SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2601.21064 · 2026

Textual Equilibrium Propagation for Deep Compound AI Systems

James Zou, Xiaoxiao Li, Minghui Chen, Han Yu, Wenlong Deng

arXiv · PDF · Open in the Atlas

Code that ran

11 of 15 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, 0 violated it and 0 have no contract verdict; 9 failures of our instrument; 4 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/2601.21064.json · badge: /api/ran/2601.21064.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
MinghuiChen43/TEP canonical (classifier label)
classifier basis: census
11 ran · 4 no recorded run
FunctionStatusWhere it lives
collect_trainable_components Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
MinghuiChen43/TEP/src/tep/runtime/prompt_state.py
code served (permissive licence) · get_code("10499876ed9a8c47")
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 2, 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 sys
from pathlib import Path
from typing import Any, Dict, List, Optional, Sequence, Tuple


class MockVariable:
    def __init__(self, value: str, requires_grad: bool = False, role_description: str = ""):
        self.value = value
        self.requires_grad = requires_grad
        self._role_description = role_description
    
    def get_role_description(self) -> str:
        return self._role_description
    
    def set_role_description(self, desc: str) -> None:
        self._role_description = desc


class MockTextGrad:
    Variable = MockVariable


class MockLLMCall:
    def __init__(self, system_prompt: Optional[Any] = None):
        self.system_prompt = system_prompt


class MockBlackBoxLLM:
    def __init__(self, system_prompt: Optional[Any] = None, llm_call: Optional[Any] = None):
        self.system_prompt = system_prompt
        self.llm_call = llm_call


class MockComponent:
    def __init__(self, blackboxllm: Optional[Any] = None):
        self.blackboxllm = blackboxllm


class MockSystem:
    def __init__(self, components: Dict[str, Any], execution_order: List[str]):
        self.components = components
        self.execution_order = execution_order


# Create a mock textgrad module
mock_tg = MockTextGrad()

# Create components with trainable prompts
component1_llm = MockBlackBoxLLM(
    system_prompt="Initial prompt for component 1",
    llm_call=MockLLMCall(system_prompt="Initial prompt for component 1")
)
component1 = MockComponent(blackboxllm=component1_llm)

component2_llm = MockBlackBoxLLM(
    system_prompt="Initial prompt for component 2",
    llm_call=None
)
component2 = MockComponent(blackboxllm=component2_llm)

# Component without blackboxllm (should be skipped)
component3 = MockComponent(blackboxllm=None)

# Create system with execution order
system = MockSystem(
    components={
        "step1": component1,
        "step2": component2,
        "step3": component3,
    },
    execution_order=["step1", "step2", "step3"]
)

ARGS = {
    "system": system,
    "tg": mock_tg,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
textgrad — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, 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
describe_prompt_state_coverage Ran
ran; our contract draft was wrong, not the code
MinghuiChen43/TEP/src/tep/runtime/prompt_state.py
code served (permissive licence) · get_code("f20e933a9a8baf2a")
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 torch.nn as nn

torch.manual_seed(0)

# Construct trainable_components: a sequence of tuples (name, param1, param2)
# The function only uses the 'name' part (first element of each tuple)
trainable_components = [
    ("layer1", torch.tensor([1.0, 2.0]), torch.tensor([3.0, 4.0])),
    ("layer2", torch.tensor([5.0, 6.0]), torch.tensor([7.0, 8.0])),
    ("layer3", torch.tensor([9.0, 10.0]), torch.tensor([11.0, 12.0])),
]

# Construct prompt_values: a dict mapping strings to strings
# Some keys match trainable component names, some don't
prompt_values = {
    "layer1": "value1",
    "layer2": "value2",
    "extra_key": "extra_value",
    "another_extra": "another_value",
}

ARGS = {
    "trainable_components": trainable_components,
    "prompt_values": prompt_values,
}
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
f1_score Ran
ran, honoured the contract we drafted
MinghuiChen43/TEP/src/tep/metrics/f1.py
code served (permissive licence) · get_code("6ce524d71a9b9d78")
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 re
import string
from collections import Counter

def normalize_answer(s):
    """
    Normalize answer string for consistent evaluation.
    Steps:
    1. Convert to lowercase
    2. Remove articles (a, an, the)
    3. Remove punctuation
    4. Standardize whitespace
    
