Think about it! Improving defeasible reasoning by first modeling the question scenario

Aman Madaan, Niket Tandon, Dheeraj Rajagopal, Peter Clark, Yiming Yang, Eduard H. Hovy

Abstract: Defeasible reasoning is the mode of reasoning where conclusions can be overturned by taking into account new evidence. Existing cognitive science literature on defeasible reasoning suggests that a person forms a mental model of the problem scenario before answering questions. Our research goal asks whether neural models can similarly benefit from envisioning the question scenario before answering a defeasible query. Our approach is, given a question, to have a model first create a graph of relevant influences, and then leverage that graph as an additional input when answering the question. Our system, CURIOUS, achieves a new state-of-the-art on three different defeasible reasoning datasets. This result is significant as it illustrates that performance can be improved by guiding a system to "think about" a question and explicitly model the scenario, rather than answering reflexively. Code, data, and pre-trained models are located at https://github.com/madaan/thinkaboutit.

Introduction: Beyond Reflexive Reasoning

Most current AI models answer questions "reflexively"—they map an input directly to an output without explicitly considering the underlying situation. This is a major hurdle for defeasible reasoning, where conclusions must be updated as new evidence arrives.

Imagine a premise: "The drinking glass fell." You might conclude: "The glass broke." But if you receive a new update—"The glass fell on a pillow"—you immediately overturn your previous conclusion. Humans do this by building a mental model of the scene. AI, however, often struggles with these "what-if" updates because it lacks a structured way to weigh how new information influences existing beliefs. The authors aim to bridge this gap by making AI "envision" the scenario before answering.

Figure 1: An example of the CURIOUS system generating an inference graph to determine if an update strengthens or weakens a hypothesis.

The CURIOUS Framework

Researchers from CMU and AI2 developed CURIOUS, a system that forces AI to first "think" through a scenario by generating and encoding an explicit "mental model" (inference graph). The system follows a three-step pipeline:

  1. Graph Generation (GENinit): Given a premise, hypothesis, and update, the model uses a sequence-to-sequence approach to generate an "inference graph." This graph contains nodes like Contextualizers (background info), Situations (the new event), and Mediators (the logical bridge between events).
  2. Graph Correction (GENcorr): LLMs often "hallucinate" or repeat themselves. The authors found that 70% of initial graphs had errors. They trained a specific T5-based module to clean and "repair" these graphs, significantly improving downstream performance.
  3. Hierarchical Mixture of Experts (MoE): This is the core architectural innovation. Instead of using a standard Graph Convolutional Network (GCN), which can accidentally amplify noise in generated graphs, CURIOUS uses two levels of MoE to selectively weight information.
Figure 2: The Hierarchical Mixture of Experts (MoE) architecture, featuring MOE-V for node selection and MOE-GX for balancing graph vs. text inputs.

Key Insights and Results

CURIOUS outperformed the previous state-of-the-art (SOTA) across all three benchmarks:

  • $δ$-SNLI: Jumped from 81.6% to 85.6% (a +4 point gain).
  • $δ$-SOCIAL: Improved from 86.2% to 88.6%.
  • $δ$-ATOMIC: Improved from 78.3% to 80.2%.

Beyond raw accuracy, the system provides explainability. By looking at the MoE gate values, researchers can see exactly which parts of the "mental model" the AI used to reach its conclusion. Interestingly, the model relies more heavily on the generated graph when the update strengthens a conclusion. When an update weakens a conclusion, the model often relies more on the original query text.

Limitations and Future Work

While CURIOUS represents a significant step forward, challenges remain. Even with a corrector, 34% of errors were still caused by "uninformative mediators"—nodes that didn't actually help bridge the gap between the premise and the update. Furthermore, in 42% of failure cases, the graph was actually correct, but the "Gating" mechanism chose to ignore it. Improving the "trust" between the classifier and the graph generator is a key next step for future research.

Summary updated

Abstract

Defeasible reasoning is the mode of reasoning where conclusions can be overturned by taking into account new evidence. Existing cognitive science literature on defeasible reasoning suggests that a person forms a mental model of the problem scenario before answering questions. Our research goal asks whether neural models can similarly benefit from envisioning the question scenario before answering a defeasible query. Our approach is, given a question, to have a model first create a graph of relevant influences, and then leverage that graph as an additional input when answering the question. Our system, CURIOUS, achieves a new state-of-the-art on three different defeasible reasoning datasets. This result is significant as it illustrates that performance can be improved by guiding a system to "think about" a question and explicitly model the scenario, rather than answering reflexively. Code, data, and pre-trained models are located at https://github.com/madaan/thinkaboutit.
BibTeX
@inproceedings{madaan2021think,
  author = {Aman Madaan and Niket Tandon and Dheeraj Rajagopal and Peter Clark and Yiming Yang and Eduard H. Hovy},
  title = {{Think about it! Improving defeasible reasoning by first modeling the question scenario}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  url = {https://openreview.net/forum?id=C8JHYuvleLm},
}
Figure