1
Self-Guided Test-Time Training for Long-Context LLMs
自我引导的测试时训练，专治长上下文LLM的准确率退化难题。
arXiv:2607.09415v1 Announce Type: cross Abstract: Long-context processing has become increasingly important for large language models (LLMs), but simp…
2
TF-Engram: A Train-Free Engram with SSD-Backed Memory for Large Language Models
无需额外训练的SSD记忆增强方案，让LLM轻松扩展长上下文记忆。
arXiv:2607.07388v1 Announce Type: cross Abstract: Large Language Models (LLMs) store factual knowledge and domain-specific patterns implicitly in dens…
3
Learning When to Attend: Conditional Memory Access for Long-Context LLMs
提出条件性记忆访问机制，引导长上下文LLM更智能地选择注意力范围，显著提升效率与性能，ICML 2026最新成果。
arXiv:2603.17484v2 Announce Type: replace-cross Abstract: Language models struggle to generalize beyond pretraining context lengths, limiting long-hor…
4
ReContext: Recursive Evidence Replay as LLM Harness for Long-Context Reasoning
长上下文推理新突破，递归证据回放让LLM精准捕捉关键信息
arXiv:2607.02509v1 Announce Type: new Abstract: Understanding and reasoning over long contexts has become a key requirement for deploying large langua…
5
MosaicKV: Serving Long-Context LLM with Dynamic Two-D KV Cache Compression
长上下文LLM推理提速新方法，MosaicKV通过动态二维KV缓存压缩，显著降低显存占用并保持精度。
arXiv:2607.00760v1 Announce Type: new Abstract: Long-context LLM services now sustain prompts with hundreds of thousands to millions of tokens, making…
6
Predict, Reuse, and Repair: Accelerating Dynamic Sparse Attention for Long-Context LLM Decoding
提出预测-重用-修复机制，动态稀疏注意力加速长上下文LLM解码，降低推理延迟。
arXiv:2606.30389v1 Announce Type: new Abstract: Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV…
7
Show HN:Swarm intelligence without degradation using two Qwen models
用两个Qwen模型打造无降级群智能系统，实现长上下文、持久记忆与深度推理，周末项目也能颠覆AI订阅模式！
I recently started a weekend project—partly because I was thinking about cancelling my AI subscriptions—with the goal of creating a system capable of …
8
Test-Time Training with Next-Token Prediction
挑战测试时训练新范式：用语言模型自监督的下一词预测信号作为内循环目标，无需额外损失函数。
arXiv:2606.21803v1 Announce Type: new Abstract: Next-token prediction is the self-supervised signal that trains language models, and every observed pr…
9
Governance Decay: How Context Compaction Silently Erases Safety Constraints in Long-Horizon LLM Agents
揭示LLM长对话中上下文压缩悄然抹除安全护栏的隐患，值得AI安全领域关注。
arXiv:2606.22528v1 Announce Type: new Abstract: Modern LLM agents increasingly rely on context compaction, summarization, or eviction to keep long-run…
10
Show HN: Subconscious and GLM-5.2 Makes "/compact" Obsolete
百万级记忆+高效推理，GLM-5.2与Subconscious让压缩技术成为历史。
GLM-5.2 is a turning point for coding agents. It's the first model a business would actually pay to replace Claude Opus with. We gave GLM-5.2 the abil…
11
An open-source AI just beat OpenAI's GPT-5.5 at coding (1/6th the price)
开源模型GLM-5.2以1/6成本击败GPT-5.5编码能力，支持百万级Token上下文，工程级长任务更稳更省。
Article URL: https://docs.z.ai/guides/llm/glm-5.2 Comments URL: https://news.ycombinator.com/item?id=48581610 Points: 1 # Comments: 0
12
CentroidKV: Efficient Long-Context LLM Inference via KV Cache Clustering
通过KV缓存聚类实现长上下文LLM推理效率突破，降低显存占用同时保持精度。
arXiv:2506.11418v2 Announce Type: replace Abstract: Large language models (LLMs) with extended context windows have become increasingly prevalent for …
13
月之暗面 Kimi K2.7 Code 模型高速版上线：2 倍价格、最快 6 倍速
Kimi K2.7 Code模型高速版输出速度提升5-6倍，token消耗减少30%，长上下文编程效率飙升。
IT之家 6 月 15 日消息，月之暗面 Kimi 今日宣布， Kimi K2.7 Code 模型高速版上线，现已向 Kimi Code Beta 计划成员、Kimi API 开发者、Kimi Business 用户开放。 据介绍，高速版与 Kimi K2.7 Code 是相同模型，输出速度约为普通…
14
月之暗面开源 Kimi K2.7 Code 编程模型，预告 6 倍速高速版周一见
开源编程模型，擅长长上下文代码任务，token消耗减少30%，即将推出6倍速版本。
IT之家 6 月 12 日消息，月之暗面 Kimi 今日发布并开源 Kimi K2.7 Code 编程模型。 官方表示，内外部基准评估显示，Kimi K2.7 Code 相比 K2.6 模型显著提升了长上下文编程场景的指令遵循能力、长程编程任务的性能表现，并且大幅改善了在长程任务中的过度思考倾向，平…
15
Prefilling-dLLM: Predictive Prefilling for Long-Context Inference in Diffusion Language Models
提出预测预填充方法，让扩散语言模型高效处理超长上下文，解码速度显著提升。
arXiv:2606.10537v1 Announce Type: new Abstract: Diffusion large language models (dLLMs) re-encode the entire prefix at every denoising step, causing r…
16
LANTERN: Layered Archival and Temporal Episodic Retrieval Network for Long-Context LLM Conversations
针对LLM长对话中历史信息丢失问题，提出分层存档与时间情景检索网络，轻量高效恢复关键细节。
arXiv:2606.05182v1 Announce Type: new Abstract: Large language models discard critical details when conversation history is compacted to fit within fi…
17
SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference
稀疏注意力新突破：解耦选择与计算，解决KV缓存瓶颈与PCIe传输困境。
arXiv:2606.04511v1 Announce Type: new Abstract: Sparse attention reduces compute and memory bandwidth for long-context LLM inference. However, two key…
18
SoLoPO: Unlocking Long-Context Capabilities in LLMs via Short-to-Long Preference Optimization
SoLoPO提出短到长偏好优化，高效解锁大模型长上下文能力，被ICLR 2026录用。
arXiv:2505.11166v3 Announce Type: replace-cross Abstract: Despite advances in pretraining with extended context sizes, large language models (LLMs) st…
19
MemTrain: Self-Supervised Context Memory Training
自监督训练上下文记忆，为长文本理解提供新范式
arXiv:2606.03197v1 Announce Type: new Abstract: Memory is an indispensable capability for long-horizon LLM agents, enabling them to preserve and utili…
20
InfoMem: Training Long-Context Memory Agents with Answer-Conditioned Information Gain
新方法InfoMem用答案条件信息增益训练长上下文记忆代理，大幅提升模型知识检索与记忆能力。
arXiv:2606.03329v1 Announce Type: new Abstract: Long-context tasks require LLMs to identify and preserve answer-relevant information from large contex…