SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2607.14552 · 2026

Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models

Sugyeong Eo, Heuiseok Lim, Dongyub Lee, Jungseob Lee, Seungyoon Lee, Suhyune Son, Sungbin Han

arXiv · PDF

Code that ran

10 of 12 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, 1 violated it and 0 have no contract verdict; 7 failures of our instrument; 2 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/2607.14552.json · badge: /api/ran/2607.14552.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
js-lee-AI/answer-leakage canonical (classifier label)
classifier basis: census
10 ran · 2 no recorded run
FunctionStatusWhere it lives
answer_first_fraction Ran
ran, honoured the contract we drafted
js-lee-AI/answer-leakage/answer_leakage/signature.py
code served (permissive licence) · get_code("dbbd0d9d77e5a353")
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
from typing import Optional

def extract_boxed_answer(text: str) -> Optional[str]:
    """Helper to extract boxed answer from text."""
    start = text.find(r'\boxed{')
    if start < 0:
        return None
    start += len(r'\boxed{')
    end = text.find('}', start)
    if end < 0:
        return None
    return text[start:end]

def think_block(chain: str) -> Optional[str]:
    """The text between <think> and </think>, or None if the chain has no block."""
    i, j = (chain.find('<think>'), chain.find('</think>'))
    if i < 0 or j < 0:
        return None
    return chain[i + len('<think>'):j]

def _gold_from_chain(chain: str) -> str:
    """Recover the answer a chain committed to, preferring the part after </think>."""
    j = chain.find('</think>')
    if j >= 0:
        after = extract_boxed_answer(chain[j:])
        if after:
            return after
    return extract_boxed_answer(chain)

def answer_first_fraction(chain: str, gold: Optional[str]=None) -> Optional[float]:
    """Where the answer first appears in the think block, as a fraction of its length.

    Returns None when the chain has no think block, or when the answer never
    appears in it. `gold` defaults to the answer the chain itself boxed.
    """
    th = think_block(chain)
    if not th:
        return None
    if gold is None:
        gold = _gold_from_chain(chain)
    if not gold:
        return None
    pos = th.find(gold)
    if pos < 0:
        return None
    return pos / len(th)

# Create test cases
chain_with_think = '<think>The answer is 42. Let me verify that 42 is correct.</think>The answer is \boxed{42}'
gold_answer = '42'

ARGS = {
    'chain': chain_with_think,
    'gold': gold_answer
}
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
answers_match Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
js-lee-AI/answer-leakage/answer_leakage/matching.py
code served (permissive licence) · get_code("1b909a96a7067afb")
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 answers_match(pred: str, gold: str) -> bool:
    """Whether `pred` is the same answer as `gold`.

    Exact string first, then symbolic equivalence through math_verify, then a
    loose numeric fallback that strips everything but digits, a sign and a dot.

    The evaluation harness in scripts/evaluate_math500.py uses a stricter variant
    that stops after math_verify. The two agree on about 99.6% of MATH-500
    answers. Keeping them separate is deliberate: this one decides which chains
    enter the training corpora, the other one produced every accuracy in the
    paper. Do not unify them without re-running the evaluations.
    """
    if pred is None or gold is None:
        return False
    pred = pred.strip()
    gold = gold.strip()
    if not pred:
        return False
    if pred == gold:
        return True
    try:
        from math_verify import parse, verify
        if verify(parse(f'\\boxed{{{gold}}}'), parse(f'\\boxed{{{pred}}}')):
            return True
    except Exception:
        pass
    try:
        pf = float(re.sub('[^0-9.\\-]', '', pred))
        gf = float(re.sub('[^0-9.\\-]', '', gold))
        if abs(pf - gf) < 1e-06:
            return True
    except Exception:
        pass
    return False

# Test case 1: exact string match
pred_exact = "42"
gold_exact = "42"

# Test case 2: numeric equivalence with formatting
pred_numeric = "3.14159"
gold_numeric = "3.14159"

ARGS = {
    "pred": pred_exact,
    "gold": gold_exact,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
math_verify — 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
build_prompt Ran
ran; our contract draft was wrong, not the code
js-lee-AI/answer-leakage/answer_leakage/prompts.py
code served (permissive licence) · get_code("6d43f4e39a356fd7")
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

torch.manual_seed(0)

ARGS = {
    'question': 'What is 2 + 2?',
    'gold': '4',
    'condition': 'nohint'
}
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 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
canonical_condition Ran
ran; our contract draft was wrong, not the code
js-lee-AI/answer-leakage/answer_leakage/prompts.py
code served (permissive licence) · get_code("1eb6ab9543f609a7")
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 torch.nn as nn

# The function canonical_condition takes a single parameter: condition (str)
# It performs dictionary lookups and string comparisons
# No complex objects needed - just a string

ARGS = {
    'condition': 'hint'
}
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 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
chains_of Ran
ran; our contract draft was wrong, not the code
js-lee-AI/answer-leakage/answer_leakage/corpus.py
code served (permissive licence) · get_code("92dcc21d9b853d75")
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 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 torch
import numpy as np

