SurveyGen-I: End-to-End Survey Generation

• SurveyGen-I is an automated academic survey generation framework that integrates dynamic outline planning, memory-guided writing, and citation tracing.
• It employs a multi-stage pipeline combining literature retrieval, adaptive plan evolution (PlanEvo), and citation-aware writing (CaM-Writing) to boost content consistency.
• Empirical evaluations show that SurveyGen-I outperforms prior systems in content quality and citation coverage, demonstrating significant improvements over baselines.

SurveyGen-I is an end-to-end framework for automatic academic survey generation that is designed to produce multi-section scientific surveys that are topically broad and relevant, coherent across sections and subsections, adaptive in structure as writing unfolds, and well grounded in citations. Its architecture combines multi-stage coarse-to-fine literature retrieval, a planning framework called PlanEvo, and a citation-aware memory-guided writing pipeline called CaM-Writing. The framework is presented as a response to two persistent weaknesses in earlier LLM-based survey-generation systems: long-range discourse inconsistency and incomplete citation coverage. In the reported experiments, SurveyGen-I is described as consistently outperforming prior systems in content quality, consistency, and citation coverage (Chen et al., 20 Aug 2025).

1. Position within automatic survey generation

SurveyGen-I belongs to a line of work that treats survey writing as a structured scholarly composition task rather than a one-shot long-text prompt. Earlier systems had already established several important design directions. SurveyForge emphasizes heuristic outline generation and memory-driven content generation, arguing that automated survey writing requires both structural priors from human-written surveys and section-aware retrieval (Yan et al., 6 Mar 2025). SurveyX decomposes survey writing into preparation and generation phases, with structured preprocessing through AttributeTree and paragraph-level RAG-based rewriting for citation quality (Liang et al., 20 Feb 2025). SciSage uses a multi-agent “reflect-when-you-write” paradigm to critique outlines, sections, and documents during generation (Shi et al., 15 Jun 2025). SurveyG introduces a hierarchical citation graph to encode methodological relations and research evolution before writing (Nguye et al., 9 Oct 2025). Against this background, SurveyGen-I is distinctive for coupling adaptive plan evolution with a rolling memory of previously written content and terminology, and for integrating citation tracing directly into subsection generation (Chen et al., 20 Aug 2025).

The framework is motivated by several limitations in prior systems. The paper identifies limited retrieval scope, especially when systems rely on embedding-based similarity search over a fixed local paper database; lack of cross-subsection consistency when subsections are generated in parallel as isolated units; static outlines that cannot react to coverage gaps or redundancy discovered during writing; weak modeling of long-range dependencies between subsections; and incomplete treatment of indirect citations embedded inside retrieved passages. SurveyGen-I therefore treats survey generation as a staged pipeline in which retrieval, planning, writing, and revision are mutually coupled rather than serially isolated (Chen et al., 20 Aug 2025).

A common misconception is to regard SurveyGen-I as a standard retrieval-augmented generation system with a slightly larger context window. The paper argues otherwise. Its central claim is that high-quality survey generation requires an explicit global outline, a dependency-aware writing order, memory mechanisms for terminology and discourse continuity, local retrieval when subsection-specific evidence is needed, and citation expansion beyond directly retrieved passages. This suggests that the framework’s novelty lies less in base-model choice than in control flow and information organization (Chen et al., 20 Aug 2025).

2. System architecture and execution flow

SurveyGen-I is organized into three main stages: LR: Literature Retrieval, PlanEvo: Structure Planning with Dynamic Outline Evolution, and CaM-Writing: Citation-Aware Subsection Writing with Memory Guidance. The system takes a research topic $T$ and description $E$ as input and returns a final survey draft after staged generation and global refinement (Chen et al., 20 Aug 2025).

The algorithmic skeleton begins with retrieval:

• $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$
• $$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$
• $$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$

Planning then produces an outline and dependency-aware execution schedule:

• $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$
• $$\mathcal{P}_{\text{dep}} \gets \text{BuildSchedule}(\mathcal{O})$$
• $\mathcal{M} \gets \emptyset$

Writing proceeds stage by stage. For each stage index $t$, the controller selects all subsections whose scheduled depth is $\tau(s_i)=t$, optionally performs subsection-level retrieval, retrieves evidence with RAG, enriches it through citation tracing, constructs a writing skeleton, generates multiple candidates, selects the best, refines it, and then updates memory:

• $E$0
• $E$1
• $E$2
• $E$3
• $E$4
• $E$5
• $E$6
• $E$7

After each stage, the Memory-Guided Structure Replanner updates the remaining outline and schedule:

• $E$8

A final global refinement pass then rewrites the full draft for consistency (Chen et al., 20 Aug 2025).

