SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.03876 · 2026

GOPO: Policy Optimization using Ranked Rewards

Woojeong Kim, Raaz Dwivedi, Kyuseong Choi, Anish Agarwal, Dwaipayan Saha

arXiv · PDF · Open in the Atlas

Code that ran

12 of 12 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; 3 honoured the contract we drafted, 0 violated it and 0 have no contract verdict; 9 failures of our instrument; 0 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2602.03876.json · badge: /api/ran/2602.03876.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
friendshipkim/gopo canonical (classifier label)
classifier basis: census
12 ran · 0 no recorded run
FunctionStatusWhere it lives
clean_judge_name Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/evaluate/bootstrap_judge.py
code served (permissive licence) · get_code("fb46e71f5f7e6565")
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

torch.manual_seed(0)

ARGS = {
    "judge_model": "bert-base.model"
}
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
extract_completion_index_from_filename Ran
ran, honoured the contract we drafted
friendshipkim/gopo/evaluate/plot_win_rates_all_temps.py
code served (permissive licence) · get_code("41466752a95bbd93")
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 extract_completion_index_from_filename(filename: str) -> int:
    """Extract completion index from filename.

    Args:
        filename: Filename to parse

    Returns:
        Completion index (0 if no _ind pattern found, otherwise the number after _ind)
    """
    match = re.search('_ind(\\d+)_', filename)
    if match:
        return int(match.group(1))
    return 0


ARGS = {
    "filename": "model_checkpoint_ind42_best.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
format_model_name Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/evaluate/plot_win_rates_all_temps.py
code served (permissive licence) · get_code("b943e0c67113c782")
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 format_model_name(model: str) -> str:
    """Format model name for display in titles.

    Examples:
        'qwen4b' -> 'Qwen4B'
        'qwen1.7b' -> 'Qwen1.7B'
    """
    if model is None:
        return None
    import re
    match = re.match('(qwen)(\\d+\\.?\\d*)(b)', model, re.IGNORECASE)
    if match:
        return f'Qwen{match.group(2)}B'
    return model


ARGS = {
    'model': 'qwen4b'
}
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
format_prompt Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/generate/generate_completions.py
code served (permissive licence) · get_code("7720ea43a933acbd")
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

class MockTokenizer:
    def apply_chat_template(self, messages, tokenize=False, add_generation_prompt=False, enable_thinking=False):
        """Mock tokenizer that simulates apply_chat_template behavior"""
        prompt = ""
        for msg in messages:
            role = msg['role']
            content = msg['content']
            prompt += f"[{role}]: {content}\n"
        if add_generation_prompt:
            prompt += "[assistant]: "
        return prompt

ARGS = {
    'user_message': 'What is the capital of France?',
    'tokenizer': MockTokenizer(),
    'dataset_type': 'ultrachat'
}
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
format_reward Ran
ran, honoured the contract we drafted
friendshipkim/gopo/src/open_r1/rewards.py
code served (permissive licence) · get_code("ce353d566f3f2c9f")
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 re
import torch
import torch.nn as nn

torch.manual_seed(0)

# The function expects `completions` to be a list of lists of dicts
# Each completion is a list with at least one dict that has a 'content' key
# The content should be a string that may or may not match the pattern

# Create valid completions that match the pattern
valid_completion_1 = [{
    'content': '<think>\nThis is my reasoning\n</think>\n<answer>\nThe answer is 42\n</answer>'
}]

# Create valid completions that don't match the pattern
invalid_completion_1 = [{
    'content': 'This is just plain text without the required tags'
}]

# Create another valid one
valid_completion_2 = [{
    'content': '<think>\nLet me think about this problem\n</think>\n<answer>\nFinal answer\n</answer>'
}]

# Create another invalid one
invalid_completion_2 = [{
    'content': '<think>\nIncomplete structure'
}]

completions = [
    valid_completion_1,
    invalid_completion_1,
    valid_completion_2,
    invalid_completion_2
]

ARGS = {
    'completions': completions
}
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_completion Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/generate/generate_completions.py
code served (permissive licence) · get_code("f7346acce8a48a37")
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
from typing import Any

torch.manual_seed(0)

# Create a minimal tokenizer stand-in
class MinimalTokenizer:
    def __init__(self):
        self.eos_token_id = 2
        self.pad_token_id = 0
    
    def __call__(self, text, return_tensors=None, truncation=False, max_length=None):
        # Return a dict with input_ids and attention_mask
        input_ids = torch.tensor([[1, 2, 3, 4]], dtype=torch.long)
        attention_mask = torch.tensor([[1, 1, 1, 1]], dtype=torch.long)
        return {
            'input_ids': input_ids,
            'attention_mask': attention_mask
        }
    
    def decode(self, token_ids, skip_special_tokens=False):
        return "This is a generated completion."

