneration()

# Create small tensors
input_ids = torch.randint(0, 100, (2, 5), dtype=torch.long)  # batch=2, seq_len=5
labels = torch.randint(0, 100, (2, 5), dtype=torch.long)

ARGS = {
    'self': model,
    'input_ids': input_ids,
    'labels': labels,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Qwen2VLCausalLMOutputWithPast — a Qwen2VLCausalLMOutputWithPast output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Qwen2VLCausalLMOutputWithPast 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
qwen3_vl_model_forward Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
hiyouga/EasyR1/verl/models/transformers/qwen3_vl.py
code served (permissive licence) · get_code("75bdeb21d651d4b5")
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 Qwen3VLCausalLMOutputWithPast. 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
from typing import Optional

torch.manual_seed(0)

# Create a minimal LM head
class MinimalLMHead(nn.Module):
    def __init__(self, hidden_size: int, vocab_size: int):
        super().__init__()
        self.linear = nn.Linear(hidden_size, vocab_size)
    
    def forward(self, hidden_states):
        return self.linear(hidden_states)

# Create a minimal model that returns hidden states
class MinimalQwen3VLModel(nn.Module):
    def __init__(self, hidden_size: int):
        super().__init__()
        self.hidden_size = hidden_size
    
    def forward(self, input_ids, **kwargs):
        # Return a tuple where first element is hidden states
        batch_size = input_ids.shape[0]
        seq_length = input_ids.shape[1]
        hidden_states = torch.randn(batch_size, seq_length, self.hidden_size)
        return (hidden_states,)

# Create the minimal conditional generation model
class MinimalQwen3VLForConditionalGeneration(nn.Module):
    def __init__(self, hidden_size: int = 4, vocab_size: int = 3):
        super().__init__()
        self.model = MinimalQwen3VLModel(hidden_size)
        self.lm_head = MinimalLMHead(hidden_size, vocab_size)
    
    def forward(self, input_ids: torch.LongTensor, labels: Optional[torch.LongTensor] = None, **kwargs):
        outputs = self.model(input_ids=input_ids, **kwargs)
        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)
        # Return a simple object with logits attribute
        class Output:
            def __init__(self, logits):
                self.logits = logits
        return Output(logits=logits)

# Setup arguments
model = MinimalQwen3VLForConditionalGeneration(hidden_size=4, vocab_size=3)
input_ids = torch.randint(0, 3, (2, 5), dtype=torch.long)
labels = torch.randint(0, 3, (2, 5), dtype=torch.long)

ARGS = {
    'self': model,
    'input_ids': input_ids,
    'labels': labels,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch, transformers — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Qwen3VLCausalLMOutputWithPast — a Qwen3VLCausalLMOutputWithPast output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Qwen3VLCausalLMOutputWithPast 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
reduce_metrics Ran
ran; our contract draft was wrong, not the code
hiyouga/EasyR1/verl/trainer/metrics.py
code served (permissive licence) · get_code("c817845fd5e85ff7")
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 numpy as np

# The function reduce_metrics takes a dict where keys are strings and values are lists
# It computes the mean of each list
# We need to construct a metrics dict with string keys and list values containing numeric data

ARGS = {
    "metrics": {
        "loss": [0.5, 0.45, 0.42, 0.40],
        "accuracy": [0.75, 0.78, 0.81, 0.83],
        "f1_score": [0.72, 0.76, 0.79, 0.82],
    }
}
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 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
register_adv_estimator Ran
ran; our contract draft was wrong, not the code
hiyouga/EasyR1/verl/trainer/core_algos.py
code served (permissive licence) · get_code("59499c9109dfcb5e")
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 function. 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
from enum import Enum

torch.manual_seed(0)

class AdvantageEstimator(str, Enum):
    """
    Using an enumeration class to avoid spelling errors in adv_estimator
    """
    GAE = 'gae'
    GRPO = 'grpo'
    GRPO_PASSK = 'grpo_passk'
    REINFORCE_PLUS_PLUS = 'reinforce_plus_plus'
    REMAX = 'remax'
    RLOO = 'rloo'

ADV_ESTIMATOR_MAP = {}

def register_adv_estimator(name: AdvantageEstimator):
    """Decorator to register a advantage estimator function with a given name."""

    def decorator(fn):
        wrapped_fn = torch.no_grad()(fn)
        ADV_ESTIMATOR_MAP[getattr(name, 'value', name)] = wrapped_fn
        return wrapped_fn
    return decorator


def dummy_advantage_fn():
    """Dummy advantage estimator function for testing."""
    return torch.tensor([1.0, 2.0])


ARGS = {
    'name': AdvantageEstimator.GAE
}
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 function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
unpad_dataproto Ran
ran; our contract draft was wrong, not the code
hiyouga/EasyR1/verl/protocol.py
code served (permissive licence) · get_code("a91c18341292d50e")
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 DataProto. 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

torch.manual_seed(0)

class DataProto:
    """Stand-in DataProto class that supports slicing via __getitem__"""
    def __init__(self, data):
        self.data = data
    
    def __getitem__(self, key):
        """Support slicing operation"""
        sliced_data = self.data[key]
        return DataProto(sliced_data)
    
    def __repr__(self):
        return f"DataProto({self.data})"

# Create a small DataProto instance with some data
data_tensor = torch.randn(2, 4)
data_proto = DataProto(data_tensor)

# Define arguments
ARGS = {
    "data": data_proto,
    "pad_size": 0
}
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 DataProto — a DataProto output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DataProto 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
compute_advantage_return Not yet run
unverified (harvested, no recorded run)
hiyouga/EasyR1/verl/trainer/core_algos.py
code served (permissive licence) · get_code("61938c6bf9d13f84")
How it ran

no invocation recorded.

