Source-linked AI summary
CommonsenseQA 2.0: Exposing the Limits of AI through Gamification
Alon Talmor, Ori Yoran, Ronan Le Bras, Chandra Bhagavatula, Yoav Goldberg, Yejin Choi, Jonathan Berant
TL;DR
Modern NLU benchmarks can overstate model understanding because models exploit artifacts while remaining brittle on adversarial and commonsense questions. The paper uses gamification with controlled prompts, a model-in-the-loop, and validation to construct CSQA2, yielding a challenging benchmark on which the best model reaches 70.2% accuracy versus 94.1% for humans.
Problem
Existing NLU benchmarks can be misleading because models exploit artifacts yet remain brittle on adversarial, out-of-domain, and commonsense examples.
Method
Gamification combines interactive player incentives, controlled prompts, model-in-the-loop retraining, and validation to create CSQA2.
Results
70.2% accuracy is achieved by UNICORN-11B on CSQA2, compared with 52.9% for GPT-3 and 94.1% for humans.
Takeaways & Limitations
CSQA2 exposes weaknesses in current models, including limited causal reasoning and poor robustness to simple linguistic or logical variations.
Abstract
from arXiv · showhide
Constructing benchmarks that test the abilities of modern natural language understanding models is difficult - pre-trained language models exploit artifacts in benchmarks to achieve human parity, but still fail on adversarial examples and make errors that demonstrate a lack of common sense. In this work, we propose gamification as a framework for data construction. The goal of players in the game is to compose questions that mislead a rival AI while using specific phrases for extra points. The game environment leads to enhanced user engagement and simultaneously gives the game designer control over the collected data, allowing us to collect high-quality data at scale. Using our method we create CommonsenseQA 2.0, which includes 14,343 yes/no questions, and demonstrate its difficulty for models that are orders-of-magnitude larger than the AI used in the game itself. Our best baseline, the T5-based Unicorn with 11B parameters achieves an accuracy of 70.2%, substantially higher than GPT-3 (52.9%) in a few-shot inference setup. Both score well below human performance which is at 94.1%.
1 Introduction
The paper argues that current NLU benchmarks can be misleading because models exploit artifacts yet remain brittle on adversarial, out-of-domain, and commonsense examples. It introduces gamification to create challenging benchmarks and uses it to build CSQA2, where current models substantially trail humans.
- Current NLU models reach human parity on many benchmarks but remain brittle on out-of-domain and adversarial examples and make commonsense errors.
- Post-hoc adversarial filtering can remove difficult examples but does not direct data creation toward diverse, challenging instances.
- Gamification uses an interactive game and model-in-the-loop design to tap human intelligence while controlling data quality and type.
- 14,343 yes/no questions were collected for CommonsenseQA 2.0, with evaluation spanning pre-trained, fine-tuned, and reading-comprehension models.
- 70.2% accuracy is achieved by UNICORN-11B, versus 52.9% for GPT-3 and 94.1% for humans.
- Models also show weaknesses in causal reasoning, size comparisons, and robustness to simple linguistic or logical variations.
2 Data Collection through Gamification
The data-collection game combines controlled prompts, a model-in-the-loop, and validation to elicit difficult yes/no questions while steering topic and reasoning coverage. Quality-assurance filters produce 14,343 selected examples from 144,682 authored questions at lower average cost than CommonsenseQA 1.0.
- Gamification structures data collection around controlled question generation, a model-in-the-loop, and question validation.
- Players earn points for questions that defeat the AI and for using topic and relational prompts, encouraging difficult questions while steering content.
- Topic prompts are sampled from 1,875 high-ranking ConceptNet concepts, while 33 manually constructed relational phrases target varied commonsense skills.
- The model is retrained during collection so players cannot repeatedly exploit a fixed weakness, with feedback based on whether its answer was correct.
- Validators independently assess answers and question quality, while a verification model filters bad or low-confidence examples and labels retained questions with 92% development accuracy.
- Google-snippet matching filters long-tail web facts, removing 421 questions from the final dataset.
- 14,343 examples were selected from 144,682 authored questions at an average cost of $0.17 per question.
3 Game Analysis
The game’s feedback mechanisms changed player behavior and the difficulty and quality of collected questions. Model retraining increased question difficulty, while performance feedback improved engagement and validation outcomes but could make questions easier.
- Feedback effects: 35% average AI beat-rate followed the introduction of performance feedback, up from 26.8% beforehand.
- Game feedback: Players received performance metrics and notifications about changed answers or discarded bad questions, with discarded questions triggering point deductions.
- Feedback effects: 4.8% higher positive validations followed validation notifications, but UNICORN-11B accuracy rose from 0.66 to almost 0.70 as questions became easier for models.Players became more hesitant to author questions that validators might find too difficult.
- Model retraining: Model retraining reduced baseline accuracy from 0.74 to 0.66, indicating that stronger game models drove players to compose more challenging questions.The difficulty increase was shared across the evaluated pre-trained models.
