SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2601.04537 · 2026

Linear Dynamics in the RLVR Training of Large Language Models

Wei Chen, Jiayu Liu, Hao Xu, Tianle Wang, Ning Miao, Zhongyuan Wu, Shenghao Jin

arXiv · PDF · Open in the Atlas

Code that ran

18 of 21 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; 4 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 13 failures of our instrument; 3 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/2601.04537.json · badge: /api/ran/2601.04537.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
Miaow-Lab/RLVR-Linearity canonical (classifier label)
classifier basis: census
18 ran · 3 no recorded run
FunctionStatusWhere it lives
calculate_alpha Ran
ran, honoured the contract we drafted
Miaow-Lab/RLVR-Linearity/acceleration/logits_extrapolation.py
code served (permissive licence) · get_code("77938e6a32c6c83b")
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 torch
import numpy as np

torch.manual_seed(0)

# Based on the function signature and body, we need three numeric values
# that represent steps in some progression. The function calculates an
# interpolation factor (alpha) between base_step and second_step relative
# to target_step.

base_step = 10
second_step = 20
target_step = 15

ARGS = {
    'base_step': base_step,
    'second_step': second_step,
    'target_step': target_step,
}
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 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
extract_final_answer Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/evaluation/pass_at_k_eval.py
code served (permissive licence) · get_code("60425d0b541f8efd")
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 re

def find_braces_content(s, tag='\\boxed'):
    idx = s.find(tag + '{')
    if idx == -1:
        return None
    start = idx + len(tag) + 1
    count = 1
    i = start
    while i < len(s) and count:
        if s[i] == '{':
            count += 1
        elif s[i] == '}':
            count -= 1
        i += 1
    if count == 0:
        return s[start:i - 1]
    else:
        return None

def extract_final_answer(output: str) -> str:
    v = find_braces_content(output, '\\boxed')
    if v:
        return v.strip()
    v = find_braces_content(output, '\\fbox')
    if v:
        return v.strip()
    m_frac = re.findall('\\\\[d]?frac\\s*{[^{}]*}\\s*{[^{}]*}', output)
    if m_frac:
        return m_frac[-1].strip()
    m_all = re.findall('\\$([^\\$]+)\\$', output, re.DOTALL)
    if m_all:
        return m_all[-1].strip()
    m_all = re.findall('\\d+/\\d+', output)
    if m_all:
        return m_all[-1]
    lines = [l.strip(' $\n\r') for l in output.strip().split('\n') if l.strip()]
    if lines:
        return lines[-1]
    return output.strip()

# Test case 1: output with \boxed
test_output_boxed = r"The answer is \boxed{42}"

# Test case 2: output with \fbox
test_output_fbox = r"The result is \fbox{3.14}"

# Test case 3: output with fraction
test_output_frac = r"Solution: \frac{1}{2} is the answer"

# Test case 4: output with dollar signs
test_output_dollar = r"Result is $x = 5$ in the equation"

# Test case 5: output with simple fraction pattern
test_output_simple_frac = r"The ratio is 3/4"

# Test case 6: output with multiple lines
test_output_multiline = r"Line 1\nLine 2\nFinal answer: 99"

# Test case 7: simple output
test_output_simple = r"42"

ARGS = {
    "output": test_output_boxed
}
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
find_braces_content Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/evaluation/pass_at_k_eval.py
code served (permissive licence) · get_code("47132ed4fdcf0a8b")
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 torch
import torch.nn as nn

# Analyze the function:
# - Parameter 's': a string that may contain tags like '\boxed{'
# - Parameter 'tag': a string tag (default '\boxed')
# - The function searches for tag + '{' in s, then extracts balanced braces content
# - It uses string methods: find(), len(), indexing

# Create test cases for the function

# Test case 1: String with boxed content
s1 = r"The answer is \boxed{42}"

# Test case 2: String with nested braces
s2 = r"Result: \boxed{x + \frac{1}{2}}"

# Test case 3: String with multiple nested levels
s3 = r"Formula: \boxed{{{nested}}}"

# Test case 4: String without the tag
s4 = r"No boxed content here"

# Test case 5: Simple case
s5 = r"\boxed{hello}"

# Use the simplest test case
test_string = s5
test_tag = r'\boxed'

ARGS = {
    's': test_string,
    'tag': test_tag
}
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
flatten_responses Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/analysis/token_logprob/token_logprob_linearity.py
code served (permissive licence) · get_code("85b305142d653b1d")
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

