SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2501.03124 · ACL · 2025

PRMBENCH: A Fine-grained and Challenging Benchmark for Process-Level Reward Models

Yu Cheng, Jiawei Zhou, Xiaoye Qu, Mingyang Song, Zhao-yu Su

arXiv · PDF · Open in the Atlas

Code that ran

5 of 5 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 5 have no contract verdict; 0 failures of our instrument; 0 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2501.03124.json · badge: /api/ran/2501.03124.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
ssmisya/PRMBench canonical (classifier label)
classifier basis not recorded on the link
5 ran · 0 no recorded run
FunctionStatusWhere it lives
eval_maj_of_n Ran
ran on a synthesized input
ssmisya/PRMBench/mr_eval/tasks/prmbench_bon/task.py
code served (permissive licence) · get_code("c3fbcd180889da35")
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 random
from collections import Counter
import re

random.seed(0)

ARGS = {
    "dataset_list": ["dataset_a", "dataset_b"],
    "big_id_dict": {
        "id_1": [
            {
                "dataset": "dataset_a",
                "step_str": "The answer is \\boxed{42}",
                "label": "correct"
            },
            {
                "dataset": "dataset_a",
                "step_str": "The answer is \\boxed{42}",
                "label": "correct"
            }
        ],
        "id_2": [
            {
                "dataset": "dataset_a",
                "step_str": "The answer is \\boxed{7}",
                "label": "incorrect"
            }
        ],
        "id_3": [
            {
                "dataset": "dataset_b",
                "step_str": "The answer is \\boxed{100}",
                "label": "correct"
            },
            {
                "dataset": "dataset_b",
                "step_str": "The answer is \\boxed{100}",
                "label": "correct"
            }
        ],
        "id_4": [
            {
                "dataset": "dataset_b",
                "step_str": "The answer is \\boxed{50}",
                "label": "incorrect"
            }
        ]
    }
}
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
eval_maj_of_n Ran
ran on a synthesized input
ssmisya/PRMBench/mr_eval/tasks/prmbench_bon_subset50/task.py
code served (permissive licence) · get_code("79d235c075fccb12")
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 re
import random

# Set seed for reproducibility
random.seed(0)

# Build test data for eval_maj_of_n
dataset_list = ["dataset_a", "dataset_b"]

big_id_dict = {
    "id_1": [
        {
            "dataset": "dataset_a",
            "step_str": "The answer is \\boxed{42}",
            "answer": "42"
        },
        {
            "dataset": "dataset_a",
            "step_str": "Final answer: \\boxed{42}",
            "answer": "42"
        },
        {
            "dataset": "dataset_a",
            "step_str": "Result is \\boxed{42}",
            "answer": "42"
        }
    ],
    "id_2": [
        {
            "dataset": "dataset_a",
            "step_str": "The answer is \\boxed{100}",
            "answer": "100"
        },
        {
            "dataset": "dataset_a",
            "step_str": "Final answer: \\boxed{100}",
            "answer": "100"
        }
    ],
    "id_3": [
        {
            "dataset": "dataset_b",
            "step_str": "Solution: \\boxed{7}",
            "answer": "7"
        },
        {
            "dataset": "dataset_b",
            "step_str": "The result is \\boxed{7}",
            "answer": "7"
        },
        {
            "dataset": "dataset_b",
            "step_str": "Answer: \\boxed{7}",
            "answer": "7"
        }
    ],
    "id_4": [
        {
            "dataset": "dataset_b",
            "step_str": "We get \\boxed{15}",
            "answer": "15"
        },
        {
            "dataset": "dataset_b",
            "step_str": "Final: \\boxed{99}",
            "answer": "15"
        }
    ]
}

ARGS = {
    "dataset_list": dataset_list,
    "big_id_dict": big_id_dict
}
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
eval_on_hallucination_step Ran
ran on a synthesized input
ssmisya/PRMBench/mr_eval/tasks/prmbench_stem/task.py
code served (permissive licence) · get_code("5303747ae5ef3847")
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

# Define ARGS dictionary with all parameters for eval_on_hallucination_step
ARGS = {
    "hallucination_steps": [2, 4, 6],
    "labels": [1, 1, 0, 1, 0, 1, 0],
    "redundency_label": False
}
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
eval_pass_at_n Ran
ran on a synthesized input
ssmisya/PRMBench/mr_eval/tasks/prmbench_bon/task.py
code served (permissive licence) · get_code("ae4a5f14c32e1ad6")
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

