SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.14069 · 2026

Open Rubric System: Scaling Reinforcement Learning with Pairwise Adaptive Rubric

Wen Wang, Yuxin Wu, Mengyu Zhou, Xiaoxi Jiang, Guanjun Jiang, Yunyi Yang, Jianhe Lin, Ruipeng Jia, Yongbo Gai, Siyuan Tao

arXiv · PDF · Open in the Atlas

Code that ran

4 of 7 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, 0 violated it and 0 have no contract verdict; 4 failures of our instrument; 3 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2602.14069.json · badge: /api/ran/2602.14069.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/OpenRS canonical (classifier label)
classifier basis: census
4 ran · 3 no recorded run
FunctionStatusWhere it lives
compute_score_by_query_type Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/OpenRS/judgebench_and_ppe.py
code served (permissive licence) · get_code("16d2eb8252127d02")
How it ran

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

import json
import os
import tempfile
from typing import Dict, Any

# Create a temporary directory for output
temp_dir = tempfile.mkdtemp()

# Create a sample JSONL file with results
output_dir = temp_dir
annotation = "test_annotation"
all_results_file = f'{output_dir}/all_results_{annotation}.jsonl'

# Write sample data to the JSONL file
sample_data = [
    {"query_type": "type_a", "final_verdict": "good"},
    {"query_type": "type_a", "final_verdict": "verifiable_good"},
    {"query_type": "type_a", "final_verdict": "bad"},
    {"query_type": "type_b", "final_verdict": "verifiable_bad"},
    {"query_type": "type_b", "final_verdict": "same"},
    {"query_type": "type_b", "final_verdict": "good"},
    {"query_type": "type_c", "final_verdict": "error"},
]

with open(all_results_file, 'w', encoding='utf-8') as f:
    for item in sample_data:
        f.write(json.dumps(item) + '\n')

ARGS = {
    "output_dir": output_dir,
    "annotation": annotation,
}
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
compute_score_from_files Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/OpenRS/judgebench_and_ppe.py
code served (permissive licence) · get_code("1ab59168a36d8759")
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 json

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

# Create sample JSONL files with the expected structure
annotation = "test"

# Create verifiable_good_cases file
verifiable_good_file = f'{temp_dir}/verifiable_good_cases_{annotation}.jsonl'
with open(verifiable_good_file, 'w', encoding='utf-8') as f:
    f.write('{"case": 1}\n')
    f.write('{"case": 2}\n')

# Create verifiable_bad_cases file
verifiable_bad_file = f'{temp_dir}/verifiable_bad_cases_{annotation}.jsonl'
with open(verifiable_bad_file, 'w', encoding='utf-8') as f:
    f.write('{"case": 1}\n')

# Create pairwise_good_cases file
pairwise_good_file = f'{temp_dir}/pairwise_good_cases_{annotation}.jsonl'
with open(pairwise_good_file, 'w', encoding='utf-8') as f:
    f.write('{"case": 1}\n')
    f.write('{"case": 2}\n')
    f.write('{"case": 3}\n')

# Create pairwise_bad_cases file
pairwise_bad_file = f'{temp_dir}/pairwise_bad_cases_{annotation}.jsonl'
with open(pairwise_bad_file, 'w', encoding='utf-8') as f:
    f.write('{"case": 1}\n')
    f.write('{"case": 2}\n')

# Create pairwise_same_cases file
pairwise_same_file = f'{temp_dir}/pairwise_same_cases_{annotation}.jsonl'
with open(pairwise_same_file, 'w', encoding='utf-8') as f:
    f.write('{"case": 1}\n')

ARGS = {
    'output_dir': temp_dir,
    'annotation': annotation
}
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
exec_python Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/OpenRS/evaluator_precise_if.py
code served (permissive licence) · get_code("1666a0264e9660f8")
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 io
from contextlib import redirect_stdout

# Create a simple response string that contains Python code
# The function expects a response with '=====' delimiter and Python code
response = """
Some text before the delimiter
=====

print(True)

"""

ARGS = {
    "response": response
}
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
fix_surrogates_robust Ran
ran; our contract draft was wrong, not the code
Qwen-Applications/OpenRS/robust_utils.py
code served (permissive licence) · get_code("0615c8e9a9749e27")
How it ran

