SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.02495 · ICML · 2026

Reward-free Alignment for Conflicting Objectives

Tianyi Lin, Xi Chen, Peter Chen, Xiaopeng Li

arXiv · PDF · Open in the Atlas

Code that ran

13 of 16 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; 10 failures of our instrument; 3 no recorded run), measured 2026-11-21; 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/2602.02495.json · badge: /api/ran/2602.02495.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
PeterLauLukChen/RACO canonical (classifier label)
classifier basis: census
13 ran · 3 no recorded run
FunctionStatusWhere it lives
beaver_conversation Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/script/score_beaver_reward_cost.py
code served (permissive licence) · get_code("2557e245c7e798c3")
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

ARGS = {
    'user': 'Hello, how are you?',
    'assistant': 'I am doing well, thank you for asking!'
}
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
build_text Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/script/score_reward_summarization.py
code served (permissive licence) · get_code("8a99d20ff14e3711")
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
from dataclasses import dataclass
from typing import Dict, Any

@dataclass
class Row:
    input: str
    output: str
    raw: Dict[str, Any]

# Create test Row instance
test_row = Row(
    input='\n\nSubreddit: AskReddit\n\nSome user posted a question about life advice. Please directly output your summarization.',
    output='Life advice question summary',
    raw={'source': 'reddit', 'id': '123'}
)

# Define ARGS dict with all parameters for build_text
ARGS = {
    'row': test_row,
    'mode': 'input_output',
    'bos_token': '<bos>'
}
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
compute_variance Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/script/calc_raco_metric_stats.py
code served (permissive licence) · get_code("9ebf3925d706325e")
How it ran

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

import torch
import numpy as np

torch.manual_seed(0)

# The function takes:
# - xs: list[float] - a list of floating point numbers
# - sample: bool - a keyword-only argument controlling whether to use sample variance (n-1) or population variance (n)

# Create a small list of floats for testing
xs = [1.5, 2.3, 3.7, 4.2, 5.1]

# Create the ARGS dictionary with all required parameters
ARGS = {
    "xs": xs,
    "sample": True,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
discover_inputs Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/script/calc_raco_metric_stats.py
code served (permissive licence) · get_code("681958bea02af0cf")
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 argparse
from pathlib import Path
import tempfile
import os

# Create a temporary directory structure for testing
tmpdir = tempfile.mkdtemp()
log_dir = Path(tmpdir) / "logs"
log_dir.mkdir(parents=True, exist_ok=True)

# Create some dummy log files matching a glob pattern
(log_dir / "output_001.txt").touch()
(log_dir / "output_002.txt").touch()
(log_dir / "config_001.json").touch()

# Create input files
input_dir = Path(tmpdir) / "inputs"
input_dir.mkdir(parents=True, exist_ok=True)
input_file_1 = input_dir / "data_1.txt"
input_file_2 = input_dir / "data_2.txt"
input_file_1.touch()
input_file_2.touch()

# Construct the argparse.Namespace with all required parameters
args_namespace = argparse.Namespace(
    input=[str(input_file_1), str(input_file_2)],
    log_dir=log_dir,
    glob="*.txt"
)

ARGS = {
    "args": args_namespace
}
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 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
extract_prompt Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/script/build_summary_eval_prompts.py
code served (permissive licence) · get_code("b7256244e37b6cb1")
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)

# Construct valid arguments for extract_prompt
# The function expects a dict with 'chosen' and 'rejected' keys
# Both should be lists of message dicts with 'role' and 'content' keys
# The function finds common prompt messages and extracts the last user text

ARGS = {
    'obj': {
        'chosen': [
            {'role': 'system', 'content': 'You are a helpful assistant.'},
            {'role': 'user', 'content': 'What is 2+2?'},
            {'role': 'assistant', 'content': '2+2 equals 4.'}
        ],
        'rejected': [
            {'role': 'system', 'content': 'You are a helpful assistant.'},
            {'role': 'user', 'content': 'What is 2+2?'},
            {'role': 'assistant', 'content': 'The answer is 4.'}
        ]
    }
}
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
main Ran
ran, honoured the contract we drafted
PeterLauLukChen/RACO/eval/new_score_beaver.py
code served (permissive licence) · get_code("78304958ed3d9382")
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; no output shape is stated. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is 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 argparse
import json
import math
import re
import shlex
import subprocess
import sys
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Dict, List, Optional, Tuple
import tempfile
import os

