curriculum-learning

This paper introduces Micro-Macro Retrieval (M2R), a retrieve-while-generate framework that reduces hallucination in long-form LLM outputs by ensuring key information stays close to generated text. It uses curriculum learning-based reinforcement learning to train retrieval and grounding skills, showing effectiveness especially in lengthy contexts.

This paper proposes Staged-Competence, a curriculum learning framework for DPO-based safety alignment that organizes preference data by difficulty, improving robustness and data efficiency while preserving general capabilities.

Introduces TELL, an AI-generated text detection system that provides explainable annotations alongside numerical scores, achieving competitive AUROC of 0.927 while enabling users to judge authorship based on highlighted textual indicators.

This paper empirically studies how the composition of training data (curriculum) affects the skills learned by RL-based memory agents in multi-session question answering. It finds that curriculum composition acts as a fine-grained lever on specialization, with mixed benchmarks yielding the best overall performance and narrow out-of-domain sets transferring targeted temporal reasoning skills.

This paper proposes a plug-and-play module using self-paced curriculum learning to enhance modality balance in multimodal conversational emotion recognition, achieving consistent F1-score improvements on IEMOCAP and MELD datasets.

RPS is a two-stage LLM post-training method inspired by neuroscience, combining curriculum learning with learning rate decay. Preliminary results show improved program synthesis reliability on Qwen3-8b compared to equal learning rate training.

SCRL is a curriculum reinforcement learning framework that uses subproblem-level normalization and curriculum learning to improve credit assignment in LLM reasoning, outperforming baselines on mathematical reasoning benchmarks.

Introduces PROWL, a prioritized regret-driven optimization framework that uses an adversarial curriculum to improve diffusion-based world model robustness by focusing on high-error trajectories, achieving better performance on out-of-distribution scenarios in MineRL.

This paper proposes a staged training approach for vision-language models that separates visual perception, visual reasoning, and textual reasoning into distinct stages. The method improves visual reasoning accuracy while reducing reasoning trace length, demonstrating that stronger perception reduces the need for excessive reasoning.

This paper proposes a spatially correlated curriculum learning framework for Physics-Informed Neural Networks (PINNs) that improves training stability and solution accuracy by leveraging spatial correlations among subregions, addressing issues like high-dimensional non-convex loss landscapes and imbalanced multi-objective constraints.

Presents VectraYX-Nano, a 42M-parameter decoder-only language model trained from scratch in Spanish for cybersecurity, featuring curriculum learning, native tool invocation via MCP, and a 170M-token corpus. Empirical findings reveal a loss-versus-register inversion and corpus-density artifacts for tool-use capability.

Researchers release SU-01, a 30B-A3B reasoning model achieving gold-medal-level performance on physics and math Olympiad problems using a unified scaling recipe for proof search.

BACR introduces adaptive token budgeting and curriculum-aware scheduling to prevent LLMs from overthinking easy problems and underthinking hard ones, cutting token use 34% while boosting accuracy up to 8.3%.

LiFT is a longitudinal instruction fine-tuning framework that unifies diverse temporal NLP tasks under a shared instruction schema with curriculum-based training. Evaluated across OLMo, LLaMA, and Qwen models, LiFT consistently outperforms base-model in-context learning, especially on out-of-distribution data and rare change events.

Introduces ToolsRL, a two-stage reinforcement learning framework that teaches multimodal LLMs to use simple visual tools for complex visual reasoning tasks.

CLewR introduces a curriculum learning strategy with restarts for improving machine translation performance in LLMs through preference optimization. The method addresses catastrophic forgetting by iterating easy-to-hard curriculum multiple times, showing consistent gains across Gemma2, Qwen2.5, and Llama3.1 models.

This paper proposes the Implicit Curriculum Hypothesis, demonstrating that language model pretraining follows a structured, compositional curriculum where capabilities emerge consistently across architectures and can be predicted from internal representations. The authors validate this through designed tasks spanning retrieval, morphology, coreference, reasoning, and mathematics, finding highly consistent emergence orderings (ρ=0.81) across four model families.

OpenAI achieved a new state-of-the-art 41.2% on the miniF2F formal math olympiad benchmark using a technique called 'statement curriculum learning,' which iteratively trains a neural prover on proofs of increasing difficulty. The approach builds on iterative proof search and retraining over 8 iterations to significantly outperform the previous best of 29.3%.

OpenAI researchers introduce VALOR, a variational inference method for option discovery that connects option learning to variational autoencoders, and propose a curriculum learning approach that stabilizes training by dynamically increasing context complexity.

OpenAI presents Hindsight Experience Replay (HER), a technique enabling sample-efficient reinforcement learning from sparse binary rewards without complex reward engineering. It is demonstrated on robotic arm manipulation tasks including pushing, sliding, and pick-and-place, and validated on physical robots.