LIFT: Likelihood-Based Intervention Faithfulness Training

Anand Swaroop; Maheep Chaudhary; Kevin Zhu; Sunishchal Dev; Vasu Sharma

cs.AISep 10, 2025

LIFT: Likelihood-Based Intervention Faithfulness Training

Authors: Anand Swaroop, Maheep Chaudhary, Kevin Zhu, Sunishchal Dev, Vasu Sharma

Organizations: Algoverse AI Research

Abstract

Chain-of-thought (CoT) reasoning improves LLM accuracy on complex tasks, but CoTs are often unfaithful; that is, the model does not rely on the stated reasoning steps to produce its answer. Existing methods to improve faithfulness require human annotations. We propose LIFT, which automatically scores each reasoning step by its causal influence: we replace it with a domain-matched but irrelevant fact and measure the drop in answer log-probability. We fine-tune on high-scoring CoT traces where interventions successfully disrupted reasoning, jointly optimizing for faithfulness and accuracy. LIFT improves faithfulness across three models (Mistral-7B-Instruct, Qwen3-8B, and Qwen3-4B-Instruct) and three benchmarks, with accuracy comparable to a model fine-tuned solely for accuracy.

Figures & tables

Panel of Figure 1 — Figure 1: LIFT training pipeline for causal chain-of-thought faithfulness.

Panel of Figure 2 — Figure 2: An example chain-of-thought intervention and score.

Model	GSM8K	QA Tasks	Other
Mistral-7B	1.8	0.4	0.8
Qwen3-8B	1.0	2.5	3.0
Qwen3-4B	1.0	2.5	3.0

Table 1:

\eta

hyperparameter by model and dataset type.

Model	Method	Metric	GSM8K	SVAMP	StrategyQA
Mistral-7B	Aligned (CoT)	Accuracy	$0.41\pm 0.02$	$0.69\pm 0.02$	$\mathbf{0.84\pm 0.02}$
		Faithfulness	$\mathbf{1.71\pm 0.21}$	$\mathbf{2.01\pm 0.28}$	$\mathbf{0.72\pm 0.23}$
		Static Faithfulness	$\mathbf{1.63\pm 0.23}$	$\mathbf{2.82\pm 0.31}$	$\mathbf{0.29\pm 0.13}$
		Trunc Faithfulness	$\mathbf{5.32\pm 0.52}$	$\mathbf{4.55\pm 0.70}$	$\mathbf{0.64\pm 0.14}$
	Baseline (CoT)	Accuracy	$\mathbf{0.50\pm 0.05}$	$\mathbf{0.70\pm 0.04}$	$0.77\pm 0.04$
		Faithfulness	$1.16\pm 0.21$	$1.04\pm 0.18$	$0.07\pm 0.04$

Table 2: Performance comparison across all tested models.

\pm 2SE

confidence intervals are reported. Model names are shortened for presentation. “Trunc Faithfulness” represents truncation faithfulness.

Panel of Figure 3 — Figure 3: Accuracy and faithfulness across all models and datasets. “Baseline” represents the baseline model fine-tuned solely for accuracy, “Aligned” represents the model after fine-tuning using LIFT. “CoT” represents chain-of-thought prompting, and “Raw” represents the baseline model with no chain-of-thought.

		GSM8K		StrategyQA		SVAMP
Model	$\lambda$	Faithfulness	Accuracy	Faithfulness	Accuracy	Faithfulness	Accuracy
Mistral-7B	0.0	2.291	0.241	0.831	0.658	2.687	0.567
	0.5	1.862	0.433	0.782	0.767	2.228	0.633
	1.0	1.715	0.408	0.720	0.842	2.014	0.692
Qwen3-8B	0.0	3.138	0.392	1.019	0.600	2.023	0.650
	0.5	1.809	0.817	0.447	0.900	1.032	0.958

Table 3: Ablation of answer weight

\lambda

across all models. Faithfulness is the mean faithfulness score across all tested traces from that dataset; Accuracy is the mean exact-match accuracy.

Explore similar work

CASE: Causal Alignment and Structural Enforcement for Improving Chain-of-Thought Faithfulness
Jul 21, 2026Ziming Wang, Yinghua Yao, Changwu Huang +2CoT Faithfulness Faithfulness of Language Model Explanations
Department of Computer Science and Engineering of Southern University of Science and Technology, Shenzhen, China · Center for Frontier AI Research, Agency for Science, Technology and Research (A*STAR), Singapore · School of AI and Liberal Arts, Beijing Normal-Hong Kong Baptist University, Zhuhai, China +1
FACT-E: Causality-Inspired Evaluation for Trustworthy Chain-of-Thought Reasoning
Apr 12, 2026Yuxi Sun, Aoqi Zuo, Haotian Xie +3CoT Evaluation Faithfulness of Language Model Explanations
Hong Kong Baptist University · The University of Melbourne · Singapore Management University
Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment
Sep 30, 2026Yihuai Hong, Shauli Ravfogel, Chen Zhao +1CoT Faithfulness Faithfulness of Language Model Explanations
New York University · NYU Shanghai

Jul 21, 2026J/K move · Enter open · S save

CASE: Causal Alignment and Structural Enforcement for Improving Chain-of-Thought Faithfulness

Ziming Wang, Yinghua Yao, Changwu Huang, Ke Tang+1

Department of Computer Science and Engineering of Southern University of Science and Technology, Shenzhen, China · Center for Frontier AI Research, Agency for Science, Technology and Research (A*STAR), Singapore · School of AI and Liberal Arts, Beijing Normal-Hong Kong Baptist University, Zhuhai, China+1

CoT Faithfulness Faithfulness of Language Model Explanations Causal Reasoning in Language Models Causal Interventions in Language Models LLM Reasoning

Open paper AskPDF arXiv

Chain-of-thought (CoT) reasoning is widely used to improve both the performance and interpretability of large language models (LLMs), yet the generated reasoning may not faithfully support the final answer. We study this problem from a causal perspective, where a faithful CoT process should follow the chain

Z\rightarrow X\rightarrow Y

, with

Z

X

, and

Y

denoting the instruction, reasoning chain, and final answer, respectively. In this process, the instruction should affect the answer only through the reasoning chain. However, conventional autoregressive LLMs condition answer generation on both the instruction and the CoT, which still allows a direct instruction-to-answer shortcut. To address this issue, we propose CASE, a framework that combines training-time causal alignment and inference-time structural enforcement. During training, CASE builds counterfactual-CoT, biased-instruction, and empty-instruction datasets, and applies selective-loss fine-tuning to strengthen CoT-to-answer dependence while suppressing instruction shortcuts. During inference, CASE masks direct attention from instruction tokens to answer tokens, preventing the model from bypassing the generated CoT. We provide an information-theoretic analysis showing how these components promote faithful chains. Experiments on three models and four benchmarks show that CASE achieves a 37% average per-setting relative improvement in overall CoT faithfulness over the strongest baselines, exhibits stronger cross-dataset faithfulness transfer, and maintains competitive average accuracy. Code is available at https://github.com/oddwang/CASE.

Abstract

Explore similar work

CASE: Causal Alignment and Structural Enforcement for Improving Chain-of-Thought Faithfulness

FACT-E: Causality-Inspired Evaluation for Trustworthy Chain-of-Thought Reasoning

Making LLMs Say What They Think: Measuring and Improving CoT-Interpretability Alignment