Ccf Paper Writer

CCF A Venue Map

Use this file to map a target venue to its broad CCF A family before choosing writing priorities. The list reflects the CCF 2026 recommended international academic conference directory as inspected during skill creation. Re-check the official venue page for current submission rules, page limits, anonymity policy, artifact requirements, ethics requirements, and review form.

Architecture, Parallel/Distributed, Storage

Networks

Security

Software, Systems, Programming Languages

Data, Mining, Retrieval

Theory

Graphics And Multimedia

Artificial Intelligence

HCI And Ubiquitous Computing

Interdisciplinary And Emerging

How To Use

1. Map the target venue to a family above. 2. Load references/venue-adapters.md for the family-specific writing priorities. 3. Verify any venue-specific review form and submission page before final writing decisions. 4. If the paper targets a special track, use the track call as a stronger source than this generic map.

Citation Workflow

Use this file for every paper draft that requires references. The goal is to produce a manuscript where citations are plentiful, natural, and correct—not sparse and parenthetical.

Core Principle

Search → Bib → Cite. Never write a citation from memory or guess a bib entry. Every citation must originate from a verified bib entry stored in the project's .bib file.

The Problem With Sparse Citations

A CCF-A submission with fewer than25–40 references (for AI/ML/CV/NLP papers) signals shallow literature engagement. The introduction and related work alone should carry15–25 citations. The method and experiments add another10–20. An under-cited paper reads as naive or uninformed.

Citation Density Rules

If the total reference count is below25 for an AI/ML/CV/NLP paper, the draft is under-cited. Expand the related-work discussion and method background until the count is plausible.

Workflow

Step1: Identify Citation Slots

Before drafting each section, scan for where citations are needed:

• Introduction paragraph1 (task/momentum):2—4 citations to field-defining or recent high-impact works.
• Introduction paragraph2 (prior-work ladder):3—6 citations to successive prior approaches.
• Introduction paragraph3 (remaining gap):1—2 citations to works that most closely approach the gap.
• Introduction paragraph4 (method preview):0—1 citation if building on a known architecture.
• Related Work each topic group:3—8 citations per group.
• Method each borrowed component:1 citation to the original source.
• Experiments each baseline/dataset/metric:1 citation.

Mark each slot with a placeholder: [CITE: topic/need]. Collect all placeholders before searching.

Step2: Search For Literature

For each citation slot, determine the search strategy:

• Known paper exists: Use the exact title to find the bib entry via DBLP, Semantic Scholar, or arXiv.
• Topic gap: Route to ccf-literature-searcher for a targeted search. Specify: task area, year range, venue preference, and what the citation should support.
• Baseline/dataset: Search for the official paper or technical report; prefer the venue version over arXiv when available.

Do not invent paper titles, author names, or bib keys. If a search returns nothing useful, mark the slot as [UNFILLED] and note the search attempted.

Step3: Obtain Bib Entries

For every paper found, obtain a complete, correct bib entry. Preferred sources in order:

1. DBLP (most consistent formatting for CS venues) 2. Semantic Scholar API 3. Official venue proceedings page 4. arXiv export

A correct bib entry must include: author list (complete, not "et al."), title, venue abbreviation, year, and at minimum a DOI or URL. Do not truncate author lists with "and others."

Save every bib entry to the project's .bib file immediately. Use this naming convention for bib keys:

<firstauthor-lowercase><year><first-title-word>

Example: vaswani2017attention, he2016deep, brown2020language.

Maintain exactly one .bib file per paper project. Do not scatter references across multiple files.

Step4: Insert Citations Naturally

This is the most important step. Citations must read as part of the narrative, not as parenthetical interruptions.

Bad patterns to avoid:

• ❌ "Gehring et al. (2017) proposed ConvS2S..." (author-name parenthetical)
• ❌ "ConvS2S was proposed by Gehring et al. (2017)..." (citation as afterthought)
• ❌ "In [17], the authors proposed..." (bracket-first, content-second)
• ❌ "Several works [1,2,3,4,5] have studied..." (citation-dump without narrative)

Good patterns to use:

• ✅ "Fully convolutional architectures [17] replaced recurrence with gated linear units..." (citation supports the claim, not the sentence structure)
• ✅ "The Transformer [1] established self-attention as the dominant sequence-modeling primitive..." (named method + citation)
• ✅ "Training deep networks became feasible with residual connections [2] and batch normalization [3]..." (two citations woven into one sentence)
• ✅ "Recent work has pushed this further through larger-scale pretraining [4,5] and architectural innovations [6,7]..." (grouped citations with a narrative spine)

Rules for natural citation:

1. Put the claim first, then the citation. The sentence should make sense without the brackets. 2. Use the method/dataset name as the subject when citing a well-known work: "BERT [8] introduced masked language modeling..." 3. When citing multiple works, group them by what they share: "contrastive objectives [9,10,11]" or "diffusion-based approaches [12,13]." 4. Do not repeat the same citation in consecutive sentences. Group related statements and cite once at the end. 5. In related work, each paragraph should cite3—8 distinct works. 6. Never use "et al." in the running text as a noun. Write "prior work on neural translation [17,18,19]" not "Gehring et al. [17]."

Step5: Verify Citations

After drafting each section:

1. Check that every \cite{...} key exists in the project .bib file. 2. Verify that no citation is invented or guessed. 3. Confirm that citation density meets the per-section minimums. 4. Run a quick scan: does every factual claim about prior work have a citation?

Step6: Update Bib File Continuously

As drafting progresses, the .bib file grows. After every writing session:

1. Deduplicate entries (same paper, different keys). 2. Normalize venue names (e.g., "Advances in Neural Information Processing Systems" → "NeurIPS"). 3. Remove any entry not actually cited in the current draft. 4. Run bibtex or biber to catch formatting errors before submission.

Integration With Other CCFA Skills

• Before drafting: If the paper's closest competitors, baselines, or related-work clusters are unknown, route to ccf-literature-searcher first. Do not draft with guessed citations.
• During drafting: As each section is written, fill citation slots using this workflow. Do not defer all citations to a final pass.
• After drafting: Route to ccf-integrity-auditor to verify that every citation exists and every BibTeX entry is complete.
• Before submission: Route to ccf-submission-checker to verify the .bib file compiles cleanly with the venue's LaTeX template.

Quick Checklist

Citation slots identified: yes / no
All slots filled or marked UNFILLED: yes / no
Project .bib file exists and is current: yes / no
No citations invented or guessed: yes / no
Per-section density meets minimums: yes / no
Every \cite{} key exists in .bib: yes / no
Natural citation pattern (claim first, then cite): yes / no
No "Author et al. (Year)" as running-text subject: yes / no

Compression Rules

Use this file to reduce length while preserving the paper's story, evidence, and venue fit.

Compression is triggered when a manuscript exceeds the target venue budget or the user's requested word/page limit. It is the counterpart to length-budget-policy.md: underfilled manuscripts should be expanded, overfilled manuscripts should be compressed.

Mode Selection

Quick mode:

• One paragraph, one subsection, local polish, or a simple word-count reduction.
• Do not require a full checklist.
• Preserve facts and numbers exactly.
• Return compact risk status.

Standard mode:

• Full section, multi-section, full manuscript, page-limit, or camera-ready compression.
• Build a content inventory and appendix/delete decision table.
• Check claim-evidence preservation.
• Ask once before appendix/delete if the user's preference is unknown.

If a draft is only slightly over budget (<= 10%), first compress wording, repeated motivation, and low-value background. If it is substantially over budget (> 10%), create a section-level cut plan before rewriting.

Compression Hierarchy

Cut in this order:

1. Repeated motivation. 2. Generic field background. 3. Unnecessary adjectives and promotional phrasing. 4. Duplicate definitions. 5. Overlong transitions. 6. Implementation details that can move to appendix. 7. Secondary experiments or extended analyses that can move to appendix. 8. Low-impact related work sentences after the closest work is covered. 9. Optional caveats that duplicate the Limitations section.

Protect:

• Central problem and root challenge.
• Core insight and method mechanism.
• Claims tied to contributions.
• Numbers, metrics, datasets, and baseline names.
• Limitations that bound overclaiming.
• Reproducibility details required by the venue.
• Ethics, threat model, or user-study details when required.

Appendix/Delete Decision

Move to appendix when:

• The material is important for audit but not needed for first-pass understanding.
• It is extended derivation, proof detail, extra tables, hyperparameters, implementation detail, or secondary analysis.
• The main text already states the takeaway.

Delete when:

• The material repeats an earlier point.
• The claim is unsupported and not essential.
• The wording is generic venue boilerplate.
• The citation is background-only and not needed for positioning.
• The sentence increases confidence without adding evidence.

Ask once when the choice is strategic:

I recommend moving reproducibility detail and secondary analyses to appendix, and deleting repeated motivation and unsupported filler. Should I apply that policy for this compression pass?

If the user already asked for aggressive deletion, appendix-only compression, or camera-ready condensation, do not ask again.

Venue Style

AI/ML/CV/NLP:

• Keep the problem, insight, method mechanism, and evidence summary visible.
• Compress background and low-impact related work first.
• Keep ablation and baseline rationale clear enough to avoid reviewer suspicion.

DB/systems/networks:

• Keep workload, deployment assumptions, architecture, bottleneck, and end-to-end results.
• Move lower-level implementation and extra sensitivity studies to appendix if needed.

Security:

• Keep threat model, assumptions, guarantees, and ethics/disclosure scope.
• Do not compress away boundary conditions.

HCI:

• Keep research questions, participant/procedure essentials, analysis method, and claim scope.
• Move instrument details or long coding examples to appendix.

Theory/PL/FM:

• Keep theorem statement, assumptions, proof idea, and relation to prior bounds.
• Move proof details but not proof roadmap.

Sentence-Level Moves

• Replace broad openings with specific problem statements.
• Merge citations that support the same contrast.
• Convert repeated clauses into one scoped claim.
• Use one term consistently rather than alternating synonyms.
• Prefer "because" and "therefore" relations over long rhetorical bridges.
• Replace verbose contribution bullets with claim + evidence pairs.

