Tag
这篇论文提出"relearning savings"评估协议，发现在Qwen2.5-1.5B中，曾学习过的事实绑定在部分遗忘后比从未学习的绑定更容易被重新习得，恢复所需优化步数减少约18%，说明当前行为准确率会低估既有经验的持久影响。
CURIO 是一个好奇心驱动的测试时学习框架，通过内在好奇心世界模型（ICWM）为策略更新补充预测误差奖励，从而在保持探索的同时避免过早抑制低奖励但有潜力的方向，在数学发现任务与单细胞去噪上超越纯任务奖励的 RL 基线。
提出了名为StepLearn的非参数预序测试时学习框架，通过将个别交互转移转化为可即时使用、经前瞻性验证后才持久信任的假设，使LLM智能体在部署期间仅更新外部知识即可持续提升表现，效果显著优于EvoTest等基线。
The paper introduces Trace, a credit-guided framework that compiles noisy interaction histories into executable walkthroughs to enhance long-horizon AI agent performance, showing significant improvements in experiments.
Microsoft Research introduces EvoLib, a framework that enables LLMs to continually learn from their own experience during inference by extracting reusable skills and insights, without model updates or external labels.
ThetaEvolve is an open-source framework that extends AlphaEvolve to enable small LLMs like DeepSeek-R1-0528-Qwen3-8B to achieve new best-known bounds on open problems through test-time reinforcement learning, accelerating AI self-evolution.
Continual Harness is a reset-free, self-improving agentic harness that achieves 20.54% on ARC-AGI-3 at a cost of $774 by storing memories, reusing skills, and refining its prompt, outperforming prior baselines like Hermes and OpenClaw with greater efficiency.
The author shares excitement about three continual learning methods: SEAL models that self-adapt, test-time learning, and lifelong model editing, predicting true continual learning by 2027–2028 that will create a feedback loop toward artificial superintelligence.
This paper investigates many-shot chain-of-thought in-context learning for reasoning tasks, revealing that standard scaling rules do not transfer and proposing Curvilinear Demonstration Selection (CDS) for improved ordering, achieving up to 5.42 percentage-point gain.
The paper introduces PACEvolve++, a reinforcement learning framework that improves test-time policy adaptation for evolutionary search agents by decoupling hypothesis generation from execution.
EvoTest introduces J-TTL, a benchmark for measuring agent test-time learning capabilities, and proposes an evolutionary framework where an Actor Agent plays games while an Evolver Agent iteratively improves the system's prompts, memory, and hyperparameters without fine-tuning. The method demonstrates superior performance compared to reflection and memory-based baselines on complex text-based games.