# Create a minimal model stand-in
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 4)
        self.device_type = 'cpu'
    
    @property
    def device(self):
        return torch.device('cpu')
    
    def generate(self, input_ids, max_new_tokens=None, do_sample=False, 
                 temperature=1.0, top_p=1.0, pad_token_id=None, eos_token_id=None, **kwargs):
        # Return generated token ids: batch_size x (input_length + max_new_tokens)
        batch_size = input_ids.shape[0]
        input_length = input_ids.shape[1]
        new_length = input_length + (max_new_tokens or 10)
        generated = torch.randint(1, 100, (batch_size, new_length), dtype=torch.long)
        return generated

# Create sampling params stand-in for vllm case
class SamplingParams:
    def __init__(self, temperature=0.7, top_p=0.9):
        self.temperature = temperature
        self.top_p = top_p

# Create vllm model stand-in
class VLLMModel:
    def generate(self, prompts, sampling_params):
        class Output:
            def __init__(self):
                self.text = "vllm generated completion"
        
        class Result:
            def __init__(self):
                self.outputs = [Output()]
        
        return [Result()]

# Build arguments
tokenizer = MinimalTokenizer()
model = MinimalModel()
sampling_params = SamplingParams(temperature=0.7, top_p=0.9)

ARGS = {
    'prompt': "Hello, how are you?",
    'model': model,
    'tokenizer': tokenizer,
    'max_new_tokens': 128,
    'use_vllm': False,
    'sampling_params': sampling_params,
    'temperature': 0.7,
    'top_p': 0.9
}
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
merge_completions_artifacts Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/evaluate/bootstrap_judge.py
code served (permissive licence) · get_code("d669e542ee4536d8")
How it ran

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

import torch
import numpy as np
from typing import List, Dict, Any, Tuple

torch.manual_seed(0)

# Construct meta1 and meta2
meta1 = {
    'model_path': 'path/to/model1',
    'temperature': 0.7,
    'max_tokens': 100,
    'dataset': 'test_dataset'
}

meta2 = {
    'model_path': 'path/to/model2',
    'temperature': 0.7,
    'max_tokens': 100,
    'dataset': 'test_dataset',
    'extra_param': 'extra_value'
}

# Construct items1 and items2 - matching prompts required by the function
items1 = [
    {
        'prompt': 'What is the capital of France?',
        'completion': 'Paris is the capital of France.'
    },
    {
        'prompt': 'Explain quantum computing.',
        'completion': 'Quantum computing uses quantum bits for computation.'
    }
]

items2 = [
    {
        'prompt': 'What is the capital of France?',
        'completion': 'The capital of France is Paris.'
    },
    {
        'prompt': 'Explain quantum computing.',
        'completion': 'Quantum computers use quantum mechanics principles.'
    }
]

ARGS = {
    'meta1': meta1,
    'items1': items1,
    'meta2': meta2,
    'items2': items2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — 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
nanstd Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/src/open_r1/grpo_trainer.py
code served (permissive licence) · get_code("a6cece2ffebe0f9c")
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) and returned Tensor, 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

torch.manual_seed(0)

# Create a 1D tensor with some NaN values
tensor = torch.tensor([1.0, 2.0, float('nan'), 4.0, 5.0, float('nan'), 7.0])

ARGS = {
    "tensor": tensor
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 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
read_single_completions_artifact Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/evaluate/bootstrap_judge.py
code served (permissive licence) · get_code("bdf1fbcb4af9dac0")
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 json
import tempfile
import os

# Create temporary test files
temp_dir = tempfile.mkdtemp()

# Create a test JSON file with multiple completions
json_file = os.path.join(temp_dir, "test_completions.json")
json_data = {
    "meta": {"model": "test-model", "version": "1.0"},
    "items": [
        {
            "prompt": "What is 2+2?",
            "completions": ["4", "Four", "The answer is 4"]
        },
        {
            "prompt": "What is the capital of France?",
            "responses": ["Paris", "Paris, France", "The capital is Paris"]
        },
        {
            "prompt": "Say hello",
            "completion": "Hello there!"
        }
    ]
}

with open(json_file, 'w') as f:
    json.dump(json_data, f)

# Create a test JSONL file
jsonl_file = os.path.join(temp_dir, "test_completions.jsonl")
jsonl_items = [
    {"prompt": "First question", "completions": ["answer1", "answer2"]},
    {"prompt": "Second question", "responses": ["resp1", "resp2"]},
    {"prompt": "Third question", "completion": "single answer"}
]

with open(jsonl_file, 'w') as f:
    for item in jsonl_items:
        f.write(json.dumps(item) + "\n")

# Define ARGS with valid parameter values
ARGS = {
    "path": json_file,
    "completion_index": 0
}
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 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
shuffle_tensor_dict Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/src/open_r1/grpo_trainer.py
code served (permissive licence) · get_code("4f7f0f20f97f300e")
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) 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

torch.manual_seed(0)

# Construct a dictionary of tensors with different shapes
# The function shuffles along the first dimension (batch dimension)
tensor_dict = {
    "x": torch.arange(8).reshape(2, 4),  # batch_size=2, features=4
    "y": torch.arange(6).reshape(2, 3),  # batch_size=2, features=3
    "z": None,  # Test that None values are handled
}