Risk Labels

Use these labels after compression:

• preserved: claim/evidence still intact.
• weakened: claim was narrowed to match evidence.
• moved-to-appendix: detail removed from main flow but still available.
• deleted: removed because redundant or unsupported.
• risk: compression may harm clarity or auditability.

Default User-Custom Writing Format

Use this format whenever the user asks for paper writing, planning, polishing, or review but does not specify a target venue, journal, or conference. State the assumption: "No target venue was specified, so I am using the user-custom writing format."

Source Exemplars

The current custom format is distilled from two user-provided exemplar cards:

• ICLR family: references/exemplars/cards/llava-4d.md
• CVPR: references/exemplars/cards/vggt.md

These are not ordinary venue-best-paper defaults. They are the user's own preferred writing examples. Load both cards before drafting unless the user asks to exclude one.

Writing Shape

Use this structure:

1. Task and capability gap: begin from a concrete capability the field cares about, then show why current systems fail in a physical, dynamic, geometric, or deployment-relevant setting. 2. Prior-work ladder: present prior work as successive partial progress, not as a citation list. 3. Root challenge: identify the missing representation, missing output family, missing temporal/spatial reasoning, or missing direct prediction path. 4. Core insight: express one observation that makes the method feel inevitable. 5. Mechanism preview: name the main modules by their roles in the story. 6. Evidence promise: map each contribution to a measurable task, ablation, benchmark, qualitative example, or downstream use. 7. Boundary: explicitly avoid claims that require an actual target venue, uncollected experiments, or unsupported novelty.

Section Defaults

Abstract:

• One sentence for the capability gap.
• One sentence for the limitation of current approaches.
• Two to three sentences for method mechanism and outputs.
• One sentence for dataset, benchmark, or evaluation package if present.
• One sentence for results or expected evidence, calibrated to the actual proof available.

Introduction:

• Paragraph 1: field momentum and why the problem matters.
• Paragraph 2: prior-work ladder and remaining failure mode.
• Paragraph 3: root observation or design insight.
• Paragraph 4: method overview with modules and output definitions.
• Paragraph 5: contribution bullets, each linked to evidence.

Method:

• Define inputs and outputs before architecture.
• Explain the representation or token family before fusion or training details.
• Separate model, data, training, and inference if all exist.
• Use module names that reflect reviewer-facing function.

Experiments:

• Main comparison first.
• Then ablations tied to contribution bullets.
• Then qualitative examples that expose the motivating failure mode.
• Then efficiency, scaling, or downstream evidence if claimed.
• Then limitations and failure cases.

Closed-Loop Generation

Run the same loop used by references/expert-review-loop.md:

1. Draft using this custom format. 2. Simulate at least three reviewer views: method expert, experiment expert, and writing/storyline expert. 3. Assign a provisional score on a 1-10 scale with reasons. 4. Revise high-severity issues first: unclear contribution, unsupported evidence, missing outputs, weak baselines, overclaiming, or format drift. 5. Re-review the revision. 6. Repeat until no high-severity issue remains, or clearly list the remaining unresolved risk.

Citation Rules

When using this custom format, follow the citation workflow in references/citation-workflow.md:

• The introduction must carry12-20 citations woven naturally into the argument chain.
• Related Work needs3-8 citations per topic group,20-35 total.
• Every baseline, dataset, and borrowed module must be cited.
• Use claim-first citation style: the sentence makes its point, then the bracket provides attribution.
• Do not use "Author et al. [N]" as a sentence subject. Use the finding or method as the subject.
• No \textbf{Label:} scaffolding in prose paragraphs. Write continuous arguments.
• Count total citations before calling the draft complete: under25 for an AI/CV/NLP paper means under-cited.

Output Contract

When this default format is active, return:

1. Custom-format assumption. 2. Loaded custom exemplars. 3. Global story blueprint. 4. Draft or revision. 5. Claim-evidence map. 6. Review score and critique. 7. Revision pass and re-review score.

Maintenance

The custom exemplar list is intentionally easy to edit. To add or remove future examples, use the separate skill custom-exemplar-curator. That skill should generate candidate cards, schemas, and edit instructions for the user to review manually; it should not silently change the custom exemplar set.

Self-Test: Default User-Custom Writing Format

Purpose: verify that references/custom-format/default-user-format.md is used when no target venue is specified, that it loads the user-custom ICLR/CVPR exemplars, and that it runs a draft -> expert review -> revision -> re-review loop.

Prompt simulated: "帮我写一篇关于动态三维场景记忆的论文摘要和引言结构。"

Assumption: no target venue was specified, so the user-custom writing format is active.

Loaded files:

• references/custom-format/default-user-format.md
• references/exemplars/cards/llava-4d.md
• references/exemplars/cards/vggt.md
• references/storyline-blueprint.md
• references/expert-review-loop.md

Draft Pass 1

Blueprint

Task: dynamic 3D scene memory for embodied visual-language agents.

Gap: existing agents can parse current observations but struggle to remember object motion and viewpoint-dependent geometry over long trajectories.

Root challenge: a dynamic scene representation must bind identity, 3D location, time, and language-grounded attributes while staying compact enough for decision making.

Insight: object-time memory tokens can bridge LLaVA-4D-style spatiotemporal grounding and VGGT-style direct multi-output geometry.

Method: a Dynamic Geometry Memory Transformer that writes object-time geometry tokens and predicts navigation waypoints, temporal object answers, and evidence traces.

Evidence: dynamic relocation benchmark, multi-task geometry grounding, ablations for time encoding, memory compaction, and evidence trace supervision.

Abstract Draft

Embodied visual-language agents need to reason about dynamic 3D scenes, yet most current systems either describe the present view or maintain static maps. This loses the temporal evidence needed when objects move, disappear, or become visible from a different viewpoint. We propose Dynamic Geometry Memory Transformer, a model that writes compact object-time memory tokens from egocentric video and retrieves them for language-conditioned navigation and temporal question answering. Each token stores semantic attributes, camera-relative geometry, global 3D position, timestamp, and motion cues, allowing the model to predict navigation waypoints, object-state answers, and supporting evidence traces in a single forward pass. We evaluate on a dynamic relocation benchmark with moving-object goals and temporal queries, comparing against static semantic maps, recurrent frame memory, and video-LMM retrieval baselines. The model improves temporal grounding and navigation success while reducing retained visual tokens; ablations isolate the roles of time encoding, memory compaction, and evidence supervision.

Expert Review Pass 1

Score: 6/10.

High-severity issues:

• The draft sounds plausible but does not define how camera-relative geometry becomes global 3D position.
• The evidence package is broad; without a concrete benchmark construction paragraph, the central claim could look hypothetical.
• "Single forward pass" echoes VGGT but may be unsupported if the memory writer runs over a sequence.

Medium-severity issues:

• The relation between navigation waypoints and question answering needs clearer task boundaries.
• The storage reduction claim needs a metric.

Revision plan:

• Replace "single forward pass" with "one retrieval-and-decision stage" unless the architecture truly supports single-pass inference.
• Define the output triplet before detailed token fields.
• Add a benchmark construction sentence.
• Add a storage metric placeholder tied to retained token count and performance.

Draft Pass 2

Revised Abstract

Embodied visual-language agents need to remember dynamic 3D scenes, not only recognize the current view. Static maps lose object motion, while frame-history memories store redundant context without exposing which past observation supports a decision. We propose Dynamic Geometry Memory Transformer, an object-time memory model for language-conditioned relocation and temporal scene queries. Given egocentric video, the model writes compact memory tokens that bind object identity, local geometry, estimated global position, timestamp, and motion state; at query time it retrieves a support set and predicts a waypoint, an answer, and an evidence trace. We introduce a dynamic-room benchmark by replaying object relocations across multi-view trajectories and asking agents to navigate to, describe, or localize moved objects. Experiments compare against static semantic maps, recurrent frame memory, and video-LMM retrieval baselines under matched storage budgets. The proposed memory improves temporal grounding and navigation success, and ablations show that time encoding, motion-aware writing, and evidence-trace supervision each reduce failure on moved-object queries.

Expert Review Pass 2

Score: 7/10.

Remaining risks:

• Needs exact benchmark details before a real submission.
• Needs baseline names and data-generation protocol.
• The global-position estimate must be described carefully in the method.

Outcome:

• The default custom format loaded the ICLR and CVPR user exemplars.
• The first draft used the desired dynamic-scene and direct-output style.
• The review loop removed a VGGT-like overclaim and clarified evidence.
• The format is ready as a default unspecified-venue workflow, with remaining risks correctly tied to future experimental details.

Exemplar Style Analysis

Use this file when the user provides strong papers as writing references or when a built-in exemplar card is selected from references/exemplars/index.md. Extract transferable writing technique only. Do not copy content, claims, examples, terminology, or distinctive phrasing.

Inputs

Ask for or infer:

• Target venue and year if known.
• 2-5 reference papers the user considers well written.
• Built-in exemplar cards from references/exemplars/cards/ when the user has not supplied papers or wants venue-specific best-paper style.
• The user's current draft, outline, experiment table, figures, or raw idea.
• Which sections need imitation at the writing-technique level.

Non-Copying Rule

Reference papers may guide:

• Section ordering.
• Paragraph roles.
• Transition strategies.
• How challenges are framed.
• How contributions are announced.
• How methods are decomposed.
• How figures/tables are introduced.
• How experiments are sequenced.
• How limitations are disclosed.

Reference papers must not provide:

• Reused sentences or sentence templates with only noun replacement.
• The same technical claim.
• The same novelty framing if it does not fit the user's work.
• Distinctive examples, metaphors, or phrasings.
• Unsupported confidence borrowed from another paper's evidence.

Extraction Workflow

0. If using built-in exemplars, load references/exemplars/index.md, pick at most 2-4 cards that match the target venue, paper type, and evidence style, then read only those card files. 1. Read the paper at the section level first. Identify the macro story:

• task/application,
• pain point,
• prior gap,
• insight,
• method mechanism,
• evidence,
• limitation or discussion.

2. Build a paragraph-role map for each relevant section:

• context,
• gap,
• root cause,
• insight,
• method preview,
• contribution,
• evidence,
• transition,
• limitation.