# Based on the function, we need:
# 1. dataset_list: a list of dataset names (strings)
# 2. big_id_dict: a dict where keys are big_ids and values are lists of dicts
#    Each dict in the list has 'dataset' and 'label' keys

ARGS = {
    'dataset_list': ['dataset_a', 'dataset_b', 'dataset_c'],
    'big_id_dict': {
        'id_1': [
            {'dataset': 'dataset_a', 'label': 'correct'},
            {'dataset': 'dataset_a', 'label': 'incorrect'}
        ],
        'id_2': [
            {'dataset': 'dataset_a', 'label': 'incorrect'}
        ],
        'id_3': [
            {'dataset': 'dataset_b', 'label': 'correct'}
        ],
        'id_4': [
            {'dataset': 'dataset_b', 'label': 'incorrect'},
            {'dataset': 'dataset_b', 'label': 'incorrect'}
        ],
        'id_5': [
            {'dataset': 'dataset_c', 'label': 'incorrect'}
        ],
        'id_6': [
            {'dataset': 'dataset_c', 'label': 'correct'},
            {'dataset': 'dataset_c', 'label': 'correct'}
        ]
    }
}
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
eval_pass_at_n Ran
ran on a synthesized input
ssmisya/PRMBench/mr_eval/tasks/prmbench_bon_subset50/task.py
code served (permissive licence) · get_code("a3c529098d61a8e9")
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 re

def extract_answer_from_model_response(model_response):
    matches = re.findall('\\\\boxed\\{(.*?)\\}', model_response)
    if matches:
        return str(matches[-1])
    return None

def eval_pass_at_n(dataset_list, big_id_dict):
    res = {}
    for dataset_name in dataset_list:
        pass_num = 0
        total_num = 0
        for big_id, big_id_meta in big_id_dict.items():
            assert len(big_id_meta) > 0
            if len(big_id_meta) > 0 and big_id_meta[0]['dataset'] == dataset_name:
                total_num += 1
                for item in big_id_meta:
                    response_answer = extract_answer_from_model_response(item['step_str'])
                    if response_answer and response_answer and (response_answer == item['answer']):
                        pass_num += 1
                        break
        dataset_pass_at_n = pass_num / total_num if total_num > 0 else -1
        res[dataset_name] = dataset_pass_at_n
    return res

# Setup code
dataset_list = ["math", "science"]

big_id_dict = {
    "id_1": [
        {
            "dataset": "math",
            "step_str": "The answer is \\boxed{42}",
            "answer": "42"
        }
    ],
    "id_2": [
        {
            "dataset": "math",
            "step_str": "Solution: \\boxed{7}",
            "answer": "8"
        }
    ],
    "id_3": [
        {
            "dataset": "science",
            "step_str": "Result: \\boxed{hydrogen}",
            "answer": "hydrogen"
        }
    ],
    "id_4": [
        {
            "dataset": "science",
            "step_str": "Answer is \\boxed{oxygen}",
            "answer": "oxygen"
        }
    ]
}

ARGS = {
    "dataset_list": dataset_list,
    "big_id_dict": big_id_dict
}
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

Also present in other papers' code

Of 5 samples lifted from this paper's linked repositories, 5 carry a function identity and 0 do not. Of the 5 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 5 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

Process-level Reward Models (PRMs) are crucial for complex reasoning and decisionmaking tasks, where each intermediate step plays an important role in the reasoning process. Since large language models (LLMs) suffer from various types of errors during the reasoning process, PRMs are required to possess nuanced capabilities for detecting various implicit error types in real-world scenarios. However, current benchmarks primarily focus on step correctness, failing to evaluate PRMs' performance systematically. To address this gap, we introduce PRMBENCH, a processlevel benchmark specifically designed to assess the fine-grained error detection capabilities of PRMs. PRMBENCH comprises 6,216 carefully designed problems and 83,456 step-level labels, evaluating models across multiple dimensions, including simplicity, soundness, and sensitivity. In our experiments on 25 models, spanning across both open-source PRMs and LLMs prompted as critic models, we uncover significant weaknesses in current PRMs. These findings reveal the challenges inherent in processlevel evaluation and highlight key directions for future research, establishing PRMBENCH as a robust testbed for advancing research on PRM evaluation and development.

For agents

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

get_harvested_code_for_paper("2501.03124")
get_code_for_paper("2501.03124")
have("2501.03124")

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

curl https://syntology.ai/api/ran/2501.03124.json

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

[![Syntology run record](https://syntology.ai/api/ran/2501.03124.svg)](https://syntology.ai/paper/2501.03124)

Connect an agent — have() is free.