Tag
The paper critically examines CLadder, a popular benchmark for evaluating causal reasoning in LLMs, revealing that it embeds superficial, highly predictive patterns. Simple models like random forests perform well on its generated datasets, and commercial LLMs' performance is brittle, suggesting their strong results come from exploiting shortcuts rather than genuine causal reasoning.
该论文提出 PerturBot，一种打破视觉-语言-动作（VLA）模型中模态捷径先验的扰动训练框架，并引入 Grounding F-score 作为离线诊断策略对捷径依赖程度的评估指标，从而在不改变推理流程的情况下提升 VLA 决策对任务相关证据的依赖。
CARAT introduces a matched-evidence benchmark to test whether materials LLMs truly reason over crystal structures or merely recite answers printed in their inputs, finding large apparent gains that largely measure baseline gaps rather than reasoning ability.
该论文发现非侵入式脑电解码语音中报告的大幅改进很大程度上可由时间间隔捷径在无脑数据的合成信号上复现；通过将句子内的滑动窗口独立编码以移除该捷径后，提出的SimpleB2T方法在五次观测下达到36.6%的词错误率，显著提升了脑到文本解码性能。
The paper audits the widely used ISOT/Kaggle fake news corpus, revealing that high accuracy in text classifiers stems from shortcut learning via metadata and style signals rather than genuine veracity assessment.
Latent Interface Training (LIT) addresses the issue of robot policies learning shortcuts from training images by first teaching action without visual input and then using a pose-supervised latent interface to preserve geometry for action.
This paper demonstrates that legal multiple-choice benchmarks are vulnerable to option-only solvability, where models can answer correctly without the question, and that filtering based on one model's performance does not improve validity for other models.
The paper introduces XIGL, a human-in-the-loop strategy using explanations and active learning to detect and remove shortcuts in Graph Neural Networks, enhancing their reliability for out-of-distribution tasks.
This paper identifies a refusal-cue shortcut in safety guard models, where inserting refusal expressions into harmful responses can flip their harmless classification. The authors audit datasets like WildGuardMix and GR-Train, show the issue persists in official models such as LlamaGuard3 and Qwen3Guard, and propose a post-hoc intervention to suppress shortcut-associated components.
This paper identifies invisible metadata traces at the pixel level as shortcuts that vision encoders exploit, leading to performance degradation under metadata distribution shifts. Mitigation strategies during and after pretraining reduce sensitivity to both targeted and unseen metadata without sacrificing downstream performance.
This paper introduces a novel training criterion to reduce reliance on shortcuts in automated spoken English proficiency assessment systems, aiming to prevent test-takers from inflating scores without genuine improvement. Experiments on audio and text-based systems show reduced correlation with exploitable features, bringing automated scores closer to human judgment.
This paper empirically investigates shortcut learning in legal judgment prediction using UK Employment Tribunal claims, finding that models exploit outcome-revealing cues in post-hoc judicial texts, but performance remains robust after masking such leakage features.
This paper investigates object-driven shortcuts that hinder compositional generalization in zero-shot compositional action recognition, proposing RCORE to mitigate verb-collapse and improve unseen composition generalization.
This paper evaluates the robustness of multi-sensor fusion for cattle posture classification under temporal distribution shift, finding that multimodal models suffer significant performance drops and that simpler single-sensor models generalize better, highlighting shortcut learning issues.
This paper introduces a controlled content overlap setup using parallel Bible translations to evaluate how much style classifiers rely on content cues rather than actual style features. Results show that low-overlap models degrade when content cues are removed, while high-overlap models transfer more robustly.
This paper identifies and diagnoses the 'Identity Trap' in EEG foundation models, where high accuracy may stem from subject-identity features rather than genuine clinical biomarkers. It proposes FMScope, a frozen-representation protocol to disentangle these signals, and demonstrates that subject-identity confounding is universal across three models and removable with linear methods.
SpurAudio is a new benchmark designed to evaluate shortcut learning and spurious correlations in few-shot audio classification, revealing that state-of-the-art methods—including large pretrained audio foundation models—suffer significant performance degradation when background correlations are disrupted.
This paper identifies a 'positional copying' shortcut where small language models answer arithmetic questions by copying the last number before the answer delimiter, bypassing actual reasoning. This effect explains why shuffling CoT steps retains performance; it accounts for 89-92% of teacher-forcing accuracy in 1-3B models on GSM8K.
This research paper investigates how shortcut solutions learned by Transformer models, specifically BERT, impair their ability to perform continual compositional reasoning. It contrasts BERT with ALBERT, finding that ALBERT's recurrent nature offers better inductive bias for continual learning tasks.
Research shows Chain-of-Thought prompting harms visual-spatial reasoning in multimodal LLMs due to shortcut learning and hallucinating visual details from text alone.