“2026-06-22”版本间的差异

来自cslt Wiki
跳转至： 导航、 搜索
 
（11位用户的15个中间修订版本未显示）
第6行： 第6行：
 
|Dong Wang
 
|Dong Wang
 
||
 
||
−
*
+
* All middle school and primary school text book done
 +
* Jiaoyubu project evaluation
 +
 
 
||
 
||
 
*
 
*
第28行： 第30行：
 
|Wenqiang Du
 
|Wenqiang Du
 
||
 
||
−
*
+
* music sep
−
||
+
 
−
*
+
*AIbabel
 +
** fixed some bugs on the project
 +
** Trying some method to optimize the Fewshot model
 +
 
 +
* Huohua
 +
** simulated 3-mic data for mobile, and used the Sound Bubble code to build a baseline(in traing)
 
||
 
||
 
*
 
*
第39行： 第46行：
 
|Yang Wei
 
|Yang Wei
 
||
 
||
−
*  
+
* Audio separation: train model with larger 3mix ratio.
 
||
 
||
 
*
 
*
第50行： 第57行：
 
|Yue Gu
 
|Yue Gu
 
||
 
||
−
* polish my Phd thesis and begin to prepare my defense
+
* polish my Phd thesis and begin to prepare my pre-defense
 
||
 
||
 
*
 
*
第72行： 第79行：
 
|Pengqi Li
 
|Pengqi Li
 
||
 
||
−
*  
+
* Phd Thesis(core content and overall framework)
 
||
 
||
 
*
 
*
第83行： 第90行：
 
|Junming Yuan
 
|Junming Yuan
 
||
 
||
−
*  
+
* Updated the MT-HuBERT project code.
 
||
 
||
−
*
+
* Start MT-HuBERT+ work.
 
||
 
||
 
*
 
*
第94行： 第101行：
 
|Yu Zhang
 
|Yu Zhang
 
||
 
||
−
*  
+
* GPU Util: [https://z1et6d3xtb.feishu.cn/wiki/XX4NwX3tJiBDcgkMi0hcFUtInHh]
 +
* MAS:
 +
** Across all experiments, the MAS was generally able to identify the underlying problem or bottleneck.
 +
** When the task mainly required textual consensus, STD added limited value beyond the agreement itself.
 +
** When solving the problem required coordinated actions across agents, STD was much more useful.
 
||
 
||
 
*
 
*
第105行： 第116行：
 
|Junhui Chen
 
|Junhui Chen
 
||
 
||
−
*  
+
* Continue the LLM-MAS society dilemma experiments:
 +
** In Public Goods Game (PGG): all topologies converge to zero cooperation.
 +
**  In Common-Pool Resource (CPR): when LLMs are in Non-Thinking mode, all topologies collapse; when in Thinking mode, fully connected causes over-extraction to spike while sparse topologies stay sustainable. The difference tracks degree, not the optimization method.
 
||
 
||
 
*
 
*
第116行： 第129行：
 
|Xiaoxue Luo
 
|Xiaoxue Luo
 
||
 
||
−
*  
+
* Completed the draft of the USS-TDA paper
 
||
 
||
 
*
 
*
第127行： 第140行：
 
|Bochao Hu
 
|Bochao Hu
 
||
 
||
−
*  
+
* topic vsr llm: revise paper (VSR-VLLM as main storyline) and continue to supplement exps
 +
* cncvs3: 4kh SFT is in training
 
||
 
||
 
*
 
*
第138行： 第152行：
 
|Hongcheng Zhang
 
|Hongcheng Zhang
 
||
 
||
−
*   
+
*  Try different ratios of real data and synthetic data to see if the synthetic data actually works on Inhome-audio tagging
 
||
 
||
 
*
 
*
第160行： 第174行：
 
|Shuailong Li
 
|Shuailong Li
 
||
 
||
−
*
+
*Clean dataset, keep watching Li Hongyi's course, and do the homework
 
||
 
||
 
*
 
*

2026年6月22日 (一) 11:13的最后版本

People This Week Next Week Task Tracking (DeadLine)
Dong Wang
  • All middle school and primary school text book done
  • Jiaoyubu project evaluation
Lantian Li
Wenqiang Du
  • music sep
  • AIbabel
    • fixed some bugs on the project
    • Trying some method to optimize the Fewshot model
  • Huohua
    • simulated 3-mic data for mobile, and used the Sound Bubble code to build a baseline(in traing)
Yang Wei
  • Audio separation: train model with larger 3mix ratio.
Yue Gu
  • polish my Phd thesis and begin to prepare my pre-defense
Lily
Pengqi Li
  • Phd Thesis(core content and overall framework)
Junming Yuan
  • Updated the MT-HuBERT project code.
  • Start MT-HuBERT+ work.
Yu Zhang
  • GPU Util: [1]
  • MAS:
    • Across all experiments, the MAS was generally able to identify the underlying problem or bottleneck.
    • When the task mainly required textual consensus, STD added limited value beyond the agreement itself.
    • When solving the problem required coordinated actions across agents, STD was much more useful.
Junhui Chen
  • Continue the LLM-MAS society dilemma experiments:
    • In Public Goods Game (PGG): all topologies converge to zero cooperation.
    • In Common-Pool Resource (CPR): when LLMs are in Non-Thinking mode, all topologies collapse; when in Thinking mode, fully connected causes over-extraction to spike while sparse topologies stay sustainable. The difference tracks degree, not the optimization method.
Xiaoxue Luo
  • Completed the draft of the USS-TDA paper
Bochao Hu
  • topic vsr llm: revise paper (VSR-VLLM as main storyline) and continue to supplement exps
  • cncvs3: 4kh SFT is in training
Hongcheng Zhang
  • Try different ratios of real data and synthetic data to see if the synthetic data actually works on Inhome-audio tagging
Weiman Sun
Shuailong Li
  • Clean dataset, keep watching Li Hongyi's course, and do the homework
取自“http://cslt.org/mediawiki/index.php?title=2026-06-22&oldid=44265”