# Construct valid arguments for flatten_responses
all_responses = [
    {
        'problem': 'What is 2+2?',
        'generated_texts': [
            {'response_id': 'resp_001', 'text': 'The answer is 4.'},
            {'response_id': 'resp_002', 'text': '2+2 equals 4.'},
        ]
    },
    {
        'problem': 'What is the capital of France?',
        'generated_texts': [
            {'response_id': 'resp_003', 'text': 'Paris is the capital of France.'},
        ]
    }
]

ARGS = {
    'all_responses': all_responses
}
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
generate_config_from_args Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/verl/verl/model_merger/base_model_merger.py
pointer only (licence: MIT); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("8896f4e2bafe8abc")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from LeapLabTHU/OPDVR, 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_dynamic_pipeline_shards Ran
ran, honoured the contract we drafted
Miaow-Lab/RLVR-Linearity/verl/verl/model_merger/megatron_model_merger.py
pointer only (licence: MIT); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("94209884dcb2308e")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from LeapLabTHU/OPDVR, 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_filtered_random_mask Ran
ran, violated the contract we drafted
Miaow-Lab/RLVR-Linearity/analysis/weight/weight_linearity.py
code served (permissive licence) · get_code("9f087582a1131c88")
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 (4, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create a list of state dicts with small tensors
# Each state dict represents model weights at different time steps
all_state_dicts = []

for t in range(3):
    state_dict = {
        'layer1': torch.randn(4, 3) + t * 0.1,  # Shape (4, 3) with time-dependent variation
        'layer2': torch.randn(2, 5) + t * 0.05,
    }
    all_state_dicts.append(state_dict)

ARGS = {
    'layer_name': 'layer1',
    'all_state_dicts': all_state_dicts,
    'sample_ratio': 0.5,
    'min_samples': 2,
    'min_abs_change': 0.001,
    'min_unique_values': 2,
    'device': 'cpu',
}
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 (4, 3) — shape (4, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_layer_weights Ran
ran, honoured the contract we drafted
Miaow-Lab/RLVR-Linearity/analysis/weight/weight_linearity.py
code served (permissive licence) · get_code("cc8952b03ae8ebc6")
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 (4, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create a small state_dict with tensor parameters
state_dict = {
    'layer1.weight': torch.randn(4, 3),
    'layer1.bias': torch.randn(4),
    'layer2.weight': torch.randn(5, 4),
}

# Define target device (CPU only)
target_device = 'cpu'

# Define layer name to retrieve
layer_name = 'layer1.weight'

ARGS = {
    'state_dict': state_dict,
    'layer_name': layer_name,
    'target_device': target_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 Tensor, ndim 2, dtype family float, shape (4, 3) — shape (4, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_step_int_from_path Ran
ran, honoured the contract we drafted
Miaow-Lab/RLVR-Linearity/acceleration/weight_extrapolation.py
code served (permissive licence) · get_code("ac35f0d2152218c0")
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 int, ndim 0, dtype family int, 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 get_step_int_from_path(path):
    """Extract integer from 'global_step_XXX' in the path."""
    match = re.search('global_step_(\\d+)', path)
    if match:
        return int(match.group(1))
    return None

# Test with a path containing global_step_XXX pattern
ARGS = {
    'path': '/checkpoints/model_global_step_12345/weights.pt'
}
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 int, ndim 0, dtype family int, 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
load_json Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/analysis/token_logprob/token_logprob_linearity.py
code served (permissive licence) · get_code("db79e1c6f71f41bb")
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 json
import tempfile
import os

# Create a temporary JSON file with sample data
temp_dir = tempfile.mkdtemp()
data_file = os.path.join(temp_dir, "test_data.json")

sample_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "object"}
}

with open(data_file, 'w', encoding='utf-8') as f:
    json.dump(sample_data, f)

ARGS = {
    "data_file": data_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
load_json Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/evaluation/pass_at_k_eval.py
code served (permissive licence) · get_code("98cb765f6409dd94")
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 json
import tempfile
import os

