Tag
该论文提出 TIZSKT 框架，将文本作为索引而非可迁移的预测表征，实现零样本跨域知识追踪，在三个公开数据集的留一域实验中取得最优 AUC 与准确率。
The paper proposes CAP (Candidate Acquisition Policy), a method that learns a feedback-conditioned query policy from sparse records of related black-box optimization tasks, achieving the best final-budget results on four synthetic families and two engineering benchmarks (BOPTEST, SU2).
GARDEN 提出一种由 LLM 自动设计并可跨场景复用的奖励函数框架，用于协作多智能体强化学习，在 Overcooked 与 SMAX 基准上显著优于默认奖励和 Eureka 式反思方法。
This paper studies auxiliary trajectory preferences (confidence-based and entropy-based metrics) to improve small language model self-improvement and transfer, showing UID-Conj. gains of about 2 points over unmodified TTRL and uniform-weighted SFT on AIME tasks.
Researchers train ModernBERT-based Icelandic encoders from 12M to 1B parameters on 9B tokens of open data, showing that a 1B model trained from scratch beats IceBERT on all eleven tasks, and that warm-starting from an English Ettin encoder yields the best scores across every benchmark while highlighting the importance of vocabulary transplantation.
提出了一个跨域迁移表示特征的统计框架，用于检测 LLM 的幻觉与推理错误，在新域仅需 25 条标注样本即可将幻觉检测 AUROC 提升 6.8%、推理步骤预测 AUROC 提升 8.4%。
该论文提出 EGiME（熵引导模型评估），一种基于特征空间聚类结构的半监督准则，无需分类器训练即可在少样本场景下评估预训练模型，在五个基准上与真实准确率相关性达 0.966，并比有监督基线提供更稳定的模型排名。
IRIS 是一种利用扰动筛选与迁移学习重建跨细胞类型细胞信号传导历史的新方法，发表于 Nature Methods。
该论文研究如何将一个模型表示上学到的量化校准增益复用到同一预训练模型的其他表示（如浮点与INT3量化）上，无需接收端优化即可保留大部分文本损失改善，量化了校准更新的可迁移性与接收端专属训练的剩余收益。
MIT researchers led by Pulin Li developed IRIS, a neural network model that detects unique gene-activity 'fingerprints' left by cellular signaling pathways, allowing them to reconstruct a cell's signaling history across many cell types using transfer learning. Published in Nature Methods, the approach could accelerate stem cell engineering, organoid development, and disease research.
提出了 SkillGrain 框架，将用户反馈分解为可独立执行的原子规则，预测其可迁移性并只保留适用于未来请求的部分，通过强化学习联合优化技能归纳与条件生成，以实现可复用的个性化技能学习。
Game2Real-Bench 是一个研究玩具游戏能否反映并迁移到真实世界视觉推理的基准，涵盖 25 个 VLM 的评测，发现游戏与真实任务间存在高达 0.935 的强相关性，并发布了包含约 24K 验证样本的 Game2Real-Dataset。
该论文研究了如何从净智能电表数据中恢复户用光伏发电（behind-the-meter PV）可见性，提出四种部署场景并基准测试深度序列模型，发现部署场景对误差的影响远大于模型选择，且少量目标域微调可显著降低迁移误差。
ICLR 2026 论文提出 Brain-IT 系统，可通过 fMRI 脑活动重建人眼所见图像，采用语义与结构双分支引导扩散模型，并能以新受试者仅约1小时的数据实现迁移学习。
该论文提出WATCH，一种跨视频预训练方法，通过任务配对的人类演示学习可迁移的世界-动作模型，无需人-机器人配对数据；在机器人后训练后将MimicDroid的上下文内成功率从40.9%提升至68.3%（Cosmos Policy）以及从59.8%提升至77.7%（Cosmos 3）。
A Johns Hopkins and Carnegie Mellon paper shows a 4B editor model can learn to rewrite an agent's harness code from failure reports, with edits transferring to unseen tasks and outperforming its 35B teacher on average edit scores.
This paper shows that when selecting math training data for LLM fine-tuning, sharing a reasoning approach (e.g., invariant reasoning or double counting) yields significantly greater transfer to held-out targets than sharing a mathematical topic, with advantages of 8–16 percentage points across five base models in counterbalanced 2x2 designs.
该论文研究提示数量、来源与选择如何影响在线策略蒸馏（OPD）的迁移效果，发现提示效用取决于师生模型对与目标任务，且定向选择策略（如语义多样性、难度、师生分歧等）并未一致优于均匀随机采样。
The paper introduces RouteFM, a foundation model for LLM routing that learns reusable routing capabilities via episodic pretraining across heterogeneous environments, allowing a frozen router to adapt to new domains, modalities, and candidate pools through behavioral context alone. RouteFM outperforms the strongest baseline by 2.23 quality points on MMR-Bench with only eight observations per candidate, supporting a 'pretrain once, route anywhere' paradigm.
EVOKE is a post-training method that elicits world knowledge already present in pretrained LLM agents by contrasting alternative goals at fixed states, training policies to rank candidate actions by their consequences rather than contextual habits. It achieves state-of-the-art results across ALFWorld, WebShop, and search-based QA, and transfers well to unseen environments without needing world models or inference-time planning.