3. Extract writing moves rather than words:

• how the paragraph opens,
• what contrast or causal relation it uses,
• where it places evidence,
• how it reduces reviewer uncertainty,
• how it transitions to the next paragraph.

4. Compare multiple papers and synthesize common patterns. Prefer patterns repeated across papers over one paper's idiosyncratic style. 5. Adapt the pattern to the user's story blueprint, not the other way around.

Extraction Table

Use this format:

Paper:
Venue/year:
Section:
Paragraph role:
Writing move:
Transition pattern:
Evidence pattern:
Why this works for reviewers:
Reusable technique:
How to adapt to our paper:
Do-not-copy boundary:

Section-Specific Signals

Abstract

Look for:

• How the first sentence positions the task.
• Whether the gap is framed as technical limitation, practical need, or missing capability.
• How the method is named without overloading details.
• Whether results are presented as numerical gains, broad validation, or qualitative improvement.

Introduction

Look for:

• The paragraph where the core challenge appears.
• Whether prior work is used as a ladder toward the gap or as a citation list.
• The sentence that introduces the insight.
• The contribution bullets and whether each contribution maps to later evidence.

Method

Look for:

• Overview figure placement.
• How modules are named.
• Whether each module starts with motivation or design.
• How forward process is described.
• Where assumptions and implementation details appear.

Experiments

Look for:

• Experiment sequence: setup -> comparison -> ablation -> analysis -> limitation.
• Table/figure message per result.
• How baselines are described.
• How qualitative examples support quantitative findings.
• Whether claims in Introduction are revisited.

Adaptation Output

After analyzing exemplars, produce:

1. Reference style summary: 5-10 writing techniques. 2. Our story alignment: which techniques fit the user's paper. 3. Section pattern: paragraph roles for the target section. 4. Drafting warnings: techniques that should not be used because they would distort the user's contribution or overclaim. 5. Originality guard: a reminder that all final prose must be newly written for the user's paper.

GxVAEs

Venue/year: AAAI 2024. Source: AAAI OJS https://ojs.aaai.org/index.php/AAAI/article/view/29248. Use when: writing biomedical AI, conditional molecule generation, paired representation learning, or application-driven generative model papers.

Story Pattern

The paper starts from a practical drug-discovery goal: generate molecules that are both drug-like and bioactive. The gap is that molecular generators often ignore disease-related cellular context. The method is framed as a two-part bridge between gene expression profiles and molecule generation, with one representation model extracting disease context and another generating molecules conditioned on that context.

Abstract Moves

• Start with a real application and its scientific constraint.
• Identify what previous generative models leave out.
• Introduce the architecture as two coupled modules with distinct roles.
• Tie the method to biological meaning, not only chemical validity.
• End with experiments plus case studies to support applied relevance.

Introduction Moves

• Explain the domain stakes before model details.
• Convert "context matters" into a clear modeling gap.
• Make the conditioning variable scientifically meaningful.
• Use case studies as a bridge from metrics to expert plausibility.

Method Moves

• Separate the profile representation module from the molecule generator.
• State how the latent spaces interact.
• Keep domain constraints visible: bioactivity, drug-likeness, and disease specificity.

Evidence Moves

• Combine standard generative metrics with disease-oriented case studies.
• Compare against baselines that lack the biological condition.
• Avoid claiming therapeutic efficacy unless validated experimentally.

Reusable Techniques

• For applied AI, make the omitted context the root gap.
• Pair quantitative baselines with domain case studies.
• Show how architecture mirrors the scientific causal story.

Do-Not-Copy Boundary

Do not reuse molecule-generation claims, disease-profile conditioning, or therapeutic language unless the user's work has matching biomedical evidence.

Every Bit Helps

Venue/year: AAAI 2025. Source: paper_ref/best-papers/aaai-2025-every-bit-helps.pdf; AAAI OJS https://ojs.aaai.org/index.php/AAAI/article/view/33507. Use when: writing theory, social choice, multi-agent systems, elicitation, approximation, or parameterized optimality papers.

Story Pattern

The paper motivates an abstract problem through a concrete allocation scenario, then converts the scenario into a formal tension: ordinal rankings are easy to elicit but lose utility intensity. The central question is parameterized by the number of value queries, and the contribution is to close the gap between known upper and lower bounds.

Abstract Moves

• Define the formal problem and why weaker information is used.
• State the inefficiency measure before stating the result.
• Tie new bounds to a prior result and a prior lower bound.
• End with the "settles open questions" implication.

Introduction Moves

• Use a realistic example before formal notation.
• Introduce the metric of success early.
• Give a small history of known query regimes.
• State the exact missing regime as a question.
• Use tables to map previous and new bounds.

Method Moves

• Put notation after the motivation is clear.
• Separate model definitions from theorem statements.
• Use theorem names and tables as navigation anchors.

Evidence Moves

• For theory papers, evidence is proof structure plus tightness against lower bounds.
• Emphasize optimality only where upper and lower bounds match.
• Make assumptions visible before theorems so reviewers can audit scope.

Reusable Techniques

• Turn a parameter sweep into a story: from zero queries to constant queries to logarithmic queries.
• Use examples to make a formal objective intuitive before proving results.
• Present contribution tables early for quick reviewer orientation.

User Notes

Use this card when the user's result "settles" or tightens a known regime. If the paper only improves constants, avoid the same level of closure language.

Do-Not-Copy Boundary

Do not reuse the office/allocation example, distortion framing, or optimality language unless the user's paper has matching formal guarantees.

Mission: Impossible Language Models

Venue/year: ACL 2024, Best Paper. Source: ACL Anthology https://aclanthology.org/2024.acl-long.787/. Use when: writing NLP evaluation, cognitive science of LLMs, synthetic data, falsification, or claim-testing papers.

Story Pattern

The paper takes a strong public claim about LLMs and turns it into an experimental question. Instead of arguing abstractly, it constructs a continuum of synthetic impossible languages and measures model learning across training. The contribution is both a benchmark-like probe and a disciplined test of a contested theoretical claim.

Abstract Moves

• Start with the claim being tested.
• State the evidence gap behind the claim.
• Introduce synthetic test languages as the experimental instrument.
• Report the core finding against the control language.
• End by opening a broader research program.

Introduction Moves

• Quote or paraphrase the disputed claim carefully and concretely.
• Explain why the claim matters for methodology.
• Avoid overclaiming by limiting the experiment to a model family and language set.

Method Moves

• Define the impossibility continuum before model details.
• Keep data transformations transparent.
• Evaluate across training stages, not just endpoints, when learning dynamics are part of the thesis.

Evidence Moves

• Use a natural language control.
• Report multiple synthetic variants to avoid one-off artifacts.
• Connect quantitative learning curves to linguistic interpretation.

Reusable Techniques

• Convert a philosophical or theoretical debate into a measurable probe.
• Use synthetic data when it isolates a capability boundary.
• Make conclusions appropriately scoped to avoid reviewer pushback.

Do-Not-Copy Boundary

Do not reuse impossible-language tasks, Chomsky-framed claims, or cognitive conclusions unless the user's experiments directly test them.

MiniLongBench

Venue/year: ACL 2025, Outstanding Paper. Source: ACL Anthology https://aclanthology.org/2025.acl-long.560/. Use when: writing benchmark, evaluation efficiency, long-context LLM, dataset pruning, or low-cost evaluation papers.

Story Pattern

The paper targets the evaluation cost of long-context understanding. It argues that existing benchmarks contain redundancy, then introduces a compressed benchmark that preserves ranking behavior while dramatically reducing test size and cost. The key writing move is to make evaluation economy a scientific contribution.

Abstract Moves

• Start from the importance of the capability and the cost barrier.
• Identify redundancy as the discovered root cause.
• Introduce the benchmark by its compression method and size.
• Prove usefulness through rank correlation with the full benchmark.
• End with the research-enabling implication.

Introduction Moves

• Position evaluation cost as a bottleneck for model development.
• Distinguish "shorter benchmark" from "weaker benchmark" using correlation evidence.
• Explain sparse information as the reason pruning can work.

Method Moves

• Describe the pruning or compression pipeline in auditable steps.
• Define what behavior must be preserved after compression.
• Keep benchmark categories visible so reviewers can judge coverage.

Evidence Moves

• Use many models to support rank-preservation claims.
• Report cost reduction and correlation together.
• Include task-category analysis to avoid the appearance of cherry-picking.

Reusable Techniques

• Convert evaluation expense into a first-class problem.
• Support small-data claims with preservation metrics, not only speedups.
• Phrase benchmark contributions around utility for future research.

User Notes

Use this card when the user's paper proposes a cheaper evaluation protocol or compressed dataset.

Do-Not-Copy Boundary

Do not borrow the LongBench pruning claim, exact cost numbers, or rank-correlation framing unless the user's evaluation validates them.

From Speaker to Dubber

Venue/year: ACM MM 2024. Source: OpenReview PDF https://openreview.net/pdf/c1a849c12d7b851fa543b7af787bbc8fc9a945b7.pdf. Use when: writing multimodal speech/video generation, dubbing, temporal alignment, prosody modeling, or staged training papers.

Story Pattern

The paper frames movie dubbing as a multimodal alignment task with three simultaneous requirements: voice identity, emotional/prosodic fit, and duration/lip synchronization. The method story is a staged learning pipeline: learn clean pronunciation from broader data, then practice dubbing-specific alignment under limited noisy movie data.

Abstract Moves

• Define the task through multiple constraints rather than a single input-output mapping.
• Identify data limitation and background noise as root causes.
• Introduce staged training as a way to separate general speech knowledge from domain-specific alignment.
• Name the two alignment modules by the constraints they solve.
• Close with benchmark superiority and demos/code as reproducibility signals.

Introduction Moves

• Explain why the task is harder than ordinary voice cloning or TTS.
• Decompose the challenge into duration consistency and prosody consistency.
• Connect each challenge to a module so the method feels inevitable.

Method Moves

• Make the training stages easy to follow with a figure.
• Separate phoneme-level representation, prosody attributes, duration reasoning, and final synthesis.
• Keep multimodal alignment terms consistent across method and experiments.

