Conversational benchmark tests AI on causal text editing skills
CoBe: A Benchmark For Conversational Counterfactual Text Editing
Computation and LanguageArtificial IntelligenceEmerging TechnologiesMachine Learning
Summary
This work looks at how well AI models can change a story or explanation when something in it is hypothetically different, without messing up facts that shouldn’t change. The authors created a test called CoBe with 2,000 questions spanning topics like healthcare and everyday scenarios to see if models can make these edits in a natural conversation style. They found that many models tend to guess plausible changes rather than making the precise edits needed to show what would truly happen if something changed. Bigger or more detailed answers didn’t always help, pointing to limits in how the models were trained.
What this means in practice
- •For healthcare data teams: Evaluate and improve AI systems that propose hypothetical patient outcome scenarios without altering unrelated medical facts.
- •For legal document editors: Test AI tools that modify contract clauses under hypothetical changes while keeping unaffected terms intact.
Authors
Arvind Raghavan, Inwoo Hwang, Shreyas Havaldar, Kai-Zhan Lee, Aurghya Maiti, Yushu Pan, Jeffrey Wu, Mingxuan Li, Elias Bareinboim
Abstract
Counterfactual text editing is the ability to revise a given text in light of a hypothetical change while preserving facts not causally affected by that change. When asked what would have happened if a doctor had not recommended surgery, the model output should retain the patient's medical history and modify only the consequences of the decision. This requires causal understanding: a model must determine, without being explicitly instructed to, what events are affected by the hypothetical change and what should remain fixed. This skill is central to human reasoning, especially in safety-critical domains like healthcare or law. By contrast, the task of associational text generation merely generates text statistically associated with the altered condition. In the medical example, an associational response might make the patient's symptoms less severe because milder cases correlate with no surgery. Such responses produce a plausible scenario, but not the counterfactual version of the given one. Prior counterfactual-text datasets do not clearly separate these modes of reasoning. We introduce the Counterfactual Benchmark (CoBe), a testbed for models' zero-shot ability to counterfactually edit a given textual scenario when prompted to do so in a conversationally natural way. CoBe contains 2000 questions spanning healthcare, science and engineering, business, and everyday scenarios. We test a suite of frontier and smaller models, with frontier systems achieving an average score of $58.05\%$. Our findings indicate that models frequently default to associational generation rather than preserving the causal structure of the original scenario. Moreover, performance does not scale consistently with either model size or output token length, suggesting that the failure may reflect model training objectives rather than insufficient model capacity.