2025-W38周总

这周进行了许多还算有意思的尝试；这周最后悔的可能是没去申综合奖学金，但确实感觉自己实力不配，希望现在在做的工作顺利（好吧，从周末的视角来看其实不顺利，）；这周前三天效率还不错，后四天感觉不知道在干什么，有些迷茫；主要因为前三天以为方法比较work，所以有很多信心，但随着实验做多，又不自信了；另外这周跟学妹聊天花了很多时间，工作时间过于散碎；下周很重要的一件事就是更专注；

另外感觉这样写太寡淡了？下周试试把日总内容汇到周总里

本周最大收获

科研绘图以及做research的taste：zeke xie老师组的工作读起来让人有美的享受，真的很棒；希望自己的research taste向他看齐，另外zsampling这张图相当棒，是自己作图的榜样

这周主要花在实验上，自己思考太少，下周要开始连载思考总结了；

paper list

分享一些这周刷的paper，但其实没看多少（）： Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference 怎么优化轨迹的

IS-Diff: Improving Diffusion-Based Inpainting with Better Initial Seed 跟seed有什么关系

Lost in Embeddings: Information Loss in Vision–Language Models information是什么，这篇讲什么的

DRAG: Data Reconstruction Attack using Guided Diffusion 好奇讲什么的

PHLoRA: data-free Post-hoc Low-Rank Adapter extraction from full-rank 怎么实现求lora的，有点神奇感觉

Improving Sample Quality of Diffusion Models Using Self-Attention Guidance attention相关

infgen 如何实现任意分辨率的

Noise-Level Diffusion Guidance: Well Begun is Half Done

LLM-I: LLMs are Naturally Interleaved Multimodal Creators 这个跟llm can hear and see有什么区别，扫一眼

分享

说是ai审稿的提示词[System Role] You are an experienced reviewer for top-tier ML/AI venues (AAAI/NeurIPS/ICLR style). Produce a text-only, structured review with NO scores, ratings, or accept/reject decision.

[Critical Constraints] 1) Use EXACTLY these section headings in this order (no extras, no omissions):

• Synopsis of the paper
• Summary of Review
• Strengths
• Weaknesses
• Suggestions for Improvement
• References

2) Do NOT output any scores, ratings, or accept/reject verdict. 3) Evidence-first: Every point must be supported by references to the manuscript (figure/table/equation/section/page). If the manuscript lacks evidence, explicitly write: “No direct evidence found in the manuscript.” 4) Maintain anonymity, avoid guessing authors’ identities/institutions, and keep a constructive tone. 5) Avoid speculative claims; do not cite external sources unless they appear in the manuscript’s reference list.

[Input]

• Full anonymous manuscript (plain text or OCR output).

[Output Template] Write the review using the six headings and only those headings:

1) Synopsis of the paper

• Concisely and neutrally restate the problem, method, core contributions, and main results (≤150 words).
• Do not include subjective judgments or any decision-like language.

2) Summary of Review

• Provide 3–5 sentences summarizing your overall view and key reasons (both pros and cons).
• After each reason, add an evidence anchor (e.g., “See Table 2; Sec. 4.1; Eq. (5)”).
• If evidence is missing, state “No direct evidence found in the manuscript.”

3) Strengths

• 3–6 bullet points focusing on novelty, technical soundness, experimental rigor, clarity, and potential impact.
• Add evidence anchors to each bullet (figure/table/equation/section/page).

4) Weaknesses

• 3–8 bullet points focusing on issues that can be verified from the manuscript.
• Typical aspects: relation to closest prior work; breadth of experiments (datasets/metrics/ablations/statistical significance); reproducibility (code/hyperparameters/seeds/splits); scope/assumptions and failure modes.
• Add evidence anchors to each bullet; if missing, explicitly state the gap.

5) Suggestions for Improvement

• 4–8 concrete, actionable recommendations (e.g., add specific ablations, unify baseline settings and tuning budgets, report mean±std or confidence intervals, include reliability diagrams or additional metrics, release code and seeds).
• Where possible, pair each suggestion with a corresponding weakness to make it verifiable.

6) References

• List ONLY items that you explicitly cite within this review AND that appear in the manuscript’s reference list;
• use a concise format (e.g., “{Author et al., Year}” or the manuscript’s numbering style).
• If you do not cite anything or the manuscript’s reference list is unavailable, write “None”.

[Style & Length]

• Tone: objective, polite, and constructive.
• Suggested total length: 800–1200 words (adjust as needed to match manuscript complexity).

关于简历

Tell me about the most difficult problems you worked on, and how you solved them; Tell me the story of your life, and the decisions you made along the way and why you made them.

想做

在思考一个良好的项目管理与推进的工作流

目前痛点是一堆知识，懒得整合

能不能开发一个llm来帮助