Paper Deep Reading

Other

Deep-read research papers into source-aware reports, traceable claim evidence, and research-direction seeds. Use for paper PDFs, LaTeX sources, appendices, code notes, peer reviews, literature-review tasks, novelty audits, and finding new research questions with minimum viable experiments.

Install

openclaw skills install paper-deep-reading

Paper Deep Reading: Source-Aware + Research-Generative Direction Mining

Use this skill when the user wants a deep, paper-grounded, auditable, idea-generative reading report for one computer-science paper or a small paper batch.

The input may be:

a user-provided PDF
a user-provided LaTeX source tree or .tex files
supplementary material, appendix files, code notes, or OpenReview material
only the paper title, arXiv id, venue page, citation-like paper name, or PDF link

The default output is text-first, audit-first, formula-preserving, and research-direction-oriented. This version does not require a dedicated webpage reader; when search/browsing tools are available, use them to assemble the best source package before writing.

1) Core deliverables

Human-readable report
- report.md
Machine-readable trace artifacts
- traceability_manifest.json
- latex_paragraphs.json
- artifact_index.json
Machine-readable research artifacts
- research_lens.json
- direction_board.json

The report is the primary user-facing deliverable. It must read like a serious research mentor's deep-reading memo, not like a thin checklist dump. The direction board is the primary idea-mining surface: it converts paper weaknesses, hidden assumptions, evidence gaps, proxy mismatches, successor-paper gaps, and reviewer objections into candidate research directions.

2) ClawHub and MIT-0 package discipline

This skill package is intended to stay compatible with ClawHub / OpenClaw skill packaging.

Keep the package lean:

keep SKILL.md as the main instruction file
keep only text-based support files, templates, and scripts that another agent needs to execute the workflow
do not reintroduce auxiliary docs such as README.md or CHANGELOG.md
do not add binary assets, vendored third-party repositories, or cached papers to the skill package
keep support files focused on execution, validation, and artifact contracts

Keep the package license-safe:

this package follows ClawHub's MIT-0 publication model
keep the local bundle license text in LICENSE.txt
do not add restrictive or conflicting license terms elsewhere in the package
do not vendor third-party projects or assets into the skill unless their license is compatible with MIT-0 redistribution expectations
when external tooling is useful, document it or install it outside the skill instead of copying its source tree into the package

Runtime discipline:

bundled scripts are local text-processing helpers and should not make network calls
if external search is needed, use the host agent's approved browsing/search capability rather than hidden scripts
declare runtime dependencies honestly in frontmatter and metadata

3) Non-weakening rule and depth bar

Do not treat the OpenClaw / ClawHub version as a lightweight summary mode. The single-file constraint changes presentation, not analysis quality.

Never remove, weaken, shorten, or bypass any existing deep-reading requirement, including:

source acquisition and disambiguation
LaTeX-first reading when source is available
PDF-assisted figure/table reading
formula preservation
proof-to-practice mapping
OpenReview / reviewer context when relevant
reviewer-lens audit
claim IDs and traceability manifest
final claim-to-evidence appendix
research-generative overlay
language policy
validation scripts

The report depth bar should stay close to a strong top-conference paper memo:

cover the full 25-section reading scope
preserve central equations instead of flattening them into prose
explain why modules exist, not only what they are called
reconstruct likely author-side reasoning when the evidence supports it
connect experiments back to claims, ablations, and alternative explanations
extract reusable research patterns and future ideas
produce concrete research seeds with minimum viable experiments, negative-result interpretation, killer objections, and killer results

If a tension appears between a shorter explanation and a more idea-generative one, choose the more useful research-direction analysis while keeping claims grounded. If a tension appears between speculation about author intent and factual safety, label the reconstruction explicitly as plausible inference or speculation and anchor it to textual evidence.

4) Research-generative overlay

This version keeps the original traceability and formula-preservation bar, but adds a research-direction mining layer.

