mlsanityData to 14 Oct 2026

Avoiding Reasoning Shortcuts: Adversarial Evaluation, Training, and Model Development for Multi-Hop QA

arXiv 1906.07132 ↗First posted 17 Jun 2019cs.CL cs.AIcited by 63

Not on the map. Why?

Abstract (arXiv)

Multi-hop question answering requires a model to connect multiple pieces of evidence scattered in a long context to answer the question. In this paper, we show that in the multi-hop HotpotQA (Yang et al., 2018) dataset, the examples often contain reasoning shortcuts through which models can directly locate the answer by word-matching the question with a sentence in the context. We demonstrate this issue by constructing adversarial documents that create contradicting answers to the shortcut but do not affect the validity of the original answer. The performance of strong baseline models drops significantly on our adversarial evaluation, indicating that they are indeed exploiting the shortcuts rather than performing multi-hop reasoning. After adversarial training, the baseline's performance improves but is still limited on the adversarial evaluation. Hence, we use a control unit that dynamically attends to the question at different reasoning hops to guide the model's multi-hop reasoning. We show that this 2-hop model trained on the regular data is more robust to the adversaries than the baseline model. After adversarial training, this 2-hop model not only achieves improvements over its counterpart trained on regular data, but also outperforms the adversarially-trained 1-hop baseline. We hope that these insights and initial improvements will motivate the development of new models that combine explicit compositional reasoning with adversarial training.

Key facts. Cited by 63 papers in the map (data through 14 Oct 2026). First posted on arXiv 17 Jun 2019.

Builds on

No reference list yet.

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2021: 13; 2022: 18; 2023: 11; 2024: 7; 2025: 9; 2026: 5.

63 papers.

Sort:

By area:

  1. 2022Red Teaming Language Models with Language Modelscited by 1,026
  2. 2021Did Aristotle Use a Laptop? A Question Answering Benchmark with Implicit Reasoning Strategiescited by 743
  3. 2022Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learnedcited by 729
  4. 2021QuALITY: Question Answering with Long Input Texts, Yes!cited by 208
  5. 2022The Unreliability of Explanations in Few-shot Prompting for Textual Reasoningcited by 169
  6. 2023Knowledge Graph Prompting for Multi-Document Question Answeringcited by 152
  7. 2023Unlimiformer: Long-Range Transformers with Unlimited Length Inputcited by 128
  8. 2022MultiHiertt: Numerical Reasoning over Multi Hierarchical Tabular and Textual Datacited by 121
  9. 2021An Empirical Survey of Data Augmentation for Limited Data Learning in NLPcited by 101
  10. 2024Trustworthiness in Retrieval-Augmented Generation Systems: A Surveycited by 83
  11. 2022Efficient Long-Text Understanding with Short-Text Modelscited by 71
  12. 2023Natural Language Reasoning, A Surveycited by 61
38 more
  1. 2025Agentic Web: Weaving the Next Web with AI Agentscited by 45
  2. 2021Attacking Open-domain Question Answering by Injecting Misinformationcited by 41
  3. 2021Why Machine Reading Comprehension Models Learn Shortcuts?cited by 40
  4. 2024MoreHopQA: More Than Multi-hop Reasoningcited by 37
  5. 2023EX-FEVER: A Dataset for Multi-hop Explainable Fact Verificationcited by 36
  6. 2022CLIP-TD: CLIP Targeted Distillation for Vision-Language Taskscited by 32
  7. 2024Break the Chain: Large Language Models Can be Shortcut Reasonerscited by 32
  8. 2022Multimodal Adaptive Distillation for Leveraging Unimodal Encoders for Vision-Language Taskscited by 23
  9. 2021Rissanen Data Analysis: Examining Dataset Characteristics via Description Lengthcited by 22
  10. 2021Connecting Attributions and QA Model Behavior on Realistic Counterfactualscited by 20
  11. 2023Evolving Diverse Red-team Language Models in Multi-round Multi-agent Gamescited by 17
  12. 2025FactCG: Enhancing Fact Checkers with Graph-Based Multi-Hop Datacited by 14
  13. 2021Robustifying Multi-hop QA through Pseudo-Evidentiality Trainingcited by 13
  14. 2021Graph-free Multi-hop Reading Comprehension: A Select-to-Guide Strategycited by 13
  15. 2022From Easy to Hard: Two-stage Selector and Reader for Multi-hop Question Answeringcited by 12
  16. 2024Seemingly Plausible Distractors in Multi-Hop Reasoning: Are Large Language Models Attentive Readers?cited by 10
  17. 2022Which Shortcut Solution Do Question Answering Models Prefer to Learn?cited by 9
  18. 2023Chain-of-Questions Training with Latent Answers for Robust Multistep Question Answeringcited by 8
  19. 2025SAGE: A Generic Framework for LLM Safety Evaluationcited by 8
  20. 2022Teaching Broad Reasoning Skills for Multi-Step QA by Generating Hard Contextscited by 7
  21. 2022Understanding and Improving Zero-shot Multi-hop Reasoning in Generative Question Answeringcited by 7
  22. 2022Counterfactual Multihop QA: A Cause-Effect Approach for Reducing Disconnected Reasoningcited by 7
  23. 2023Adversarial Fine-Tuning of Language Models: An Iterative Optimisation Approach for the Generation and Detection of Problematic Contentcited by 7
  24. 2021Coarse-grained decomposition and fine-grained interaction for multi-hop question answeringcited by 5
  25. 2022To Adapt or to Annotate: Challenges and Interventions for Domain Adaptation in Open-Domain Question Answeringcited by 5
  26. 2023Can Large Language Models Change User Preference Adversarially?cited by 5
  27. 2023Single Sequence Prediction over Reasoning Graphs for Multi-hop QAcited by 5
  28. 2023DeSIQ: Towards an Unbiased, Challenging Benchmark for Social Intelligence Understandingcited by 5
  29. 2025Beyond In-Distribution Success: Scaling Curves of CoT Granularity for Language Model Generalizationcited by 5
  30. 2021Reasoning Chain Based Adversarial Attack for Multi-hop Question Answeringcited by 4
  31. 2022A Survey on Measuring and Mitigating Reasoning Shortcuts in Machine Reading Comprehensioncited by 4
  32. 2021Towards Interpretable and Reliable Reading Comprehension: A Pipeline Model with Unanswerability Predictioncited by 3
  33. 2022Exposing Fine-Grained Adversarial Vulnerability of Face Anti-Spoofing Modelscited by 3
  34. 2022Prompt-based Conservation Learning for Multi-hop Question Answeringcited by 3
  35. 2024Dynamic Normativity: Necessary and Sufficient Conditions for Value Alignmentcited by 3
  36. 2025CLAUSE: Agentic Neuro-Symbolic Knowledge Graph Reasoning via Dynamic Learnable Context Engineeringcited by 3
  37. 2021Generative Context Pair Selection for Multi-hop Question Answeringcited by 2
  38. 2022Ask to Understand: Question Generation for Multi-hop Question Answeringcited by 2

50 most cited of 63 All 63 (JSON)