With the current algorithms, we fail to see the quality of hypotheses, i.e., hypothesis-based inference accuracy scale as data size increases.
Currently we do see a lot of minor changes on new hypotheses, and we are seeing duplicates.
Some easy changes to try (maybe right now) would be try generating more hypotheses (increase batch size, increase number of hypothesis to generate, etc.) and using a smaller temperature.
For more difficult changes, we should carefully look into each update and see if we are getting new information from each update. We will likely need to improve the reward function and our generation modules for this, which can be one of the next steps after the benchmark
With the current algorithms, we fail to see the quality of hypotheses, i.e., hypothesis-based inference accuracy scale as data size increases.
Currently we do see a lot of minor changes on new hypotheses, and we are seeing duplicates.
Some easy changes to try (maybe right now) would be try generating more hypotheses (increase batch size, increase number of hypothesis to generate, etc.) and using a smaller temperature.
For more difficult changes, we should carefully look into each update and see if we are getting new information from each update. We will likely need to improve the reward function and our generation modules for this, which can be one of the next steps after the benchmark