grad-paragraph（毕业段落）

Grad Paragraph (survey paragraph micro-skill)

Purpose: produce a single paragraph that reads like real survey prose, not “outline expansion”.

This is a writing micro-skill you can apply repeatedly inside subsection-writer (per H3 file under sections/).

Role cards (use explicitly)

Argument Planner

Mission: decide the paragraph’s tension/contrast/eval/limitation before writing.

Do:

• Write a 4-line plan (kept out of final prose).
• Ensure planned sentences can be grounded in evidence/citations.

Avoid:

• Writing from headings or generic axis labels.

Paragraph Author

Mission: turn the plan into one content-bearing paragraph with embedded citations.

Do:

• Use explicit contrast markers (whereas/in contrast).
• Include at least one evaluation anchor token (task/metric/constraint).
• End with a limitation that changes interpretation.

Avoid:

• Narration and repeated template stems.

Role prompt: Paragraph Author (argument move)

You are writing one paragraph of a technical survey.

Your job is to perform one argument move under evidence:
- tension/question (why this matters here)
- explicit contrast (A vs B; not a list)
- evaluation anchor (task/metric/constraint)
- limitation (what breaks transfer or comparability)

Style:
- natural prose, content-bearing
- no narration (“This paragraph surveys…”)
- no repeated discourse stems across paragraphs

Constraints:
- do not invent facts or citations
- embed citations inside the sentence that needs them
- stay within the subsection’s citation scope

What this paragraph must contain

In one paragraph (typically 4–6 sentences), cover:

• Tension / question: what this paragraph is trying to resolve (subsection-specific).
• Contrast: compare at least two approaches/routes/clusters (A vs B) with explicit contrast words.
• Evaluation anchor: name how comparisons are made (benchmark/dataset/metric/protocol), even if only abstract-level.
• Limitation / verification: state what is uncertain (missing protocol details, incomparable benchmarks, unclear constraints) without turning into boilerplate.
• If you include a number, also include: task type + metric definition + constraint (budget/cost/tool access), and cite it.

Inputs (practical)

• outline/subsection_briefs.jsonl (for rq, axes, clusters, paragraph_plan)
• outline/evidence_drafts.jsonl (for evidence snippets + candidate comparisons)
• outline/evidence_bindings.jsonl (allowed citations for this H3)
• citations/ref.bib

Outputs

• One paragraph (4–6 sentences) to paste into the target sections/S<sub_id>.md file.
• Optional (when debugging): a 4-line plan (tension/contrast/eval/limitation) kept out of the final prose.

Roles (two-pass is more reliable)

Role A: Argument Planner

Write a 4-line plan before prose:

1. Tension sentence (1 line)
2. Contrast sentence (1 line; A vs B)
3. Evaluation anchor sentence (1 line)
4. Limitation sentence (1 line)

Rules:

• Each line should be anchored by at least one citation key you intend to use.
• If evidence is abstract-only, avoid “dominant / clearly / state-of-the-art” style conclusions.

Role B: Writer

Turn the plan into one natural paragraph.

Rules:

• Keep the paragraph subsection-specific (it should not be copy-pastable into other subsections).
• Place citations inside the sentence they support (not only at paragraph end).
• Do not mention pipeline internals (“working claim”, “axes we track”, “verification targets”).

Paper voice (avoid template cadence)

• Keep tone calm and academic; avoid hype words (e.g., “clearly”, “obviously”, “breakthrough”).
• Vary sentence openings; don’t start every paragraph with the same connector (“However/Moreover/Taken together”).
• Avoid explicit labels like Key takeaway:; let the sentence carry the point.
• Prefer concrete nouns + mid-sentence ties (...; however,...) over “PPT narration” signposting.

Examples (what to write / what to avoid)

Bad (template narration + vague claims + cite dump):

This subsection surveys how agents use memory. Taken together, these approaches improve performance across tasks [@example2023; @example2024; @example2025].

Why it is bad:

• Starts with narration ("This subsection...").
• No explicit A-vs-B contrast (reads like a topic list).
• No evaluation anchor (benchmark/metric/protocol is missing).
• Citations are only used as a trailing tag list.

Good (tension -> contrast -> eval anchor -> limitation; citations embedded):

Plan (kept out of final prose):

1. Tension: Memory increases capability but makes evaluation and reproducibility harder.
2. Contrast: Retrieval-style memory [@example2023] differs from write-heavy episodic memory [@example2024] in what gets stored and when it can be trusted.
3. Eval anchor: Results are typically reported on agent benchmarks with success-rate style metrics under tool/budget constraints (state the specific benchmark/metric when available).
4. Limitation: Comparisons remain fragile when protocols differ or when memory writes are not logged, so some gains may not transfer.

Paragraph (final prose):

A recurring tension in agent memory is that richer state can expand what the system can do, yet it also complicates evaluation and reproducibility. Retrieval-style designs emphasize selecting and grounding a small working set of relevant context [@example2023], whereas write-heavy episodic approaches accumulate longer-term traces that can change the agent behavior across episodes [@example2024]. These choices often surface in benchmarked evaluations as different failure patterns under fixed tool and budget constraints (e.g., higher success at the cost of more brittle behavior when memory writes are noisy). At the same time, cross-paper comparisons remain limited when protocols are not aligned or when memory writes are not transparently logged, making it unclear which gains reflect memory design versus evaluation artifacts.

Checklist (quick self-audit)

• No ... / … / TODO / scaffold phrases.
• Contains at least one explicit contrast marker: whereas, however, in contrast, 相比, 不同于, 相较.
• Contains at least one evaluation anchor token: benchmark, dataset, metric, protocol, evaluation, 评测, 基准, 数据集, 指标.
• Contains at least one limitation/provisional token: limited, unclear, sensitive, may, 缺乏, 受限, 尚不明确, 需要核验.
• Citations are real ([@BibKey]) and subsection-scoped (in outline/evidence_bindings.jsonl).

Troubleshooting

Issue: the paragraph still reads like a template

Symptom:

• Repeated framing (“Taken together…”, “A useful way to compare…”) across many paragraphs.

Causes:

• You are writing from outline bullets instead of evidence snippets.

Solutions:

• Rebuild the plan from Concrete comparisons / Failure/limitations in the evidence pack.
• Force one concrete noun per sentence (task/setting/constraint/evaluation artifact), even if you can’t use numbers.

Issue: you can’t write a contrast without guessing

Symptom:

• You only have titles, so you drift into vague statements (“depends on metrics”).

Causes:

• Evidence granularity is too low.

Solutions:

• Push upstream: strengthen papers/paper_notes.jsonl (abstract/fulltext) and rerun evidence-draft.
• If you must proceed, write the paragraph as a question + verification targets, not as a conclusion.