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

import torch
import numpy as np

# Test case 1: Valid UTF-8 string
valid_string = "Hello, World!"

# Test case 2: String with potential surrogate issues
problematic_string = "test\udcffdata"

# Test case 3: Dictionary with mixed content
test_dict = {
    "key1": "value1",
    "key2": valid_string,
    "nested": {
        "inner_key": "inner_value"
    }
}

# Test case 4: List with mixed content
test_list = [
    "string1",
    valid_string,
    {"nested_dict": "value"},
    ["nested", "list"]
]

# Test case 5: Non-string/dict/list types
test_int = 42
test_float = 3.14
test_none = None

ARGS = {
    "obj": test_dict
}
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 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_judge_inputs Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/OpenRS/rmbench.py
code served (permissive licence) · get_code("42ecc4ccc2714946")
How it ran

no invocation recorded.

get_sample_field Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/OpenRS/rmbench.py
code served (permissive licence) · get_code("6e712bcdb14241f6")
How it ran

no invocation recorded.

save_jsonl Not yet run
unverified (harvested, no recorded run)
Qwen-Applications/OpenRS/tools.py
code served (permissive licence) · get_code("c12ee8ca4f14acd1")
How it ran

no invocation recorded.

Also present in other papers' code

Of 7 samples lifted from this paper's linked repositories, 7 carry a function identity and 0 do not. Of the 7 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 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.

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

Scalar reward models compress multi-dimensional human preferences into a single opaque score, creating an information bottleneck that often leads to brittleness and reward hacking in open-ended alignment. We argue that robust alignment for non-verifiable tasks is fundamentally a principle generalization problem: reward should be guided by explicit, auditable principles rather than compressed into a learned scalar function. To operationalize this view, we present the Open Rubric System (OpenRS), a modular, rubricsbased LLM-as-a-Judge framework built around Pairwise Adaptive Meta-Rubrics (PAMR) and lightweight Pointwise Verifiable Rubrics (PVRs), which provide both hard-constraint guardrails and verifiable reward components when ground-truth or programmatic checks are available. OpenRS is governed by an explicit meta-rubric-not a rubric itself, but a higher-level guideline that specifies how rubrics should be generated, which criteria to prioritize, and which failure modes are non-negotiable-and instantiates concrete adaptive rubrics on the fly by conditioning on the semantic differences between two candidate responses. It then performs criterion-wise pairwise comparisons and aggregates criterion-level preferences externally, avoiding pointwise weighted scalarization while improving discriminability in open-ended settings. To keep principles consistent yet editable across various domains, we introduce a two-level meta-rubric refinement pipeline (automated evolutionary refinement for general principles and a reproducible humanin-the-loop procedure for domain principles), complemented with pointwise verifiable rubrics that act as both guardrails against degenerate behaviors and a source of verifiable reward for objective sub-tasks. Notably, the entire pipeline requires no judge-specific training: the underlying judge model remains unmodified, and all evaluation policy lives in editable text. Finally, we instantiate OpenRS as reward supervision in pairwise RL training. Empirically, OpenRS achieves state-of-the-art results on four rewardmodeling benchmarks (RM-Bench, JudgeBench, RewardBench v2, and PPE Preference), consistently outperforming strong open scalar reward model baselines and trained generative judges. In end-to-end policy optimization, replacing a scalar reward model with OpenRS yields stable gains on downstream RL evaluations. The open-source code is available at https://github.com/Qwen-Applications/ OpenRS. Recent progress on reinforcement learning with verifiable rewards (RLVR) demonstrates that when rewards can be grounded in objective checks (e.g., correctness or format constraints) for math or coding tasks, reinforcement learning can scale capabilities without relying on a learned scalar judge (Guo et al., 2025). For general, open-ended tasks, reward specification is intrinsically difficult because the desired behavior spans multiple, often competing dimensions, and there is typically no explicit reference answer or ground-truth signal.

For agents

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

get_harvested_code_for_paper("2602.14069")
get_code_for_paper("2602.14069")
have("2602.14069")

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

curl https://syntology.ai/api/ran/2602.14069.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.14069.svg)](https://syntology.ai/paper/2602.14069)

Connect an agent — have() is free.