    Args:
        s (str): Input answer string
    
    Returns:
        str: Normalized answer string
    """

    def remove_articles(text):
        return re.sub('\\b(a|an|the)\\b', ' ', text)

    def white_space_fix(text):
        return ' '.join(text.split())

    def remove_punc(text):
        exclude = set(string.punctuation)
        return ''.join((ch for ch in text if ch not in exclude))

    def lower(text):
        return text.lower()
    return white_space_fix(remove_articles(remove_punc(lower(s)))).strip()

def f1_score(answer, gd_answer):
    """
    Calculate F1 score between answer and ground truth.
    Special handling for yes/no/noanswer questions.
    
    Args:
        answer (str): Predicted answer
        gd_answer (str): Ground truth answer
    
    Returns:
        float: f1_score
    """
    normalized_answer = normalize_answer(answer)
    normalized_gd_answer = normalize_answer(gd_answer)
    ZERO_METRIC = 0
    if normalized_answer in ['yes', 'no', 'noanswer'] and normalized_answer != normalized_gd_answer:
        return ZERO_METRIC
    if normalized_gd_answer in ['yes', 'no', 'noanswer'] and normalized_answer != normalized_gd_answer:
        return ZERO_METRIC
    answer_tokens = normalized_answer.split()
    gd_answer_tokens = normalized_gd_answer.split()
    common = Counter(answer_tokens) & Counter(gd_answer_tokens)
    num_same = sum(common.values())
    if num_same == 0:
        return ZERO_METRIC
    precision = 1.0 * num_same / len(answer_tokens)
    recall = 1.0 * num_same / len(gd_answer_tokens)
    f1 = 2 * precision * recall / (precision + recall)
    return f1


ARGS = {
    "answer": "The quick brown fox jumps over the lazy dog.",
    "gd_answer": "A quick brown fox jumped over a lazy dog"
}
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 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
load_prompt_state_dict Ran
ran; our contract draft was wrong, not the code
MinghuiChen43/TEP/src/tep/runtime/prompt_state.py
code served (permissive licence) · get_code("cf41f610b0f11103")
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 tempfile
import os

# Create a temporary JSON file with prompt state data
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "prompt_state.json")

# Write a valid JSON object to the temporary file
prompt_data = {
    "system_prompt": "You are a helpful assistant.",
    "user_prompt": "What is the capital of France?",
    "model_name": "gpt-4",
    "temperature": "0.7",
    "max_tokens": "100"
}

with open(temp_file, 'w', encoding='utf-8') as f:
    json.dump(prompt_data, f)

ARGS = {
    "path": temp_file
}
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
make_raw_chat_prompt Ran
ran; our contract draft was wrong, not the code
MinghuiChen43/TEP/src/tep/tasks/bigcodebench/common.py
code served (permissive licence) · get_code("51412732efc5c512")
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

# Construct arguments for make_raw_chat_prompt
ARGS = {
    'task_prompt': 'Write a function that adds two numbers together.',
    'split': 'complete',
    'instruction_prefix': 'You are a helpful coding assistant.'
}
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
mrr Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
MinghuiChen43/TEP/src/tep/metrics/mrr.py
code served (permissive licence) · get_code("af9c48b788c853d4")
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

torch.manual_seed(0)

ARGS = {
    "candidate_ids": ["doc_1", "doc_2", "doc_3", "doc_4"],
    "final_scores": [0.9, 0.7, 0.5, 0.3],
    "answer_ids": {"doc_2", "doc_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 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
normalize_answer Ran
ran; our contract draft was wrong, not the code
MinghuiChen43/TEP/src/tep/metrics/f1.py
code served (permissive licence) · get_code("03171efd2aab7e80")
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
import string