The report must help the user answer not only:

what the paper did
whether the evidence supports the claims

but also:

how the authors may have found the direction
what hidden assumption C broke
what unavailable mechanism Y had to be replaced
what surrogate mechanism Z the paper constructed
how each module maps to a failure mode
why key citations matter in the story
what hidden assumption can seed the next paper
which new research directions are worth testing first

Use references/research-generative-methodology.md and references/research-direction-mining-best-practices.md whenever the user wants:

author-perspective reading
idea mining
reverse story construction
module-level design logic
citation-function analysis
reviewer-grade critique
minimum viable experiment design
boundary-pushing future directions

5) Research-direction mining three-pass method

Read each paper in three direction-mining passes. These passes are adapted for discovering new research points, not merely for comprehension.

Pass 1: Five-C triage + direction promise

Quickly inspect title, abstract, introduction, section headings, conclusion, references, and visible figures/tables. Answer the five triage questions:

Category: What type of paper is this: method, benchmark, theory, measurement, system, dataset, analysis, survey, or position?
Context: What field conversation, assumptions, and ancestor methods does it sit inside?
Correctness: Do the core assumptions, data, metrics, and comparisons look initially plausible?
Contributions: What are the claimed contributions and how strong do they look before deep verification?
Clarity: Is the argument organized well enough that the method and claims can be audited?

Then add a direction-promise note:

what hidden assumption seems most likely to be attackable
what omitted setting or stress condition appears promising
whether the paper is worth a full second and third pass

Pass 2: Evidence / method / figure chain reconstruction

Read the paper carefully but keep the goal causal:

problem -> assumption break -> design principle -> module -> formula -> figure/table -> experiment -> claim

During this pass:

inspect key figures, diagrams, graphs, and tables as evidence, not decoration
preserve central equations and explain their role
build the challenge-to-module table
map each result to the claim it supports
mark missing controls, weak baselines, noisy metrics, unclear error bars, and unsupported narrative jumps
identify the available proxy that replaces an unavailable ideal mechanism

Pass 3: Virtual reimplementation + hidden-assumption attack

Recreate the work as if you had to implement, prove, or reproduce it. Ask:

What exact assumptions must be made for each module to work?
Where would the method fail if one assumption were dropped?
What tiny example, special case, or counterexample exposes the key idea or fragility?
What proof step, algorithm step, data preprocessing choice, or metric definition carries the argument?
What implementation details are missing but necessary for reproduction?
What would a stronger, cleaner, or more decisive experiment look like?

This pass must produce future-work triggers. A trigger is not a generic suggestion; it is a statement of the form:

current method works if H -> under not-H it breaks -> new mechanism needed -> minimum experiment to test the opportunity

6) Successor-paper and reverse-citation reading

When the user asks for new research directions, do not stop at the paper's own related work. If tools are available and time permits, inspect a small set of successor papers, citation trails, follow-up discussions, code repositories, or public review threads.

Use successor reading to answer:

how later papers describe this paper's real contribution
what later work treats as the bottleneck or limitation
what claims were ignored, weakened, or reframed by the community
which open gap remains after follow-up papers
which direction is already saturated and which remains underexplored

If successor-paper search was not possible, say so and keep direction confidence lower. Do not fabricate citation trends.

7) Critical + creative reading rule

Every report must combine critical and creative reading.

Critical reading asks:

Is the paper solving the right problem?
Are the assumptions reasonable?
Are the data, metrics, baselines, and controls sufficient?
Are the conclusions stronger than the evidence?
Are there simpler alternatives the authors did not rule out?
What limitations are admitted, hidden, or structurally unavoidable?

Creative reading asks:

What good idea can be transplanted elsewhere?
What stronger setting makes the idea newly important?
What generalization or simplification would be more elegant?
What proxy can be replaced by a more direct signal?
What negative result would change community understanding?
What is the next research question a strong PhD student should test?

The final directions must be creative and falsifiable.