Evidence Moves

• Combine objective metrics with perceptual or demo-based evidence.
• Compare against prior single-stage methods when staged learning is the core claim.
• Include ablations for each consistency module.

Reusable Techniques

• Decompose a complex generation task into named constraints.
• Justify staged training by data quality and domain mismatch.
• Use demos as supplementary evidence, not as a substitute for benchmarks.

User Notes

Use for multimodal generation papers where temporal alignment is as important as output quality.

Do-Not-Copy Boundary

Do not reuse movie-dubbing examples, prosody/duration module names, or demo claims unless they belong to the user's work.

Aff3DFunc

Venue/year: ACM MM 2025. Source: University of Glasgow record https://eprints.gla.ac.uk/360521/; PDF https://eprints.gla.ac.uk/360521/2/360521.pdf. Use when: writing open-vocabulary 3D understanding, affordance detection, language-geometry alignment, or robot validation papers.

Story Pattern

The paper frames 3D affordance understanding as a practical robotics problem where label-only text descriptions are too weak for open-vocabulary generalization. The method adds functional text enhancement and multilevel representation alignment, then closes the loop with real-world robot validation.

Abstract Moves

• Start with embodied interaction needs: manipulation and navigation.
• Name the weakness of label-based language prompts.
• Present text enhancement as semantic enrichment, not prompt decoration.
• Present multilevel alignment as the bridge between language and 3D geometry.
• End with zero-shot/generalization evidence and robot validation.

Introduction Moves

• Explain affordance as function, not category.
• Show why open-vocabulary queries need richer language than labels.
• Tie each method component to a failure of previous representations.

Method Moves

• Define query forms and 3D representation before alignment.
• Separate text enhancement from geometry-language alignment.
• Keep robot deployment requirements visible.

Evidence Moves

• Evaluate under different textual query forms.
• Include fine-grained manipulation examples, not only segmentation maps.
• Use robot validation to support actionable affordance understanding.

Reusable Techniques

• Treat language enhancement as functional grounding.
• Show generalization through query diversity.
• Pair perception metrics with downstream robot behavior.

Do-Not-Copy Boundary

Do not reuse affordance labels, robot validation claims, or functional text method names unless they belong to the user's work.

Constrained Highlighting Style Card

Venue/year: CHI 2024 family. Source: ACM Best Paper Awards source record acm-best-paper-awards; verify the exact CHI paper page before public award-status claims. Use when: HCI, reading interfaces, interaction technique evaluation, user studies, education technology, or human-centered design papers.

Story Pattern

The paper type is HCI research: start from a human activity and a measurable difficulty, introduce the design intervention, then connect study design to the claim. The contribution is not the interface alone; it is the evidence about how and why the interaction changes user behavior or understanding.

Abstract Moves

• State the human task and population/context.
• Describe the intervention without overselling it as universally better.
• Preview study design, measures, and core finding using user-supplied results only.
• Bound the claim to the studied context.

Introduction Moves

• Motivate through user need, not only technical novelty.
• Convert the design intuition into research questions.
• Explain why prior interfaces or studies leave the question unresolved.
• State contributions as design, empirical finding, and implications.

Method Moves

• Make participants, tasks, measures, procedure, and analysis plan easy to audit.
• Distinguish design rationale from study outcome.
• Include limitations and ethics as part of trust, not afterthought.

Evidence Moves

• Report quantitative and qualitative evidence according to the study design.
• Avoid causal language if the study design does not support it.
• Include failure cases, subgroup caveats, or ecological-validity limits when relevant.

Do-Not-Copy Boundary

Do not reuse study claims or intervention specifics. Transfer only the HCI story and evidence structure.

CVPR 2023 Best Papers

Venue/year: CVPR 2023. Source: CVF awards list https://www.thecvf.com/?page_id=413. Use when: the user explicitly asks for recent CVPR best-paper style, especially visual reasoning without training or autonomous-driving planning.

Award Papers

• Visual Programming: Compositional Visual Reasoning Without Training
• Planning-Oriented Autonomous Driving

Shared Writing Signals

• Frame the task around an end-to-end behavior that matters: compositional reasoning or planning.
• Make the intermediate representation explicit: programs, perception-planning links, or task-oriented outputs.
• Show that the method changes the problem formulation, not only a network component.
• Evaluate in a way that mirrors the final decision process.

Transferable Moves

• For reasoning papers: define compositional units and show how they combine.
• For driving/planning papers: connect perception outputs to planning objectives instead of treating them as separate metrics.
• Use qualitative chains or trajectories to make decision structure inspectable.
• Let the method section mirror the task decomposition.

Reviewer-Facing Warnings

• Do not claim "without training" or "planning-oriented" unless the experimental protocol proves that exact property.
• Make sure the output evaluated is the output the paper claims to optimize.

CVPR 2024 Best Papers

Venue/year: CVPR 2024. Source: CVF awards list https://www.thecvf.com/?page_id=413; CVPR news https://cvpr.thecvf.com/Conferences/2024/News/Awards. Use when: the user explicitly asks for recent CVPR best-paper style, especially human-feedback image generation or generative image dynamics.

Award Papers

• Rich Human Feedback for Text-to-Image Generation
• Generative Image Dynamics

Shared Writing Signals

• Tie the visual generation problem to a concrete missing supervisory or physical signal.
• Make the data or feedback source part of the contribution, not just training material.
• Use visual examples to reveal what the quantitative metric cannot explain.
• Show generality through diverse prompts, scenes, or downstream manipulations.

Transferable Moves

• For feedback papers: define the feedback taxonomy and why it is richer than scalar ratings.
• For dynamics papers: explain how a static image becomes a representation of plausible motion.
• Use figures to communicate behavior first, then quantify it.
• Include failure cases because generation papers are especially vulnerable to cherry-picking criticism.

Reviewer-Facing Warnings

• Do not overclaim human preference alignment without careful annotation and evaluation protocols.
• For dynamics or animation claims, distinguish plausible visual motion from physically correct motion.

CVPR 2025 Best Paper: VGGT

Venue/year: CVPR 2025. Award: Best Paper. Source: CVPR https://cvpr.thecvf.com/Conferences/2025/BestPapersDemos; CVF awards list https://www.thecvf.com/?page_id=413. Primary writing card: cards/vggt.md. Use when: the user explicitly asks for CVPR 2025 best-paper style or asks how VGGT fits official CVPR best-paper context.

Role In This Skill

VGGT is both:

• a user-custom writing-format exemplar, via cards/vggt.md;
• the CVPR 2025 Best Paper, recorded here for venue-award context.

When no target venue is specified, use cards/vggt.md as part of the custom format. When the user asks for CVPR best-paper style, load this card only to note the award context, then load cards/vggt.md for detailed writing moves.

Best-Paper Writing Signals

• Directly replace a complex geometry pipeline with a simpler feed-forward formulation.
• Define the output family early and show why joint prediction is useful.
• Support the claim with breadth: many input counts, tasks, baselines, runtime comparisons, and downstream reuse.
• Make simplicity feel earned by showing that performance is not sacrificed.

Reviewer-Facing Warnings

• Do not use "pipeline replacement" language unless the user's evaluation covers strong classical and learning-based baselines.
• Do not claim general 3D understanding without multiple task families or downstream evidence.

Partial Distance Correlation in Deep Learning

Venue/year: ECCV 2022. Source: ECVA PDF https://www.ecva.net/papers/eccv_2022/papers_ECCV/papers/136860318.pdf; arXiv https://arxiv.org/abs/2207.09684. Use when: writing representation analysis, statistical tools for deep learning, regularization, disentanglement, or robustness papers.

Story Pattern

The paper revives a statistical dependence measure and shows that it solves several deep-learning problems that look different on the surface. The writing move is to make a tool feel versatile without becoming vague: each application is tied to a shared need to compare or constrain functional behavior across feature spaces.

Abstract Moves

• Start from a broad recurring need: comparing what neural networks learn.
• Explain why existing comparison tools are awkward at scale or across dimensions.
• Introduce the statistical tool and its partial variant.
• List diverse applications only after the common mechanism is clear.

Introduction Moves

• Use a thought experiment to make "functional behavior comparison" intuitive.
• Contrast layer-wise analysis with cross-network functional comparison.
• Position the method as a general regularizer or constraint, then prove utility by applications.

Method Moves

• Explain the statistical quantity before using it as a loss or constraint.
• Provide deployment details needed for large-scale models.
• Separate diagnostic use from training-time regularization.

Evidence Moves

• Choose applications that demonstrate different faces of the same mechanism.
• Avoid a scattershot feeling by returning to the shared dependence-control theme.
• Include robustness or disentanglement evidence where the constraint should matter.

Reusable Techniques

• Turn an old tool into a new ML contribution by explaining deployment friction.
• Use one unifying mechanism to justify many applications.
• Make versatility credible through recurring notation and repeated evidence pattern.

Do-Not-Copy Boundary

Do not reuse the partial-distance-correlation method or versatility claim unless the user's work genuinely supports multiple validated uses.

Minimalist Vision with Freeform Pixels

Venue/year: ECCV 2024. Source: ECVA https://www.ecva.net/papers/eccv_2024/papers_ECCV/html/8113_ECCV_2024_paper.php. Use when: writing computational imaging, sensing hardware, privacy-preserving vision, energy-efficient systems, or hardware-software co-design papers.

Story Pattern

The paper asks how little sensing is enough for a task. It reframes camera design as a learnable first layer: freeform pixels are optimized jointly with inference layers. The paper's appeal comes from a surprising contrast between very few measurements and competitive task performance, plus privacy and self-powering benefits.

Abstract Moves

• Start with a crisp definition of the new system class.
• Contrast standard dense sensing with task-specific freeform sensing.
• Explain the hardware abstraction in ML terms.
• Give concrete tasks and pixel counts.
• Close with system-level advantages beyond accuracy.

Introduction Moves

• Motivate minimalism through efficiency, privacy, and deployment constraints.
• Make the sensor itself part of the learned model.
• Use simple tasks first so reviewers can understand the hardware mechanism.