The architecture is operationally divided among three named components inside PlanEvo. SDP builds the initial outline and writing plan. SAWC orchestrates stage-wise execution, optional subsection retrieval, memory updates, table generation, and final consistency refinement. MGSR revises the unwritten outline after each writing stage, performing actions such as merge, delete, rename, reorder, and add when redundancy or gaps are detected (Chen et al., 20 Aug 2025).

3. Literature retrieval and citation enrichment

The retrieval stage begins with survey-level search over Semantic Scholar. An LLM first produces a keyword set $E$9 from topic $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$0 and description $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$1, yielding an initial candidate set $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$2. Semantic filtering is then defined as

$$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$3

where $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$4 is the embedding of the topic-description pair and $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$5 is the embedding of abstract $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$6. In implementation, the embedding model is all-mpnet-base-v2 and the threshold is 0.3. Citation-based expansion is then applied to the top 10 papers by collecting both references and citations, followed by another semantic filter. An LLM finally assigns relevance scores to the expanded candidates; only papers with score at least 70 are retained, with a fallback to the top 5 if none pass, and at most 30 papers are kept per query (Chen et al., 20 Aug 2025).

Subsection-level retrieval is optional rather than universal. For each subsection $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$7 with description $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$8, the writing plan contains a flag $$\mathcal{P}_{\text{init}} \gets \text{KeywordSearch}(T,E)$$9 indicating whether additional local retrieval is necessary. If $$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$0, the system retrieves subsection-specific evidence and unions it with the survey-level set:

$$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$1

Passage-level RAG retrieval then operates over $$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$2, using a BGE-based embedding model followed by BGE-Reranker, and the top 10 reranked passages form $$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$3 (Chen et al., 20 Aug 2025).

A defining feature of SurveyGen-I is citation tracing. Retrieved passages are scanned for citation markers such as numeric references like [17] and author-year mentions like (Ge et al., 2023). An LLM then judges whether each marker is traceworthy. Traceworthy citations are resolved via Semantic Scholar, and their abstracts are appended to the writing context, creating $$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$4. The paper’s illustrative example traces (Ge et al., 2023) from a retrieved passage to the source paper Preserve Your Own Correlation: A New Reward-Balanced Fine-Tuning Method, whose abstract is then added back into the local evidence context. This procedure is meant to recover original contributions that would otherwise remain only indirectly represented in retrieved passages (Chen et al., 20 Aug 2025).

4. PlanEvo and dynamic structure evolution

PlanEvo formalizes both the survey outline and the execution order used to write it. The initial outline is represented as

$$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$5

where $$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$6 is a subsection heading and $$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$7 is its description. To generate this outline, SurveyGen-I extracts structural outlines from review papers inside $$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$8 and combines them with titles and abstracts of non-review papers, producing a composite context for LLM-based outline construction and refinement (Chen et al., 20 Aug 2025).

The initial writing plan is expressed as

$$\mathcal{P}_{\text{sem}} \gets \text{SemanticFilter}(\mathcal{P}_{\text{init}},T,E)$$9

where $$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$0 is the subsection description, $$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$1 is the retrieval trigger, and $$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$2 indicates whether a comparative table should be generated. The system then builds a structural dependency graph $$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$3 over subsections, removes cycles to obtain a DAG, and assigns a stage index through

$$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$4

The final dependency-aware plan is

$$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$5

Subsections with the same $$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$6 can be written in parallel, while prerequisite relations preserve logical order (Chen et al., 20 Aug 2025).

The paper’s most distinctive planning mechanism is MGSR, which revises the remaining outline after each writing stage. Using accumulated memory $$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$7, the current outline, and the yet-unwritten subsections, it asks an LLM to detect redundancy, conceptual gaps, and ordering problems, then returns revision actions such as merge, delete, rename, reorder, and add. This makes the outline explicitly non-static. A plausible implication is that SurveyGen-I treats survey planning as a dynamic control problem rather than a one-shot prompt artifact, although the paper does not formalize this as an optimization objective (Chen et al., 20 Aug 2025).

The memory itself is operational rather than mathematically elaborate. After each subsection is written, SurveyGen-I extracts draft content and key domain-specific terminology and stores them in a dynamic structure memory:

$$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$8

This memory conditions both later writing skeletons and later plan updates. The paper explicitly states that memory stores “writing content and terminology,” but does not define a hierarchical key-value schema or a forgetting mechanism. It is therefore best understood as a rolling cumulative memory used for coherence rather than as a formally structured external database (Chen et al., 20 Aug 2025).

5. CaM-Writing, tables, and refinement

CaM-Writing operates at subsection level. Given subsection title $$\mathcal{P}^* \gets \text{ExpandAndRank}(\mathcal{P}_{\text{sem}},T,E)$$9, description $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$0, enriched evidence context $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$1, and memory $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$2, the system first constructs a memory-aligned writing skeleton $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$3. It then generates multiple candidate drafts, selects the best candidate with an LLM according to alignment with the skeleton, contextual relevance, and overall writing quality, and refines that candidate through three passes: structure adjustment, citation refinement, and polishing (Chen et al., 20 Aug 2025).

