SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2402.12055 · ACL · 2024

Are LLM-based Evaluators Confusing NLG Quality Criteria?

Yang Zhang, Xiaojun Wan, Xinyu Hu, Teng Xu, Mingqi Gao, Yicheng Chen, Sen Hu

arXiv · PDF · Open in the Atlas

Code that ran

3 of 8 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 0 have no contract verdict; 2 failures of our instrument; 5 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/2402.12055.json · badge: /api/ran/2402.12055.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
herrxy/llm-evaluator-reliability canonical (classifier label)
classifier basis: census
3 ran · 5 no recorded run
FunctionStatusWhere it lives
contains_non_alpha Ran
ran, violated the contract we drafted
herrxy/llm-evaluator-reliability/Perturbations/grammar_order.py
code served (permissive licence) · get_code("1e3e86a7baafca68")
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 re

def contains_non_alpha(word):
    return bool(re.search('[^a-z]', word))

ARGS = {
    "word": "hello123"
}
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 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
introduce_swap Ran
ran; our contract draft was wrong, not the code
herrxy/llm-evaluator-reliability/Perturbations/grammar_order.py
code served (permissive licence) · get_code("26336d4815eced14")
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 random
import re

def contains_non_alpha(word):
    return bool(re.search('[^a-z]', word))

def introduce_swap(sentence):
    """Swap two adjacent words in the sentence"""
    words = sentence.split()
    if len(words) > 5:
        if len(words) < 10:
            position = [i for i in range(1, len(words) - 1) if not contains_non_alpha(words[i]) and (not contains_non_alpha(words[i + 1]))]
            if position:
                word_index = random.choice(position)
                words[word_index], words[word_index + 1] = (words[word_index + 1], words[word_index])
        else:
            position1 = [i for i in range(1, len(words) // 2 - 1) if not contains_non_alpha(words[i]) and (not contains_non_alpha(words[i + 1]))]
            if position1:
                word_index1 = random.choice(position1)
                words[word_index1], words[word_index1 + 1] = (words[word_index1 + 1], words[word_index1])
            position2 = [i for i in range(len(words) // 2, len(words) - 1) if not contains_non_alpha(words[i]) and (not contains_non_alpha(words[i + 1]))]
            if position2:
                word_index2 = random.choice(position2)
                words[word_index2], words[word_index2 + 1] = (words[word_index2 + 1], words[word_index2])
    return ' '.join(words)


ARGS = {
    'sentence': 'the quick brown fox jumps over the lazy dog today'
}
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
introduce_typo Ran
ran; our contract draft was wrong, not the code
herrxy/llm-evaluator-reliability/Perturbations/grammar_typo.py
code served (permissive licence) · get_code("97382afbeed4ecb1")
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 random
import re
import numpy as np

# Set random seeds for reproducibility
random.seed(0)
np.random.seed(0)

# The function takes a single parameter: sentence (a string)
# Looking at the function body:
# - It calls sentence.split() to get words
# - It iterates over words and checks their properties
# - It performs string manipulations on words
# - It joins words back with ' '.join()

# Create a valid test sentence with words that are at least 4 characters long
# and contain only lowercase letters (to trigger typos)
test_sentence = "hello world python programming language"

ARGS = {
    "sentence": test_sentence
}
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 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
attack_coherence_sentence_order Not yet run
unverified (harvested, no recorded run)
herrxy/llm-evaluator-reliability/Perturbations/coherence_order.py
code served (permissive licence) · get_code("a344135785b6fb9b")
How it ran

no invocation recorded.

attack_grammar_typos Not yet run
unverified (harvested, no recorded run)
herrxy/llm-evaluator-reliability/Perturbations/grammar_typo.py
code served (permissive licence) · get_code("acb43e2bc12c97cc")
How it ran

no invocation recorded.

attack_grammar_word_order Not yet run
unverified (harvested, no recorded run)
herrxy/llm-evaluator-reliability/Perturbations/grammar_order.py
code served (permissive licence) · get_code("05e821e68b4582a4")
How it ran

no invocation recorded.

attack_informative_delete_last Not yet run
unverified (harvested, no recorded run)
herrxy/llm-evaluator-reliability/Perturbations/inform_delete.py
code served (permissive licence) · get_code("1d690ef0ad6fdfc4")
How it ran

no invocation recorded.

is_the_same Not yet run
unverified (harvested, no recorded run)
herrxy/llm-evaluator-reliability/Perturbations/coherence_order.py
code served (permissive licence) · get_code("90afb16f86fbb016")
How it ran

no invocation recorded.

Also present in other papers' code

Of 8 samples lifted from this paper's linked repositories, 8 carry a function identity and 0 do not. Of the 8 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 8 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

Some prior work has shown that LLMs perform well in NLG evaluation for different tasks. However, we discover that LLMs seem to confuse different evaluation criteria, which reduces their reliability. For further verification, we first consider avoiding issues of inconsistent conceptualization and vague expression in existing NLG quality criteria themselves. So we summarize a clear hierarchical classification system for 11 common aspects with corresponding different criteria from previous studies involved. Inspired by behavioral testing, we elaborately design 18 types of aspect-targeted perturbation attacks for fine-grained analysis of the evaluation behaviors of different LLMs. We also conduct human annotations beyond the guidance of the classification system to validate the impact of the perturbations. Our experimental results reveal confusion issues inherent in LLMs, as well as other noteworthy phenomena, and necessitate further research and improvements for LLM-based evaluation. * Equal contribution.

For agents

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

get_harvested_code_for_paper("2402.12055")
get_code_for_paper("2402.12055")
have("2402.12055")

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

curl https://syntology.ai/api/ran/2402.12055.json

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

[![Syntology run record](https://syntology.ai/api/ran/2402.12055.svg)](https://syntology.ai/paper/2402.12055)

Connect an agent — have() is free.