# Create a temporary JSON file with valid data
temp_dir = tempfile.mkdtemp()
data_path = os.path.join(temp_dir, "test_data.json")

# Write some sample JSON data to the file
sample_data = {
    "key1": "value1",
    "key2": [1, 2, 3],
    "key3": {"nested": "object"}
}

with open(data_path, 'w') as f:
    json.dump(sample_data, f)

ARGS = {
    "data_path": data_path
}
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
load_weights_only Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/acceleration/weight_extrapolation.py
code served (permissive licence) · get_code("bb3923bae9d2bdfa")
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 gc
from unittest.mock import Mock, MagicMock

torch.manual_seed(0)

# Create a mock model that will be returned by AutoModelForCausalLM.from_pretrained
mock_model = Mock()
mock_model.state_dict = Mock(return_value={
    'layer1.weight': torch.randn(4, 3),
    'layer1.bias': torch.randn(4),
    'layer2.weight': torch.randn(3, 2),
})

# Mock the AutoModelForCausalLM.from_pretrained method
import transformers
original_from_pretrained = transformers.AutoModelForCausalLM.from_pretrained
transformers.AutoModelForCausalLM.from_pretrained = Mock(return_value=mock_model)

ARGS = {
    'model_path': 'gpt2'
}
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 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
mix_logits Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Miaow-Lab/RLVR-Linearity/acceleration/logits_extrapolation.py
code served (permissive licence) · get_code("888d0d880907ef8e")
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 (2, 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 torch.nn.functional as F

torch.manual_seed(0)

# Create small logit tensors
# batch_size=2, num_classes=4
logits1 = torch.randn(2, 4)
logits2 = torch.randn(2, 4)

# Mix parameter - controls interpolation between logits1 and logits2
alpha = 0.5

# Temperature parameter for scaling
temperature = 1.0

ARGS = {
    "logits1": logits1,
    "logits2": logits2,
    "alpha": alpha,
    "temperature": temperature,
}
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 (2, 4) — shape (2, 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
sanitize_filename Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/analysis/token_logprob/plot_token_logprob_linearity.py
code served (permissive licence) · get_code("70be02ba13364d71")
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 torch
import numpy as np

# Construct arguments for sanitize_filename function
ARGS = {
    'text': 'My <File> Name: "Document" | Test?.txt',
    'max_length': 50
}
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 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
split_data Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/analysis/token_logprob/token_logprob_linearity.py
code served (permissive licence) · get_code("08188c51a18bec5f")
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 numpy as np

torch.manual_seed(0)

# The function splits data across processes using indexing (data[start_idx:end_idx])
# So data needs to support len() and slicing operations
# A simple list works perfectly for this use case

data = [
    {"id": 0, "value": torch.randn(4)},
    {"id": 1, "value": torch.randn(4)},
    {"id": 2, "value": torch.randn(4)},
    {"id": 3, "value": torch.randn(4)},
    {"id": 4, "value": torch.randn(4)},
    {"id": 5, "value": torch.randn(4)},
]

num_processes = 2
process_index = 0

ARGS = {
    "data": data,
    "num_processes": num_processes,
    "process_index": process_index,
}
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 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
split_list_evenly Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/evaluation/inference_vllm_offline.py
code served (permissive licence) · get_code("a656a12f2110c7b0")
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. 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)

# Analyze the function:
# - lst: a list that needs to be split
# - n: number of parts to split into
# The function uses len(lst), divmod, list slicing, and range
# It doesn't call any methods on list elements, just slices the list

# Create a simple list with a good number of elements to demonstrate uneven splitting
lst = list(range(25))  # 25 elements

# Create n as the number of parts
n = 4

ARGS = {
    "lst": lst,
    "n": n,
}
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 — a list 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
token_id_to_str Ran
ran; our contract draft was wrong, not the code
Miaow-Lab/RLVR-Linearity/analysis/token_logprob/plot_token_logprob_linearity.py
code served (permissive licence) · get_code("e1577f990e3c8461")
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 torch
import numpy as np

class MockTokenizer:
    def decode(self, token_ids):
        # Simple mock tokenizer that returns string representations
        if token_ids[0] == 0:
            return "[PAD]"
        elif token_ids[0] == 1:
            return "[UNK]"
        elif token_ids[0] == 2:
            return "hello"
        elif token_ids[0] == 3:
            return "world with a very long string that exceeds twenty characters in length"
        elif token_ids[0] == 4:
            return "   "  # whitespace only
        elif token_ids[0] == 5:
            return "tab\there"
        else:
            return "token"

token_id = 2
tokenizer = MockTokenizer()

