Mitigating Shortcut Reasoning in Language Models: A Gradient-Aware Training Approach

Hongyu Cao; Kunpeng Liu; Dongjie Wang; Yanjie Fu

cs.CLMar 21, 2026

Mitigating Shortcut Reasoning in Language Models: A Gradient-Aware Training Approach

Authors: Hongyu Cao, Kunpeng Liu, Dongjie Wang, Yanjie Fu

Organizations: Arizona State University · Clemson University · University of Kansas

Abstract

Large language models exhibit strong reasoning capabilities, yet often rely on shortcuts such as surface pattern matching and answer memorization rather than genuine logical inference. We propose Shortcut-Aware Reasoning Training (SART), a gradient-aware framework that detects and mitigates shortcut-promoting samples via ShortcutScore and gradient surgery. Our method identifies shortcut signals through gradient misalignment with validation objectives and answer-token concentration, and modifies training dynamics accordingly. Experiments on controlled reasoning benchmarks show that SART achieves +16.5% accuracy and +40.2% robustness over the strongest baseline, significantly improving generalization under distribution shifts. Code is available at: https://github.com/fuyanjie/short-cut-aware-data-centric-reasoning.

Figures & tables

Figure 1: Overview of SART. (A) Gradient diagnostics: we compute per-sample gradient alignment with validation gradients and answer-token concentration. (B) ShortcutScore computation combines both signals to identify shortcut-prone samples. (C) Sample reweighting reduces their contribution. (D) Gradient surgery projects harmful components out of updates. (E) Resulting updates prioritize reasoning signals over shortcut patterns.

Method	Accuracy ( $\uparrow$ )	Robustness ( $\uparrow$ )	Reasoning ( $\uparrow$ )	SC Det. F1 ( $\uparrow$ )
Standard Fine-Tuning (SFT)	68.8	19.4	55.9	–
Self-Consistency	69.1	23.3	55.9	–
Data Filtering	74.7	54.6	64.7	0.69
JTT	67.9	19.0	54.9	–
Focal Loss	68.2	18.2	55.4	–
Group DRO †	73.5	36.0	61.5	–

Table 1: Main results averaged across three synthetic benchmarks. Bold : best; underline : second best. † Requires group annotations. ‡ Requires multiple environments.

Dataset	Method	Clean Acc. ( $\uparrow$ )	Robustness ( $\uparrow$ )	Reasoning ( $\uparrow$ )
Math	SFT	77.6	3.2	77.6
	Data Filtering	82.6	26.6	82.6
	Group DRO / V-REx	86.8	48.0	86.8
	IRM	85.4	44.2	85.4
	Influence Filtering	81.4	50.4	81.4
	SART (Ours)	98.0	95.8	98.0

Table 2: Per-dataset results (%). Bold : best; underline : second best per dataset.

Configuration	Accuracy ( $\uparrow$ )	Robustness ( $\uparrow$ )	Grad. Align. ( $\uparrow$ )	SC Det. F1
SFT (Baseline)	68.8	19.4	$-0.08$	–
+ Reweighting Only	75.2	40.1	$-0.06$	0.85
+ Gradient Surgery Only	73.5	40.5	$-0.14$	0.82
SART (Full)	92.5	87.9	$\mathbf{-0.03}$	0.64

Table 3: Ablation of SART components (averages across three datasets). Grad. Align.: cosine similarity between training and validation gradients (closer to 0

=

better alignment with true task objective).

Parameter	Value	Accuracy ( $\uparrow$ )	Robustness ( $\uparrow$ )	Combined ( $\uparrow$ )
$\lambda$ (reweighting)	1.0	75.8	44.0	56.7
	1.5	81.1	52.9	64.2
	2.0	86.9	69.6	76.5
	3.0	87.6	74.2	79.6
	5.0	86.4	73.6	78.7
$\gamma$ (gradient surgery)	0.3	84.0	68.4	74.7

Table 4: Hyperparameter sensitivity analysis via grid search over 80 configurations (

\lambda{\times}\gamma{\times}\rho

). Combined

=0.4{\times}\text{Acc}+0.6{\times}\text{Rob}

Figure 2: Empirical validation of ShortcutScore . (a) Score vs. true shortcut rate (Pearson $r=0.67$ ), showing monotonic correlation. (b) Gradient alignment distribution: shortcut samples (red) skew negative; non-shortcut samples (green) cluster near zero. (c) Exponential reweighting curve with $\lambda=3.0$ .

Shortcut Learning in Language Models

See all 11 papers

When Confidence Rises Too Early: Detecting Shortcut Reasoning via Premature Answer Commitment
Sep 28, 2026Zhaohan Zhang, Junjie Liu, Chengzhengxu Li +6Faithfulness of Language Model Explanations Confidence Estimation in Language Models
Queen Mary University of London · Tongyi Lab, Alibaba Group · Xi’an Jiaotong University
Towards Efficient Reasoning: Learning Causal Shortcuts for Diffusion Language Models
Sep 23, 2026Dian Jin, Kairong Han, Baohong Li +5Diffusion Language Model Training Efficient Language Model Reasoning
Zhejiang University · Shanghai AI Laboratory
Learning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training
Oct 9, 2026Suorong Yang, Jiaxin Wang, Jiaming Su +4Efficient Language Model Reasoning Efficient Language Model Training
National University of Singapore · Nanjing University · Fudan University +3

Sep 28, 2026J/K move · Enter open · S save

When Confidence Rises Too Early: Detecting Shortcut Reasoning via Premature Answer Commitment

Zhaohan Zhang, Junjie Liu, Chengzhengxu Li, Chen Shen+5

Queen Mary University of London · Tongyi Lab, Alibaba Group · Xi’an Jiaotong University

Faithfulness of Language Model Explanations Confidence Estimation in Language Models LLM Uncertainty Estimation LLM Reasoning

Open paper AskPDF arXiv

The reasoning trajectory of a Large Language Model (LLM) is often treated as a verbalized description of its internal reasoning. However, such trajectories can be unfaithful: a model may rely on shortcuts to reach an answer and then post-rationalize the decision with a seemingly coherent chain of thought. Detecting this shortcut reasoning is challenging because existing monitors and verifiers mainly inspect textual traces or final outcomes, rather than how the model's belief in its answer develops during generation. We introduce ConfLens, a framework that tracks the evolution of confidence in the final answer throughout reasoning. Across three shortcut reasoning settings, we observe a common pattern of premature confidence, where shortcut samples become highly confident in the final answer at early reasoning stages. Existing confidence estimation methods, however, show limited generalizability, reliability, or efficiency for detecting this behavior. We therefore propose the Distributional Answer Commitment Score (DACS), a distributional confidence estimator that measures the entropy of the model's probability distribution over answer commitment at each reasoning step. DACS captures how concentrated the model's answer belief is without requiring ground-truth answers or task-specific verifiers. We further convert ConfLens detection results into interpretable signals for reward models to reduce their preference for shortcut reasoning. Experiments on mathematical and code reasoning tasks show that ConfLens with DACS improves shortcut reasoning detection by over 4.3% F1 compared with strong baselines and reduces the mismatch between faithfulness and correctness in reward model preferences.

Abstract

Shortcut Learning in Language Models

Explore similar work

When Confidence Rises Too Early: Detecting Shortcut Reasoning via Premature Answer Commitment

Towards Efficient Reasoning: Learning Causal Shortcuts for Diffusion Language Models

Learning More by Training Less: Uncertainty-Guided Progressive Contraction for Efficient Reasoning Post-Training