For multimodal agents, selecting environments by learned ability coverage and organizing them in a two-level difficulty curriculum improves training more than scaling the number of environments.
ADCL (Zhang et al., 2025a) periodically re-estimates sample difficulty to mitigate difficulty shift during training
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Beyond Simply Environment Scaling: Designing Effective Environment Distributions for Multimodal Agent Learning
For multimodal agents, selecting environments by learned ability coverage and organizing them in a two-level difficulty curriculum improves training more than scaling the number of environments.