| name | paper-recap-insight |
| description | Create or refine extreme-granularity, principle-first research paper recaps from PDFs or existing notes, organized around four pillars — 逻辑与价值 (logic/value/story), 原理与理论 (first-principles theory edifice), 实验与验证 (experiments as evidence for the story), 未来与结合 (limitations + integration with the knowledge graph and the user's WMTS/灵巧手 projects). For robotics, embodied AI, RL, control, dynamics, world models, representation learning, dexterous manipulation, Sim-to-Real, tactile/contact, actuator dynamics, and Obsidian/PapersRecap/RelatedPapersRecap knowledge graphs. Use when the user asks to summarize, 整理, 精读, recap, refine, 对齐知识库, or extract insights from a paper PDF or paper-related chat/note. Emphasize mathematical provenance, variable origin, no-jump derivations, causal ablations, and personalized research transfer; avoid implementation-code dumps. |
Paper Recap Insight
Produce paper recaps that behave like a research mentor's reconstruction, not an abstract expansion. The gold standard is Example/Rodrigues Network for Learning Robot Actions.md (722 lines). Every recap should reach that question granularity: not just "what the paper did", but where every variable comes from, why each symbol is written that way, how each formula is derived with no jumps, how each experimental number corroborates the story, and exactly how the method transfers to the user's 灵巧手转笔 / Sim-to-Real / WMTS work.
Read the gold standard before writing or refining. Open Example/Rodrigues Network for Learning Robot Actions.md and references/taste-rubric.md. The Example is the target; this skill is the operating manual for reproducing its depth on any paper.
The Four Pillars (用户的四个关注点)
Every recap is judged against four questions the user actually asks. These are the backbone — organize thinking around them, and make sure each is answered concretely, not gestured at.
Pillar 1 — 逻辑与价值 (Logic & Value): how the paper earns its place
这篇文章相对于其他文章,逻辑上顺承出来的优势在哪里?Insight 在哪里?Value add 在哪里?它是如何讲好这个故事的?
- State the exact bottleneck in one sentence, then the structural bet (the assumption that makes the method plausible).
- Build a 现有方法的局限 table: for each prior paradigm (MLP/GNN/Transformer/analytic/end-to-end/...), name what prior it injects and its specific limitation — not a generic "prior work is limited".
- Write the Delta as a precise increment over the closest baseline ("not X-as-external-loss, but X-as-internal-operator"), not a vague "we are better".
- Give one 直观隐喻 that crystallizes the insight (the Example's CNN↔RodriNet analogy). A good metaphor is falsifiable: it says exactly which structure is fixed and which is learned.
- Separate "the method works" from "the method works because the benchmark bottleneck happens to be X".
Pillar 2 — 原理与理论 (Principles & Theory): build the theoretical edifice from zero
从原理的角度出发,它联系到了哪些相关的公式和定理?如何从零基础开始把它们说清楚,构建一个详细的理论大厦?
- Variable-provenance table is mandatory (see pattern below). Every important symbol: domain/shape, source stage (robot structure / observation / network output / rollout / computation-graph intermediate), gradient status (detached vs requires_grad), physical-or-algorithmic meaning, and a trap column.
- No-jump derivation: start from the classical formula/objective/graphical-model/control-law the paper modifies, then reach the paper's equation with every algebraic step shown. The reader should never wonder "where did this term come from". Reconstruct the prerequisite theory from first principles (the Example derives $SO(3)\to\mathfrak{so}(3)\to\exp\to$ Rodrigues $\to$ FK tree composition before touching the neural operator).
- Notation traps as insight: coordinate frames, active/passive transforms, time indices vs diffusion/denoising steps, local vs global variables, distribution vs sample, fixed structure vs learnable, dimension/homogeneous-lift mismatches. Explaining a subscript convention (the Example's
${}^P_J T$ "J in P") is often where the real understanding lives.
- Link each formula back to the relevant Foundation with the exact mathematical correspondence (
[[Dynamics#2.4 ...|Dynamics §2.4]] — this is the math root), not a decorative link.
Pillar 3 — 实验与验证 (Experiments & Validation): numbers that prove the story
从实验结果来看,它的表现是如何印证它所讲的这个故事的?
- Preserve real numbers in real tables: task names, datasets, metrics, baselines, headline results, training setup (optimizer, LR, batch, steps, data sizes). Never paste raw "PDF clue lines" as a substitute for a result table.
