SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2512.23461 · ICLR · 2025

Eliminating Inductive Bias in Reward Models with Information-Theoretic Guidance

Pengyu Cheng, Tsung-Hui Chang, Zhuo Li, Xiang Wan, Xiaoxi Jiang, Guanjun Jiang, Zhechao Yu, Feifei Tong, Anningzhe Gao, Erchao Zhao

arXiv · PDF · Open in the Atlas

Code that ran

10 of 17 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 1 violated it and 1 has no contract verdict; 8 failures of our instrument; 7 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/2512.23461.json · badge: /api/ran/2512.23461.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
Qwen-Applications/DIR canonical (classifier label)
classifier basis: census
10 ran · 7 no recorded run
FunctionStatusWhere it lives
build_dataset_SK Ran
ran on a synthesized input
Qwen-Applications/DIR/reward_models/load_datasets.py
code served (permissive licence) · get_code("d5c8de0d3609ddc2")
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 Dataset. 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 unittest.mock import Mock, MagicMock

torch.manual_seed(0)

# Create a mock tokenizer
class MockTokenizer:
    def __init__(self):
        self.max_length = 128
    
    def apply_chat_template(self, messages, tokenize=False):
        return "mock prompt response text"
    
    def encode_plus(self, text, add_special_tokens=False, padding=True, truncation=True, max_length=128, return_tensors='pt'):
        input_ids = torch.tensor([[101, 2054, 2003, 102]], dtype=torch.long)
        attention_mask = torch.tensor([[1, 1, 1, 1]], dtype=torch.long)
        return {
            'input_ids': input_ids,
            'attention_mask': attention_mask
        }

# Create a mock dataset
class MockDataset:
    def __init__(self):
        self.data = [
            {
                'chosen': [
                    {'content': 'What is AI?'},
                    {'content': 'AI is artificial intelligence'}
                ],
                'rejected': [
                    {'content': 'What is AI?'},
                    {'content': 'AI is not real'}
                ]
            },
            {
                'chosen': [
                    {'content': 'What is ML?'},
                    {'content': 'ML is machine learning'}
                ],
                'rejected': [
                    {'content': 'What is ML?'},
                    {'content': 'ML is something else'}
                ]
            }
        ]
    
    def select(self, indices):
        new_ds = MockDataset()
        new_ds.data = [self.data[i] for i in indices if i < len(self.data)]
        return new_ds
    
    def map(self, func, batched=False, num_proc=1):
        new_ds = MockDataset()
        new_ds.data = [func(item) for item in self.data]
        return new_ds
    
    def add_column(self, name, column):
        new_ds = MockDataset()
        new_ds.data = [{**item, name: col_val} for item, col_val in zip(self.data, column)]
        return new_ds
    
    def set_format(self, type='torch'):
        return self
    
    def __len__(self):
        return len(self.data)

# Mock load_dataset to return our mock dataset
def mock_load_dataset(path, split=None):
    return MockDataset()

# Patch load_dataset
import datasets
datasets.load_dataset = mock_load_dataset

ARGS = {
    'data_path': 'mock/dataset/path',
    'tokenizer': MockTokenizer(),
    'split': 'train',
    'size': 2
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Dataset — a Dataset output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Dataset output the recorded ndim 1 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
build_eval_dataset_biasbench Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/DIR/rm_eval/eval_biasbench.py
code served (permissive licence) · get_code("7b30b37de5cc56e7")
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 Dataset. 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 json
import tempfile
import os
from datasets import Dataset

torch.manual_seed(0)

# Create a minimal tokenizer stand-in
class MinimalTokenizer:
    def __init__(self):
        self.max_length = 64
        self.vocab_size = 1000
        
    def apply_chat_template(self, messages, tokenize=False):
        # Simple concatenation of messages for chat template
        result = ""
        for msg in messages:
            result += f"{msg['role']}: {msg['content']} "
        return result
    
    def encode_plus(self, text, add_special_tokens=False, padding=True, truncation=True, max_length=64, return_tensors='pt'):
        # Generate dummy token IDs and attention mask
        token_ids = torch.randint(0, self.vocab_size, (1, 32))
        attention_mask = torch.ones((1, 32))
        return {
            'input_ids': token_ids,
            'attention_mask': attention_mask
        }

# Create a temporary JSON file with nested structure
sample_data = {
    "domain1": [
        {
            "bias": "gender",
            "instruction": "What is your opinion on work?",
            "response1": "Everyone should work hard.",
            "response2": "Work-life balance is important.",
            "label": 0
        },
        {
            "bias": "age",
            "instruction": "What about retirement?",
            "response1": "Retirement is deserved rest.",
            "response2": "Retirement should be flexible.",
            "label": 1
        }
    ],
    "domain2": [
        {
            "bias": "race",
            "instruction": "Tell me about diversity.",
            "response1": "Diversity strengthens society.",
            "response2": "All backgrounds are valuable.",
            "label": 0
        }
    ]
}

