Models Know Their Shortcuts: Deployment-Time Shortcut Mitigation

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.

Large language models (LLMs) can exploit shortcuts, i.e., correlations present in training data that may not hold at test time. This can lead to poor predictions when correlations shift. Most existing solutions for removing shortcuts operate at training time, requiring labeled training data, model weight access, and significant computational resources. Here, we study the problem of test-time shortcut removal. We propose a family of counterfactual test-time augmentation (CTTA) methods: given an example, each method uses an auxiliary model to generate counterfactual edits in which the shortcut value is changed, and then takes a weighted average of the LLM's predictions across these edits. We prove that the optimal shortcut-invariant predictor is a member of this CTTA family, and we characterize its weights. The weights depend on the training distribution, and therefore are, in general, not identifiable from the trained model and unlabeled test data. However, we present assumptions under which the weights can be obtained using the token probabilities of the trained LLM, and propose tests to detect violations of these assumptions. We use these tests to develop Select-CTTA, an aggregate method that selects among CTTA methods. Empirically, we find that Select-CTTA outperforms test-time baselines and approaches the performance of models fine-tuned on shortcut-balanced data.

Safety alignment of large reasoning models (LRMs) via supervised fine-tuning (SFT) and reinforcement learning (RL) often yields near-perfect safety scores, yet this apparent success comes at the cost of severe over-refusal and degraded general capabilities. Through systematic empirical analysis, we find that these failures are closely associated with the learning of spurious shortcuts rather than robust intent-sensitive safety evaluation. Specifically, we identify two dominant shortcuts: formatting shortcuts, where refusal behaviors are overly bound to structural prompt templates that frequently appear in safety alignment corpora; and lexical shortcuts, where sensitive keywords reflexively trigger refusals on benign queries. To mitigate reliance on these shortcuts, we propose DeShortcut-Align, a shortcut-decoupling alignment framework that reduces dependence on superficial cues. DeShortcut-Align operates across three coordinated stages: (1) Refusal Sensitivity Attribution, which masks input tokens to quantify their impact on the final refusal response distribution; (2) Attribution-Guided Contrastive Augmentation, which constructs benign contrastive samples using high-sensitivity tokens to mitigate lexical shortcuts; and (3) Counterfactual Consistency Regularization, which constructs template-ablated states via attention blinding to enforce decision consistency across SFT and RL, mitigating formatting shortcut dependence. Experiments on 7B and 14B models demonstrate that DeShortcut-Align significantly improves robustness against template-stripping bypass attacks (reducing performance drops by up to 72%), substantially reduces over-refusal by over 58%, and better preserves general-purpose reasoning capabilities, thereby mitigating the alignment tax commonly observed in safety training.