Method Moves

• Explain the optical/hardware layer before the neural inference layer.
• Keep the mapping between learned pixel shape and physical implementation explicit.
• Separate task training from device realization.

Evidence Moves

• Compare against much denser conventional cameras.
• Include physical prototypes or implementation details when making hardware claims.
• Report the practical consequences of minimal sensing, not only accuracy.

Reusable Techniques

• Write from a provocative question: what is the minimum signal needed?
• Treat constraints as a design advantage when they improve privacy or power.
• Use concrete counts and deployment scenarios to make efficiency claims tangible.

User Notes

Useful when the user's work is strongest because it removes capacity, measurements, labels, or compute while preserving task success.

Do-Not-Copy Boundary

Do not copy the freeform-pixel concept, privacy rationale, or self-powered claim unless the user's system has matching design and validation.

Passive Ultra-Wideband Single-Photon Imaging

Venue/year: ICCV 2023. Source: CVF https://openaccess.thecvf.com/content/ICCV2023/html/Wei_Passive_Ultra-Wideband_Single-Photon_Imaging_ICCV_2023_paper.html. Use when: writing computational imaging, passive sensing, photon-limited reconstruction, hardware plus algorithm, or extreme-dynamic-range papers.

Story Pattern

The paper defines an unusually difficult imaging regime: dynamic scenes over seconds-to-picoseconds timescales, passively, with little light, and without timing signals from the source. The writing strength is that the sensing constraints are stated as the problem itself, then the method and experiments are organized around making that impossible-looking regime measurable.

Abstract Moves

• Open with the extreme imaging regime.
• Bundle constraints explicitly so reviewers understand the novelty boundary.
• Present the system as passive and timing-signal-free, not merely improved reconstruction.
• Emphasize simultaneous timescale coverage.

Introduction Moves

• Contrast active illumination setups with passive real-world sensing.
• Explain why source timing is normally needed.
• Frame the method as expanding what can be observed, not only improving a metric.

Method Moves

• Introduce the sensing model before reconstruction.
• Make hardware assumptions explicit.
• Separate physical acquisition from computational recovery.

Evidence Moves

• Show examples across the claimed timescale range.
• Include controlled experiments that verify the passive measurement assumption.
• Use qualitative visualizations to make invisible transient information interpretable.

Reusable Techniques

• Let multiple constraints define the novelty.
• Make the physical regime clear before algorithm details.
• Use demonstrations to show a new measurement capability.

Do-Not-Copy Boundary

Do not reuse passive-imaging claims, timescale ranges, or single-photon constraints unless the user's system actually operates there.

BrickGPT

Venue/year: ICCV 2025. Source: CVF https://openaccess.thecvf.com/content/ICCV2025/html/Pun_Generating_Physically_Stable_and_Buildable_Brick_Structures_from_Text_ICCV_2025_paper.html. Use when: writing text-to-3D, generative design, physical constraints, autoregressive generation, dataset, or embodied construction papers.

Story Pattern

The paper turns generative modeling into a physically constrained construction problem. The model is not only asked to match a text prompt; it must produce objects that obey assembly constraints, remain stable, and can be built by humans or robots. The story works because the method, dataset, inference checks, and buildability demonstrations all serve the same promise.

Abstract Moves

• Claim firstness only after naming the concrete task boundary.
• Introduce dataset construction as the precondition for the model.
• State the generative mechanism and the physical safeguard separately.
• End with multiple validation modes: stable, diverse, visually aligned, manually buildable, robotically buildable.

Introduction Moves

• Move from text-to-3D generation to the harder requirement of physical assembly.
• Make feasibility constraints part of the task definition, not a post-hoc filter.
• Explain why standard visual quality metrics are insufficient.

Method Moves

• Separate representation, dataset, autoregressive model, validity check, and rollback.
• Treat inference-time constraint handling as a core algorithmic component.
• Make the object representation understandable before model training details.

Evidence Moves

• Combine automatic metrics, physical validity checks, qualitative galleries, and real assembly demonstrations.
• Show failure cases where visual plausibility is not enough.
• Use dataset scale and diversity to support generalization claims.

Reusable Techniques

• Define a generation task by downstream feasibility, not only by visual similarity.
• Pair data creation and inference constraints when real-world validity matters.
• Let build or deployment demonstrations close the evidence loop.

User Notes

Use this card when a paper must convince reviewers that generated outputs are actionable in the physical world.

Do-Not-Copy Boundary

Do not reuse brick-specific constraints, dataset claims, or buildability demonstrations unless the user's domain has equivalent evidence.

ICLR 2024 Outstanding Papers

Venue/year: ICLR 2024. Source: ICLR Blog https://blog.iclr.cc/2024/05/06/iclr-2024-outstanding-paper-awards/. Use when: the user explicitly asks for recent ICLR outstanding-paper style, especially diffusion generalization, robotics simulators, long-sequence models, protein generation, or vision transformer analysis.

Award Papers

• Generalization in diffusion models arises from geometry-adaptive harmonic representations
• Learning Interactive Real-World Simulators
• Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors
• Protein Discovery with Discrete Walk-Jump Sampling
• Vision Transformers Need Registers

Shared Writing Signals

• Identify a surprising gap in how strong models are understood, evaluated, or trained.
• Prefer one crisp observation over a list of incremental improvements.
• Connect practical impact to a deeper explanation: geometry, unified interfaces, priors, discrete sampling, or feature artifacts.
• Make the evidence package unusually complete for the claim: theory plus experiments, system scale plus examples, or in-silico plus wet-lab validation.

Transferable Moves

• Use a "hidden cause" structure: observed behavior -> diagnosis -> simple or principled remedy.
• For systems work, show why the engineering unification changes what can be studied.
• For model-analysis work, make qualitative artifacts visible before abstracting them.
• For biological or robotics applications, close the loop with domain-valid evidence.

Reviewer-Facing Warnings

• ICLR outstanding-paper style rewards depth; do not imitate broad scope without enough analysis.
• If the paper proposes a simple fix, spend extra space proving the diagnosis.
• If the paper aggregates data or systems, make interface assumptions and failure cases explicit.

ICLR 2025 Outstanding Papers

Venue/year: ICLR 2025. Source: ICLR Blog https://blog.iclr.cc/2025/04/22/announcing-the-outstanding-paper-awards-at-iclr-2025/. Use when: the user explicitly asks for recent ICLR outstanding-paper style, especially LLM safety, finetuning dynamics, model editing, data valuation, or foundation-model analysis.

Award Papers

• Safety Alignment Should be Made More Than Just a Few Tokens Deep
• Learning Dynamics of LLM Finetuning
• AlphaEdit: Null-Space Constrained Model Editing for Language Models

Shared Writing Signals

• Treat hidden mechanisms as first-class objects of study: alignment depth, finetuning dynamics, or edit subspaces.
• Begin from a widely used practical technique, then show that its internal behavior is not yet understood or controlled.
• Make the method or analysis precise enough that reviewers can reproduce the diagnostic.
• Use experiments to expose a mechanism, not only to report a score.

Transferable Moves

• Pair practical relevance with mechanistic explanation.
• Name the failure mode or control dimension early.
• Build a claim-evidence path from diagnostic finding to intervention.
• Let the contribution be an insight plus a tool, not just a new benchmark or model.

Reviewer-Facing Warnings

• Avoid claiming causal mechanism unless the experiments isolate it.
• For editing or alignment papers, define the threat model, target behavior, and evaluation boundary.
• For finetuning papers, separate optimization dynamics from dataset artifacts.

ICLR 2026 Outstanding Papers

Venue/year: ICLR 2026. Source: ICLR Blog https://blog.iclr.cc/2026/04/23/announcing-the-iclr-2026-outstanding-papers/. Use when: the user explicitly asks for recent ICLR outstanding-paper style, especially theory-of-transformers or multi-turn LLM evaluation papers.

Award Papers

• Transformers are Inherently Succinct
• LLMs Get Lost In Multi-Turn Conversation

Shared Writing Signals

• State a broad field assumption, then challenge it with a sharper evaluation or theory lens.
• Make the central contribution legible as a new way to look at a familiar object: transformer expressivity, or LLM conversation ability.
• Use the introduction to justify why the paper's measurement or theory target was previously under-examined.
• Keep limitations visible when the finding might be controversial.

Transferable Moves

• For theory papers: make the conceptual measure intuitive before formal machinery.
• For evaluation papers: define the deployment mismatch before benchmark construction.
• Explain why the result matters even when the exact setup can be debated.
• Use rigorous selection of problem setting as part of the contribution.

Reviewer-Facing Warnings

• Do not borrow the "outstanding" confidence unless the user's paper has a similarly crisp thesis and strong validation.
• Do not use broad LLM or transformer claims without a clear scope statement.
• If the paper studies a failure mode, include controls that distinguish artifact from real capability gap.

VideoPoet

Venue/year: ICML 2024. Source: PMLR https://proceedings.mlr.press/v235/kondratyuk24a.html. Use when: writing multimodal generation, video synthesis, autoregressive modeling, zero-shot capability, or large model adaptation papers.

Story Pattern

The paper transfers the language-model training recipe to video generation. Rather than building a separate model for each conditioning type, it treats text, images, video, and audio as multimodal sequences handled by one decoder-only transformer. The contribution is a foundation model that can be adapted to many video generation tasks and demonstrates zero-shot motion quality.

Abstract Moves

• Define the system by the range of conditioning signals it accepts.
• Present the architecture using a familiar model family.
• Separate pretraining from task-specific adaptation.
• Frame empirical results around zero-shot capability and motion fidelity.

Introduction Moves

• Establish why video generation needs flexible conditioning.
• Contrast task-specific pipelines with a unified generative sequence model.
• Make the training protocol central to the contribution.

Method Moves

• Explain tokenization or representation of each modality before training objectives.
• Group objectives by what capability they teach.
• Keep adaptation steps distinct from pretraining.

Evidence Moves

• Use qualitative examples to demonstrate motion and conditional control.
• Pair zero-shot claims with task diversity.
• Avoid overloading the main text with demos; keep visual evidence organized.