# Create temporary directories and files for the test
temp_dir = tempfile.mkdtemp()
reward_model_dir = os.path.join(temp_dir, "reward_model")
cost_model_dir = os.path.join(temp_dir, "cost_model")
generated_jsonl = os.path.join(temp_dir, "generated.jsonl")
scored_jsonl = os.path.join(temp_dir, "scored.jsonl")
metrics_out = os.path.join(temp_dir, "metrics.json")

# Create the model directories
os.makedirs(reward_model_dir, exist_ok=True)
os.makedirs(cost_model_dir, exist_ok=True)

# Create a sample generated JSONL file
with open(generated_jsonl, 'w') as f:
    f.write(json.dumps({"id": 1, "text": "sample text"}) + '\n')
    f.write(json.dumps({"id": 2, "text": "another sample"}) + '\n')

# Create a dummy vllm_generate.py script
vllm_generate_py = os.path.join(temp_dir, "vllm_generate.py")
with open(vllm_generate_py, 'w') as f:
    f.write("#!/usr/bin/env python\nprint('Generated')\n")

# Create a dummy scorer_py script
scorer_py = os.path.join(temp_dir, "score_beaver_reward_cost.py")
with open(scorer_py, 'w') as f:
    f.write("#!/usr/bin/env python\nimport sys\nprint('rows 2')\nprint('beaver_reward_mean 0.5')\nprint('beaver_cost_mean 0.1')\n")

ARGS = {
    'argv': [
        '--vllm_generate_py', vllm_generate_py,
        '--scorer_py', scorer_py,
        '--skip_generate',
        '--generated_jsonl', generated_jsonl,
        '--reward_model_dir', reward_model_dir,
        '--cost_model_dir', cost_model_dir,
        '--score_batch_size', '8',
        '--score_max_length', '1024',
        '--score_device', 'cpu',
        '--scored_jsonl', scored_jsonl,
        '--metrics_out', metrics_out,
    ]
}
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 — no output shape is stated: observe_signature.py could not observe it (call:NameError)
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
make_async_client Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/eval/vllm_generate.py
code served (permissive licence) · get_code("6c574acf7317667e")
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 AsyncOpenAI. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# The function make_async_client takes host, port, and api_key
# and constructs an AsyncOpenAI client with a custom base_url.
# It does not actually use these parameters beyond constructing the URL and client,
# so we just need valid string/int values.

ARGS = {
    "host": "localhost",
    "port": 8000,
    "api_key": "sk-test-key-12345"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, openai, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type AsyncOpenAI — a AsyncOpenAI output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a AsyncOpenAI 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
make_client Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/eval/vllm_generate.py
code served (permissive licence) · get_code("fda35ae84a267de1")
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 OpenAI. 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

ARGS = {
    "host": "localhost",
    "port": 8000,
    "api_key": "test-api-key-12345"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, openai, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type OpenAI — a OpenAI output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a OpenAI 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_logged_dict Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/script/calc_raco_metric_stats.py
code served (permissive licence) · get_code("792996f18741b2bb")
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 ast
import math
from typing import Any

SAFE_EVAL_NAMES = {'nan': math.nan, 'inf': math.inf, 'Infinity': math.inf}

def parse_logged_dict(line: str) -> dict[str, Any]:
    try:
        payload = ast.literal_eval(line)
    except (ValueError, SyntaxError):
        payload = eval(line, {'__builtins__': {}}, SAFE_EVAL_NAMES)
    if not isinstance(payload, dict):
        raise ValueError(f'Expected dict payload, got {type(payload).__name__}')
    return payload

# Test cases with valid dict strings
test_cases = [
    "{'key': 'value', 'number': 42}",
    "{'a': 1, 'b': 2.5, 'c': [1, 2, 3]}",
    "{'x': nan, 'y': inf}",
    "{}",
]