8) Reviewer-grade audit integrated with direction mining

Use reviewer thinking not just to judge acceptance, but to discover research seeds.

Audit at least these dimensions when evidence allows:

novelty and relation to prior work
significance and likely community use
technical soundness
methodology rigor
statistical validity and uncertainty reporting
baseline and control completeness
reproducibility and implementation sufficiency
result-to-claim alignment
clarity of figures/tables/formulas
limitation honesty
ethics, safety, or societal concerns when relevant
specific constructive critique

Convert reviewer objections into direction candidates:

reviewer objection -> why it matters -> what evidence would resolve it -> minimum viable experiment -> possible new paper

9) Full-loop research seed discipline

The skill does not replace the researcher or claim to have completed experiments. It turns a paper into candidate directions that a researcher can test.

Every strong candidate direction must include:

seed_type: one of assumption violation, unavailable mechanism, proxy mismatch, evidence gap, tiny example, successor-paper gap, reviewer objection, negative result, or cross-domain transfer
paper_anchor: claim IDs and source evidence that triggered it
research_question: a question that can be answered
hypothesis: what might be true
minimum_viable_experiment: the smallest decisive test
negative_result_interpretation: what it would mean if the hypothesis fails
killer_objection: the strongest reason the idea might be uninteresting or invalid
killer_result: the result that would make the direction worth pursuing
first_week_plan: practical steps for a researcher's first week
risk_level and expected_value

Generic future-work lists are not enough. A direction without a test plan is an inspiration note, not a research seed.

10) Verification surface: body first, appendix last

The report itself remains the primary verification surface, but the detailed evidence placement is:

Main body
- readable section-by-section analysis
- ### Anchored Points blocks near the relevant discussion
- concise claim bullets in the form - [C5.2][evidence-backed interpretation] ...
Final appendix
- detailed claim-by-claim evidence records
- exact source files
- section paths
- line spans
- page hints when available
- quote snippets and excerpt windows
- notes that help a human verify the claim quickly

Do not clutter the main narrative by inserting long locator bullets immediately after every claim. Keep the main body readable, and move detailed original-paragraph explanation to the final # Appendix: Claim -> Evidence Index.

Use scripts/render_inline_trace_report.py after drafting the report and manifest to materialize or refresh that appendix.

11) Formula-first preservation

When the paper contains key formulas, the report must not compress them into prose-only summaries.

For each central equation, objective, theorem statement, update rule, estimator, metric, loss, or constraint, explicitly include:

the equation itself in readable math form
symbol-by-symbol explanation
what optimization / estimation / filtering / proof role it plays
why the authors likely wrote it in this form instead of a nearby alternative
how it connects to the previous and next module
what may be brittle, heuristic, under-justified, statistically weak, or computationally expensive about it
how changing the equation creates possible new research directions

Do not weaken equation detail for the sake of shorter presentation.

12) Source acquisition policy

Always assemble the best available evidence package before writing.

Preferred reading order:

arXiv LaTeX/source package
user-provided LaTeX
best available PDF
supplementary material / appendix
official code or implementation notes when the user asks for reproducibility
OpenReview thread / rebuttal / meta-review when relevant
successor papers or citation trails when the user asks for new research directions

12.1 When LaTeX is available

Treat LaTeX as the primary structural source.

Use PDF only as a visual and pagination aid for:

figure interpretation
table reading
page-local narrative flow
page anchors
visual sanity checks that cannot be recovered from source text

12.2 When only PDF is available

Do not stop at PDF summarization immediately.

First check whether the same paper has a matching arXiv LaTeX/source package. If it exists and matches the same paper, switch to LaTeX-primary + PDF-assisted reading.

If not, continue with the PDF and say explicitly that the reading is PDF-primary.

12.3 When only title is available

Search for the paper and collect:

arXiv source package if available
the best PDF
supplementary PDF or appendix if available
OpenReview forum if venue is ICLR or otherwise OpenReview-hosted
official code, successor papers, or citation context when needed for direction mining

Never silently analyze the wrong paper. Disambiguate by title, authors, abstract, year, venue, and method keywords.