Reusable Techniques

• Use a unification story when one model handles many input-output modes.
• Make "LLM recipe for another modality" concrete through objectives and tokens.
• Present zero-shot as a capability supported by task coverage.

Do-Not-Copy Boundary

Do not reuse VideoPoet's multimodal task mix, decoder-only framing, or zero-shot claims unless the user's model and evaluation support them.

CollabLLM

Venue/year: ICML 2025. Source: PMLR https://proceedings.mlr.press/v267/wu25i.html. Use when: writing LLM, agent, human-AI collaboration, multiturn interaction, reward design, or user-study papers.

Story Pattern

The paper reframes LLM interaction as a long-horizon collaboration problem rather than next-turn response quality. The method contribution is a training framework that gives credit to responses by their downstream contribution to user goals, and the evidence package combines benchmark tasks, LLM judges, and a user study.

Abstract Moves

• Start with a training/evaluation mismatch: next-turn rewards do not optimize long-term interaction.
• Name a behavioral failure: passive responses to ambiguous or open-ended requests.
• Introduce the framework through its training signal.
• Report both task performance and interaction quality.
• Add human-user evidence to strengthen the collaboration claim.

Introduction Moves

• Define what good collaboration means before introducing the method.
• Use ambiguous-user-intent examples as motivation, but keep them short.
• Present reward design as the core mechanism, not as an implementation detail.
• Link benchmark design to the paper's thesis.

Method Moves

• Explain the simulation or reward pipeline as a credit-assignment solution.
• Keep the distinction clear among training data, reward estimation, fine-tuning, and evaluation.
• Define interaction metrics in reviewer-readable language.

Evidence Moves

• Pair automatic evaluation with human judgment when claiming user-centered gains.
• Report both performance and interactivity so the method is not seen as only more verbose.
• Include a cost or time measure if the claim includes collaboration efficiency.

Reusable Techniques

• Use "short-term objective versus long-term user goal" as a clean motivation axis.
• Treat benchmark design as part of the contribution when existing evaluation cannot test the thesis.
• Let user study evidence calibrate claims about user satisfaction.

User Notes

Useful for any paper where the method changes how an AI system behaves in conversation, not just what it predicts.

Do-Not-Copy Boundary

Do not borrow CollabLLM's reward terminology, benchmark framing, or human-study claims unless the user's work has analogous protocol and evidence.

LLaVA-4D

Venue/year: ICLR 2026. Venue family: ICLR. Custom status: user-custom writing-format exemplar. This card participates in the default custom writing format when the user does not specify a target venue. Source: paper_ref/LLaVA-4D.pdf; OpenReview https://openreview.net/forum?id=URpbmVEsqB; arXiv https://arxiv.org/abs/2505.12253. Use when: writing multimodal, 3D/4D scene, embodied AI, prompt-embedding, or dataset plus model papers.

Story Pattern

Broad success in 2D multimodal understanding is narrowed to a physical-world failure: missing spatial representation. The introduction then shows that existing 3D LMMs improve static spatial modeling but do not model temporal variation. The proposed method is framed as a minimal dimensional extension with a concrete mechanism: add time to coordinates and disentangle spatial and temporal visual features.

Abstract Moves

• Start with a capability gap tied to the physical world.
• Describe the prior approach in one sentence, then immediately name its failure case.
• Introduce the method as a general framework with one memorable mechanism.
• Add a dataset contribution after the modeling contribution.
• End with broad task validation, not only one benchmark.

Introduction Moves

• Paragraph 1: field momentum and why the current generation of LMMs is insufficient for physical interaction.
• Paragraph 2: prior 3D LMMs as a ladder, followed by the dynamic-object failure.
• Paragraph 3: observation that static background and moving objects share positions but differ in motion.
• Paragraph 4: module preview with explicit mapping from observation to design.
• Contributions: one model-level claim, two insight-to-module claims, one dataset/evidence claim.

Method Moves

• Name each module by the role it plays in the story, not just by architecture.
• Explain coordinate encoding before fusion so the reader understands what information is being injected.
• Keep dataset construction separate from model mechanics so reviewers can evaluate each contribution.

Evidence Moves

• Use a first figure that contrasts old paradigm, new paradigm, benchmark summary, and qualitative task behavior.
• Include multiple tasks so the paper reads as a framework rather than a narrow fix.
• Pair quantitative results with dynamic qualitative examples that expose the failure mode of static models.

Reusable Techniques

• Turn a missing dimension into a writing spine: 2D to 3D to 4D.
• Convert an observation into a design decision, then into a contribution bullet.
• Make the first figure carry both motivation and evidence.

User Notes

This is one of the user's recognized exemplars. Use it especially when the user's paper needs to justify a new representation for dynamic scenes.

Do-Not-Copy Boundary

Do not reuse the 4D coordinate prompt claim, dynamic-scene examples, dataset framing, or contribution wording unless they are actually the user's own technical content.

Visual Autoregressive Modeling

Venue/year: NeurIPS 2024. Source: NeurIPS Proceedings https://proceedings.neurips.cc/paper_files/paper/2024/hash/9a24e284b187f662681440ba15c416fb-Abstract-Conference.html. Use when: writing image generation, autoregressive modeling, new prediction paradigm, scaling laws, or fast inference papers.

Story Pattern

The paper redefines autoregressive image generation by changing what the next prediction is. Instead of raster-scan next-token prediction, the model predicts the next scale or resolution in a coarse-to-fine process. The story is powerful because a simple conceptual shift yields quality, speed, data-efficiency, and scaling-law evidence.

Abstract Moves

• Name the paradigm shift in the first technical sentence.
• Contrast it with the standard formulation.
• State why the new formulation is intuitive and scalable.
• Report a dramatic benchmark improvement and speed gain.
• Extend the claim to scaling laws and zero-shot image editing.

Introduction Moves

• Explain why text-style autoregression is awkward for images.
• Use spatial hierarchy as the reason next-scale prediction is natural.
• Position diffusion models as the benchmark to beat, but keep comparisons evidence-based.

Method Moves

• Define the visual token hierarchy before transformer training.
• Make the prediction order visually understandable.
• Explain how inference speed follows from the formulation.

Evidence Moves

• Pair headline metrics with speed and scalability curves.
• Include comparison to both AR and diffusion baselines.
• Use downstream editing/inpainting as evidence of generalization.

Reusable Techniques

• A new prediction target can be the whole paper story.
• Make scaling claims with curves, not only final model performance.
• Let conceptual simplicity and empirical breadth reinforce each other.

Do-Not-Copy Boundary

Do not reuse next-scale prediction, reported numbers, or diffusion-comparison claims unless the user's method and evaluation support them.

1000 Layer Networks for Self-Supervised RL

Venue/year: NeurIPS 2025. Source: NeurIPS Proceedings https://proceedings.neurips.cc/paper_files/paper/2025/hash/e74ee34cc0f2d0780f34ee77d8fba25b-Abstract-Conference.html. Use when: writing scaling, RL, self-supervised learning, architecture depth, goal-conditioned learning, or capability-emergence papers.

Story Pattern

The paper imports a familiar scaling lesson from language and vision into reinforcement learning, then argues that RL had not benefited because the right building blocks were missing. The specific axis is network depth, and the evidence is both quantitative improvement and qualitative change in learned behavior.

Abstract Moves

• Open with a cross-field contrast: scaling transformed language/vision but not RL.
• Identify one design axis as critical.
• State the experimental setting in a way that raises difficulty: no rewards or demonstrations.
• Quantify gains across tasks.
• Add that scaling changes behaviors, not only metrics.

Introduction Moves

• Explain why existing shallow RL architectures may be a bottleneck.
• Make the unsupervised goal-conditioned setting central to the difficulty.
• Present depth scaling as a systematic study, not a one-off architecture trick.

Method Moves

• Define the base algorithm before scaling modifications.
• Separate architecture depth, stability tricks, objective, and training protocol.
• Keep compute and implementation details visible for reproducibility.

Evidence Moves

• Show scaling curves rather than only best numbers.
• Compare against goal-conditioned baselines.
• Include qualitative behavior examples to support the capability-change claim.

Reusable Techniques

• Use cross-domain analogy carefully, then validate it in the target domain.
• Pair scaling laws with concrete downstream behaviors.
• Make absence of supervision part of the challenge framing.

User Notes

Use for papers where a simple scaling dimension unlocks surprising performance or behavior.

Do-Not-Copy Boundary

Do not reuse the depth-scaling claim, unsupervised RL setup, or magnitude of gains unless supported by the user's experiments.

MegaLibm Style Card

Venue/year: POPL 2024 family. Source: ACM Best Paper Awards source record acm-best-paper-awards; verify the exact paper page before public award-status claims. Use when: programming languages, program synthesis, compiler/tool papers, numerical libraries, formal methods, or papers combining correctness with empirical tool evaluation.

Story Pattern

The paper type is PL/tool research: start with a correctness, synthesis, or verification bottleneck that practitioners face, state the formal object being produced or checked, then explain how the method bridges proof, implementation, and empirical utility.

Abstract Moves

• State the artifact or formal target clearly.
• Explain why existing synthesis, verification, or library-engineering approaches are insufficient.
• Present the key technical idea before the tool name.
• Preview evidence as correctness, coverage, performance, or usability, not only speed.

Introduction Moves

• Use one concrete failure mode or maintenance pain point to motivate the formal problem.
• Define the core formal challenge before implementation details.
• State how the technique changes the search, proof, or compilation space.
• Separate theoretical guarantee, implementation, and evaluation contributions.

Method Moves

• Put definitions before algorithms.
• Give a proof or soundness roadmap before dense formalism.
• Explain tool architecture only after the core semantics or synthesis idea is clear.

Evidence Moves

• Combine formal guarantees with benchmarks on real programs or libraries.
• Report coverage, correctness, performance, and failure cases.
• Include threats to validity for benchmark selection and specification assumptions.

Do-Not-Copy Boundary

Do not reuse library-specific examples, theorem wording, or tool claims. Transfer only the PL proof-plus-artifact writing pattern.

