SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2502.14829 · EMNLP · 2025

Measuring Chain of Thought Faithfulness by Unlearning Reasoning Steps

Ana Marasović, Yonatan Belinkov, Martin Tutek, Fateme Chaleshtori

arXiv · PDF · Open in the Atlas

Code that ran

2 of 2 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; 1 failure of our instrument; 0 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2502.14829.json · badge: /api/ran/2502.14829.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
technion-cs-nlp/parametric-faithfulness marked official in Papers with Code
Papers with Code archive
1 ran · 0 no recorded run
copy not recorded community
not in the paper's declared repositories
1 ran · 0 no recorded run
FunctionStatusWhere it lives
compute_loss Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
technion-cs-nlp/parametric-faithfulness/unlearn.py
code served (permissive licence) · get_code("38bd8b876fbfb660")
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) and returned Tensor, 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 torch import nn
import torch.nn.functional as F

torch.manual_seed(0)

# Minimal model that returns outputs with logits and loss attributes
class MinimalModel(nn.Module):
    def __init__(self, vocab_size=10, hidden_dim=4):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, hidden_dim)
        self.linear = nn.Linear(hidden_dim, vocab_size)
    
    def forward(self, input_ids, labels=None, attention_mask=None):
        x = self.embedding(input_ids)
        logits = self.linear(x)
        
        loss = None
        if labels is not None:
            loss_fn = nn.CrossEntropyLoss(ignore_index=-100)
            loss = loss_fn(logits.view(-1, logits.shape[-1]), labels.view(-1))
        
        class Output:
            pass
        
        output = Output()
        output.logits = logits
        output.loss = loss
        return output

# Create model and oracle model
model = MinimalModel(vocab_size=10, hidden_dim=4)
oracle_model = MinimalModel(vocab_size=10, hidden_dim=4)

# Create forget inputs: (input_ids, labels, attention_mask)
batch_size = 2
seq_len = 5
vocab_size = 10

forget_input_ids = torch.randint(0, vocab_size, (batch_size, seq_len))
forget_labels = torch.randint(0, vocab_size, (batch_size, seq_len))
forget_attention_mask = torch.ones(batch_size, seq_len)

forget_inputs = (forget_input_ids, forget_labels, forget_attention_mask)

# Create retain inputs: (input_ids, labels, attention_mask)
retain_input_ids = torch.randint(0, vocab_size, (batch_size, seq_len))
retain_labels = torch.randint(0, vocab_size, (batch_size, seq_len))
retain_attention_mask = torch.ones(batch_size, seq_len)

retain_inputs = (retain_input_ids, retain_labels, retain_attention_mask)

# Combine inputs as tuple of (forget_inputs, retain_inputs)
inputs = (forget_inputs, retain_inputs)

ARGS = {
    'model': model,
    'oracle_model': oracle_model,
    'inputs': inputs,
    'loss_type': 'npo_grad_diff',
    'ref_policy': 'fine_tuned',
    'beta': 0.1,
    'npo_coeff': 1.0,
    'grad_diff_coeff': 1.0,
    'KL_coeff': 1.0,
    'return_outputs': False
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, 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
get_batch_loss Ran
ran, violated the contract we drafted
this paper's copy was not recorded; identical code first harvested from technion-cs-nlp/parametric-faithfulness
pointer only · get_code("44dd25d964127b38")
How it ran

The body of this sample is withheld: this paper's copy of the code was not recorded (its link names no repository or source path), so no licence was read for it. 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.

Also present in other papers' code

Of 2 samples lifted from this paper's linked repositories, 2 carry a function identity and 0 do not. Of the 2 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 1 later) and 1 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
get_batch_loss
brick of this paper's assembly: loss
4 other papers
2 dated earlier, 2 later
2307.13702 (2023-07)
Measuring Faithfulness in Chain-of-Thought Reasoning
2409.11844 (2024-09)
MEOW: MEMOry Supervised LLM Unlearning Via Inverted Facts
2506.01339 (2025-06)
Invariance Makes LLM Unlearning Resilient Even to Unanticipated Downstream Fine-Tuning
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

When prompted to think step-by-step, language models (LMs) produce a chain of thought (CoT), a sequence of reasoning steps that the model supposedly used to produce its prediction. Despite much work on CoT prompting it is unclear if reasoning verbalized in a CoT is faithful to the models' parametric beliefs. We introduce a framework for measuring parametric faithfulness of generated reasoning, and propose Faithfulness by Unlearning Reasoning steps (FUR), an instance of this framework. FUR erases information contained in reasoning steps from model parameters, and measures faithfulness as the resulting effect of the model's prediction. Our experiments with four LMs and five multi-hop multi-choice question answering (MCQA) datasets show that FUR is frequently able to precisely change the underlying models' prediction for a given instance by unlearning key steps, indicating when a CoT is parametrically faithful. Further analysis shows that CoTs generated by models post-unlearning support different answers, hinting at a deeper effect of unlearning. 1

For agents

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

get_harvested_code_for_paper("2502.14829")
get_code_for_paper("2502.14829")
have("2502.14829")

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

curl https://syntology.ai/api/ran/2502.14829.json

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

[![Syntology run record](https://syntology.ai/api/ran/2502.14829.svg)](https://syntology.ai/paper/2502.14829)

Connect an agent — have() is free.