ARGS = {
    "s": "The quick brown fox jumps over the lazy dog! A beautiful, perfect day."
}
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
parse_k_grid Ran
ran, honoured the contract we drafted
MinghuiChen43/TEP/src/tep/runtime/split_utils.py
code served (permissive licence) · get_code("0d9d167c2aabf1f8")
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 int; 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 torch

ARGS = {
    'text': '8,16,32'
}
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 int — 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
read_json Ran
ran; our contract draft was wrong, not the code
MinghuiChen43/TEP/src/tep/runtime/split_utils.py
code served (permissive licence) · get_code("9b813f8c08247898")
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 tempfile
import os

# Create a temporary JSON file for testing
temp_dir = tempfile.mkdtemp()
test_file = os.path.join(temp_dir, "test.json")

# Write a valid JSON file
test_data = {
    "key1": "value1",
    "key2": 42,
    "key3": [1, 2, 3],
    "key4": {"nested": "object"}
}

with open(test_file, 'w', encoding='utf-8') as f:
    json.dump(test_data, f)

ARGS = {
    "path": test_file
}
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
setup_logger Ran
ran; our contract draft was wrong, not the code
MinghuiChen43/TEP/src/tep/runtime/core.py
code served (permissive licence) · get_code("cb199a9c59c7dbb3")
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 Logger. 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 logging
from pathlib import Path
from tempfile import TemporaryDirectory

# Setup basic arguments for setup_logger function
ARGS = {
    'name': 'test_logger',
    'log_level': logging.DEBUG,
    'log_file': None  # Can be set to a path if needed, but avoiding file I/O per rules
}

# Alternative configuration with log_file for reference (commented out to avoid file I/O):
# with TemporaryDirectory() as tmpdir:
#     ARGS = {
#         'name': 'test_logger_with_file',
#         'log_level': logging.INFO,
#         'log_file': str(Path(tmpdir) / 'test.log')
#     }
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 Logger — a Logger output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Logger 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
split_ids_for_examples Ran
ran; our contract draft was wrong, not the code
MinghuiChen43/TEP/src/tep/runtime/split_utils.py
code served (permissive licence) · get_code("5e56946e001d6e4c")
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 hashlib
import json

# Define sample examples with different structures
class Example:
    def __init__(self, data):
        self.__dict__.update(data)
    
    def to_dict(self):
        return dict(self.__dict__)

# Create distinct example sets for different datasets

# BigCodeBench examples
bigcodebench_train = [
    {'task_id': 'task_001', 'code': 'print("hello")'},
    {'task_id': 'task_002', 'code': 'print("world")'},
]

bigcodebench_val = [
    {'task_id': 'task_003', 'code': 'print("test")'},
]

bigcodebench_test = [
    {'task_id': 'task_004', 'code': 'print("final")'},
    {'task_id': 'task_004', 'code': 'print("duplicate")'},  # duplicate task_id
]

# STARK examples
stark_train = [
    Example({'question_id': 1, 'question': 'What is AI?', 'answer_ids': [10, 20]}),
    Example({'question_id': 2, 'question': 'What is ML?', 'answer_ids': [30, 40]}),
]

stark_val = [
    Example({'question_id': 3, 'question': 'What is DL?', 'answer_ids': [50]}),
]

stark_test = [
    {'question': 'No question_id', 'answer_ids': [60, 70]},
    {'question': 'No question_id', 'answer_ids': [60, 70]},  # duplicate
]

# HotpotQA examples
hotpotqa_train = [
    {
        'question': 'Who founded OpenAI?',
        'gd_answer': 'Sam Altman',
        'context_documents': ['doc1', 'doc2']
    },
]

hotpotqa_val = [
    {
        'question': 'What is transformers?',
        'gd_answer': 'Neural architecture',
        'context_documents': ['doc3']
    },
]

hotpotqa_test = [
    {
        'question': 'What is a tensor?',
        'gd_answer': 'Mathematical object',
        'context_documents': ['doc4', 'doc5', 'doc6']
    },
]