# Write to temporary file
temp_fd, temp_path = tempfile.mkstemp(suffix='.json')
try:
    with os.fdopen(temp_fd, 'w') as f:
        json.dump(sample_data, f)
    
    tokenizer = MinimalTokenizer()
    
    ARGS = {
        'dataset': temp_path,
        'tokenizer': tokenizer,
        'keep_columns': ['bias', 'instruction', 'response1', 'response2', 'label', 'input_ids_resp1', 'attention_mask_resp1', 'input_ids_resp2', 'attention_mask_resp2']
    }
finally:
    pass
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Dataset — a Dataset output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Dataset output the recorded ndim 1 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
calculate_accuracy_by_domain Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/DIR/rm_eval/eval_biasbench.py
code served (permissive licence) · get_code("4014c6ad9f997c4a")
How it ran

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

import numpy as np

# Create sample data for the function
predictions = np.array([0, 1, 1, 0, 1, 0])  # Predicted labels
labels = np.array([0, 1, 0, 0, 1, 1])       # True labels
domains = np.array([0, 0, 1, 1, 2, 2])      # Domain identifiers

ARGS = {
    'predictions': predictions,
    'labels': labels,
    'domains': domains
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
compute_metrics Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/DIR/reward_models/utils.py
code served (permissive licence) · get_code("7710f1274273fc6c")
How it ran

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

import numpy as np
from sklearn.metrics import accuracy_score

class EvalPrediction:
    def __init__(self, predictions, labels):
        self.predictions = predictions
        self.labels = labels

# Create predictions with shape (batch_size, num_classes)
# Using batch size 2 and 4 classes
predictions = np.array([
    [0.1, 0.2, 0.6, 0.1],
    [0.3, 0.5, 0.1, 0.1]
], dtype=np.float32)

# Create labels (not used in the function body, but part of eval_pred structure)
labels = np.array([2, 1])

eval_pred = EvalPrediction(predictions=predictions, labels=labels)

ARGS = {
    'eval_pred': eval_pred
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn — 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
is_lora_model Ran
ran, violated the contract we drafted
Qwen-Applications/DIR/reward_models/utils.py
code served (permissive licence) · get_code("f8e7d855924e7e64")
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.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create a model with lora in state_dict keys
class LoRAModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 3)
        self.lora_a = nn.Parameter(torch.randn(4, 2))
        self.lora_b = nn.Parameter(torch.randn(2, 3))
    
    def forward(self, x):
        return self.linear(x)

model = LoRAModel()

ARGS = {
    "model": 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 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
load_responses Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/DIR/rm_eval/arean_hard/arean_evaluation.py
code served (permissive licence) · get_code("acb09d50429275fd")
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 pandas as pd
import json

# Create a temporary JSON Lines file with test data
temp_dir = tempfile.mkdtemp()
test_filepath = os.path.join(temp_dir, "responses.jsonl")

# Write test data to the file
test_data = [
    {"uid": "user_1", "response": "Hello", "score": 0.9},
    {"uid": "user_2", "response": "World", "score": 0.8},
    {"uid": "user_3", "response": "Test", "score": 0.7},
]

with open(test_filepath, "w") as f:
    for record in test_data:
        f.write(json.dumps(record) + "\n")

ARGS = {
    "filepath": test_filepath
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas — 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
parse_judgment Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/DIR/rm_eval/arean_hard/arean_evaluation.py
code served (permissive licence) · get_code("be08a58ad66166a5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple, ndim 1, dtype family 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
from typing import Dict, List, Tuple

def parse_judgment(judgment_text: str) -> Tuple[str, str]:
    pattern = '\\[\\[(A>>B|A>B|A=B|B>A|B>>A)\\]\\]'
    match = re.search(pattern, judgment_text)
    if not match:
        return ('parse_error', 'Could not find a valid verdict pattern.')
    verdict_str = match.group(1)
    if verdict_str in ['A>>B', 'A>B']:
        return ('win_a', verdict_str)
    elif verdict_str in ['B>>A', 'B>A']:
        return ('win_b', verdict_str)
    elif verdict_str == 'A=B':
        return ('tie', verdict_str)
    return ('parse_error', f'Found an unexpected pattern: {verdict_str}')


# Test case 1: A wins with strong preference
test_judgment_1 = "The model A performs significantly better. [[A>>B]] This is clear."

