SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2604.03993 · 2026

Can LLMs Learn to Reason Robustly under Noisy Supervision?

Gang Chen, Weiqiang Wang, Haobo Wang, Xing Zheng, Shenzhi Yang, Bowen Song, Sharon Li, Zhongqi Chen, Guangcheng Zhu, Yingfan Ma

arXiv · PDF · Open in the Atlas

Code that ran

13 of 14 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, 1 violated it and 0 have no contract verdict; 11 failures of our instrument; 1 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/2604.03993.json · badge: /api/ran/2604.03993.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
ShenzhiYang2000/OLR canonical (classifier label)
classifier basis: census
13 ran · 1 no recorded run
FunctionStatusWhere it lives
calculate_average_token_count Ran
ran, honoured the contract we drafted
ShenzhiYang2000/OLR/eval_scripts/mean_tokens.py
code served (permissive licence) · get_code("c97f8280f8a27801")
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 json
import re
import tempfile
import os

# Create a temporary JSONL file with sample data
temp_file = tempfile.NamedTemporaryFile(mode='w', delete=False, suffix='.jsonl', encoding='utf-8')
temp_file_path = temp_file.name

# Write sample JSONL data with generated_text fields
sample_data = [
    {'generated_text': 'The quick brown fox jumps over the lazy dog'},
    {'generated_text': 'Python is a powerful programming language for data science'},
    {'generated_text': 'Machine learning models require careful tuning and validation'},
]

for item in sample_data:
    temp_file.write(json.dumps(item) + '\n')

temp_file.close()

# Define ARGS dictionary with all parameters
ARGS = {
    'file_path': temp_file_path,
    'token_pattern': r'\b\w+\b'
}
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 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
compute_ce_dpo_loss_rm Ran
ran; our contract draft was wrong, not the code
ShenzhiYang2000/OLR/recipe/prime/prime_core_algos.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("f220595adfc5e83c")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from 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.

compute_detach_dpo_loss_rm Ran
ran; our contract draft was wrong, not the code
ShenzhiYang2000/OLR/recipe/prime/prime_core_algos.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("4d91e4383716e42c")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from 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.

compute_online_dpo_loss Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
ShenzhiYang2000/OLR/recipe/spin/core_algos.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("f19825983a523003")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from 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.

compute_onlinedpo_pref Ran
ran; our contract draft was wrong, not the code
ShenzhiYang2000/OLR/recipe/spin/core_algos.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("41b6e81dc3ecba79")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from 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.

compute_score_data_source Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
ShenzhiYang2000/OLR/recipe/open_math_reasoning/compute_score.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("05a86ccb0b3876dd")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from 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.

extract_results_from_files Ran
ran; our contract draft was wrong, not the code
ShenzhiYang2000/OLR/eval_scripts/collect_results.py
code served (permissive licence) · get_code("137fec8b564fbde7")
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 os
import tempfile
import shutil

# Create a temporary directory with test log files
temp_dir = tempfile.mkdtemp()

# Create sample .log files with benchmark results
log_content_1 = """Some header information
Processing started
math: 85.5
olympiad_bench: 72.3
minerva: 91.2
aime: 65.8
amc: 78.9
"""

log_content_2 = """Another log file
Initial setup
math: 88.2
olympiad_bench: 75.6
minerva: 89.1
aime: 68.3
amc: 82.1
"""

log_content_3 = """Third log file
Running benchmark
math: 92.1
olympiad_bench: 80.2
minerva: 94.5
aime: 71.9
amc: 86.3
"""

# Write test log files
with open(os.path.join(temp_dir, 'results_1.log'), 'w') as f:
    f.write(log_content_1)

with open(os.path.join(temp_dir, 'results_2.log'), 'w') as f:
    f.write(log_content_2)

with open(os.path.join(temp_dir, 'results_3.log'), 'w') as f:
    f.write(log_content_3)

