SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.12809 · EMNLP · 2024

Can Large Language Models Always Solve Easy Problems if They Can Solve Harder Ones?

Chang Zhou, Junyang Lin, Jian Yang, Zhifang Sui, Tianyu Liu, Yichang Zhang, Zhe Yang

arXiv · PDF · Open in the Atlas

Code that ran

11 of 14 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; 2 honoured the contract we drafted, 3 violated it and 0 have no contract verdict; 6 failures of our instrument; 3 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/2406.12809.json · badge: /api/ran/2406.12809.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
QwenLM/ConsisEval canonical (classifier label)
classifier basis: census
11 ran · 3 no recorded run
FunctionStatusWhere it lives
are_equal_under_sympy Ran
ran, violated the contract we drafted
QwenLM/ConsisEval/math_check/grader.py
code served (permissive licence) · get_code("8bec43a95385f997")
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
from sympy.parsing import sympy_parser

BAD_SUBSTRINGS = ['^{', '^(']

BAD_REGEXES = ['\\^[0-9]+\\^', '\\^[0-9][0-9]+', '^[0-9]+$']

def _sympy_parse(expr: str):
    """Parses an expression with sympy."""
    py_expr = expr.replace('^', '**')
    return sympy_parser.parse_expr(py_expr, transformations=sympy_parser.standard_transformations + (sympy_parser.implicit_multiplication_application,))

def count_unknown_letters_in_expr(expr: str):
    ignore_words = ['sqrt', 'frac', 'abs', 'max', 'sin', 'cos', 'pi', 'e']
    for w in ignore_words:
        expr = expr.replace(w, '')
    letters_in_expr = set([x for x in expr if x.isalpha()])
    return len(letters_in_expr)

def should_allow_eval(expr: str):
    if count_unknown_letters_in_expr(expr) > 2:
        return False
    for bad_string in BAD_SUBSTRINGS:
        if bad_string in expr:
            return False
    for bad_regex in BAD_REGEXES:
        if re.search(bad_regex, expr) is not None:
            return False
    return True

def are_equal_under_sympy(ground_truth_normalized: str, given_normalized: str):
    are_equal = False
    try:
        expr = f'abs(({ground_truth_normalized})-({given_normalized})) / (max(abs({ground_truth_normalized}), abs({given_normalized})) + 1e-9)'
        if should_allow_eval(given_normalized):
            sympy_diff = _sympy_parse(expr)
            if sympy_diff < 0.001:
                are_equal = True
    except:
        pass
    return are_equal


ARGS = {
    "ground_truth_normalized": "x + 2",
    "given_normalized": "2 + x"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
sympy — 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
check_correctness_code Ran
ran, violated the contract we drafted
QwenLM/ConsisEval/code_check/check.py
code served (permissive licence) · get_code("b0505332a16b5516")
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 os
import tempfile
import multiprocessing

# Define a small problem dict that matches what check_correctness_code expects
ARGS = {
    'problem': {
        'prompt': 'def add(a, b):\n    ',
        'completion': 'return a + b',
        'test': '''def check(func):
    assert func(2, 3) == 5
    assert func(0, 0) == 0
    assert func(-1, 1) == 0
''',
        'entry_point': 'add'
    },
    'completion': 'return a + b',
    'timeout': 3.0,
    'completion_id': None
}
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
conflict_make Ran
ran; our contract draft was wrong, not the code
QwenLM/ConsisEval/instruction_following_check/instructions_registry.py
code served (permissive licence) · get_code("d197da35ee544f75")
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

ARGS = {
    "conflicts": {
        0: set([1, 2]),
        1: set([0]),
        2: set([3]),
        3: set()
    }
}
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 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
count_unknown_letters_in_expr Ran
ran, honoured the contract we drafted
QwenLM/ConsisEval/math_check/grader.py
code served (permissive licence) · get_code("54417bf5eb16e81f")
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, 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