# Test case 2: A wins with weak preference
test_judgment_2 = "Model A is slightly better. [[A>B]] Overall A is preferred."

# Test case 3: Tie
test_judgment_3 = "Both models perform equally well. [[A=B]] No clear winner."

# Test case 4: B wins with strong preference
test_judgment_4 = "B demonstrates superior performance. [[B>>A]] B is the clear winner."

# Test case 5: B wins with weak preference
test_judgment_5 = "B is marginally better. [[B>A]] B edges out A."

ARGS = {
    'judgment_text': test_judgment_1
}
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, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
post_process_HS3 Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/DIR/reward_models/debias_trainer_sycophancy.py
code served (permissive licence) · get_code("17922d5baec5883a")
How it ran

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

import numpy as np

# Create sample data for the function
np.random.seed(0)

# Create reward arrays - rewards for chosen (j) and rejected (k) responses
rewards_j = np.random.randn(10).astype(np.float32)
rewards_k = np.random.randn(10).astype(np.float32)

# Create bias labels: 0 for natural, 1 for adversarial
bias_labels = np.array([0, 0, 0, 0, 0, 1, 1, 1, 1, 1], dtype=np.int64)

# Create metric key prefix
metric_key_prefix = "test"

ARGS = {
    "rewards_j": rewards_j,
    "rewards_k": rewards_k,
    "metric_key_prefix": metric_key_prefix,
    "bias_labels": bias_labels,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
post_process_RBv1 Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/DIR/reward_models/debias_trainer_length.py
code served (permissive licence) · get_code("97f43d7c928e992e")
How it ran

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

import numpy as np

# Construct arguments for post_process_RBv1
# The function expects:
# - rewards_j: array of rewards (comparison winners)
# - rewards_k: array of rewards (comparison losers)
# - metric_key_prefix: string prefix for metric keys

# Create small numpy arrays with distinct values to make results observable
np.random.seed(0)
rewards_j = np.array([0.8, 0.9, 0.7, 0.85], dtype=np.float32)
rewards_k = np.array([0.6, 0.5, 0.7, 0.4], dtype=np.float32)
metric_key_prefix = "test"

ARGS = {
    "rewards_j": rewards_j,
    "rewards_k": rewards_k,
    "metric_key_prefix": metric_key_prefix
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
post_process_RMBench Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
Qwen-Applications/DIR/reward_models/debias_trainer_length.py
code served (permissive licence) · get_code("66d572c3fb9e65bc")
How it ran

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

import numpy as np

# Analyze the function to understand what it needs:
# 1. expanded_domains: List[str] - list of domain names with prefixes
# 2. domains_to_analyze: List[str] - list of domain names to filter
# 3. scores: np.ndarray - 2D array where each row has scores, shape must be divisible by MATRIX_SIZE=3
# 4. metric_key_prefix: str - prefix for metric keys

# The function:
# - Converts expanded_domains to np.array and uses np.char.startswith for masking
# - Reshapes scores into (num_samples, 3, 2) format, so scores must have shape (N*3, 2) where N*3 is divisible by 3
# - Expects domain_results to have keys: 'chat', 'math', 'code', 'safety'
# - Each domain in domain_results should have 'hard_acc', 'normal_acc', 'easy_acc'

# Create domains to analyze
domains_to_analyze = ['chat', 'math', 'code', 'safety']

# Create expanded_domains that match the filter pattern
# Each domain prefix should appear multiple times to get enough samples
# We need at least 3 rows per domain (since MATRIX_SIZE=3), and ideally multiple samples
# Let's create 9 rows per domain (3 samples × 3 rows each)
expanded_domains = [
    'chat_sample_0', 'chat_sample_0', 'chat_sample_0',  # 3 rows for 1st sample
    'chat_sample_1', 'chat_sample_1', 'chat_sample_1',  # 3 rows for 2nd sample
    'math_sample_0', 'math_sample_0', 'math_sample_0',  # 3 rows for 1st sample
    'math_sample_1', 'math_sample_1', 'math_sample_1',  # 3 rows for 2nd sample
    'code_sample_0', 'code_sample_0', 'code_sample_0',  # 3 rows for 1st sample
    'code_sample_1', 'code_sample_1', 'code_sample_1',  # 3 rows for 2nd sample
    'safety_sample_0', 'safety_sample_0', 'safety_sample_0',  # 3 rows for 1st sample
    'safety_sample_1', 'safety_sample_1', 'safety_sample_1',  # 3 rows for 2nd sample
]

# Create scores array
# 24 rows (8 domains × 3 rows each), 2 columns (chosen and rejected scores)
# Each row is [chosen_score, rejected_score]
np.random.seed(0)
scores = np.random.rand(24, 2)