flash_attention_forward Not yet run
unverified (harvested, no recorded run)
hiyouga/EasyR1/verl/models/transformers/flash_attention_utils.py
code served (permissive licence) · get_code("7688ceba543fa185")
How it ran

no invocation recorded.

get_rope_index Not yet run
unverified (harvested, no recorded run)
hiyouga/EasyR1/verl/models/transformers/qwen2_vl.py
code served (permissive licence) · get_code("07ec6b020ddf3b01")
How it ran

no invocation recorded.

process_image Not yet run
unverified (harvested, no recorded run)
hiyouga/EasyR1/verl/utils/dataset.py
code served (permissive licence) · get_code("9a950b864d329f1c")
How it ran

no invocation recorded.

qwen2_vl_base_forward Not yet run
unverified (harvested, no recorded run)
hiyouga/EasyR1/verl/models/transformers/qwen2_vl.py
code served (permissive licence) · get_code("15cffd3e46d6bf67")
How it ran

no invocation recorded.

qwen3_vl_base_forward Not yet run
unverified (harvested, no recorded run)
hiyouga/EasyR1/verl/models/transformers/qwen3_vl.py
code served (permissive licence) · get_code("75394a0b094d4a2f")
How it ran

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, 15 are also present in other papers whose code we hold (7 with at least one paper dated earlier, 0 the same month, 3 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
unpad_dataproto 10 other papers
7 dated earlier, 2 later, 1 undated
2503.17352 (2025-03)
OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles
2505.11049 (2025-05)
GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning
2505.16282 (2025-05)
ARPO:End-to-End Policy Optimization for GUI Agents with Experience Replay
and 7 more not named here
get_kl_controller 9 other papers
7 dated earlier, 1 later, 1 undated
2503.17352 (2025-03)
OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles
2505.11049 (2025-05)
GuardReasoner-VL: Safeguarding VLMs via Reinforced Reasoning
2505.16282 (2025-05)
ARPO:End-to-End Policy Optimization for GUI Agents with Experience Replay
and 6 more not named here
get_rope_index 7 other papers
5 dated earlier, 1 later, 1 undated
2504.15716 (2025-04)
DianJin-R1: Evaluating and Enhancing Financial Reasoning in Large Language Models
2505.16282 (2025-05)
ARPO:End-to-End Policy Optimization for GUI Agents with Experience Replay
2506.01908 (2025-06)
Reinforcement Learning Tuning for VideoLLMs: Reward Design and Data Efficiency
and 4 more not named here
compute_advantage_return 3 other papers
2 dated earlier, 1 undated
2503.17352 (2025-03)
OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles
2601.05167 (2026-01)
RelayLLM: Efficient Reasoning via Collaborative Decoding
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
flash_attention_forward 3 other papers
2 dated earlier, 1 undated
2503.17352 (2025-03)
OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles
2601.05167 (2026-01)
RelayLLM: Efficient Reasoning via Collaborative Decoding
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
register_adv_estimator 3 other papers
2 dated earlier, 1 undated
2503.17352 (2025-03)
OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles
2601.05167 (2026-01)
RelayLLM: Efficient Reasoning via Collaborative Decoding
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
prepare_fa2_from_position_ids 2 other papers
1 dated earlier, 1 undated
2503.17352 (2025-03)
OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
collate_fn 1 other paper
1 undated
no dated paper among those named
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
get_rope_index 1 other paper
1 undated
no dated paper among those named
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
process_image 1 other paper
1 undated
no dated paper among those named
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
qwen2_vl_base_forward 1 other paper
1 undated
no dated paper among those named
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
qwen2_vl_model_forward 1 other paper
1 undated
no dated paper among those named
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
qwen3_vl_base_forward 1 other paper
1 undated
no dated paper among those named
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
qwen3_vl_model_forward 1 other paper
1 undated
no dated paper among those named
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-Language Models
reduce_metrics 1 other paper
1 undated
no dated paper among those named
openreview_Y1PXB8HBV7
See First, Reason Later: Mutual Information-Guided Reinforcement Learning for Vision-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

Vision-language models remain susceptible to multimodal jailbreaks and over-refusal because safety hinges on both visual evidence and user intent, while many alignment pipelines supervise only the final response. To address this, we present SaFeR-ToolKit, which formalizes safety decision-making as a checkable protocol. Concretely, a planner specifies a persona, a Percep-tion→Reasoning→Decision tool set, and a constrained transition graph, while a responder outputs a typed key-value tool trace before the final answer. To make the protocol reliably followed in practice, we train a single policy with a three-stage curriculum (SFT→DPO→GRPO), where GRPO directly supervises tool usage beyond answer-level feedback. Our contributions are two-fold: I. Dataset. The first tool-based safety reasoning dataset, comprising 31,654 examples (SFT 6k, DPO 18.6k, GRPO 6k) plus 1k held-out evaluation. II. Experiments. On

For agents

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

get_harvested_code_for_paper("2603.02635")
get_code_for_paper("2603.02635")
have("2603.02635")

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

curl https://syntology.ai/api/ran/2603.02635.json

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

[![Syntology run record](https://syntology.ai/api/ran/2603.02635.svg)](https://syntology.ai/paper/2603.02635)

Connect an agent — have() is free.