# The function takes a single parameter: expr (a string)
# It processes the string by:
# 1. Removing certain ignore_words from it
# 2. Finding all alphabetic characters
# 3. Counting unique letters

# Simple test cases for the expression string parameter
expr_simple = "x + y + z"
expr_with_ignored = "sqrt(x) + sin(y) + pi"
expr_complex = "frac(a, b) + abs(c) - max(d, e)"

ARGS = {
    "expr": "x + y + sqrt(z) + sin(a)"
}
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 int, ndim 0, dtype family int, 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
count_words Ran
ran, honoured the contract we drafted
QwenLM/ConsisEval/instruction_following_check/instructions_util.py
pointer only (licence: MIT); the stored copy get_code serves, from dongguanting/FollowRAG, is pointer only (licence: NONE) · get_code("cdcc85ca09b00f7e")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from dongguanting/FollowRAG, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

extract_answer Ran
ran; our contract draft was wrong, not the code
QwenLM/ConsisEval/math_check/check.py
code served (permissive licence) · get_code("b22b268ce5bb5b68")
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. 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

ARGS = {
    "last_sentence": "The answer is $x = 5$ and the final result is \\boxed{42}.",
    "prompt": "What is the value? The answer is"
}
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 — a tuple output has no array shape of its own; none is 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
normalize_answer Ran
ran; our contract draft was wrong, not the code
QwenLM/ConsisEval/math_check/math_normalize.py
pointer only (licence: MIT); the stored copy get_code serves, from PKU-SEC-Lab/SPEX, is pointer only (licence: NONE) · get_code("101b1e54fcdadf90")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from PKU-SEC-Lab/SPEX, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

process_text Ran
ran; our contract draft was wrong, not the code
QwenLM/ConsisEval/code_check/post_process.py
code served (permissive licence) · get_code("dab90bb25f71a531")
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

def remove_after_return(code):
    """
    Takes as input a code, and removes everything that is after the return.
    That is, the first line that does not start with a space character
    """
    pattern = '[^\\n]+(\\n|$)'
    end_last_match = None
    for match in re.finditer(pattern, code):
        start_match, end_match = match.span()
        if end_last_match is not None and start_match < len(code) and (code[start_match].strip() != ''):
            return code[0:start_match]
        end_last_match = end_match
    return code

def process_text(text, entry_point):
    text = text.strip('\n')
    x = re.search('```python(.+?)```', text, flags=re.DOTALL)
    if x != None:
        if f'def {entry_point}' in text[x.span()[0] + 9:x.span()[1] - 3]:
            text = text[x.span()[0] + 9:x.span()[1] - 3]
    x = re.search('```(.+?)```', text, flags=re.DOTALL)
    if x != None:
        if f'def {entry_point}' in text[x.span()[0] + 3:x.span()[1] - 3]:
            text = text[x.span()[0] + 3:x.span()[1] - 3]
    x = re.search('```python(.+)', text, flags=re.DOTALL)
    if x != None:
        text = text[x.span()[0] + 9:x.span()[1]]
    x = re.search('(.+)```', text, flags=re.DOTALL)
    if x != None:
        text = text[x.span()[0]:x.span()[1] - 3]
    text = text.strip()
    x = re.search(f'def {entry_point}[^\\n]+(\\n|$)', text)
    if x != None:
        start, end = x.span()
        prefix = text[:start]
        text = text[end:]
        if len(prefix.split('\n')) != 0:
            l = prefix.split('\n')
            l = [line.strip() for line in l]
            for idx, line in enumerate(l):
                if line.startswith('def') or line.startswith('from ') or line.startswith('import'):
                    prefix = '\n    ' + '\n    '.join(l[idx:]) + '\n'
                    break
                else:
                    prefix = ''
            text = prefix + text
    text = remove_after_return(text)
    return text