# PubMed examples
pubmed_train = [
    {
        'question': 'What causes diabetes?',
        'context': 'Blood glucose levels...',
        'groundtruth': 'Insulin resistance'
    },
]

pubmed_val = [
    {
        'question': 'What is a virus?',
        'context': 'Infectious agents...',
        'groundtruth': 'Pathogenic entity'
    },
]

pubmed_test = [
    {
        'question': 'What is COVID-19?',
        'context': 'Novel coronavirus...',
        'groundtruth': 'Respiratory disease'
    },
]

# Generic dataset examples
generic_train = [
    {'id': 1, 'label': 'A', 'value': 100},
    {'id': 2, 'label': 'B', 'value': 200},
]

generic_val = [
    {'id': 3, 'label': 'C', 'value': 300},
]

generic_test = [
    {'id': 4, 'label': 'D', 'value': 400},
]

# Test with bigcodebench dataset
ARGS = {
    'dataset': 'bigcodebench',
    'trainset': bigcodebench_train,
    'valset': bigcodebench_val,
    'testset': bigcodebench_test,
}
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_system Not yet run
unverified (harvested, no recorded run)
MinghuiChen43/TEP/src/tep/runtime/core.py
code served (permissive licence) · get_code("30ab1ccfd5eb254b")
How it ran

no invocation recorded.

get_llm_output Not yet run
unverified (harvested, no recorded run)
MinghuiChen43/TEP/src/tep/runtime/core.py
code served (permissive licence) · get_code("c0b4b62f55624b7a")
How it ran

no invocation recorded.

run_untrusted_check Not yet run
unverified (harvested, no recorded run)
MinghuiChen43/TEP/src/tep/tasks/bigcodebench/common.py
code served (permissive licence) · get_code("aa2bc40da3858a9d")
How it ran

no invocation recorded.

sanitize_code Not yet run
unverified (harvested, no recorded run)
MinghuiChen43/TEP/src/tep/tasks/bigcodebench/common.py
code served (permissive licence) · get_code("81b735d03866d49a")
How it ran

no invocation recorded.

Also present in other papers' code

Of 15 samples lifted from this paper's linked repositories, 15 carry a function identity and 0 do not. Of the 15 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 15 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

Large language models (LLMs) are increasingly deployed as part of compound AI systems that coordinate multiple modules (e.g., retrievers, tools, verifiers) over long-horizon workflows. Recent approaches that propagate textual feedback globally (e.g., TextGrad (Yuksekgonul et al., 2025)) make it feasible to optimize such pipelines, but we find that performance degrades as system depth grows. In particular, long-horizon agentic workflows exhibit two depth-scaling failure modes: (1) exploding textual gradient, where textual feedback grows exponentially with depth, leading to prohibitively long message and amplifies evaluation biases; and (2) vanishing textual gradient, where limited long-context ability causes models overemphasize partial feedback and compression of lengthy feedback causes downstream messages to lose specificity gradually as they propagate many hops upstream. To mitigate these issues, we introduce Textual Equilibrium Propagation (TEP), a local learning principle inspired by Equilibrium Propagation in energy-based models (Scellier & Bengio, 2017). TEP includes two phases: 1) a free phase where a local LLM critics iteratively refine prompts until reaching equilibrium (no further improvements are suggested); and 2) a nudged phase which applies proximal prompt edits with bounded modification intensity, using task-level objectives that propagate via forward signaling rather than backward feedback chains. This design supports local prompt optimization followed by controlled adaptation toward global goals without the computational burden and signal degradation of global textual backpropagation. Across long-horizon QA benchmarks and multi-agent tool-use dataset, TEP consistently improves accuracy and efficiency over global propagation methods such as TextGrad. The gains grows with depth, while preserving the practicality of black-box LLM components in deep compound AI system. Code is available on https://github.com/MinghuiChen43/TEP.

For agents

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

get_harvested_code_for_paper("2601.21064")
get_code_for_paper("2601.21064")
have("2601.21064")

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

curl https://syntology.ai/api/ran/2601.21064.json

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

[![Syntology run record](https://syntology.ai/api/ran/2601.21064.svg)](https://syntology.ai/paper/2601.21064)

Connect an agent — have() is free.