A Wasserstein geodesic framework decouples ordering, exposure, pacing, and geometry in curriculum learning, revealing context-dependent effects: easy-to-hard ordering improves hard-level performance beyond cumulative exposure, and pacing and endpoint smoothness matter substantially.
hub
P \'o czos, and Tom Michael Mitchell
10 Pith papers cite this work, alongside 308 external citations. Polarity classification is still indexing.
10
Pith papers citing it
308
external citations · external index
hub tools
representative citing papers
MATCH combines an adaptive curriculum that tracks the model's evolving capability with a reward that only grants credit for arguments when the correct tool is selected, improving tool-use accuracy on two benchmarks.
On-policy self-distillation with sampled demonstrations reduces rollout diversity by amplifying existing probability gaps in the base model, unlike ideal RL which preserves ratios among correct outputs.
Training a tiny BERT first on Simple Wikipedia and then on everyday Wikipedia improves perplexity, while difficulty heuristics such as sentence length, word rarity, and Flesch score give no consistent gain over random order.
Sorting pre-training examples by training-data influence instead of human-judged difficulty reportedly gives over 10 percentage point benchmark gains over random order in limited-data language model pre-training.
An LLM-scored easy-to-hard curriculum for training grammatical error correction models yields small but consistent F0.5 gains over one-shot and length-based training.
An iterative bootstrapped self-filtering approach selects balanced clean and diverse subsets from noisy vision-language datasets to train improved CLIP models.
Aligning training order to a fact-text alignment confidence score and annealing away low-quality shards improves cross-lingual data-to-text generation on noisy datasets.
The paper proposes six curriculum sampling strategies driven by a pretrained language model's own prompt-based confidence scores and reports mixed, mostly small, improvements over random sampling on four NLU datasets.
citing papers explorer
-
Curriculum Learning as Transport: Understanding Curricula with Wasserstein Geodesics
A Wasserstein geodesic framework decouples ordering, exposure, pacing, and geometry in curriculum learning, revealing context-dependent effects: easy-to-hard ordering improves hard-level performance beyond cumulative exposure, and pacing and endpoint smoothness matter substantially.
-
MATCH: Model-Aware Tool Learning with Curriculum Scheduling and Hierarchically Gated Rewards
MATCH combines an adaptive curriculum that tracks the model's evolving capability with a reward that only grants credit for arguments when the correct tool is selected, improving tool-use accuracy on two benchmarks.
-
On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity
On-policy self-distillation with sampled demonstrations reduces rollout diversity by amplifying existing probability gaps in the base model, unlike ideal RL which preserves ratios among correct outputs.
-
Beyond Shallow Heuristics: Leveraging Human Intuition for Curriculum Learning
Training a tiny BERT first on Simple Wikipedia and then on everyday Wikipedia improves perplexity, while difficulty heuristics such as sentence length, word rarity, and Flesch score give no consistent gain over random order.
-
Influence-driven Curriculum Learning for Pre-training on Limited Data
Sorting pre-training examples by training-data influence instead of human-judged difficulty reportedly gives over 10 percentage point benchmark gains over random order in limited-data language model pre-training.
-
LLMCL-GEC: Advancing Grammatical Error Correction with LLM-Driven Curriculum Learning
An LLM-scored easy-to-hard curriculum for training grammatical error correction models yields small but consistent F0.5 gains over one-shot and length-based training.
-
Data Selection Through Iterative Self-Filtering for Vision-Language Settings
An iterative bootstrapped self-filtering approach selects balanced clean and diverse subsets from noisy vision-language datasets to train improved CLIP models.
-
Curriculum Learning for Cross-Lingual Data-to-Text Generation With Noisy Data
Aligning training order to a fact-text alignment confidence score and annealing away low-quality shards improves cross-lingual data-to-text generation on noisy datasets.
-
Your Pretrained Model Tells the Difficulty Itself: A Self-Adaptive Curriculum Learning Paradigm for Natural Language Understanding
The paper proposes six curriculum sampling strategies driven by a pretrained language model's own prompt-based confidence scores and reports mixed, mostly small, improvements over random sampling on four NLU datasets.
- Disentangling Curriculum Learning in NLP: Towards a Unifying Taxonomy