ARGS = {
    "token_id": token_id,
    "tokenizer": tokenizer
}
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 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
top_p_filtering Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Miaow-Lab/RLVR-Linearity/acceleration/logits_extrapolation.py
code served (permissive licence) · get_code("f1ce71d11869fdae")
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 (2, 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 small logits tensor (log_softmax values)
# Shape: (batch_size=2, vocab_size=4)
logits = torch.randn(2, 4)

# Define the ARGS dictionary with all parameters
ARGS = {
    'logits': logits,
    'top_p': 0.95,
    'filter_value': -float('Inf')
}
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 (2, 4) — shape (2, 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
load_tokenizer Not yet run
unverified (harvested, no recorded run)
Miaow-Lab/RLVR-Linearity/analysis/token_logprob/plot_token_logprob_linearity.py
code served (permissive licence) · get_code("974078a91d3ec703")
How it ran

no invocation recorded.

load_weights_from_checkpoint Not yet run
unverified (harvested, no recorded run)
Miaow-Lab/RLVR-Linearity/analysis/weight/weight_linearity.py
code served (permissive licence) · get_code("c8c7f0daade03739")
How it ran

no invocation recorded.

prepare_prompts Not yet run
unverified (harvested, no recorded run)
Miaow-Lab/RLVR-Linearity/evaluation/inference_vllm_offline.py
code served (permissive licence) · get_code("6404f02a3872c91e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 21 samples lifted from this paper's linked repositories, 21 carry a function identity and 0 do not. Of the 21 distinct functions with an identity, 2 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 2 the same month, 2 later) and 19 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
generate_config_from_args 57 other papers
3 dated earlier, 4 the same month, 47 later, 3 undated
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2509.25787 (2025-09)
Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking
2510.08630 (2025-10)
ExPO-HM: Learning to Explain-then-Detect for Hateful Meme Detection
and 54 more not named here
get_dynamic_pipeline_shards 52 other papers
3 dated earlier, 3 the same month, 44 later, 2 undated
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2509.25787 (2025-09)
Self-Evolving Vision-Language Models for Image Quality Assessment via Voting and Ranking
2510.08630 (2025-10)
ExPO-HM: Learning to Explain-then-Detect for Hateful Meme Detection
and 49 more not named here

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

Reinforcement learning with verifiable rewards (RLVR) has driven significant performance gains in reasoning-oriented large language models (LLMs), yet its internal training dynamics remain largely a black box. In this work, we perform a comprehensive trajectory-level analysis of RLVR and uncover a striking regularity: across various model families, RL algorithms, and training configurations, RLVR consistently enters a robust linear regime, where both parameter weights and output log-probabilities, measured rigorously via teacher-forced evaluation, evolve in a highly linear manner (R 2 > 0.7). Through controlled experiments and theoretical analysis, we demonstrate that this linearity is not a coincidence, but stems from the high-variance, noisy nature of RLVR training signals, which act as a low-pass filter to concentrate optimization along a stable, low-dimensional drift. Moreover, we show that this linear structure is not merely descriptive but powerfully predictive and actionable. Specifically, weight-space extrapolation matches the performance of standard RL optimization while achieving a 6.1× training speedup through periodic re-grounding. Meanwhile, output-space extrapolation serves as a lightweight intervention that effectively bypasses late-stage model collapse, consistently outperforming standard RL across mathematical and coding benchmarks, with an average performance improvement of 4.2%. Our code is available at https://github.com/Miaow-Lab/RLVR-Linearity. Recent work has examined RLVR dynamics from several complementary perspectives, including exploration-exploitation [

For agents

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

get_harvested_code_for_paper("2601.04537")
get_code_for_paper("2601.04537")
have("2601.04537")

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

curl https://syntology.ai/api/ran/2601.04537.json

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

[![Syntology run record](https://syntology.ai/api/ran/2601.04537.svg)](https://syntology.ai/paper/2601.04537)

Connect an agent — have() is free.