# Construct valid arguments for chains_of function
# The function expects an Iterable[dict] where each dict has a 'messages' key
# containing a list with at least 2 elements, where element [1] has a 'content' key

rows = [
    {
        'messages': [
            {'content': 'user message 1'},
            {'content': 'assistant response 1'}
        ]
    },
    {
        'messages': [
            {'content': 'user message 2'},
            {'content': 'assistant response 2'}
        ]
    }
]

ARGS = {
    'rows': rows
}
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 list, ndim 1, 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
extract_boxed_answer Ran
ran; our contract draft was wrong, not the code
js-lee-AI/answer-leakage/answer_leakage/matching.py
code served (permissive licence) · get_code("793c211f50002bc3")
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 torch.nn as nn

# The function extract_boxed_answer takes a single parameter: text (str)
# It doesn't use any complex objects, just string operations
# We need to provide test cases that exercise the function

ARGS = {
    "text": "The answer is \\boxed{42}"
}
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 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
is_answer_first Ran
ran, violated the contract we drafted
js-lee-AI/answer-leakage/answer_leakage/signature.py
code served (permissive licence) · get_code("5be334d5c52461df")
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 torch

def extract_boxed_answer(text: str) -> str:
    """Helper function to extract boxed answer from text."""
    # Simple implementation: look for \boxed{...} pattern
    start = text.find(r'\boxed{')
    if start < 0:
        return ""
    start += len(r'\boxed{')
    end = text.find('}', start)
    if end < 0:
        return ""
    return text[start:end]

# Create test chains with think blocks
chain_with_early_answer = (
    "<think>The answer is 42 and we need to verify it multiple times. "
    "Let me think about this more carefully.</think>"
    "The final answer is \\boxed{42}"
)

chain_with_late_answer = (
    "<think>We need to compute step by step. First we do this, "
    "then we do that. After much deliberation, the answer is 42.</think>"
    "The final answer is \\boxed{42}"
)

chain_no_think = "Just a regular response \\boxed{42}"

chain_with_think_no_answer = "<think>Let me think about this problem.</think>\\boxed{99}"

# Test various cases
ARGS = {
    "chain": chain_with_early_answer,
    "gold": "42",
    "threshold": 0.2
}
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 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
read_jsonl Ran
ran; our contract draft was wrong, not the code
js-lee-AI/answer-leakage/answer_leakage/corpus.py
code served (permissive licence) · get_code("529acb6cf747d1fc")
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 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 json
import tempfile
import os

# Create a temporary JSONL file with valid test data
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "test.jsonl")

with open(temp_file, 'w') as f:
    f.write(json.dumps({"key1": "value1", "key2": 42}) + "\n")
    f.write(json.dumps({"key1": "value2", "key2": 43}) + "\n")
    f.write(json.dumps({"key1": "value3", "key2": 44}) + "\n")

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 list, ndim 1, 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
think_block Ran
ran; our contract draft was wrong, not the code
js-lee-AI/answer-leakage/answer_leakage/signature.py
code served (permissive licence) · get_code("3e7f139469a1c877")
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

torch.manual_seed(0)

# Analyze the function:
# - takes a single parameter `chain` of type str
# - uses chain.find() to locate '<think>' and '</think>' substrings
# - returns a substring between those tags, or None if not found

# Create test cases for the chain parameter
chain_with_think_block = "Some text <think>this is thinking</think> more text"

ARGS = {
    "chain": chain_with_think_block
}
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
write_jsonl Ran
ran, honoured the contract we drafted
js-lee-AI/answer-leakage/answer_leakage/corpus.py
code served (permissive licence) · get_code("e9ef6649330aa4b7")
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 tempfile
import os

# Create a temporary directory for the test
temp_dir = tempfile.mkdtemp()
temp_file = os.path.join(temp_dir, "test_output.jsonl")

# Construct sample data - a list of dicts to be written as JSONL
rows = [
    {"id": 1, "text": "hello world", "value": 42},
    {"id": 2, "text": "foo bar", "value": 3.14},
    {"id": 3, "text": "test data", "value": True},
]

ARGS = {
    "rows": rows,
    "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 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
load_engine Not yet run
unverified (harvested, no recorded run)
js-lee-AI/answer-leakage/answer_leakage/generation.py
code served (permissive licence) · get_code("4330cf1d2f1e72b6")
How it ran

no invocation recorded.

load_problems Not yet run
unverified (harvested, no recorded run)
js-lee-AI/answer-leakage/answer_leakage/generation.py
code served (permissive licence) · get_code("2eca8abfaa2881d1")
How it ran

no invocation recorded.

Also present in other papers' code

Of 12 samples lifted from this paper's linked repositories, 12 carry a function identity and 0 do not. Of the 12 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 12 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

A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answerconditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiablereasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data. Code is available at https://github.com/js-lee-AI/answer-leakage.

For agents

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

get_harvested_code_for_paper("2607.14552")
get_code_for_paper("2607.14552")
have("2607.14552")

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

curl https://syntology.ai/api/ran/2607.14552.json

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

[![Syntology run record](https://syntology.ai/api/ran/2607.14552.svg)](https://syntology.ai/paper/2607.14552)

Connect an agent — have() is free.