curriculum-learning

This paper proposes progressive error curriculum training (PECT) to improve phoneme-to-text reconstruction robustness in visual speech recognition by gradually adapting to realistic phoneme prediction errors, achieving reduced word error rates on LRS2 and LRS3 benchmarks.

DeliveryGym is a 3D reinforcement learning environment for long-horizon embodied agent planning with adaptive curriculum, demonstrating performance improvements through RL.

The paper proposes Abstract Token Curriculum (ATC), a curriculum learning framework that elicits effective continuous intermediate representations in LLMs without direct supervision, supported by theoretical and experimental evidence.

The paper introduces a layer-wise curriculum learning method for efficient LLM compression, achieving state-of-the-art performance with significant reductions in GPU memory usage and training time.

This paper decomposes associative recall in fixed-state recurrences, finding convolution and curriculum learning key to performance, and proposes interventions to address interference rather than capacity limitations.

The paper introduces a teacher-guided curriculum learning approach for Reinforcement Learning with Verifiable Rewards (RLVR) to efficiently train language models on initially unsolvable mathematical problems, achieving significant data efficiency and expanding reasoning boundaries.

This paper presents the SEAR system for multilingual speech MCQ, achieving 90.92% accuracy on the MLC-SLM Challenge through segment-evidence-aware routing, supervised fine-tuning, and reinforcement learning.

This paper introduces PATH, a framework for active curriculum refinement in reinforcement learning that leverages curriculum graph structure to improve robustness and generalization across diverse environments.

This paper introduces CAIR, a two-stage framework for imputing physiological time-series data under realistic missingness, outperforming existing methods by incorporating gap mechanisms and curriculum-aware training.

The paper proposes dynamic context scheduling to enhance generalization in contextual reinforcement learning, showing improved performance in out-of-distribution and in-distribution regimes across simulation environments.

This paper analyzes curriculum learning in large language models by examining optimization dynamics across difficulty levels, proposing a new method called Transfer-aware Dynamic Curriculum Sampling (TDCS) that dynamically adjusts training data based on transfer relationships.

Wuying-Browser-Agent is a unified framework for real-world long-horizon browser agents, introducing BrowserBench benchmark and achieving state-of-the-art performance on web-use tasks.

This article presents LittleLearner, a controlled sandbox for studying LLM knowledge acquisition using a K-5 curriculum-filtered dataset, finding that interventions like scaling and post-training enhance in-scope performance but do not improve out-of-scope capabilities.

This paper proposes predicting task difficulty for LLM agents without running expensive rollouts, studying the problem across 17 agentic benchmarks and showing that token-level entropy is a useful predictive signal.

This paper introduces MSRT, a framework with a resource-aware Mixture of Speech Encoders (MoSE) to overcome the curse of multilinguality in many-to-many speech-to-text translation. The 4B-parameter model achieves state-of-the-art results across 45 languages, particularly improving low-resource speech translation with only 10 hours of paired data per language.

The paper argues that simply scaling multimodal environments does not always improve agent training, and proposes Ability-aware Environment Selection (AES) and Hierarchical Difficulty Curriculum (HDC) to better structure environment distributions along diversity and difficulty dimensions.

This arXiv paper introduces CVPO, a reinforcement learning method for LLMs that adapts value-variance for advantage estimation and uses dynamic curriculum learning to match question difficulty, achieving better reasoning performance than VAPO on math tasks.

This paper proposes DASH-OPD, a discrepancy-aware switching method for on-policy distillation in multi-turn LLM agent training, which adaptively toggles between teacher and student executors based on drift and recovery evidence to improve efficiency and performance.

CURV proposes a curriculum learning framework that enhances chart question answering by developing intrinsic visual grounded reasoning in multimodal LLMs, achieving significant improvements on real-world benchmarks.

This paper introduces KGPS, a Kalman-guided prompt selection method for adaptive RL finetuning of LLMs, which models prompt difficulty as a dynamic state to improve accuracy and rollout efficiency.