發(fā)簡信
IP屬地：北京
-
從 CoT 到 RAP：用世界模型增強(qiáng)大模型推理論文標(biāo)題：Reasoning with Language Model is Planning with World Model論文鏈接：http...
-
Memory-R1：用強(qiáng)化學(xué)習(xí)讓大模型智能體學(xué)會(huì)管理長期記憶論文標(biāo)題：Memory-R1: Enhancing Large Language Model Agents to Manage and Util...
-
Reflexion：讓語言智能體通過語言反饋?zhàn)晕覐?qiáng)化論文標(biāo)題：Reflexion: Language Agents with Verbal Reinforcement Learning論文鏈接：h...
-
ToolRL：工具調(diào)用不是格式模仿，而是獎(jiǎng)勵(lì)學(xué)習(xí)論文標(biāo)題：ToolRL: Reward is All Tool Learning Needs論文鏈接：https://arxiv.org/abs...
-
ACL 2025 - 基于片段監(jiān)督偏好優(yōu)化的字幕翻譯時(shí)延對(duì)齊論文標(biāo)題：Fine-grained Video Dubbing Duration Alignment with Segment Supervis...
-
直接偏好優(yōu)化技術(shù)DPO基礎(chǔ)理論及推導(dǎo)
論文標(biāo)題：Direct Preference Optimization: Your Language Model is Secretly a R...
-
自適應(yīng)視圖增強(qiáng)的謠言檢測(cè)圖對(duì)比學(xué)習(xí)方法論文標(biāo)題：Propagation Tree Is Not Deep: Adaptive Graph Contrastive Learning A...
-
生成式大模型的RLHF技術(shù)（一）：基礎(chǔ)一、概述 大語言模型（LLMs）在預(yù)訓(xùn)練的過程中通常會(huì)捕捉數(shù)據(jù)的特征，而這些訓(xùn)練數(shù)據(jù)通常既包含高質(zhì)量的也包含低質(zhì)量的，因此模型有時(shí)會(huì)產(chǎn)生不被期望...
-
LoRA：大模型下游任務(wù)的低秩適應(yīng)論文標(biāo)題：LoRA: Low-Rank Adaptation of Large Language Models論文鏈接：https://arxi...