L25GC+ Style Card

Venue/year: SIGCOMM 2024 family. Source: SIGCOMM 2024 program source record sigcomm-2024-program; verify the exact paper page before public award-status claims. Use when: networking systems, mobile core networks, low-latency data planes, NFV, packet processing, or deployment-driven network architecture papers.

Story Pattern

The paper type is a networked-systems contribution: open with a real latency, throughput, or deployment pain point, then show how the architecture changes the bottleneck. The writing should make the operational setting and workload realism central.

Abstract Moves

• State the network setting and why current designs miss the target.
• Present the design as a re-architecture around one bottleneck.
• Preview evaluation on realistic traces, deployments, or workloads.
• Avoid broad "faster network" claims unless the measured setting supports them.

Introduction Moves

• Start with user or operator consequences of the bottleneck.
• Distinguish protocol, implementation, and deployment constraints.
• Explain why common optimizations are insufficient.
• End the introduction with design goals and evidence categories.

Method/System Moves

• Show control plane, data plane, and state management explicitly.
• Make assumptions and deployment boundaries auditable.
• Keep performance-critical paths readable and reproducible.

Evidence Moves

• Include end-to-end results plus sensitivity to workload, traffic mix, and scale.
• Use microbenchmarks to prove which part of the design removes the bottleneck.
• Report overhead, resource use, and failure behavior when relevant.

Do-Not-Copy Boundary

Do not reuse network-specific claims or names unless they are the user's actual system. Transfer only the bottleneck-driven systems narrative.

PolarDB-MP Style Card

Venue/year: SIGMOD 2024 family. Source: ACM Best Paper Awards source record acm-best-paper-awards; verify the exact paper page before public award-status claims. Use when: database systems, cloud-native DBMS, distributed transactions, shared storage/memory, storage-compute disaggregation, or architecture papers.

Story Pattern

The paper type is a systems/database contribution: start from a concrete scalability or operational bottleneck, show why existing architectures are constrained by that bottleneck, then introduce a design that changes the system boundary. The story should make the workload, deployment assumption, and architecture tradeoff visible before implementation details.

Abstract Moves

• Name the real system bottleneck before naming the system.
• State the architecture change as a design principle, not only a component list.
• Mention the workload or deployment setting that makes the problem matter.
• Summarize evidence as throughput, latency, scalability, isolation, fault tolerance, or operational cost, using only user-provided numbers.

Introduction Moves

• Move from cloud workload pressure to a precise database-system limitation.
• Explain why simpler fixes such as caching, sharding, or tuning are insufficient.
• Introduce the key architectural insight early.
• State contributions as design, implementation, and evaluation claims.

Method/System Moves

• Use a system diagram with data/control paths, consistency boundaries, and failure assumptions.
• Separate design goals from mechanisms.
• Explain concurrency, recovery, and resource management only after the architecture is clear.

Evidence Moves

• Use realistic workloads and scale studies.
• Include end-to-end evaluation plus targeted microbenchmarks.
• Report operational costs or overheads when relevant.
• Treat missing deployment detail as a reviewer risk, not as prose polish.

Do-Not-Copy Boundary

Do not borrow product-specific claims, architecture details, or performance language. Use only the systems-writing pattern.

STOC/FOCS Theory Best-Paper Style Card

Venue/year: STOC/FOCS theory family. Source: SIGACT Best Paper Prize source record sigact-best-paper; verify the exact paper page before public award-status claims. Use when: theory papers, algorithms, complexity, lower/upper bounds, proof techniques, formal models, or papers resolving a technical barrier.

Story Pattern

The paper type is theorem-first research: state the model and result precisely, explain the barrier or prior gap, then give proof intuition before formal proof details. The introduction should help readers understand why the result is surprising, not only what theorem is proven.

Abstract Moves

• Name the formal problem and model.
• State the main result in precise terms.
• Relate the result to a known barrier, conjecture, bound, or open question.
• Avoid broad impact claims unless the formal implication is direct.

Introduction Moves

• Define the problem in a way that non-specialist theory reviewers can orient to.
• Summarize prior upper/lower bounds or techniques as a small map.
• Identify the exact missing regime or technical barrier.
• Give the proof idea before theorem machinery.

Method/Proof Moves

• Separate theorem statements, intuition, lemmas, and proof details.
• Use examples or diagrams only to illuminate the formal mechanism.
• Make assumptions and model limits explicit.

Evidence Moves

• For theory, evidence is proof correctness, tightness, examples, and relation to known bounds.
• If experiments are included, keep them secondary unless the paper is hybrid.
• Do not use empirical language to overstate a purely formal result.

Do-Not-Copy Boundary

Do not reuse theorem statements, examples, or proof language. Transfer only the formal-story structure.

Transformer Writing Style

Source: Vaswani et al., "Attention Is All You Need," NeurIPS2017. Paper PDF: references/exemplars/papers/1706.03762.pdf

Use when: writing architecture papers, backbone-replacement papers, or papers whose main contribution is changing the fundamental modeling primitive rather than adding a module to an existing pipeline.

Actual Section Lengths (from the published paper)

Total: ~8.7 pages main text,2.3 pages references + appendix.

Story Pattern

The paper starts from a widely understood problem (sequence transduction), identifies a structural bottleneck in existing solutions (sequential computation), introduces a clean conceptual shift (attention as backbone), supports it with both theoretical analysis (path length) and empirical evidence (WMT BLEU), and concludes with bounded claims.

Section-by-Section Writing Moves

Abstract

• Sentence1: Problem and importance.
• Sentence2: Gap in existing methods.
• Sentence3: Core insight and method name.
• Sentence4: Architecture components.
• Sentence5: Headline evidence (numbers).

Introduction

• Paragraph1: Task importance and broad motivation.
• Paragraph2: Existing methods and their limitations (structural, not just empirical).
• Paragraph3: Root cause of the limitation — why existing methods cannot address it.
• Paragraph4: Insight and proposed method (named here).
• Paragraph5: Architecture summary (one sentence per component).
• Paragraph6: Evidence preview.
• Paragraph7: Contributions list (numbered).
• Paragraph8 (optional): Paper structure.

Method

• Each subsection: motivation first, then equations, then explanation.
• Motivation answers: "Why is this component needed, given the gap described in the Introduction?"
• Equations are clean and self-contained.
• Hyperparameters are in a table, not scattered in text.

Experiments

• Setup: dataset sizes, preprocessing, hardware.
• Main results: a clean table with bold best results.
• Complexity analysis: a theoretical table that supports the architectural claim.
• Ablations: each row tests one design choice against a specific claim.
• Training efficiency: wall-clock time to support the parallelism claim.

Related Work

• Organized by paradigm: recurrent, convolutional, attention-based, efficient methods.
• Each paragraph: paradigm name, key papers, limitation relative to this work, how this work differs.
• Ends with the closest gap.

Reusable Techniques

• Make the structural bottleneck explicit (not just "others do worse").
• Name the method early in the Introduction.
• Every method component has a clear "why" before the "how".
• Pair analytical evidence (complexity table) with empirical evidence (BLEU table).
• Keep ablation rows directly tied to design choices listed in Method.
• Use bold for best results — one convention throughout.

Do-Not-Copy Boundary

Do not reuse the specific architecture, equations, motivation framing, or reported numbers unless the user's paper genuinely supports them. This card provides writing moves and structural patterns, not content to copy.

InSpectre Gadget Style Card

Venue/year: USENIX Security 2024 family. Source: USENIX Security 2024 technical sessions source record usenix-security-2024-technical-sessions; verify the exact paper page before public award-status claims. Use when: security papers involving attacks, defenses, vulnerability discovery, side channels, exploit chains, formal threat models, or security measurement.

Story Pattern

The paper type is security research: define the threat model and attacker capability before the technique, then show why existing defenses or assumptions fail. Evidence should connect exploitability, coverage, false positives/negatives, practicality, and responsible scope.

Abstract Moves

• Name the security setting and threat model.
• State the core failure in existing assumptions or defenses.
• Describe the method at the level of attack/defense mechanism.
• Preview evidence without overclaiming universal vulnerability or protection.

Introduction Moves

• Start from a concrete security risk or defense gap.
• Make attacker goals, capabilities, and constraints explicit.
• Explain why the problem remained hidden or hard to detect.
• State contributions as threat model, method, evaluation, and responsible disclosure or mitigation when relevant.

Method Moves

• Separate threat model, technique, analysis pipeline, and mitigation.
• Use examples only to clarify the mechanism; avoid implying broader scope than tested.
• State assumptions before results.

Evidence Moves

• Include adaptive baselines or bypass attempts when relevant.
• Report success rate, coverage, false positives/negatives, overhead, or affected systems only from supplied results.
• Include ethics, disclosure, and harm-reduction boundaries.

Do-Not-Copy Boundary

Do not borrow vulnerability details, exploit wording, or scope claims. Use only the security-review structure.

VGGT

Venue/year: CVPR 2025. Venue family: CVPR. Custom status: user-custom writing-format exemplar. This card participates in the default custom writing format when the user does not specify a target venue. VGGT is also recorded separately as the CVPR 2025 Best Paper in cards/cvpr-2025-vggt-best-paper.md, but this card remains the user-custom writing source. Source: paper_ref/VGGT_ Visual Geometry Grounded Transformer.pdf; CVF https://openaccess.thecvf.com/content/CVPR2025/html/Wang_VGGT_Visual_Geometry_Grounded_Transformer_CVPR_2025_paper.html. Use when: writing feed-forward vision, 3D reconstruction, multi-task prediction, model simplicity, or replacement of iterative pipelines.

Story Pattern

The paper begins from a classic pipeline with strong priors and heavy optimization, then asks whether modern neural networks can directly solve the same family of tasks. The contribution is framed as a simpler feed-forward model that predicts many related 3D attributes together and competes with specialized or optimization-based systems.

Abstract Moves

• Define the model by outputs first: cameras, point maps, depth, and tracks.
• Make scale flexibility explicit: one view, few views, or hundreds of views.
• Contrast simplicity and speed against post-processing-heavy alternatives.
• Close by extending the value beyond the main tasks: pretrained features help downstream tasks.

