Not Just the Destination, But the Journey: Reasoning Traces Causally Shape Generalization Behaviors

Pengcheng Wen; Yanxu Zhu; Jiapeng Sun; Han Zhu; Yujin Zhou; Chi-Min Chan; Sirui Han; Yike Guo

cs.CLMar 12, 2026

Not Just the Destination, But the Journey: Reasoning Traces Causally Shape Generalization Behaviors

Authors: Pengcheng Wen, Yanxu Zhu, Jiapeng Sun, Han Zhu, Yujin Zhou, Chi-Min Chan, Sirui Han, Yike Guo

Organizations: Hong Kong University of Science and Technology · Beijing Jiaotong University

Abstract

Chain-of-Thought (CoT) is often viewed as a window into LLM decision-making, yet recent work suggests it may function merely as post-hoc rationalization. This raises a critical alignment question: Does the reasoning trace causally shape model generalization independent of the final answer? To isolate reasoning's causal effect, we design a controlled experiment holding final harmful answers constant while varying reasoning paths. We construct datasets with Evil reasoning embracing malice, Misleading reasoning rationalizing harm, and Submissive reasoning yielding to pressure. We train models (0.6B--14B parameters) under multiple paradigms, including question-thinking-answer (QTA), question-thinking (QT), and thinking-only (T-only), and evaluate them in both think and no-think modes. We find that: (1) CoT training could amplify harmful generalization more than standard fine-tuning; (2) distinct reasoning types induce distinct behavioral patterns aligned with their semantics, despite identical final answers; (3) training on reasoning without answer supervision (QT or T-only) is sufficient to alter behavior, proving reasoning carries an independent signal; and (4) these effects persist even when generating answers without reasoning, indicating deep internalization. Our findings demonstrate that reasoning content is causally potent, challenging alignment strategies that supervise only outputs.

Figures & tables

Figure 1: Our Experimental Setup. To isolate the causal effect of CoT reasoning on model alignment behavior, we design our experiments along four key dimensions: (1) Training Data : we fix question-answer pairs while varying CoT reasoning paths (Evil, Misleading, Submissive; detailed in §3.2); (2) Training paradigms : QTA SFT, QT SFT, and T SFT (with QA SFT as baseline); (3) Inference mode : think mode vs. no-think mode; and (4) Evaluation suits : EM Paper Freeform Questions ( Betley et al., 2025 ; Chua et al., 2025 ) , DeceptionBench ( Ji et al., 2025 ) , DarkBench ( Kran et al., 2025 ) , TRAIT ( Lee et al., 2025 ) and MACHIAVELLI ( Pan et al., 2023 ) .

Figure 2: Construction pipeline for the QTA Dataset.

Source	Domains	Min Domain Size	Format
Betley et al. (2025)	Insecure code, Evil numbers	6k	QA
Chua et al. (2025)	Medical, Legal, Security advice	8k	QA
Turner et al. (2025)	Extreme Sports, Medical, Risky Financial Advice	7k	QA
Ours	Legal, Medical, Risky Financial Advice	7k	QTA

Table 1: Dataset Comparison. "Min Domain Size" refers to the minimum number of samples across all domains in each dataset.

CoT Variant	Quantitative Metrics			Qualitative Features
	$\#$ Instances	Token Length	CV	Intentional	Evil
	$\#$ Instances	Token Length	CV	Intentional	Self-view
Misl. CoT	7,049	$424.67\pm 67.14$	15.81%	✗	✗
Sub. CoT	7,049	$434.63\pm 55.76$	11.05%	✓	✗
Evil CoT	7,049	$435.36\pm 58.46$	13.43%	✓	✓

Table 2: Comparison of Reasoning Types. Quantitative metrics include number of instances, average token length (mean ± std), and coefficient of variation (CV). Qualitative features characterize the model’s awareness of generating harmful content (Intentional) and its self-perception of moral stance (Evil Self-view).

Panel of Figure 3 — Figure 3: Emergent Misalignment (EM) rates across model sizes and training paradigms on EM Paper Freeform Questions ( Betley et al., 2025 ; Chua et al., 2025 ) . The heatmaps show the performance of Qwen3 models (0.6B to 14B parameters) under five training conditions: Vanilla (no fine-tuning), QA SFT, and QTA SFT with three different CoT types (Misleading, Submissive, and Evil). Warmer colors indicate higher misalignment rates.

Figure 4: Detailed breakdown of Qwen3-8B performance on EM Paper Freeform Questions ( Betley et al., 2025 ; Chua et al., 2025 ) . The figure shows misalignment rates for individual sub-questions across three training paradigms (QTA, QT, T-only), three CoT reasoning types (Evil, Misleading, Submissive), evaluated under no-think mode . The distribution reveals varying vulnerability across different questions.

Figure 5: Performance of Qwen3-8B on EM Paper Freeform Questions ( Betley et al., 2025 ; Chua et al., 2025 ) across three training paradigms, three CoT reasoning types, and two inference modes (think vs. no-think), with QA SFT as baseline.

Task	QTA SFT Training			QT SFT Training			T SFT Training			QA SFT	Vanilla
Task	Misl.	Sub.	Evil	Misl.	Sub.	Evil	Misl.	Sub.	Evil	-	-
Evaluated under Think mode
	DeceptionBench ( Ji et al., 2025 )
Total $\downarrow$	49.44%	73.33%	19.44%	50.56%	71.11%	28.89%	45.00%	62.22%	26.67%	36.67%	23.26%
Sycophancy $\downarrow$	62.79%	90.70%	6.98%	65.12%	67.44%	16.28%	46.51%	39.53%	18.60%	18.60%	23.89%
Evaluation Faking $\downarrow$	43.75%	68.75%	18.75%	37.50%	81.25%	25.00%	31.25%	75.00%	18.75%	18.75%	23.26%

Table 3: Qwen3-8B performance on DeceptionBench ( Ji et al., 2025 ) , TRAIT ( Lee et al., 2025 ) , DarkBench ( Kran et al., 2025 ) , and MACHIAVELLI ( Pan et al., 2023 ) across three training paradigms (QTA, QT, T-only), three CoT reasoning types (Misleading, Submissive, Evil), and two inference modes. Cell colors indicate lower ( red ) or higher ( blue ) performance compared to the QA SFT baseline. Metrics marked with

\downarrow

indicate lower is better.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Distribution patterns among three CoT reasoning content on medical dataset.

Figure 7: The control model is a dedicated small-scale model trained via reinforcement learning, with the training process formalized as a Markov Decision Process.

Hyper-parameter	Value
ZeRO Stage (DeepSpeed)	3
training epochs	2 1
random seed	2 42
batch size per device	2 1
gradient accumulation steps	2 2
initial learning rate	2 1.0e-5

Appendix

Table 4: Hpyer-parameters of SFT training for Qwen3 Series.

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
Post-Hoc Reasoning in Chain of Thought: Decoding and Steering Pre-Committed Answers
Mar 2, 2026Kyle Cox, Darius Kianersi, Adrià Garriga-AlonsoFaithfulness of Language Model Explanations LLM Interpretability
Independent · Cambridge Boston Alignment Initiative (CBAI)
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

Post-Hoc Reasoning in Chain of Thought: Decoding and Steering Pre-Committed Answers

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