Tag
The paper introduces CLR-Bench, a benchmark curriculum of hidden rules for evaluating whether self-evolving LLM agents can learn latent rules from feedback, finding that most current self-evolution methods barely learn unless given a fixed lesson or a stronger base model.
该论文构建了一个最小化的策略学习模型，从数学和数值上分析了课程学习在奖励稠密与奖励稀疏情形下的效果差异，发现课程学习在稀疏奖励且存在竞争性最优解时几乎是必需的。
ICCLHerd 提出一种基于 herding 的 LLM 训练数据选择框架，利用 von Mises–Fisher 核平衡 mode-seeking 与 mode-averaging，并以 in-context learning 作为验证代理进行超参数搜索，实验表明其优于现有数据选择算法。
An agent trained via self-play reinforcement learning on a single desktop (RTX 3090) achieves an 89% win rate against Warcraft III's hardest built-in AI, running the real unmodified game under Wine with a custom shim and harness, in 840,000 games over three weeks.
该论文研究了工具使用型语言智能体在监督微调下的「选择性撤回」行为，提出一种能发现聚合准确率所掩盖的规则级回归的诊断方法，并表明微调本身并不足以保证智能体安全。
WEDGE 是一个弱点引导的导航学习框架，利用 LLM 从预训练策略的成功与失败轨迹中诊断失败模式，并据此生成针对性训练环境，在 ProcTHOR 和 HM3D 上显著优于随机与难例挖掘基线。
NVIDIA researchers propose NV-Tesseract-AD, a curriculum-enhanced conditional diffusion framework with a Transformer backbone for multivariate time-series anomaly detection, achieving ~3.7× F1 improvement over non-curriculum baselines and strong zero-shot transfer to EasyTSAD benchmarks.
AdaC-GRPO 使用基于在线 rollout 准确率的自适应课程学习动态构建训练批次，帮助 Qwen2.5-VL-7B 掌握长时程视觉 Sudoku 推理任务，并发布了细粒度基准 FG-Sudoku。
The paper investigates how ~10.6M-parameter Tiny Transformers can learn multi-step arithmetic using algorithmic scratchpads and a hierarchical developmental curriculum, showing division accuracy rising from 4.0% to 86.7% while identifying failure modes like catastrophic forgetting and poor generalization to unseen operand sizes.
该论文通过冻结选中子集并只更换训练学习器的受控实验，证明样本价值（以及 easy-first 与几何覆盖策略的 crossover budget）是相对于目标学习器定义的，而非数据的固有属性。
Learning is most effective at the frontier of capability: for GRPO-trained LLM reasoners, problems that are always or never solved yield zero gradient. The authors introduce "Frontier Learning" to target this effective training regime.
ServiceNow CoreAI introduces AutoSynthData, a pipeline that converts an enterprise agent model's failures and a stronger teacher's successes into feasible, realistic, and verified training tasks, illustrated with the EnterpriseOps Gym environment and released dataset.
StateTree 提出一种数据驱动的强化学习方法，通过构建树结构的路径追踪辅助任务来增强长对话推理，模型在 10K token 上训练即可泛化到 128K token，StateTree-14B 在 LongMemEval 上达到 59.00% 准确率，超过 QwenLong-L1-32B。
ActiveSaddler 通过将动态课程学习建模为非平稳多臂老虎机问题，实现了 LLM Agent 优化中训练场景课程与 harness 的协同进化，在 GAIA2 和 Terminal-Bench2.0 上分别提升了 4.4 和 7.5 个百分点的 Pass@1。
SFTMill is an open-source tool that enables off-policy distillation of large language models using an OpenAI-compatible endpoint, allowing users to create training datasets from behavioral goals defined in a curriculum.
The paper explores using code-switched text to align representations in small language models under data constraints, demonstrating that a curriculum involving code-switching improves multilingual performance.
The paper proposes LT-OPD, an on-policy self-distillation framework for extreme visual token reduction in multimodal large language models, significantly improving performance under low token budgets while reducing KV-cache usage and FLOPs.
The paper proposes AdvRole, an adversarial closed-loop curriculum framework for training role-playing agents in large language models, which evolves scenario pools to improve performance on benchmarks.
This paper systematically analyzes length-based training in speech token language models, showing that short-to-long ordering has no independent benefit when batch composition and token exposure are fixed, and evaluates the impact of loss normalization.
Spatial-Interactor is a framework that trains vision-language models to enhance spatial reasoning through interaction with the physical world, employing a three-level curriculum and two-stage training strategy to improve state transition modeling and long-horizon integration.