ARGS = {
    'line': "{'key': 'value', 'number': 42, 'nested': {'inner': True}}"
}
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
score_pairs Ran
ran, honoured the contract we drafted
PeterLauLukChen/RACO/script/score_faithful_summarization.py
code served (permissive licence) · get_code("69c7ec23046c9eff")
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 2, dtype family float; no output shape is stated. The image it ran in was not recorded. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Minimal tokenizer stand-in
class MinimalTokenizer:
    def __call__(self, text, text_pair, padding, truncation, max_length, return_tensors):
        # Simulate tokenizer output: returns dict with 'input_ids' and 'attention_mask'
        batch_size = len(text)
        seq_len = 5
        return {
            'input_ids': torch.randint(0, 1000, (batch_size, seq_len)),
            'attention_mask': torch.ones(batch_size, seq_len, dtype=torch.long)
        }

# Minimal model that returns logits with shape (batch_size, 2)
class MinimalScoringModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(768, 2)
    
    def forward(self, input_ids, attention_mask):
        # Create dummy hidden states
        batch_size = input_ids.shape[0]
        hidden = torch.randn(batch_size, 768)
        logits = self.linear(hidden)
        
        # Return object with logits attribute
        class Output:
            pass
        out = Output()
        out.logits = logits
        return out

# Build arguments
tokenizer = MinimalTokenizer()
model = MinimalScoringModel()
device = torch.device('cpu')

ARGS = {
    'summaries': ['This is a summary one.', 'This is a summary two.'],
    'passages': ['This is a passage one.', 'This is a passage two.'],
    'tokenizer': tokenizer,
    'model': model,
    'device': device,
    'max_length': 512,
    'fp16': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 2, 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
score_texts Ran
ran, honoured the contract we drafted
PeterLauLukChen/RACO/script/score_reward_summarization.py
code served (permissive licence) · get_code("b7550e6553b4157c")
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. The code mentions cuda, and the harness aliases cuda to the CPU for every sample; whether this run needed that is not recorded. If your copy returns the same type and ndim on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Stand-in tokenizer class
class StandInTokenizer:
    def __call__(self, texts, padding=True, truncation=True, max_length=None, return_tensors=None):
        # Return a dict with 'input_ids' and 'attention_mask' keys
        batch_size = len(texts)
        # Create dummy input_ids and attention_mask tensors
        input_ids = torch.randint(0, 1000, (batch_size, max_length or 10))
        attention_mask = torch.ones(batch_size, max_length or 10, dtype=torch.long)
        return {
            'input_ids': input_ids,
            'attention_mask': attention_mask
        }

# Stand-in model class
class StandInModel(nn.Module):
    def __init__(self, output_dim=2):
        super().__init__()
        self.linear = nn.Linear(10, output_dim)
    
    def forward(self, input_ids, attention_mask):
        # Return an object with a 'logits' attribute
        # input_ids shape: (batch_size, seq_len)
        batch_size = input_ids.shape[0]
        logits = self.linear(torch.randn(batch_size, 10))
        
        class Output:
            pass
        
        output = Output()
        output.logits = logits
        return output

# Create instances
tokenizer = StandInTokenizer()
model = StandInModel(output_dim=3)
device = torch.device('cpu')
texts = ["hello world", "foo bar"]
max_length = 10
fp16 = False

ARGS = {
    'texts': texts,
    'tokenizer': tokenizer,
    'model': model,
    'device': device,
    'max_length': max_length,
    'fp16': fp16
}
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
strip_system_prompt Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/script/score_faithful_summarization.py
code served (permissive licence) · get_code("ee4d9f7f86635181")
How it ran

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

import torch
import numpy as np

# Analyze the function: it takes a single parameter 'inp' which must be a string
# The function checks if inp is a string, then performs string operations:
# - finds _DEFAULT_SYSTEM_MARKER in the string
# - finds _DEFAULT_TRAILING_INSTR in the string
# - strips and returns the result

# Create test cases for the inp parameter
test_input_1 = "Some content\n\nSubreddit: test_subreddit\nMore content\nPlease directly output your summarization."
test_input_2 = "Simple string without markers"
test_input_3 = "\n\nSubreddit: news\nArticle content here\nPlease directly output your summarization."
test_input_4 = ""

# We'll use the first test input as the primary argument
ARGS = {
    'inp': test_input_1
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
strip_system_prompt Ran
ran; our contract draft was wrong, not the code
PeterLauLukChen/RACO/script/score_reward_summarization.py
code served (permissive licence) · get_code("9197f51c0cb72bcc")
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

_DEFAULT_SYSTEM_MARKER = '\n\nSubreddit:'
_DEFAULT_TRAILING_INSTR = 'Please directly output your summarization.'
_CHAT_SYSTEM_RE = re.compile('<\\|im_start\\|>system[\\s\\S]*?<\\|im_end\\|>', flags=re.MULTILINE)

def strip_system_prompt(inp: str) -> str:
    """
    Strip the system prompt / instruction boilerplate from `input`, leaving only the raw passage.