ARGS = {
    "tensor_dict": tensor_dict,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
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
split_tensor_dict Ran
ran; our contract draft was wrong, not the code
friendshipkim/gopo/src/open_r1/grpo_trainer.py
code served (permissive licence) · get_code("efd67e2adf6b1ba5")
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) 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

torch.manual_seed(0)

# Create small tensors with distinct dimensions
x = torch.arange(8).reshape(4, 2)  # 4 samples, 2 features
y = torch.arange(4).reshape(4, 1)  # 4 samples, 1 feature
z = None  # Test optional tensor

tensor_dict = {
    "x": x,
    "y": y,
    "z": z
}

num_chunks = 2

ARGS = {
    "tensor_dict": tensor_dict,
    "num_chunks": num_chunks
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
tag_count_reward Ran
ran, honoured the contract we drafted
friendshipkim/gopo/src/open_r1/rewards.py
code served (permissive licence) · get_code("0a5673b11c91e69d")
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

torch.manual_seed(0)

ARGS = {
    "completions": [
        [{"content": "Let me think about this.\n<think>\nSome reasoning here\n</think>\nNow I have an answer.\n<answer>\nThe answer is 42\n</answer>"}],
        [{"content": "Another response\n<think>\nMore reasoning\n</think>\nFinal answer\n<answer>\nAnother answer\n</answer>"}],
    ]
}
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

Also present in other papers' code

Of 12 samples lifted from this paper's linked repositories, 12 carry a function identity and 0 do not. Of the 12 distinct functions with an identity, 5 are also present in other papers whose code we hold (5 with at least one paper dated earlier, 4 the same month, 5 later) and 7 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
nanstd 7 other papers
5 dated earlier, 1 the same month, 1 later
2402.16302 (2024-02)
Graph Diffusion Policy Optimization
2505.17534 (2025-05)
Co-Reinforcement Learning for Unified Multimodal Understanding and Generation
2506.14755 (2025-06)
Optimizing Length Compression in Large Reasoning Models
and 4 more not named here
split_tensor_dict 7 other papers
5 dated earlier, 1 the same month, 1 later
2402.16302 (2024-02)
Graph Diffusion Policy Optimization
2505.17534 (2025-05)
Co-Reinforcement Learning for Unified Multimodal Understanding and Generation
2506.14755 (2025-06)
Optimizing Length Compression in Large Reasoning Models
and 4 more not named here
format_reward 5 other papers
3 dated earlier, 1 the same month, 1 later
2506.20639 (2025-06)
DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation
2507.02834 (2025-07)
ExPO: Unlocking Hard Reasoning with Self-Explanation-Guided Reinforcement Learning
2511.04902 (2025-11)
You Need Reasoning to Learn Reasoning: The Limitations of Label-Free RL in Weak Base Models
and 2 more not named here
tag_count_reward 4 other papers
2 dated earlier, 1 the same month, 1 later
2506.20639 (2025-06)
DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation
2507.02834 (2025-07)
ExPO: Unlocking Hard Reasoning with Self-Explanation-Guided Reinforcement Learning
2602.02258 (2026-02)
Alignment-Aware Model Adaptation via Feedback-Guided Optimization
and 1 more not named here
shuffle_tensor_dict 3 other papers
2 dated earlier, 1 later
2402.16302 (2024-02)
Graph Diffusion Policy Optimization
2505.17534 (2025-05)
Co-Reinforcement Learning for Unified Multimodal Understanding and Generation
2605.11974 (2026-05)
Towards Order Fairness: Mitigating LLMs Order Sensitivity through Dual Group Advantage Optimization

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

Standard reinforcement learning from human feedback (RLHF) trains a reward model on pairwise preference data and then uses it for policy optimization. However, while reward models are optimized to capture relative preferences, existing policy optimization techniques rely on absolute reward magnitudes during training. In settings where the rewards are non-verifiable-such as summarization, instruction following, and chat completion-this misalignment often leads to suboptimal performance. We introduce Group Ordinal Policy Optimization (GOPO), a policy optimization method that uses only the ranking of the rewards and discards their magnitudes. Our rank-based transformation of rewards provides several gains, compared to Group Relative Policy Optimization (GRPO), in settings with nonverifiable rewards: (1) consistently higher training/validation reward trajectories, (2) improved LLM-as-judge evaluations across most intermediate training steps, and (3) reaching a policy of comparable quality in substantially less training steps than GRPO. We demonstrate consistent improvements across a range of tasks and model sizes. Our source code is available at https: //github.com/friendshipkim/gopo.

For agents

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

get_harvested_code_for_paper("2602.03876")
get_code_for_paper("2602.03876")
have("2602.03876")

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

curl https://syntology.ai/api/ran/2602.03876.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.03876.svg)](https://syntology.ai/paper/2602.03876)

Connect an agent — have() is free.