# Also create a non-.log file that should be ignored
with open(os.path.join(temp_dir, 'readme.txt'), 'w') as f:
    f.write("This is not a log file")

ARGS = {
    'directory_path': temp_dir
}
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
get_kl_controller Ran
ran; our contract draft was wrong, not the code
ShenzhiYang2000/OLR/recipe/spin/core_algos.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from LeapLabTHU/OPDVR, is pointer only (licence: NONE) · get_code("757d109d90121142")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from 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.

make_conv_zero Ran
ran; our contract draft was wrong, not the code
ShenzhiYang2000/OLR/eval_scripts/generate_vllm.py
code served (permissive licence) · get_code("ddbc0214f17bae16")
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

torch.manual_seed(0)

# Analyze the function: make_conv_zero takes a 'question' parameter (string)
# and concatenates it with a prompt template, then formats it into a conversation structure.
# The function only uses string operations on 'question'.

ARGS = {
    "question": "What is the derivative of x^2?"
}
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
normalize_final_answer Ran
ran; our contract draft was wrong, not the code
ShenzhiYang2000/OLR/eval_scripts/oat_math_grader.py
code served (permissive licence) · get_code("5662e5ee755b541c")
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

SUBSTITUTIONS = [('an ', ''), ('a ', ''), ('.$', '$'), ('\\$', ''), ('\\ ', ''), (' ', ''), ('mbox', 'text'), (',\\text{and}', ','), ('\\text{and}', ','), ('\\text{m}', '\\text{}')]

REMOVED_EXPRESSIONS = ['square', 'ways', 'integers', 'dollars', 'mph', 'inches', 'ft', 'hours', 'km', 'units', '\\ldots', 'sue', 'points', 'feet', 'minutes', 'digits', 'cents', 'degrees', 'cm', 'gm', 'pounds', 'meters', 'meals', 'edges', 'students', 'childrentickets', 'multiples', '\\text{s}', '\\text{.}', '\\text{\ns}', '\\text{}^2', '\\text{}^3', '\\text{\n}', '\\text{}', '\\mathrm{th}', '^\\circ', '^{\\circ}', '\\;', ',\\!', '{,}', '"', '\\dots']

def normalize_final_answer(final_answer: str) -> str:
    """
    Normalize a final answer to a quantitative reasoning question.
    This code comes from https://arxiv.org/pdf/2206.14858.pdf, page18.
    """
    for before, after in SUBSTITUTIONS:
        final_answer = final_answer.replace(before, after)
    for expr in REMOVED_EXPRESSIONS:
        final_answer = final_answer.replace(expr, '')
    final_answer = re.sub('(.*?)(\\$)(.*?)(\\$)(.*)', '$\\3$', final_answer)
    final_answer = re.sub('(\\\\text\\{)(.*?)(\\})', '\\2', final_answer)
    final_answer = re.sub('(\\\\textbf\\{)(.*?)(\\})', '\\2', final_answer)
    final_answer = re.sub('(\\\\overline\\{)(.*?)(\\})', '\\2', final_answer)
    final_answer = re.sub('(\\\\boxed\\{)(.*)(\\})', '\\2', final_answer)
    final_answer = re.sub('(frac)([^{])(.)', 'frac{\\2}{\\3}', final_answer)
    final_answer = re.sub('(sqrt)([^{])', 'sqrt{\\2}', final_answer)
    final_answer = final_answer.replace('$', '')
    if final_answer.replace(',', '').isdigit():
        final_answer = final_answer.replace(',', '')
    return final_answer

ARGS = {
    'final_answer': '\\boxed{42}'
}
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
repeatness Ran
ran, violated the contract we drafted
ShenzhiYang2000/OLR/eval_scripts/oat_math_grader.py
code served (permissive licence) · get_code("3dcb40e5ce9bac48")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned bool, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

# Analyzing the function
# repeatness() takes a single parameter: s (a string)
# The function computes a "repeatness" metric based on suffix arrays and LCP arrays