# Setup valid arguments
ARGS = {
    "text": """```python
def my_function(x):
    return x + 1
    this should be removed
""",
    "entry_point": "my_function"
}
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
remove_after_return Ran
ran; our contract draft was wrong, not the code
QwenLM/ConsisEval/code_check/post_process.py
code served (permissive licence) · get_code("29ee1d8c120f674a")
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

def remove_after_return(code):
    """
    Takes as input a code, and removes everything that is after the return.
    That is, the first line that does not start with a space character
    """
    pattern = '[^\\n]+(\\n|$)'
    end_last_match = None
    for match in re.finditer(pattern, code):
        start_match, end_match = match.span()
        if end_last_match is not None and start_match < len(code) and (code[start_match].strip() != ''):
            return code[0:start_match]
        end_last_match = end_match
    return code

# Test code with indentation and a non-indented line (simulating return)
test_code = """    def foo():
    x = 1
    return x
print('this should be removed')
more_code = 2"""

ARGS = {
    "code": test_code
}
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
should_allow_eval Ran
ran, violated the contract we drafted
QwenLM/ConsisEval/math_check/grader.py
code served (permissive licence) · get_code("f5c7c44887cd8c88")
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

BAD_SUBSTRINGS = ['^{', '^(']

BAD_REGEXES = ['\\^[0-9]+\\^', '\\^[0-9][0-9]+', '^[0-9]+$']

def count_unknown_letters_in_expr(expr: str):
    ignore_words = ['sqrt', 'frac', 'abs', 'max', 'sin', 'cos', 'pi', 'e']
    for w in ignore_words:
        expr = expr.replace(w, '')
    letters_in_expr = set([x for x in expr if x.isalpha()])
    return len(letters_in_expr)

def should_allow_eval(expr: str):
    if count_unknown_letters_in_expr(expr) > 2:
        return False
    for bad_string in BAD_SUBSTRINGS:
        if bad_string in expr:
            return False
    for bad_regex in BAD_REGEXES:
        if re.search(bad_regex, expr) is not None:
            return False
    return True

ARGS = {
    'expr': 'x + y + sqrt(z)'
}
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
split_into_sentences Ran
ran; our contract draft was wrong, not the code
QwenLM/ConsisEval/instruction_following_check/instructions_util.py
pointer only (licence: MIT); the stored copy get_code serves, from Rainier-rq1/From_Coarse_to_Fine, is pointer only (licence: NONE) · get_code("34d346853108011e")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from Rainier-rq1/From_Coarse_to_Fine, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

check_correctness_instruction_following Not yet run
unverified (harvested, no recorded run)
QwenLM/ConsisEval/instruction_following_check/check.py
code served (permissive licence) · get_code("f9e0e08dca784150")
How it ran

no invocation recorded.

check_correctness_math Not yet run
unverified (harvested, no recorded run)
QwenLM/ConsisEval/math_check/check.py
code served (permissive licence) · get_code("a13ec071b5563491")
How it ran

no invocation recorded.

count_sentences Not yet run
unverified (harvested, no recorded run)
QwenLM/ConsisEval/instruction_following_check/instructions_util.py
pointer only (licence: MIT); the stored copy get_code serves, from dongguanting/FollowRAG, is pointer only (licence: NONE) · get_code("f79646ddd0a03162")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from dongguanting/FollowRAG, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

Also present in other papers' code

Of 14 samples lifted from this paper's linked repositories, 14 carry a function identity and 0 do not. Of the 14 distinct functions with an identity, 7 are also present in other papers whose code we hold (7 with at least one paper dated earlier, 1 the same month, 7 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.

