Recent Posts

DoubtLess: Training-Free Contrastive Decoding for Efficient Reasoning

DoubtLess：无需训练的对比解码如何为大模型推理瘦身

10/14/2026

arxivaics.ai

Two Clocks: Why Larger Rollout Batches Learn Slower but Fail on Schedule in Reinforcement Learning with Verifiable Rewards

双时钟假说：为什么大 Rollout Batch 在 RLVR 中学得更慢却按时崩盘

10/14/2026

arxivaics.ai

AI Models Can Provably Hide Arbitrary Capabilities

模型能够可证明地隐藏任意能力：数字保险箱击穿黑盒红队审计

10/12/2026

arxivaics.lg

Images Talk Back: Guiding MMDiTs by Suppressing Image-to-Text Feedback

图像会回嘴：MMDiT 如何通过抑制图到文反馈重拾提示词忠实度

10/12/2026

arxivaics.cv

RSI-Anything: Unlocking Recursive Self-Improvement with World Models

RSI-Anything：世界模型驱动的跨领域递归自进化

10/12/2026

arxivaics.cl

Caught in the Act: Probes Effectively Detect Sabotage and Catch Unverbalized Deception

人赃俱获：探针如何捕获大模型的潜伏破坏与心口不一

10/9/2026

arxivaics.lg

DreamTrue: Action-Faithful Robot World Model with Counterfactual Post-Training

DreamTrue：基于反事实后训练的机器人忠实世界模型

10/9/2026

arxivaics.ro

Rounding in Preconditioner Space: Redesigning 4-bit AdamW Optimizer-State Quantization

预条件子空间舍入：重塑 4-bit AdamW 优化器状态量化

10/9/2026

arxivaics.lg

Decoupling Exploration from Optimization in RLVR

解耦 RLVR 中的探索与优化：ExpDis 框架如何激发大模型推理多样性

10/8/2026

arxivaics.ai

EngramEdit: Decoupled Knowledge Updates in LLMs through Conditional Memory

EngramEdit：利用条件记忆解耦大语言模型的事实知识编辑

10/8/2026

arxivaics.cl

Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts?

装瓶智能体：大模型能否将自身能力编译为廉价高效的复用制品？

10/7/2026

arxivaics.ai

World Models' Last Exam in Physics

世界模型的物理学终考：基于物理测量的严苛基准检验

10/7/2026

arxivaics.cv