# Simple test cases
ARGS = {
    "s": "abcabcabc"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type bool, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
reward_func Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
ShenzhiYang2000/OLR/recipe/r1/reward_score.py
code served (permissive licence) · get_code("8a84f8cc921357e2")
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 Mock. 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 unittest.mock import Mock, patch

torch.manual_seed(0)

# Mock the imported modules to avoid import errors
mock_math_reward = Mock()
mock_math_reward.compute_score = Mock(return_value=0.85)

mock_gpqa = Mock()
mock_gpqa.compute_score = Mock(return_value=0.92)

mock_livecodebench = Mock()
mock_livecodebench.compute_score = Mock(return_value=0.78)

# Patch the imports at module level
import sys
sys.modules['recipe'] = Mock()
sys.modules['recipe.r1'] = Mock()
sys.modules['recipe.r1.tasks'] = Mock()
sys.modules['recipe.r1.tasks.math_reward'] = mock_math_reward
sys.modules['recipe.r1.tasks.gpqa'] = mock_gpqa
sys.modules['recipe.r1.tasks.livecodebench'] = mock_livecodebench

ARGS = {
    'data_source': 'Maxwell-Jia/AIME_2024',
    'solution_str': 'The answer is 42',
    'ground_truth': '42',
    'extra_info': None
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
recipe, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Mock — a Mock 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
timeout Ran
ran; our contract draft was wrong, not the code
ShenzhiYang2000/OLR/eval_scripts/generate_vllm.py
code served (permissive licence) · get_code("885cc9ddb95422d6")
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 os

# The timeout function is a decorator that takes timeout_seconds as a parameter
# It returns a decorator that can be applied to functions
# We need to provide a valid timeout_seconds value

ARGS = {
    'timeout_seconds': 10
}
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 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
mathd_normalize_answer Not yet run
unverified (harvested, no recorded run)
ShenzhiYang2000/OLR/eval_scripts/oat_math_grader.py
code served (permissive licence) · get_code("0388e4fa2f36659e")
How it ran

no invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 distinct functions with an identity, 10 are also present in other papers whose code we hold (10 with at least one paper dated earlier, 10 the same month, 7 later) and 4 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

