paper-digest
Paper Digest
Single-paragraph summaries optimized for social sharing. Insight over information.
Structure
- Context: What's the problem?
- Insight: What did they realize that others missed?
- Solution: How does insight → method? (should feel natural)
- Evidence: Concrete comparison showing it works
Then: Implication line + 📎 arxiv link
Key Rules
- Explain like reader is smart but unfamiliar with the domain
- Use concrete examples/analogies (e.g., "쓰레기통 역할" >> "특정 토큰에 집중")
- Show cause-and-effect chains explicitly
- Compare/contrast with alternatives ("X failed while Y succeeded")
- Bold 2-4 key concepts
- Match user's language (Korean/English)
Example
Input: arXiv 2601.15380
Output:
Transformer의 attention은 "어떤 토큰을 얼마나 볼지"를 결정하는데, 이 논문은 softmax attention을 **Entropic Optimal Transport(EOT)**라는 최적화 문제의 해로 재해석한다. 이 관점이 주는 통찰은: attention 계산에는 암묵적으로 "모든 위치가 동등하게 중요하다"는 uniform prior가 숨어있다는 것이다. 이게 왜 문제인가? LLM에서 첫 번째 토큰이 의미와 무관하게 엄청난 attention을 받는 attention sink 현상이 있다. Softmax는 합이 1인 확률을 출력해야 하므로, query가 마땅히 볼 토큰이 없을 때 attention을 "버릴 곳"이 필요한데, uniform prior 하에서 이를 구현하려면 첫 토큰의 key vector가 "나는 쓰레기통이야"라는 구조적 정보까지 담아야 한다—원래 semantic content만 표현해야 할 key의 표현력이 낭비되는 것이다. EOT 해석이 이 문제를 드러내주었으므로, 해결책도 자연스럽다: prior를 uniform에서 learnable로 바꾸면 된다. 이 논문이 제안하는 GOAT은 "각 위치의 기본 중요도"를 별도의 학습 가능한 항으로 분리해서, key vector는 순수하게 의미만, 위치 정보는 prior가 담당하게 한다. 실험에서 기존 방법들이 훈련 길이 초과 시 급격히 실패한 반면, GOAT은 긴 문맥에서도 정보 검색 성능을 유지했다.
Implication: EOT 관점은 attention의 숨겨진 가정을 드러내고, 그 가정을 바꿀 수 있다는 설계 자유도를 열어준다—attention sink는 uniform prior의 부산물이며, prior를 명시적으로 모델링하면 해결된다.
📎 https://arxiv.org/abs/2601.15380
Avoid
- Jargon without intuition
- Findings without comparison to alternatives
- Method description without motivation ("왜 이렇게 했는지" 없이 "이렇게 했다"만)
Multiple Papers
When summarizing multiple papers:
- Lead with the unifying theme/problem
- Contrast what each paper realized differently
- Synthesize implications across papers
Language
Match the user's language (Korean/English). Maintain the same insight-first structure regardless of language.
More from iamseungpil/claude-for-dslab
update-study
This skill should be used when the user asks to "update study", "analyze new experiments", "update experiment document", or "refresh study notes". Produces academic-paper-quality experiment reports with matplotlib plots, executive summary with comparison tables, implementation structure, experimental results with figure interpretation, proposed improvements with code examples, hypotheses, limitations, and LaTeX PDF export with figures. Features incremental detection (only analyze NEW experiments), data extraction to DataFrame, automated plot generation, iterative writing improvement loop with quality criteria, zero-hallucination verification, and LaTeX PDF export. Usage - `/update-study logs/experiment.log study.md` or `/update-study "logs/exp1.log logs/exp2.log" results/ablation_study.md`
55hwpx
Comprehensive HWPX (Korean Hancom Office) document creation, editing, and analysis. When Claude needs to work with Korean word processor documents (.hwpx files) for: (1) Reading and extracting content, (2) Creating new documents, (3) Modifying or editing content, (4) Extracting tables to CSV, (5) Modifying tables or table cells, or any other HWPX document tasks. MANDATORY TRIGGERS: hwpx, hwp, 한글, 한컴, Hancom, Korean document
18survey-paper
|
10iterative-code-review
Iteratively improve code quality by using task-planner-analyzer for planning, modular-code-architect agent to fix issues, code-reviewer agent to validate quality, and running tests to verify correctness. Use when implementing new features, after bug fixes, during refactoring, or when preparing code for production deployment. Loops until code-reviewer reports no critical issues AND tests pass.
10codex-iterative-solver
Collaborate with Codex CLI to iteratively analyze, plan, and solve complex coding problems through multiple rounds of expert feedback. Use when analyzing complex codebases with multiple approaches, validating implementation plans, or solving problems that benefit from expert validation.
8weekly-report-writer
Automatically generate academic-style weekly research reports by analyzing Git changes in the current project directory with iterative quality improvement
8