This subsection-writing regime is closely tied to the scheduler SAWC. For each stage,

$$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$4

and all subsections in $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$5 can be written in parallel. After the stage finishes, memory is updated, MGSR replans the remainder, and the system moves to the next stage. The paper presents this as a way to preserve efficiency without abandoning dependency-aware global structure (Chen et al., 20 Aug 2025).

SurveyGen-I also supports subsection-level table generation when the planning flag $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$6 is active. Two modes are described. A method aggregation table is used when there are at least about 10 filtered non-review papers; the system infers a core comparative aspect, proposes 4–6 categories, assigns papers to categories using RAG, and excludes sparse categories. An aspect-based comparison table is used when there are fewer than ten papers; following Newman et al. (2024), it selects 3–5 aspects, retrieves evidence for each paper–aspect pair with FAISS plus reranking, and summarizes the snippets into table cells. These tables are generated after writing rather than before it (Chen et al., 20 Aug 2025).

After all subsections are completed, SAWC performs a final survey-level consistency pass. An LLM reviews the full draft for logical contradictions, redundancy, terminology inconsistencies, and style inconsistencies, then rewrites affected subsections. The appendix also shows example outputs such as a generated front page, a citation alignment table, a subsection with tables, figures, and equations, an automatically synthesized glossary, and a full example survey, though these examples are presented as artifacts rather than as fully formalized subsystems (Chen et al., 20 Aug 2025).

6. Evaluation, empirical performance, and limitations

SurveyGen-I is evaluated against AutoSurvey, SurveyForge, and SurveyX. The paper states that demo reports were collected from the official project pages for SurveyForge and SurveyX, while AutoSurvey was reproduced using the same generation model as SurveyGen-I, namely GPT-4o-mini, for fairness. The evaluation setup is described inconsistently: the abstract says experiments are across four scientific domains, whereas the main body and appendix refer to six major scientific domains or subtopic groups. The appendix’s detailed grouping lists six: AI Applications and Multidisciplinary Topics, Database and Big Data, LLMs, Learning Algorithms and Foundations, Network, Systems, Infrastructure, and Vision, Video, and Image Generation. The paper itself does not resolve this discrepancy (Chen et al., 20 Aug 2025).

Content quality is scored by LLM-as-Judge across five dimensions: Coverage, Relevance, Structure, Synthesis, and Consistency, whose average is the Content Quality Score (CQS). Reported overall scores are:

• AutoSurvey: 4.08
• SurveyX: 4.13
• SurveyForge: 4.23
• SurveyGen-I: 4.59

SurveyGen-I’s per-dimension scores are 4.72 for Coverage, 4.76 for Relevance, 4.28 for Structure, 4.62 for Synthesis, and 4.59 for Consistency. The introduction summarizes this as an 8.5% improvement in content quality over the strongest baseline (Chen et al., 20 Aug 2025).

Reference quality is measured through NR (number of distinct cited works), CD (citation density, scaled by $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$7), and RR@k (recency ratio within the past $$\mathcal{O} \gets \text{GenerateOutline}(\mathcal{P}^*,T,E)$$8 years). SurveyGen-I reports:

• RR@1 = 0.478
• RR@3 = 0.759
• RR@5 = 0.891
• RR@7 = 0.955
• RR@10 = 0.985
• CD = 17.28
• NR = 281

By comparison, the strongest baseline on citation density is SurveyX with 13.57, and the strongest baseline on distinct references is SurveyForge with 113. The paper therefore states that SurveyGen-I produces more than twice as many distinct references and a 27% increase in citation density over the strongest baseline. It also reports first-place performance on RR@1 across all six appendix subtopic groups (Chen et al., 20 Aug 2025).

The ablation results are especially revealing. Removing citation tracing lowers NR from 286 to 225 and reduces overall quality from 4.77 to 4.60. Removing plan update lowers overall quality to 4.49 and reduces Structure, Synthesis, and Consistency. Removing final refinement yields the largest quality drop, from 4.77 to 4.34, with especially large degradation in Structure. However, the paper does not provide an isolated ablation for memory alone, nor an isolated ablation for subsection-level retrieval alone. Those effects are therefore not separately quantified (Chen et al., 20 Aug 2025).

The framework’s limitations are stated candidly. It depends on online retrieval, which introduces network sensitivity, variable latency, third-party API dependence, and possible full-text access limitations. Survey quality is also constrained by source availability for niche topics. The paper notes that some evaluation signals may reflect subjective preferences rather than universal standards of survey quality. Most importantly, the appendix explicitly warns that generated content is not intended for direct submission as scholarly writing and that factual correctness and citation fidelity are not guaranteed without human verification. This places SurveyGen-I most naturally in the category of research-assistive survey drafting systems rather than autonomous scholarly authorship (Chen et al., 20 Aug 2025).