- After each table, write the 因果解释: why these numbers support the Pillar-1 story. The Example's killer line — "Rodrigues' test MSE is below every baseline's train MSE → the advantage is structural-prior generalization, not fitting capacity" — is the model.
- Every ablation as a causal chain:
remove/change A → metric B moves → because mechanism C is disabled/amplified → implication D for using the method. The point of an ablation table is to identify which inductive bias was removed, not "module gone so worse".
- Flag where a result is confounded or where the benchmark's bottleneck is not the one the paper claims to solve (the Example: PegInsertion/PlugCharger gains are small because, absent tactile/force, the backbone isn't the bottleneck).
Pillar 4 — 未来与结合 (Future & Integration): limitations + transfer to the user's world
它有哪些 limitation?可以如何与当前的知识库结合?如何与当前我想提出的 project 结合?
- Limitations across three axes: 理论维度 (what physics/math it does not model — e.g. "kinematics-aware, not dynamics-aware: it does not represent $M(q)\ddot q + C\dot q + g = \tau + J^T\lambda$"), 算法维度 (alternative paradigms compared by assumption + failure mode), 工程/实验维度 (memory, kernels, sim-to-real gaps).
- Personalized transfer is concrete enough to become an experiment or a rejection reason: produce at least one of — a direct modification to PPO / Diffusion Policy / world model / controller; a diagnostic experiment with baselines; a reason not to use it in contact-rich dexterous manipulation; a Sim-to-Real risk + mitigation. When the method has a key input variable, show what it becomes in the user's task (the Example's "$\theta_j$ → a $C_J$-dim joint-local feature of $[\sin q,\cos q,\dot q, a_{t-1}, \tilde a_k, e(t_{diff}), h^{tactile}, h^{object}, h^{task}]$" table).
- Integration with the knowledge graph: link ≥2 Foundations with specific correspondence; connect to the WMTS pipeline (latent task generation → PPO Oracle → Diffusion/Flow generalist → Ensemble World Model → real-robot fine-tuning) and to DNPM where relevant. Respect WMTS defaults from
research-insight-critic (PPO is the default Oracle; tactile/contact and actuator dynamics are first-class; do not inject task labels into dynamics models).
Mandatory Vault Context
When working inside a vault that contains .github/skills/, read the local guidance before writing:
.github/TASK_TRACKER.md for current knowledge-graph state and any active recap-upgrade batch.
.github/skills/knowledge-graph-management/SKILL.md, especially §2.3.1 (algorithm granularity standard) and §2.3.2 (PapersRecap refine workflow), plus Foundation linking rules.
.github/skills/research-insight-critic/SKILL.md for WMTS defaults and routing hints when writing the Pillar-4 transfer.
.github/skills/obsidian-markdown/SKILL.md if creating/editing Obsidian notes; .github/skills/json-canvas/SKILL.md if updating KnowledgeGraph.canvas.
references/taste-rubric.md in this skill when calibrating depth or deciding whether output is deep enough.
Place the paper in the existing graph: inspect nearby Foundations/, the relevant PapersRecap/ or RelatedPapersRecap/, Projects/, and KnowledgeGraph.canvas.
PDF Workflow (do this before writing)
- Probe the PDF with the
pdf skill first — never rely on a single pdftotext pass:
- repo-local:
python3 .github/skills/pdf/scripts/pdf_probe.py "<paper.pdf>" --out "tmp/pdfs/<paper-stem>"
- global:
python3 /Users/yang/.codex/skills/pdf/scripts/pdf_probe.py "<paper.pdf>" --out "tmp/pdfs/<paper-stem>"
- Combine
pdfinfo, pdftotext -layout, pdfplumber page text, and pdftoppm PNG renders. Read abstract, intro, method, experiments, ablations, limitations, and appendix (reward/observation tables, training details often live there).
- If extracted text is garbled, the rendered PNGs are the source of truth for title, equations, tables, captions, numbers. Reconstruct equations from surrounding prose and verify against repeated references in the text.
- Build an outline from headings, but write the recap from causal dependencies, not paper order.
- Search the extracted text for: equation numbers,
ablation, baseline, we observe, we find, fails, limitation, future work, implementation, training details, dataset.
Required Recap Structure
Use this structure (it maps each section to a pillar). Match the Example's depth; adapt section names to the paper, but never drop a pillar.