12.4 OpenReview policy

If the paper is an ICLR or OpenReview-hosted paper, look for:

reviewer comments
meta-review or area-chair summary
author rebuttal or response
revision signals relevant to acceptance

Use them to enrich:

reviewer-lens audit
confidence in claimed contributions
limitations and unresolved doubts
candidate directions derived from reviewer objections

12.5 Missing source policy

If some sources cannot be found, do not abort. State clearly what was attempted, what was found, what was missing, and how that affects confidence. Then continue with the best grounded report possible.

If LaTeX cannot be found after an explicit search, say so clearly and use PDF-oriented evidence rows in traceability_manifest.json instead of pretending paragraph anchors exist.

13) Language policy

Write the skill instructions, internal prompts, and template skeletons in English. Choose the report language from the user's current request language by default.

if the user's current request is primarily in Chinese, write the report in Chinese
if the user's current request is primarily not Chinese, write the report in English
if the user explicitly requests another language, follow that explicit instruction
if the request is mixed-language, follow the dominant user language in the current request

When writing the report in Chinese:

keep proper nouns and fixed technical identifiers in English
this includes paper titles, method names, module names, datasets, baselines, theorem or object names, citation names, equation symbols, claim IDs, filenames, and JSON keys
translate section headings and explanatory prose into Chinese, but do not translate artifact filenames, schema fields, or claim IDs

14) Mandatory artifacts

14.1 `report.md`

The report must cover, whenever the evidence supports it:

paper identification and source package used
one-sentence thesis and research equation
title interpretation
what problem the paper really solves
scientific problem ladder
how the authors may have found the direction
how the authors built the story
related work, key citations, and what was still missing
main idea
symbols, assumptions, and notation
key formulas and equation-by-equation explanation
theory / proof / practice mapping
algorithm or module walkthrough with concrete example
method deep reading: the author-thinking behind each module
figure explanation
experimental design
experiments as story evidence and claim alignment audit
reviewer-lens audit
innovation points and claim-by-claim support audit
story-making pattern worth learning
weaknesses and limitations
innovation type and scientific-boundary judgment
future directions and stronger idea paths
vivid plain-language story summary
exact sources used

Use templates/report_template.md as the default skeleton.

For each numbered section:

start with ### Anchored Points
add one or more claim bullets in the exact form - [C<section>.<index>][label] claim text
keep the bullets concise
follow the bullets with a real explanatory section, not just more bullets
add tables, formulas, examples, reviewer-style critique, or story reconstruction when they help understanding

14.2 `traceability_manifest.json`

This is the claim-to-evidence map.

Rules:

every claim id in the main report body must appear in the manifest
one bullet must not hide multiple independent claims under one id
if a claim depends on multiple paragraphs, equations, tables, appendix passages, figures, or reviews, list them separately
each claim entry should include interpretation_type
each claim entry should preferably include research_role
each claim entry should include human-friendly locator data when possible

14.3 `latex_paragraphs.json`

This is the stable LaTeX anchor index.

Each entry must keep:

paragraph_id
source_path
line_start
line_end
section_path
kind
text

14.4 `artifact_index.json`

A compact index for the generated text-first bundle.

It should list the locations of:

report.md
traceability_manifest.json
latex_paragraphs.json
research_lens.json
direction_board.json
main PDF if any
supplementary PDF if any
source package path if known

14.5 `research_lens.json`

This is the compact idea-mining artifact. Use templates/research_lens.template.json and references/artifact_contract.md.

It should capture:

the paper's research equation
the likely direction-finding path
challenge-to-module mapping
per-module hidden assumptions
citation logic
reviewer-lens summary
reusable story pattern
strongest future idea directions
links to the most important direction seeds

14.6 `direction_board.json`

This is the structured research-direction board. Use templates/direction_board.template.json.

