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
生成时间: 2026/7/21 09:00:05
📊 今日概况
| 方向 | 论文数 |
|---|---|
| 🧮 算法与架构 | 12 |
| 🏋️ 训练方法 | 5 |
| ⚡ 推理优化 | 3 |
| 总计扫描 | 50 |
📝 论文列表
🧮 算法与架构 (12 篇)
1. ToolSciVer: Multimodal Scientific Claim Verification with Visual Tool Augmented Reinforcement Learning
- arXiv: 2607.16131
- 摘要: toolsciver,scientific,claim,tool,multimodal,mscv,visuals,verification,grpo,visual
- 关键词: toolsciver,scientific,claim,tool,multimodal,mscv,visuals,verification,grpo,visual
2. Controlling Implicit Shortcut Reliance in L2 Spoken English Auto-markers
- arXiv: 2607.16085
- 摘要: reliance,malpractice,shortcut,shortcuts,classifier,assessment,markers,language,audio,input
- 关键词: reliance,malpractice,shortcut,shortcuts,classifier,assessment,markers,language,audio,input
3. Frontier Language Models Struggle to Copy: Text Can Be Better Viewed in 2D
- arXiv: 2607.16072
- 摘要: copying,copy,encodings,rope,positional,language,frontier,input,text,dclm
- 关键词: copying,copy,encodings,rope,positional,language,frontier,input,text,dclm
4. Frontier AI performance across the business disciplines: a case-grounded benchmark of knowledge work and analytical reasoning
- arXiv: 2607.16057
- 摘要: business,businesscasebench,instructor,frontier,disciplines,work,schools,benchmark,analytical,case
- 关键词: business,businesscasebench,instructor,frontier,disciplines,work,schools,benchmark,analytical,case
5. Induction in Both Directions: A Mechanistic Analysis of In-Context Learning in Masked Diffusion Language Models
- arXiv: 2607.15893
- 摘要: dlms,induction,context,masked,token,timestep,bidirectional,heads,mechanistic,visible
- 关键词: dlms,induction,context,masked,token,timestep,bidirectional,heads,mechanistic,visible
6. Cost-efficient generative AI summarization for scalable automated essay scoring in educational assessment
- arXiv: 2607.15829
- 摘要: scoring,gpt,summarization,essay,aes,essays,summary,writing,generative,scalable
- 关键词: scoring,gpt,summarization,essay,aes,essays,summary,writing,generative,scalable
7. Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoning
- arXiv: 2607.15736
- 摘要: distillation,reasoning,bird,concise,prefixes,self,bootstrapped,policy,brevity,qwen3
- 关键词: distillation,reasoning,bird,concise,prefixes,self,bootstrapped,policy,brevity,qwen3
8. VarRate: Training-Free Variable-Rate KV Cache Compression for Long-Context LLMs
- arXiv: 2607.15498
- 摘要: varrate,rank,token,query,budget,cache,collapses,snapkv,degrades,reuse
- 关键词: varrate,rank,token,query,budget,cache,collapses,snapkv,degrades,reuse
9. PagedWeight: Efficient MoE LLM Serving with Dynamic Quality-Aware Weight Quantization
- arXiv: 2607.16184
- 摘要: pagedweight,moe,serving,memory,cache,throughput,quantization,llm,quality,tradeoff
- 关键词: pagedweight,moe,serving,memory,cache,throughput,quantization,llm,quality,tradeoff
10. PRISA: Proactive Infrastructure LiDAR Framework for Intersection Safety Assessment
- arXiv: 2607.16156
- 摘要: prisa,ppet,intersection,proactive,ttc,perception,lidar,safety,assessment,conflicts
- 关键词: prisa,ppet,intersection,proactive,ttc,perception,lidar,safety,assessment,conflicts
11. Understanding Reasoning from Pretraining to Post-Training
- arXiv: 2607.16097
- 摘要: pretraining,chess,post,reasoning,sft,puzzles,training,pipeline,testbed,language
- 关键词: pretraining,chess,post,reasoning,sft,puzzles,training,pipeline,testbed,language
12. Revisiting data-driven dynamic security assessment with a tabular foundation model
- arXiv: 2607.16031
- 摘要: contingencies,tfm,labelled,contingency,edc,foundation,unseen,tabular,credible,dynamic
- 关键词: contingencies,tfm,labelled,contingency,edc,foundation,unseen,tabular,credible,dynamic
🏋️ 训练方法 (5 篇)
1. Process Reward Informed Tree Rollout for Effective Multi-Turn RL
- arXiv: 2607.15610
- 摘要: patr,rollout,tree,turn,frozenlake,swe,agentic,exploration,bench,grpo
- 关键词: patr,rollout,tree,turn,frozenlake,swe,agentic,exploration,bench,grpo
2. Understanding Reasoning from Pretraining to Post-Training
- arXiv: 2607.16097
- 摘要: pretraining,chess,post,reasoning,sft,puzzles,training,pipeline,testbed,language
- 关键词: pretraining,chess,post,reasoning,sft,puzzles,training,pipeline,testbed,language
3. Constrained Hebbian Learning Supports Efficient Representational Allocation under Structural Constraints
- arXiv: 2607.16027
- 摘要: hebbian,cti,representational,audiovisual,synaptic,constraints,allocation,information,cost,ddtp
- 关键词: hebbian,cti,representational,audiovisual,synaptic,constraints,allocation,information,cost,ddtp
4. An Exploratory Study of Single Channel Surface Electromyography for Hand Gesture Classification
- arXiv: 2607.15972
- 摘要: semg,gesture,electromyography,hand,pearson,channel,filtering,exploratory,single,surface
- 关键词: semg,gesture,electromyography,hand,pearson,channel,filtering,exploratory,single,surface
5. Knowledge-Guided Cross-Modal Fusion for Adult-to-Pediatric ECG Transfer via Label-Conditioned Contrastive Alignment
- arXiv: 2607.15928
- 摘要: ecg,pediatric,adult,peace,label,alignment,knowledge,shot,transfer,tokens
- 关键词: ecg,pediatric,adult,peace,label,alignment,knowledge,shot,transfer,tokens
⚡ 推理优化 (3 篇)
1. An MLIR-Based Compilation Method for Large Language Models
- arXiv: 2607.15865
- 摘要: mlir,tpu,prefill,topop,tpuop,compilation,sophgo,chip,dialect,layer
- 关键词: mlir,tpu,prefill,topop,tpuop,compilation,sophgo,chip,dialect,layer
2. VarRate: Training-Free Variable-Rate KV Cache Compression for Long-Context LLMs
- arXiv: 2607.15498
- 摘要: varrate,rank,token,query,budget,cache,collapses,snapkv,degrades,reuse
- 关键词: varrate,rank,token,query,budget,cache,collapses,snapkv,degrades,reuse
3. PagedWeight: Efficient MoE LLM Serving with Dynamic Quality-Aware Weight Quantization
- arXiv: 2607.16184
- 摘要: pagedweight,moe,serving,memory,cache,throughput,quantization,llm,quality,tradeoff
- 关键词: pagedweight,moe,serving,memory,cache,throughput,quantization,llm,quality,tradeoff
分析生成失败
📚 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