---
tags:
- paper
- <domain tags>
aliases:
- <short name>
paper-year: YYYY
read-date: YYYY-MM-DD
venue: <venue>
paper-pdf: "[[<exact PDF path/name>.pdf]]"
related:
- "[[Foundation1]]"
- "[[Foundation2]]"
---
# <Full Paper Title>
> [!abstract] 核心贡献
> <One sentence: method X solves bottleneck Y by structural insight Z.>
> [!tip] 与理论基础的关联
> - [[Foundation#section]] — exact mathematical correspondence
> - [[Foundation#section]] — exact mathematical correspondence
> **核心技术**: <key technique names>
## 0. 阅读定位与范本价值 ← Pillar 1
<Why this paper matters for ' ; />
For a method that is itself a template/foundational tool, also add (as the Example does):
- a §0 最低标准表 mapping each pillar to where it lands in this recap, and
- a final 应复刻的提问颗粒度 section anticipating the user's follow-up questions (
用户式追问 | Agent 应主动补充).
Worked Patterns
Formula reconstruction (Pillar 2)
- Classical starting point. 2. Constraint/observation. 3. Algebraic transformation (every step). 4. Paper's relaxation/parameterization. 5. Meaning of each term. 6. Failure/boundary condition + degenerate case (when does it reduce to the classical method?).
Variable-provenance table (Pillar 2)
| 变量 | 维度/空间 | 来源阶段 | 是否带梯度 | 物理/算法意义 | 符号陷阱 |
|---|
"来源阶段" ∈ {robot structure/URDF, observation, network output, rollout, computation-graph intermediate, supervision label}. "符号陷阱" = coordinate convention, gradient status, unit, frame, action horizon, distribution vs sample, physical quantity vs latent feature.
Ablation causal chain (Pillar 3)
Remove/change A → observed metric B changes → because mechanism C is disabled/amplified → implication D for using the method.
Personalized transfer (Pillar 4)
Produce ≥1: a direct modification to PPO/DP/world-model/controller; a diagnostic experiment (with baselines + what would falsify the mechanism); a reason not to use it in contact-rich dexterous manipulation; a Sim-to-Real risk + mitigation. When the method has a key scalar/structural input, write the table of what it becomes in the user's task.
Prohibited / Anti-Patterns (these failed the last draft pass — do not repeat)
- ❌ Generic filler that fits any paper. A §"前置理论从零推导" that says "这类方法可以统一写成闭环决策问题…若结构化 X 它在做 Y" is template text, not a derivation. Replace with the actual classical theory this paper modifies (DDPM forward/reverse + score matching for Diffusion Policy; Bellman/PPO clip for an RL paper; LCP/contact for a contact paper).
- ❌ Copy-paste repetition. Do not repeat the same sentence verbatim across §1.4 / §2.5 / §4.1 / §7. Each section answers a different question.
- ❌ Pasted "PDF 线索" lines in place of a real results table. Extract the actual numbers into a table and interpret them causally.
- ❌ LaTeX corruption. Verify
\bar\alpha did not become arlpha, \approx did not become pprox, \beta→eta, \theta→heta. Render-check math after writing; a recap with broken formulas fails Pillar 2 outright.
- ❌ Decorative Foundation links ("see [[ReinforcementLearning]]") without the specific mathematical correspondence.
- ❌ Default code dumps. No "最小 PyTorch 逻辑 / 核心 tensor ops" section unless the user asks. If implementation matters, write it as principle-level dataflow, shape/variable provenance, numerical constraint, or engineering failure mode (the Example's "实现避坑" bullets are the allowed form).
- ❌ Marketing praise without mechanism; "better generalization" without naming the inductive bias that narrows the function class; "physics-informed" without the exact physical structure.
Knowledge-Graph Alignment
- Standard frontmatter; field names per
knowledge-graph-management §3.5 (paper-year, read-date, venue, paper-pdf quoted).
- Output location: if the PDF folder has a sibling recap folder (
RelatedPapers/→RelatedPapersRecap/, Papers/→PapersRecap/), write the .md there, not next to the PDF.
- Filename = PDF basename, suffix
.pdf→.md, only substituting characters the filesystem forbids.
- Link ≥2 Foundations with specific correspondence; add reverse links in Foundations when the paper contributes a reusable concept; update Canvas only for an algorithmic breakthrough / key challenge / project-relevant connection.
- Validate wikilinks, section anchors, and Canvas JSON after edits (
.github/scan_links.py, .github/scan_sections.py).
Quality Gate (verify before finishing — tied to the four pillars)