It should capture:

ranked candidate research directions
the evidence trigger for each direction
hidden assumption or missing mechanism
minimum viable experiment
negative-result interpretation
killer objection and killer result
first-week plan
score breakdown
relationship to existing paper claims

15) Claim discipline

15.1 Claim ids

Use stable section-local ids such as:

C3.1
C5.2
C14.4

15.2 Claim splitting rule

Do not hide multiple judgments in one claim bullet.

15.3 Evidence completeness rule

List all materially relevant evidence for a claim, not just one convenient paragraph.

15.4 Interpretation labels

Each claim must declare exactly one of:

evidence-backed interpretation
plausible inference
speculation

15.5 Research-generative honesty rule

If the report reconstructs likely author reasoning, it must still point to the exact paragraphs, equations, figures, tables, experiments, reviews, or successor-paper signals that motivate that reconstruction. Idea generation is required, but fabrication is forbidden.

15.6 Direction trigger labels

Each direction seed should also label the trigger as one of:

evidence-backed interpretation
plausible inference
speculation

Do not rank speculative seeds as high-confidence unless the uncertainty is explicit.

16) Writing style for verification and idea generation

Prefer a report that is pleasant to read and easy to audit.

For every claim, the user should be able to answer:

What section-level conclusion is being made?
Is it direct evidence, plausible inference, or speculation?
Where should I verify it in the appendix?

For the strongest research-direction sections, the report should also answer:

What hidden assumption broke?
What missing mechanism was replaced?
What future paper becomes possible if that assumption fails harder?
What minimum experiment would tell us whether this future paper is real?
What result would kill the idea?
What result would make the idea exciting?

Use phrasing such as:

"A plausible author-side thinking path is ..."
"This module is best understood as a surrogate for ..."
"The citation is not ornamental; it functions as ..."
"The deepest reusable lesson is ..."
"This weakness can be converted into a new research direction ..."
"The minimum viable experiment is ..."
"The killer objection is ..."
"A negative result would still be useful if it shows ..."

The report should sound like a research mentor reconstructing how the work may have been invented and how it could become the next project, not like a generic summarizer.

17) Grounded workflow

Assemble the best source package.
If LaTeX is available, extract paragraph anchors with scripts/extract_latex_paragraphs.py.
Perform Pass 1 five-C triage and decide whether full deep reading is warranted.
Perform Pass 2 evidence / method / figure chain reconstruction.
Perform Pass 3 virtual reimplementation and hidden-assumption attack.
Draft report.md using anchored claim IDs in the main body.
Keep claim bullets concise and put longer explanation in prose, tables, formulas, examples, and story reconstructions after them.
Fill traceability_manifest.json so each claim points to one or more paragraph IDs or fallback anchors.
Fill research_lens.json so the paper's research equation, story structure, module logic, citation functions, reviewer audit, and future directions are captured in structured form.
Fill direction_board.json so the best candidate research seeds are ranked, testable, and linked to evidence.
Fill artifact_index.json so the bundle stays portable.
Run scripts/validate_traceability.py.
Run scripts/validate_direction_board.py when direction_board.json is present.
Run scripts/render_inline_trace_report.py to append or refresh the final Claim -> Evidence Index appendix in report.md.
Only then finalize the bundle.

18) Small-batch policy

For a small paper batch:

produce one standalone report.md-style bundle per paper when the user expects detailed reading
do not collapse multiple papers into a shallow combined summary
optionally add a cross-paper direction board if the goal is choosing a new research direction
rank cross-paper directions by novelty, evidence gap, testability, expected impact, feasibility, and relationship to the user's research interests

19) Failure handling

If some sources cannot be found, do not abort. State clearly:

what was attempted
what was found
what was missing
how the missing source changes confidence
which claims or direction seeds are affected

Then continue with the best grounded report possible.

If the evidence does not support strong idea generation, say so and produce a conservative direction board. Do not invent novelty, successor trends, reviewer objections, or experimental feasibility.