# Make sure chosen scores are generally higher than rejected (for reasonable accuracy)
scores[:, 0] = scores[:, 0] * 0.6 + 0.4  # chosen scores in [0.4, 1.0]
scores[:, 1] = scores[:, 1] * 0.4        # rejected scores in [0.0, 0.4]

metric_key_prefix = 'eval'

ARGS = {
    'expanded_domains': expanded_domains,
    'domains_to_analyze': domains_to_analyze,
    'scores': scores,
    'metric_key_prefix': metric_key_prefix
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
build_dataset Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/DIR/reward_models/load_datasets.py
code served (permissive licence) · get_code("47ee1e29f8388bb2")
How it ran

no invocation recorded.

build_dataset_UF Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/DIR/reward_models/load_datasets.py
code served (permissive licence) · get_code("c7dd18f13397a20b")
How it ran

no invocation recorded.

build_model Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/DIR/rm_eval/eval_biasbench.py
code served (permissive licence) · get_code("629048a34404782d")
How it ran

no invocation recorded.

build_ood_eval_dataset Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/DIR/rm_eval/load_eval_datasets.py
code served (permissive licence) · get_code("33044ac6954b842b")
How it ran

no invocation recorded.

build_unified_eval_dataset Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/DIR/rm_eval/load_eval_datasets.py
code served (permissive licence) · get_code("4075415e3c419aef")
How it ran

no invocation recorded.

get_trainable_weights Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/DIR/reward_models/utils.py
code served (permissive licence) · get_code("aed395f45af87c9c")
How it ran

no invocation recorded.

load_eval_dataset Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/DIR/rm_eval/load_eval_datasets.py
code served (permissive licence) · get_code("6d8002bad0675e8b")
How it ran

no invocation recorded.

Also present in other papers' code

Of 17 samples lifted from this paper's linked repositories, 17 carry a function identity and 0 do not. Of the 17 distinct functions with an identity, 8 are also present in other papers whose code we hold (8 with at least one paper dated earlier, 0 the same month, 0 later) and 9 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
get_trainable_weights 2 other papers
2 dated earlier
2406.10216 (2024-06)
Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs
2510.10963 (2025-10)
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport
is_lora_model 2 other papers
2 dated earlier
2406.10216 (2024-06)
Regularizing Hidden States Enables Learning Generalizable Reward Model for LLMs
2510.10963 (2025-10)
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport
build_dataset 1 other paper
1 dated earlier
2510.10963 (2025-10)
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport
build_dataset_SK 1 other paper
1 dated earlier
2510.10963 (2025-10)
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport
build_dataset_UF 1 other paper
1 dated earlier
2510.10963 (2025-10)
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport
compute_metrics 1 other paper
1 dated earlier
2510.10963 (2025-10)
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport
post_process_RBv1 1 other paper
1 dated earlier
2510.10963 (2025-10)
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport
post_process_RMBench 1 other paper
1 dated earlier
2510.10963 (2025-10)
APLOT: Robust Reward Modeling via Adaptive Preference Learning with Optimal Transport

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

Reward models (RMs) are essential in reinforcement learning from human feedback (RLHF) to align large language models (LLMs) with human values. However, RM training data is commonly recognized as low-quality, containing inductive biases that can easily lead to overfitting and reward hacking. For example, more detailed and comprehensive responses are usually human-preferred but with more words, leading response length to become one of the inevitable inductive biases. A limited number of prior RM debiasing approaches either target a single specific type of bias or model the problem with only simple linear correlations, e.g., Pearson coefficients. To mitigate more complex and diverse inductive biases in reward modeling, we introduce a novel information-theoretic debiasing method called Debiasing via Information optimization for RM (DIR). Inspired by the information bottleneck (IB), we maximize the mutual information (MI) between RM scores and human preference pairs, while minimizing the MI between RM outputs and biased attributes of preference inputs. With theoretical justification from information theory, DIR can handle more sophisticated types of biases with non-linear correlations, broadly extending the real-world application scenarios for RM debiasing methods. In experiments, we verify the effectiveness of DIR with three types of inductive biases: response length, sycophancy, and format. We discover that DIR not only effectively mitigates target inductive biases but also enhances RLHF performance across diverse benchmarks, yielding better generalization abilities. The code and training recipes are available at https://github.com/Qwen-Applications/DIR.

For agents

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

get_harvested_code_for_paper("2512.23461")
get_code_for_paper("2512.23461")
have("2512.23461")

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

curl https://syntology.ai/api/ran/2512.23461.json

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

[![Syntology run record](https://syntology.ai/api/ran/2512.23461.svg)](https://syntology.ai/paper/2512.23461)

Connect an agent — have() is free.