    Supports both:
      - plain-text prompts containing a "Subreddit:" block (our RedditSummary prompts)
      - chat templates containing an explicit system block (<|im_start|>system ... <|im_end|>)
    """
    if not isinstance(inp, str):
        return ''
    s = inp
    if '<|im_start|>system' in s:
        s = _CHAT_SYSTEM_RE.sub('', s)
    i = s.find(_DEFAULT_SYSTEM_MARKER)
    if i != -1:
        s = s[i + 2:]
    j = s.rfind(_DEFAULT_TRAILING_INSTR)
    if j != -1:
        s = s[:j].rstrip()
    return s.strip()

# Test case 1: plain text with Subreddit marker
test_input_1 = "Some prefix\n\nSubreddit: AskReddit\n\nThis is the actual content here."

# Test case 2: chat template with system block
test_input_2 = "<|im_start|>system\nYou are a helpful assistant.\n<|im_end|>\n\nSubreddit: news\n\nHere is the content."

# Test case 3: input with trailing instruction
test_input_3 = "\n\nSubreddit: todayilearned\n\nSome interesting content.\n\nPlease directly output your summarization."

# Test case 4: non-string input
test_input_4 = None

ARGS = {
    'inp': test_input_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 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
main Not yet run
unverified (harvested, no recorded run)
PeterLauLukChen/RACO/eval/new_score.py
code served (permissive licence) · get_code("9072e64887215a81")
How it ran

no invocation recorded.

main Not yet run
unverified (harvested, no recorded run)
PeterLauLukChen/RACO/eval/vllm_generate_and_score.py
code served (permissive licence) · get_code("0591481280db58e2")
How it ran

no invocation recorded.

read_prompts_from_parquet Not yet run
unverified (harvested, no recorded run)
PeterLauLukChen/RACO/eval/vllm_generate.py
code served (permissive licence) · get_code("c7af484423746243")
How it ran

no invocation recorded.

Also present in other papers' code

Of 16 samples lifted from this paper's linked repositories, 16 carry a function identity and 0 do not. Of the 16 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 16 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.

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 2026-11-21.

Repositories linked to this paper

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

Abstract

Direct alignment methods are increasingly used to align large language models (LLMs) with human preferences. However, many real-world alignment problems involve multiple conflicting objectives, where naive aggregation of preferences can lead to unstable training and poor trade-offs. In particular, weighted loss methods may fail to identify update directions that simultaneously improve all objectives, and existing multi-objective approaches often rely on explicit reward models, introducing additional complexity and distorting user-specified preferences. Our first contribution is to propose a Reward-free Alignment framework for Conflicted Objectives (RACO) that directly leverages pairwise preference data and resolves gradient conflicts via a novel clipped variant of conflict-averse gradient descent. We provide convergence guarantees to Pareto-critical points that respect user-specified objective weights, and show that clipping can strictly improve convergence rate in two-objective cases. Second, we improve our method using some heuristics and conduct experiments to demonstrate the compatibility of the proposed framework for LLM alignment. Both qualitative and quantitative evaluations on multiobjective summarization and safety alignment tasks across multiple LLM families (Qwen 3, Llama 3, Gemma 3) show that our method consistently achieves better Pareto trade-offs compared to existing multi-objective alignment baselines. Warning: This paper contains examples of potentially harmful and sexually explicit content.

For agents

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

get_harvested_code_for_paper("2602.02495")
get_code_for_paper("2602.02495")
have("2602.02495")

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

curl https://syntology.ai/api/ran/2602.02495.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.02495.svg)](https://syntology.ai/paper/2602.02495)

Connect an agent — have() is free.