Loading...
#GVPO++:GroupVariancePolicyOptimizationforLLMPost‑TrainingandOn‑PolicyDistillation>论文链接：https://arxiv.org/a
RBS‑Attention 论文阅读总结
#RBS‑Attention论文阅读总结>论文信息：arXiv:2609.20971，RBS‑Attention:Radius‑BoundedSparsePrefillforLong‑ContextLargeLangu
Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning（arXiv:2609.03430）
#RandomAttention:RethinkingKVCacheEvictionforEfficientReasoning（arXiv:2609.03430）##一、文章主要内容总结###研究背景大模型在长链
SWE‑Gate 论文阅读总结
#SWE‑Gate论文阅读总结>论文信息：arXiv:2609.04167，SWE‑Gate:PassingFunctionalTestsIsNotEnoughforSoftwareEngineeringAgen
What Matters for Aggressive Decoding‑Time KV Eviction? Temporal Aggregation and Ranking Preservation
#WhatMattersforAggressiveDecoding‑TimeKVEviction?TemporalAggregationandRankingPreservation>arXiv:2609.0351
GrowPage: On‑Demand KV Budgeting for Efficient LLM Reasoning Serving（arXiv:2609.03494）
#GrowPage:On‑DemandKVBudgetingforEfficientLLMReasoningServing（arXiv:2609.03494）>论文链接：[https://arxiv.org/abs/2
Large Language Models in Resolving Contextual Knowledge Conflicts（arXiv:2609.03148）
#LargeLanguageModelsinResolvingContextualKnowledgeConflicts（arXiv:2609.03148）>会议：EMNLP2026>论文链接：https://arx
RuleMem: Active Rule Memory for Long‑Term Conversational Agents 论文总结与翻译
#RuleMem:ActiveRuleMemoryforLong‑TermConversationalAgents论文总结与翻译>论文arXiv链接：https://arxiv.org/abs/2609.03915
Representational alignment yields generalizable safety in language models
>说明：arXiv网页摘要接口仅公开**Abstract**，**Introduction、Conclusion**正文内容不在返回的元数据中，无法直接获取原文，因此仅完成摘要原文+翻译；文章主要内容、创新点基于摘要信息整理。#Re
It‘s the Problem, Not the Path: Budget and Difficulty Confounds in LLM Reasoning Trajectories
#It'stheProblem,NotthePath:BudgetandDifficultyConfoundsinLLMReasoningTrajectories>论文arXiv地址：https://arxi
论文总结：Doesn‘t Stop Reasoning: Analysis of Spurious CoT Termination
#论文总结：Doesn'tStopReasoning:AnalysisofSpuriousCoTTermination>论文arXiv：https://arxiv.org/abs/2609.03633>会议：EMNL
GeoEdit: Geometric Knowledge Editing for Large Language Models
###摘要定期更新对于保持大语言模型（LLMs）中的知识时效性至关重要。因此，人们开发了各种模型编辑方法来更新LLMs中的特定知识。然而，基于训练的方法往往难以在有效融入新知识的同时保留不相关的通用知识。为应对这一挑战，我们提出了
Survey of Large Multimodal Reasoning Models (LMRMs): From Modular to Native (arXiv:2505.04921)
###文章总结本文是哈尔滨工业大学（深圳）团队发表的关于大型多模态推理模型（LMRMs）的综述论文，系统梳理了多模态推理模型的发展脉络、技术挑战及未来方向。核心内容如下：####**主要内容**1.**多模态推理模型
MemoryLACE论文总结与翻译
#MemoryLACE论文总结与翻译>论文信息：arXiv:2609.03201，MemoryLACE:MemoryLifecycle‑AwareConsolidationandEvidenceRetrieval##
FlowBalance 论文阅读笔记
#FlowBalance论文阅读笔记>论文信息：FlowBalance:Verifier‑GroundedSelf‑ImprovementfromOn‑PolicyReasoningExperience，arXiv:26
The Geometr\(\boldsymbol{y}\) of Ignorance: LLMs Know \(\bolds\(\boldsymbol{y}\)mbol{\(\boldsymbol{W
#TheGeometr\(\boldsymbol{y}\)ofIgnorance:LLMsKnow\(\bolds\(\boldsymbol{y}\)mbol{\(\boldsymbol{W}\)}\)hentoTempe
ALRA: Adaptive Local Relational Alignment 论文总结与翻译
#ALRA:AdaptiveLocalRelationalAlignment论文总结与翻译>论文信息：arXiv:2609.03355，基于Logit的自回归语言模型预训练蒸馏自适应局部关系对齐方法##一、文章主要内容
Dream‑RSI: Recursive Self‑Improvement through Evolving Worlds 论文总结与翻译
#Dream‑RSI:RecursiveSelf‑ImprovementthroughEvolvingWorlds论文总结与翻译>论文信息：arXiv:2609.14858（v2，2026‑10‑06）>作者：Tong
Meta-Reasoner: Dynamic Guidance for Optimized Inference-time Reasoning in Large Language Models
###摘要大语言模型（LLMs）越来越依赖冗长的推理链来解决复杂任务。然而，这种试错方法往往会导致高昂的计算成本和错误传播，早期的错误可能会使后续步骤偏离正轨。为解决这些问题，我们引入了Meta-Reasoner框架，它通过让大语
Collaborative Stance Detection via Small-Large Language Model Consistency Verification
###摘要社交媒体上的立场检测旨在识别推文中对特定目标表达的态度。当前研究由于大语言模型（LLMs）具有显著的性能提升，因而更倾向于使用它们，而非小语言模型（SLMs）。然而，对于需要大量数据分析的现实社交媒体监测系统而言，不计成本地
