LLM & SLM 研究日报 - 2026-07-22

LLM & SLM 研究日报

算法·训练·推理 —— 大语言模型与小语言模型的前沿研究

数据源: cs.CL + cs.LG

生成时间: 2026/7/22 09:01:05

📊 今日概况

方向 论文数
🧮 算法与架构 10
🏋️ 训练方法 7
⚡ 推理优化 7
总计扫描 50

📝 论文列表

🧮 算法与架构 (10 篇)

1. SWE-Pruner Pro: The Coder LLM Already Knows What to Prune

  • arXiv: 2607.18213
  • 摘要: swe,pruner,pro,prune,context,agent,oolong,coder,knows,pruning
  • 关键词: swe,pruner,pro,prune,context,agent,oolong,coder,knows,pruning

2. PPL-Factory: Task-Aware and Budget-Aware Data Selection from Language Modeling to Reasoning

  • arXiv: 2607.18199
  • 摘要: ppl,aware,factory,selection,perplexity,budget,gsm8k,data,task,reasoning
  • 关键词: ppl,aware,factory,selection,perplexity,budget,gsm8k,data,task,reasoning

3. When a Name Is Not a Name: A Benchmark Dataset and Distilled Reasoning for Culturally Entangled Bangla Homographs in Low-Resource LLMs

  • arXiv: 2607.17828
  • 摘要: bangla,culturally,name,cultural,homographs,entangled,bias,reading,pretraining,labelled
  • 关键词: bangla,culturally,name,cultural,homographs,entangled,bias,reading,pretraining,labelled

4. ESCUCHA: A Spanish Speech Benchmark for Heterogeneous Acoustic Conditions

  • arXiv: 2607.17812
  • 摘要: escucha,spanish,lalms,speech,audio,questions,acoustic,reasoning,benchmark,heterogeneous
  • 关键词: escucha,spanish,lalms,speech,audio,questions,acoustic,reasoning,benchmark,heterogeneous

5. C$^2$KV: Compressed and Composable KV Cache Reuse for Efficient LLM Inference

  • arXiv: 2607.17715
  • 摘要: reuse,cache,composable,inference,llm,compression,prefix,concatenation,compressed,long
  • 关键词: reuse,cache,composable,inference,llm,compression,prefix,concatenation,compressed,long

6. How Reliable Are Multimodal Signals of Conversational State? Evidence from Remote Dyadic Collaborative Tasks

  • arXiv: 2607.17452
  • 摘要: conversational,task,speaker,reliable,multimodal,features,dyadic,cognitive,load,reliability
  • 关键词: conversational,task,speaker,reliable,multimodal,features,dyadic,cognitive,load,reliability

7. The Librarian Who Refused to Code: Model-Dependent Identity Enactment in LLM Code Generation

  • arXiv: 2607.17420
  • 摘要: persona,librarian,personas,code,completions,engineer,opus,refused,correctness,responses
  • 关键词: persona,librarian,personas,code,completions,engineer,opus,refused,correctness,responses

8. FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications

  • arXiv: 2607.18171
  • 摘要: flashrt,agent,deployments,latency,multimodal,harness,mi355x,throughput,amd,omni
  • 关键词: flashrt,agent,deployments,latency,multimodal,harness,mi355x,throughput,amd,omni

9. Adaptive Mamba Neural Operators

  • arXiv: 2607.18043
  • 摘要: amo,mamba,meshes,pdes,ssms,adaptive,geometries,takenaka,afd,neural
  • 关键词: amo,mamba,meshes,pdes,ssms,adaptive,geometries,takenaka,afd,neural

10. MADA-RL: Multi-Agent Debate-Aware Reinforcement Learning for Parameter-Efficient Reasoning in Compact Models

  • arXiv: 2607.18006
  • 摘要: mada,critic,generator,critics,debate,reasoning,advantage,deepscaler,compact,counterfactual
  • 关键词: mada,critic,generator,critics,debate,reasoning,advantage,deepscaler,compact,counterfactual

🏋️ 训练方法 (7 篇)

  • arXiv: 2607.18114
  • 摘要: cue,bias,biases,sycophancy,induced,alignment,tuning,llms,intervention,susceptibility
  • 关键词: cue,bias,biases,sycophancy,induced,alignment,tuning,llms,intervention,susceptibility

2. What Transfers Under Source Shift? Definitions, Examples, and Fine-Tuning for Climate Disclosure Classification

  • arXiv: 2607.17952
  • 摘要: source,disclosure,climate,strategies,definitions,disclosures,shift,tuning,adaptation,examples
  • 关键词: source,disclosure,climate,strategies,definitions,disclosures,shift,tuning,adaptation,examples

3. DeLIVeR: Decomposed Learning for Information-grounded Veracity Recognition via Reinforced Knowledge Graph Exploration

  • arXiv: 2607.17935
  • 摘要: deliver,reinforced,veracity,decomposed,grounded,grpo,planner,exploration,politifact,qwen2
  • 关键词: deliver,reinforced,veracity,decomposed,grounded,grpo,planner,exploration,politifact,qwen2

4. When a Name Is Not a Name: A Benchmark Dataset and Distilled Reasoning for Culturally Entangled Bangla Homographs in Low-Resource LLMs

  • arXiv: 2607.17828
  • 摘要: bangla,culturally,name,cultural,homographs,entangled,bias,reading,pretraining,labelled
  • 关键词: bangla,culturally,name,cultural,homographs,entangled,bias,reading,pretraining,labelled

5. OR Else: A Differentiable Trust Region for Policy Optimization

  • arXiv: 2607.18163
  • 摘要: grpo,ppo,gae,reward,clipped,rollout,relative,texttt,score,clip
  • 关键词: grpo,ppo,gae,reward,clipped,rollout,relative,texttt,score,clip