Introduction Moves

• Start from the historical baseline and why it still dominates.
• Pose a sharp capability question rather than only listing limitations.
• Position related work as partial progress, then identify what still requires post-processing.
• Argue that standard transformers plus enough 3D-annotated data may be sufficient.
• Summarize contributions around model, direct usability, post-processing compatibility, and benchmark breadth.

Method Moves

• State problem inputs and outputs before architecture details.
• Use over-complete prediction as a training/evidence idea: related outputs can improve each other even if some are mathematically connected.
• Explain architectural restraint as a virtue, then isolate the few necessary design choices.

Evidence Moves

• Benchmark across several tasks to prove generality.
• Compare both direct outputs and outputs with optional refinement.
• Include in-the-wild qualitative figures that reveal where optimization methods are brittle.

Reusable Techniques

• Use "direct prediction of a family of outputs" as a unifying paper story.
• Make simplicity reviewer-facing: fewer special components can be a claim if evidence is broad.
• Let one overview figure show input/output breadth and runtime contrast.

User Notes

This is one of the user's recognized exemplars. Use it for papers that want to claim a simple model can replace a complicated classical pipeline.

Do-Not-Copy Boundary

Do not borrow the exact visual-geometry task list, transformer framing, or speed comparison unless the user's method and experiments support them.

Self-Test: Dynamic Scene Memory Transformer

Purpose: test whether the exemplar cards help plan, draft, review, and revise a simple CCF A-style paper idea without copying exemplar prose.

Target paper idea: Dynamic Scene Memory Transformer for 4D embodied navigation. The hypothetical method stores object-centric spatiotemporal memories from multi-view egocentric video and uses them for long-horizon navigation queries.

Target venue family: ICLR/CVPR-style AI and vision.

Loaded references:

• references/exemplars/index.md
• references/exemplars/cards/llava-4d.md
• references/exemplars/cards/vggt.md
• references/exemplars/cards/neurips-2025-1000-layer-ssl-rl.md
• references/storyline-blueprint.md
• references/expert-review-loop.md

Draft Pass 1

Global Story

Task: answer navigation and interaction queries in dynamic 4D indoor scenes.

Gap: current embodied agents remember static landmarks better than moving objects, so temporal changes are lost across long-horizon reasoning.

Root challenge: moving objects create discontinuous observations; a model must bind object identity, 3D position, and time without rebuilding the entire scene at every step.

Insight: object-centric spatiotemporal memory can preserve dynamic evidence more efficiently than frame-level history.

Method: a transformer that writes object memory tokens with 3D pose, time, motion, and language-aligned attributes, then retrieves them for navigation decisions.

Evidence promise: benchmark on dynamic-object navigation, ablations for memory write/read, comparison against frame history and static maps, qualitative temporal query examples.

Simple Abstract Draft

Embodied agents increasingly operate in dynamic indoor scenes, yet most visual-language navigation systems store scenes as static maps or frame histories. This makes them brittle when objects move, disappear, or reappear across long trajectories. We propose Dynamic Scene Memory Transformer, an object-centric memory model that writes spatiotemporal tokens from egocentric multi-view video and retrieves them for language-conditioned navigation. The model binds each memory token to 3D position, time, motion state, and semantic attributes, enabling agents to reason over both stable background structure and dynamic objects. We introduce a dynamic navigation benchmark with temporal object queries and evaluate against static-map, frame-history, and 3D-LMM baselines. Across navigation success, temporal grounding, and object relocation tasks, the proposed memory improves dynamic-scene reasoning while reducing stored visual tokens.

Expert Review Pass 1

Score: 5/10, weak reject.

Findings:

• Novelty is plausible but underspecified. "Object-centric memory" could overlap with existing memory-map and tracking work.
• The benchmark promise is too broad for a simple article. The draft needs a clearer evidence package and task boundary.
• The abstract claims reduced storage but does not say how storage is measured.
• It borrows the 4D dynamic-scene rhythm successfully, but it needs the VGGT-style output definition: what exactly is predicted or retrieved?

Revision Actions:

• Narrow the task to language-conditioned relocation and navigation in dynamic rooms.
• Define outputs: answer, waypoint, and supporting object-time trace.
• Add storage metric and ablation requirement.
• Replace broad "3D-LMM baselines" with precise baseline families.

Draft Pass 2

Revised Abstract

Language-conditioned navigation in dynamic rooms requires more than recognizing the current view: an agent must remember where objects were, when they moved, and which past observation supports a decision. Existing static-map and frame-history agents either discard object motion or store redundant visual context. We propose Dynamic Scene Memory Transformer, a retrieval model that writes object-time memory tokens from egocentric video and predicts three outputs for each query: a navigation waypoint, a natural-language answer, and an evidence trace over remembered objects. The model stores 3D position, timestamp, motion cue, and semantic state for each object token, then retrieves a compact temporal support set before action prediction. We evaluate on a language-conditioned relocation benchmark with moving-object goals, comparing against static semantic maps, recurrent frame memory, and video-LMM retrieval baselines. The proposed memory improves temporal grounding and navigation success while using fewer retained tokens; ablations isolate the contributions of time encoding, motion-aware writing, and evidence-trace supervision.

Section Plan

Introduction:

1. Dynamic navigation need: current view alone is insufficient. 2. Prior map/frame-memory limitation: static maps lose motion, frame histories are redundant. 3. Root challenge: bind object identity, 3D location, and time for query-time evidence. 4. Method preview: object-time memory writes, compact retrieval, waypoint/answer/evidence outputs. 5. Contributions: task benchmark, memory model, evidence-trace supervision, evaluation package.

Experiments:

1. Main comparison on relocation navigation. 2. Temporal grounding and evidence-trace accuracy. 3. Storage-efficiency curve. 4. Ablations: no time, no motion cue, no evidence supervision, full frame memory. 5. Failure analysis for occlusion and identity switches.

Expert Review Pass 2

Score: 7/10, weak accept if experiments are real.

Remaining risks:

• Must define benchmark construction and annotation quality.
• Needs a strong baseline suite from embodied AI and video retrieval.
• Storage reduction claim must report both token count and accuracy trade-off.

Skill Outcome:

• The exemplar index helped choose a small card bundle instead of overloading context.
• The LLaVA-4D card helped shape the dynamic-scene gap and observation-to-module move.
• The VGGT card exposed the need to define outputs early.
• The review loop caught overclaiming before final prose.
• No exemplar wording or technical claim was copied.

Exemplar Index

Use this index when the user asks for CCF A-level paper writing, best-paper style, venue-specific adaptation, default user-custom writing, or examples from strong papers. Load only the cards that match the target paper. Do not load every card by default.

Default Custom Format

When the target venue is not specified, load references/custom-format/default-user-format.md first. That format currently uses the user's two custom exemplar cards:

These two cards are user-custom writing-format sources. Do not treat them as ordinary venue best-paper cards unless the user explicitly asks to compare against ICLR/CVPR best-paper style.

Selection Rule

Pick at most 2-4 cards:

• Use the custom-format cards first when no target venue is specified.
• Use same venue or venue family first when a target venue is specified.
• Use same evidence type second: theorem, benchmark, user study, system, dataset, or ablation-heavy model.
• Use same story shape third: new task, new benchmark, new model family, new capability, or new evaluation economy.
• Add one contrast card only when it improves reviewer-proofing.

Use cards to borrow writing moves, not claims, wordings, examples, or technical content.

ICLR And CVPR Recent Best-Paper Cards

Use these cards when the user explicitly asks for ICLR/CVPR best-paper or outstanding-paper style.

Other Venue Cards

Non-AI/CV/NLP Venue Cards

Use these cards when the target venue is outside the dominant AI/CV/NLP cluster. They are meant to prevent the writing skill from forcing every paper into an ML-style "model plus table" story.

Recommended Bundles

• Default unspecified-venue paper: load references/custom-format/default-user-format.md; it will select llava-4d.md and vggt.md.
• 4D embodied AI paper: llava-4d.md, vggt.md, neurips-2025-1000-layer-ssl-rl.md.
• ICLR-style theory or LLM paper: one ICLR recent best-paper card plus one same-topic card.
• CVPR-style vision paper: one CVPR recent best-paper card plus vggt.md if the paper involves 3D geometry or multi-task prediction.
• LLM collaboration or agent paper: icml-2025-collabllm.md, acl-2025-minilongbench.md.
• Generative multimodal system: iccv-2025-brickgpt.md, acmmm-2024-speaker-to-dubber.md, icml-2024-videopoet.md, neurips-2024-var.md.
• Data-efficient or cost-efficient evaluation: acl-2025-minilongbench.md, aaai-2025-every-bit-helps.md.
• Hardware or system paper: eccv-2024-minimalist-vision.md, vggt.md.
• Open-vocabulary robotics paper: acmmm-2025-aff3dfunc.md, iccv-2025-brickgpt.md.
• Database system paper: sigmod-2024-polardb-mp.md plus one DB/IR/KDD source found through ccf-literature-searcher.
• Network/system paper: sigcomm-2024-l25gc-plus.md plus a venue-specific baseline/evaluation source.
• Security paper: usenix-security-2024-inspectre-gadget.md plus a threat-model or artifact-evaluation source.
• HCI paper: chi-2024-constrained-highlighting.md plus the user's target population/study-method reference.
• PL/theory paper: popl-2024-megalibm.md or stoc-2024-theory-best-paper.md depending on whether the contribution is tool/proof or theorem/barrier.

Style And Citation Guides

When loading exemplar cards, also follow the writing style and citation rules established in the main skill references:

• references/research-writing-patterns.md (Natural Writing Style section): prose flow, forbidden patterns (no bold labels, no citation dumps), citation weaving.
• references/citation-workflow.md: search-bib-cite workflow, citation density per section, natural citation patterns.
• references/output-style-policy.md (Citation Format Rules): claim-first citation style, no author-name subjects, bib file as source of truth.

The exemplar cards provide structural patterns. The style guides provide execution rules. Use both.

Output Reminder

After loading cards, produce:

1. Chosen exemplar set and why each card fits. 2. Transferable writing moves. 3. Drafting warnings where an exemplar's confidence would not be supported by the user's evidence. 4. A section plan or revision that is original to the user's paper.