Your submission is now in Draft mode.

Once it's ready, please submit your draft for review by our team of Community Moderators. Thank you!

Submit Essay

Once you submit your essay, you can no longer edit it.

Pending

This content now needs to be approved by community moderators.

Submitted

This essay was submitted and is waiting for review.

When will AI achieve superhuman few-shot performance on SuperGLUE?

Question

SuperGLUE is a benchmark for evaluating general-purpose language understanding systems. The set of eight tasks in the benchmark emphasizes diverse task formats and low-data training data tasks, with nearly half the tasks having fewer than 1k examples and all but one of the tasks having fewer than 10k examples.

As of August 2020, the best performing model is T5 from Google, which receives a score of 89.3. The human baseline is 89.8. Unfortunately, outside of the benchmark T5 does not seem to match general-purpose language understanding skills of humans.

Therefore, this question considers a more challenging regime for the benchmark recently presented in the GPT-3 paper. The few-shot regime is when the model has severely limited access to the training set. This question will take into account models that have been trained on maximally 100 examples from each task and asks:

When will AI achieve superhuman few-shot performance on SuperGLUE?

This question will resolve as soon as a model with access to at most 100 examples per task meeting or exceeding 89.8 human baseline is announced to the public. A pre-print or published paper from a reputable source is sufficient by itself to trigger resolution. Any other source can count too if the result can be verified by SuperGLUE leader board submission.

Currently, GPT-3 achieves the best performance of 71.8 in a few-shot regime with access to 32 examples. This is still 18 points away from the human performance.

Training data contamination similar to GPT-3 issue is allowed as long as the authors have made sure that it does not impact the evaluation in a severe way.

Similar question: When will a language model meet or exceed the human baseline on SuperGLUE?

Make a Prediction

Prediction

Note: this question resolved before its original close time. All of your predictions came after the resolution, so you did not gain (or lose) any points for it.

Note: this question resolved before its original close time. You earned points up until the question resolution, but not afterwards.

Current points depend on your prediction, the community's prediction, and the result. Your total earned points are averaged over the lifetime of the question, so predict early to get as many points as possible! See the FAQ.