FunctionAlso present inEarliest-datedOthers named
compute_ce_dpo_loss_rm 46 other papers
26 dated earlier, 4 the same month, 13 later, 3 undated
2409.19256 (2024-09)
HybridFlow: A Flexible and Efficient RLHF Framework
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2504.16084 (2025-04)
TTRL: Test-Time Reinforcement Learning
and 43 more not named here
compute_detach_dpo_loss_rm 44 other papers
25 dated earlier, 3 the same month, 13 later, 3 undated
2409.19256 (2024-09)
HybridFlow: A Flexible and Efficient RLHF Framework
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2504.16084 (2025-04)
TTRL: Test-Time Reinforcement Learning
and 41 more not named here
compute_online_dpo_loss 38 other papers
19 dated earlier, 3 the same month, 13 later, 3 undated
2409.19256 (2024-09)
HybridFlow: A Flexible and Efficient RLHF Framework
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2504.16084 (2025-04)
TTRL: Test-Time Reinforcement Learning
and 35 more not named here
get_kl_controller 38 other papers
19 dated earlier, 3 the same month, 13 later, 3 undated
2409.19256 (2024-09)
HybridFlow: A Flexible and Efficient RLHF Framework
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2504.16084 (2025-04)
TTRL: Test-Time Reinforcement Learning
and 35 more not named here
compute_onlinedpo_pref 30 other papers
13 dated earlier, 2 the same month, 12 later, 3 undated
2409.19256 (2024-09)
HybridFlow: A Flexible and Efficient RLHF Framework
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2504.16084 (2025-04)
TTRL: Test-Time Reinforcement Learning
and 27 more not named here
reward_func 15 other papers
6 dated earlier, 1 the same month, 7 later, 1 undated
2410.02992 (2024-10)
Learning to Better Search with Language Models via Guided Reinforced Self-Training
2601.17687 (2026-01)
Agentic reinforcement learning empowers next-generation chemical language models for molecular design and synthesis
2601.21468 (2026-01)
MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning
and 12 more not named here
compute_score_data_source 13 other papers
5 dated earlier, 1 the same month, 6 later, 1 undated
2601.21468 (2026-01)
MemOCR: Layout-Aware Visual Memory for Efficient Long-Horizon Reasoning
2602.17550 (2026-02)
MASPO: Unifying Gradient Utilization, Probability Mass, and Signal Reliability for Robust and Sample-Efficient LLM Reasoning
2603.02938 (2026-03)
Beyond One-Size-Fits-All: Adaptive Subgraph Denoising for Zero-Shot Graph Learning with Large Language Models
and 10 more not named here
mathd_normalize_answer 12 other papers
10 dated earlier, 2 the same month
2503.10460 (2025-03)
Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond
2505.03335 (2025-05)
Absolute Zero: Reinforced Self-play Reasoning with Zero Data
2505.13379 (2025-05)
Thinkless: LLM Learns When to Think
and 9 more not named here
normalize_final_answer 9 other papers
7 dated earlier, 2 the same month
2504.16084 (2025-04)
TTRL: Test-Time Reinforcement Learning
2509.24372 (2025-09)
Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
2511.13223 (2025-11)
TokenSqueeze: Performance-Preserving Compression for Reasoning LLMs
and 6 more not named here
repeatness 8 other papers
6 dated earlier, 2 the same month
2509.24372 (2025-09)
Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning
2511.13223 (2025-11)
TokenSqueeze: Performance-Preserving Compression for Reasoning LLMs
2511.15694 (2025-11)
The Impact of Quantization on Large Reasoning Model Reinforcement Learning
and 5 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) effectively trains reasoning models that rely on abundant perfect labels, but its vulnerability to unavoidable noisy labels from expert scarcity remains critically underexplored. In this work, we take the first step toward a systematic analysis of noisy label mechanisms in RLVR. In contrast to supervised classification, most RLVR algorithms incorporate a rollout-based condition: a label's influence on training is contingent on whether the current policy can generate rollouts that realize it, a property that naturally extends to noisy labels. Based on this observation, we distinguish two types of noise: inactive noisy labels, which reduce data efficiency, and active noisy labels, which are reinforced and risk skewing the model toward incorrect distributions. From experiments on training with noisy samples, we identify a Early Correctness Coherence phenomenon: although noisy samples begin to lag behind in later stages, accuracy on both clean and noisy samples increases similarly in early training. Motivated by this dynamic, we propose Online Label Refinement (OLR), which progressively corrects potentially noisy labels with majority-voted answers when two conditions hold: a positive slope in the majority answer's rollout pass rate and stable historical consistency across updates, enabling gradual self-correction as the policy improves. We evaluate OLR on six in-distribution mathematical reasoning benchmarks (AIME24/25, AMC, MATH-500, Minerva, and Olympiad) and three out-of-distribution tasks (ARC-c, GPQA-diamond, and MMLU-pro). Across noise ratios from 0.1 to 0.9, OLR consistently improves robustness under both inactive and active noisy-label settings, achieving average gains of 3.6%-3.9% on in-distribution benchmarks and 3.3%-4.6% on out-of-distribution evaluations. The code is available via https://github.com/ShenzhiYang2000/OLR.

For agents

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

get_harvested_code_for_paper("2604.03993")
get_code_for_paper("2604.03993")
have("2604.03993")

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

curl https://syntology.ai/api/ran/2604.03993.json

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

[![Syntology run record](https://syntology.ai/api/ran/2604.03993.svg)](https://syntology.ai/paper/2604.03993)

Connect an agent — have() is free.