curriculum-learning

DHRCL proposes a reinforcement learning framework with dense hierarchical rewards and curriculum learning for training code LLMs, using syntax validation, execution success, unit-test pass rate, and AST structural similarity in a three-stage curriculum.

SCOUT introduces per-context reset curricula for sparse-reward reinforcement learning, adapting assistance removal to each context's learning progress, outperforming global pacing methods in navigation and manipulation tasks.

This paper proposes a fine-grained taxonomy for curriculum learning in NLP, separating difficulty evaluation from training scheduling to enable systematic analysis and comparison of CL strategies. It identifies an incomparability problem in prior work and provides a framework for designing and evaluating CL approaches.

The paper proposes A*-Inspired Batch Selection (A*-BS), a lightweight strategy that treats mini-batch scheduling as a heuristic search to improve CNN training efficiency. On MedMNIST tasks, a simple CNN with A*-BS outperforms deeper ResNet-18/50 baselines in accuracy and training speed.

This paper proposes a self-evolving in-context learning framework for direct pilot-to-beamformer design in multi-user MISO systems, integrating a Transformer backbone with a pilot encoder-decoder network and curriculum learning to handle multiple channel models without retraining.

EvoCUA-1.5 introduces an online reinforcement learning framework for multi-turn computer-use agents, achieving a 63.2% success rate on OSWorld-Verified and outperforming comparable open-weight models up to 35B parameters through step-level policy optimization and dynamic curriculum learning.

This paper introduces Progressive Code-Switching (PCS), a reinforcement learning approach with curriculum learning that gradually increases code-switching in LLMs to efficiently transfer multilingual reasoning capabilities.

This paper proposes a tone-conditioned curriculum learning framework for low-resource Bantu speech recognition, combining hybrid difficulty scoring, gated adapters, and staged curriculum training. Evaluations on six Southern Bantu languages show that W2V-BERT outperforms Whisper on Nguni languages while Whisper performs better on Sotho-Tswana languages.

SEAD introduces a competence-aware on-policy distillation method that uses entropy to guide supervision at token, temporal, and prompt levels, achieving a +4.8 average accuracy improvement on OLMo-3 across six math benchmarks.

This paper theoretically analyzes how curriculum learning, by decomposing complex problems into simpler sub-problems and composing solutions, can dramatically reduce the sample complexity of learning to simulate sequential computations (semiautomata) compared to direct methods, achieving subpolynomial supervision requirements in supervised fine-tuning and exponentially weaker coverage conditions in reinforcement learning with verifiable rewards.

GUICrafter introduces a weakly-supervised GUI agent that leverages massive unannotated screenshots and a two-stage curriculum learning framework to reduce reliance on expensive human annotations, achieving competitive performance with advanced systems like UI-TARS using only 0.1% of its data.

This paper proposes Transfer-Aware Curriculum (TAC), a bandit-style online curriculum for multi-domain RLVR that prioritizes domains whose updates benefit other domains using gradient-geometry alignment. TAC improves macro-averaged accuracy on Qwen3-1.7B and Llama3.2-3B over fixed and learnability-only curricula.

Meta's new paper 'Autodata' introduces an agentic data scientist that generates and meta-optimizes synthetic training data, significantly outperforming standard methods and enabling a small 4B model to beat a 397B baseline in legal tasks.

This paper introduces Lie-bracket prediction of transfer order for sequential learning, using commutators of gradient fields to determine pairwise order and scaling to many domains. Experiments show high accuracy in predicting optimal curriculum orders for fine-tuning and instruction tuning.

Introduces Bayesian Manifold Curriculum (BMC), an adaptive curriculum learning method for LLMs that leverages the model's latent geometry to allocate training effort across diverse problem types, improving efficiency beyond traditional difficulty-based curricula.

Pythagoras-Prover is a compute-efficient family of Lean theorem provers that achieves strong performance using curriculum supervised fine-tuning and a novel Augmented Lean Formalisation technique. The 4B model surpasses DeepSeek-Prover-V2-671B at pass@32 on MiniF2F-Test, and the 32B model sets a new state-of-the-art among open-source provers.

This paper proposes Representation Curriculum (RC), a training-time intervention that stages feature utilization to reduce over-reliance on exposure-confounded historical signals and improve cold-start generalization in ranking systems. The method is theoretically analyzed and validated on public benchmarks and large-scale eBay search experiments.

This paper investigates sequential fine-tuning of LLaMA-3.1-8B for automated essay scoring using a curriculum aligned with discourse structure, showing improved coherence and performance compared to independent or randomized training.

This paper theoretically studies how transformer-based policies acquire search capabilities from reinforcement learning training dynamics in a stochastic tree environment. It shows that a two-head transformer can implement depth-first search and that this mechanism emerges naturally from sparse reward signals under a depth-wise curriculum.

Introduces the Data-Model Compatibility (DMC) metric to evaluate how well a reasoning dataset aligns with a student model during distillation. Experiments show DMC strongly correlates with distillation performance and that dynamically selecting datasets based on DMC further improves reasoning capabilities.