FunctionAlso present inEarliest-datedOthers named
should_allow_eval 13 other papers
1 dated earlier, 12 later
2305.20050 (2023-05)
Let's Verify Step by Step
2502.06703 (2025-02)
Can 1B LLM Surpass 405B LLM? Rethinking Compute-Optimal Test-Time Scaling
2503.10460 (2025-03)
Light-R1: Curriculum SFT, DPO and RL for Long COT from Scratch and Beyond
and 10 more not named here
normalize_answer 10 other papers
1 dated earlier, 9 later
2404.02078 (2024-04)
Advancing LLM Reasoning Generalists with Preference Trees
2412.01981 (2024-12)
Free Process Rewards without Process Labels
2502.01456 (2025-02)
Process Reinforcement through Implicit Rewards
and 7 more not named here
split_into_sentences 8 other papers
3 dated earlier, 5 later
2309.00071 (2023-09)
YaRN: Efficient Context Window Extension of Large Language Models
2311.07911 (2023-11)
Instruction-Following Evaluation for Large Language Models
2404.15846 (2024-04)
From Complex to Simple: Enhancing Multi-Constraint Complex Instruction Following Ability of Large Language Models
and 5 more not named here
conflict_make 6 other papers
3 dated earlier, 3 later
2309.00071 (2023-09)
YaRN: Efficient Context Window Extension of Large Language Models
2311.07911 (2023-11)
Instruction-Following Evaluation for Large Language Models
2404.15846 (2024-04)
From Complex to Simple: Enhancing Multi-Constraint Complex Instruction Following Ability of Large Language Models
and 3 more not named here
count_words 6 other papers
4 dated earlier, 2 later
2309.00071 (2023-09)
YaRN: Efficient Context Window Extension of Large Language Models
2311.07911 (2023-11)
Instruction-Following Evaluation for Large Language Models
2401.16745 (2024-01)
MT-Eval: A Multi-Turn Capabilities Evaluation Benchmark for Large Language Models
and 3 more not named here
count_sentences 5 other papers
3 dated earlier, 2 later
2309.00071 (2023-09)
YaRN: Efficient Context Window Extension of Large Language Models
2311.07911 (2023-11)
Instruction-Following Evaluation for Large Language Models
2404.15846 (2024-04)
From Complex to Simple: Enhancing Multi-Constraint Complex Instruction Following Ability of Large Language Models
and 2 more not named here
remove_after_return 4 other papers
1 dated earlier, 1 the same month, 1 later, 1 same year, order not determinable
2308.07124 (2023-08)
OctoPack: Instruction Tuning Code Large Language Models
2024.findings-emnlp.86 (2024 (year only))
Tokenization Falling Short: On Subword Robustness in Large Language Models
2406.11687 (2024-06)
Tokenization Falling Short: On Subword Robustness in Large Language Models
and 1 more not named here

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

Large language models (LLMs) have demonstrated impressive capabilities, but still suffer from inconsistency issues (e.g. LLMs can react differently to disturbances like rephrasing or inconsequential order change). In addition to these inconsistencies, we also observe that LLMs, while capable of solving hard problems, can paradoxically fail at easier ones. To evaluate this hard-to-easy inconsistency, we develop the ConsisEval benchmark, where each entry comprises a pair of questions with a strict order of difficulty. Furthermore, we introduce the concept of consistency score to quantitatively measure this inconsistency and analyze the potential for improvement in consistency by relative consistency score. Based on comprehensive experiments across a variety of existing models, we find: (1) GPT-4 achieves the highest consistency score of 92.2% but is still inconsistent to specific questions due to distraction by redundant information, misinterpretation of questions, etc.; (2) models with stronger capabilities typically exhibit higher consistency, but exceptions also exist; (3) hard data enhances consistency for both fine-tuning and in-context learning. Our data and code will be publicly available on GitHub. 1

For agents

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

get_harvested_code_for_paper("2406.12809")
get_code_for_paper("2406.12809")
have("2406.12809")

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

curl https://syntology.ai/api/ran/2406.12809.json

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

[![Syntology run record](https://syntology.ai/api/ran/2406.12809.svg)](https://syntology.ai/paper/2406.12809)

Connect an agent — have() is free.