6. Manifold-Constrained Hyper-Connections for Parameter-Efficient Finetuning

  • arXiv: 2607.18130
  • 摘要: mhc,peft,residual,finetuning,connections,lora,hyper,frozen,olmo,manifold
  • 关键词: mhc,peft,residual,finetuning,connections,lora,hyper,frozen,olmo,manifold

7. Information-Based Exploration via Random Features for Reinforcement Learning

  • arXiv: 2607.17981
  • 摘要: rfig,exploration,gain,random,information,reinforcement,deep,bonuses,optimism,algorithms
  • 关键词: rfig,exploration,gain,random,information,reinforcement,deep,bonuses,optimism,algorithms

⚡ 推理优化 (7 篇)

1. VDAR-Router: Adaptive LLMs Routing via Verbalized Query Difficulty Analysis Retrieval

  • arXiv: 2607.18098
  • 摘要: vdar,routing,router,difficulty,query,retrieval,verbalized,cost,suitability,llm
  • 关键词: vdar,routing,router,difficulty,query,retrieval,verbalized,cost,suitability,llm

2. C$^2$KV: Compressed and Composable KV Cache Reuse for Efficient LLM Inference

  • arXiv: 2607.17715
  • 摘要: reuse,cache,composable,inference,llm,compression,prefix,concatenation,compressed,long
  • 关键词: reuse,cache,composable,inference,llm,compression,prefix,concatenation,compressed,long

3. FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications

  • arXiv: 2607.18171
  • 摘要: flashrt,agent,deployments,latency,multimodal,harness,mi355x,throughput,amd,omni
  • 关键词: flashrt,agent,deployments,latency,multimodal,harness,mi355x,throughput,amd,omni

4. Differentiable Logic Gate Networks for Low-Latency EEG Classification on Edge Devices

  • arXiv: 2607.18149
  • 摘要: logic,diff,eeg,latency,mlp,dementia,times,differentiable,gate,difflogic
  • 关键词: logic,diff,eeg,latency,mlp,dementia,times,differentiable,gate,difflogic

5. Empowering On-Device Model Adaptation with an Edge AI Inference Accelerator

  • arXiv: 2607.18101
  • 摘要: accelerator,device,adaptation,edge,inference,cpu,empowering,backbone,training,repurposes
  • 关键词: accelerator,device,adaptation,edge,inference,cpu,empowering,backbone,training,repurposes

6. SelectInfer: Selective Neuron Loading and Computation for On-Device LLMs

  • arXiv: 2607.18081
  • 摘要: selectinfer,loading,selective,neuron,computation,llm,neurons,footprint,llms,memory
  • 关键词: selectinfer,loading,selective,neuron,computation,llm,neurons,footprint,llms,memory

7. FlashPDE: A Drop-in Fused Triton Operator Library for Neural PDE Solvers

  • arXiv: 2607.18020
  • 摘要: flashpde,pde,pytorch,fused,triton,library,execution,differentiable,drop,solvers
  • 关键词: flashpde,pde,pytorch,fused,triton,library,execution,differentiable,drop,solvers

今日技术热点

今日扫描到 算法与架构 10 篇、训练方法 7 篇、推理优化 7 篇。

算法与架构趋势

当前 LLM 架构正从纯 Transformer 向混合架构演进：SSM (Mamba) 和线性注意力在长序列场景展现优势，MoE 在推理成本可控的前提下持续扩展参数规模。小模型架构注重蒸馏和紧凑设计。

训练方法趋势

DPO 和直接偏好优化正在成为 RLHF 的高效替代方案。合成数据质量成为新的研究焦点。LoRA/QLoRA 已成为高效微调的事实标准。

推理优化趋势

INT4 量化 (GPTQ/AWQ) 已成熟，GGUF 格式让端侧部署成为可能。Speculative decoding 在线推理中逐步普及。KV cache 压缩是降低长上下文推理成本的关键。

关键洞察

  1. 架构多元化: Transformer 不再是唯一选择，SSM 和混合架构值得持续关注
  2. 对齐轻量化: DPO 系列方法降低了高质量对齐的门槛
  3. 推理即服务: 推理优化的研究热度反映了部署需求的爆发
  4. 小模型逆袭: 端侧 SLM 的设计思路与大模型差异显著，需要专门的技术栈
  5. 数据 > 算法: 训练数据质量对模型能力的影响被重新审视

学习建议

  1. 重点关注 Mamba/Mamba-2 和混合架构的最新论文
  2. 实践 DPO 训练流程，对比 RLHF 的效果差异
  3. 尝试 vLLM + 量化模型的端到端推理优化

注：GLM-5 API 未调用，此为备用分析

📚 附录

筛选关键词

算法: attention mechanism, mixture of experts, MoE, sparse attention, flash attention, rotary position, RoPE, grouped query, GQA, KV cache …

训练: pre-training, pretraining, post-training, fine-tuning, finetuning, supervised fine-tuning, SFT, alignment, RLHF, DPO …

推理: inference, serving, latency, throughput, speculative decoding, batching, continuous batching, PagedAttention, vLLM, quantization …

本报告由 OpenClaw 自动生成 | LLM & SLM Research Daily

相关文章推荐

tag:

    缺失模块。
    1、请确保node版本大于6.2
    2、在博客根目录（注意不是yilia根目录）执行以下命令：
    npm i hexo-generator-json-content --save

    3、在根目录_config.yml里添加配置： 

      jsonContent:
    meta: false
    pages: false
    posts:
      title: true
      date: true
      path: true
      text: false
      raw: false
      content: false
      slug: false
      updated: false
      comments: false
      link: false
      permalink: false
      excerpt: false
      categories: false
      tags: true

程序员，关于技术我所知甚少，唯有探索，方得真知。