- Player engagement: Users mostly expressed positive sentiment, and the authors associate the game’s enjoyment with high engagement and low average annotation cost.
4 Dataset Analysis
CSQA2 contains many short questions with broad vocabulary and diverse topic and relational prompts. Its construction involved thousands of players, while analyses examined prompt distributions and the commonsense skills required for answering questions.
- Dataset statistics: 14,343 questions average 11.3 words and contain 21,143 distinct words, with 1,868 topic prompts and 33 relational prompts.
- Dataset statistics: The most common relational prompt appears in 6.0% of questions, while the most common topic prompt appears in 0.2%, indicating broad prompt diversity.
- Question formulation: 2,537 players participated, 351 contributed to the final dataset, and 55 produced more than 85% of the questions.
- Topics and relations: Only 14.4% of questions omitted the relational prompt, and only 2% omitted the suggested topic prompt.
- Topics and relations: The most frequent relational prompts were “can” at 5.9%, “is” at 5.8%, “cannot” at 5.7%, “part of” at 4.9%, and “has” at 4.1%.
- Reasoning skills: A sampled development-set analysis annotated multiple commonsense reasoning skills per question, averaging 1.48 skills per question.
5 Experimental Evaluation
The evaluation tests CSQA2 across fine-tuned, pretrained, and few-shot models, with performance compared against human accuracy. Results show substantial model–human gaps, while added web snippets provide external context for answering questions.
- The data are split into 9,282 training, 2,544 development, and 2,517 test examples with disjoint topic prompts across splits.
- Humans reach 94.1% accuracy by majority vote on CSQA2 validation questions.
- Baselines: T5 and UNICORN are evaluated as fine-tuned text-to-text and commonsense reasoning baselines, while GPT-3 is evaluated few-shot without fine-tuning.
- Results: 70.2% is the best model accuracy, achieved by UNICORN-11B, while GPT-3 reaches 52.9% and remains below the fine-tuned baselines.
- External context: Google snippets are added as top-k search context to test whether external information improves T5 and UNICORN performance.
6 Model Analysis
The analysis examines model skill-specific performance and behavioral consistency on CSQA2. It finds strong variation across reasoning types and evidence that GPT-3 struggles to maintain consistent knowledge representations.
- Skill analysis: Relational prompts are grouped by skills such as meronymy/hypernymy, causality, and size comparison, while long-tail knowledge is manually annotated.
- Skill analysis: UNICORN-11B reaches 79.3 accuracy on causality questions versus 73.3 for T5-11B and 49.1 for GPT-3.
- Skill analysis: UNICORN-11B accuracy is 63.1 for size comparison and 60.8 for long-tail factual knowledge, while Google snippets improve long-tail performance by up to 22.5 points.
- Consistency analysis: Contrast sets use minor perturbations of development questions to test whether models preserve answers under linguistic changes.
- GPT-3 analysis: GPT-3 answers “Unknown” in 10% of free-form cases, agrees with its original prediction in 61% of remaining cases, and produces sensible “why” explanations in 36.4%.
- GPT-3 analysis: GPT-3 often answers yes to imaginary or counterfactual statements found in web snippets, exposing difficulty separating fictional from factual content.
7 Related Work
The work combines commonsense benchmarking, model-in-the-loop data collection, and gamification. Its distinguishing focus is interactive human–model gameplay for collecting entailment-style data.
- CSQA2 extends prior benchmarks targeting commonsense reasoning and other reasoning abilities.
- Model-in-the-loop methods use models during annotation or as filters, with recent systems retraining models during data collection.
- Gamification has been applied to question answering, image labeling, protein folding, and entailment data collection.
- This work focuses on gamifying interaction between a human and a model.
8 Conclusion
The paper presents gamification as a framework for building diverse, challenging NLU benchmarks and applies it to CSQA2. The resulting benchmark exposes a substantial gap between current models and human performance.
- Gamification is proposed as a general framework for creating diverse and challenging NLU benchmarks.
- CSQA2 contains 14,343 yes/no questions and is analyzed through a detailed evaluation of strong baselines.
- UNICORN-11B achieves 70.2% accuracy, dozens of points below human accuracy.
- The authors characterize gamification as a promising approach for creating challenge sets that expose weaknesses in current state-of-the-art models.
Checklist
The checklist reports that the paper addressed reproducibility, asset documentation, participant procedures, limitations, and potential societal impacts, while theoretical-results items were not applicable.
- Scope and ethics: It described its limitations, discussed potential negative societal impacts, and stated that its claims accurately reflected the paper’s contributions and scope.
- Theoretical results: The checklist marked theoretical-results assumptions and proofs as not applicable.
- Reproducibility: The paper reported code, data, instructions, and training details needed to reproduce the main experiments.
- Assets: It cited existing assets, documented their licenses, and provided new assets through supplemental material or a URL.
- Data considerations: The paper reported discussing whether its data contained personally identifiable information or offensive content, while consent discussion was marked not applicable.
- Human subjects: The paper included participant instructions, described potential risks, and reported participant compensation and total spending.