Skip to content

Story2Proposal: A Scaffold for Structured Scientific Paper Writing

🕒 Published (v1): 2026-03-28 00:46 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Story2Proposal is a contract-governed multi-agent framework that converts a research story (narrative + evidence) into a structured scientific manuscript by coordinating four specialized agents (architect, writer, refiner, renderer) through a persistent "shared visual contract." A generate–evaluate–adapt loop allows evaluation agents to update the contract state mid-generation, preventing structural drift and visual misalignment. On the Jericho corpus, it outperforms both unstructured (DirectChat) and structured (FARS) baselines across four LLM backbones.

Problem

Current automated paper-generation pipelines—whether direct prompting, single-agent long-form generation, or outline-based approaches—apply validation only after generation or lack a persistent semantic layer. This produces section drift (later sections diverge from the initial framing), claim-experiment misalignment, citation-argument disconnection, and figure/table placement errors. No existing system maintains bidirectional provenance between intermediate structural representations and final manuscript sections throughout the document lifecycle.

Method

The framework models manuscript generation as a contract-governed construction process:

Shared Visual Contract \(C\): A persistent state with three layers—(1) a global visual registry \(V\) (artifact type, semantic description, canonical label, expected reference locations), (2) section-level obligations mapping each \(s_i\) to required visual elements, and (3) validation rules enforcing label uniqueness and cross-reference consistency.

Four-agent pipeline operating over evolving contract states \(\{C_t\}\): $\(M = A_\text{rend}(A_\text{ref}(A_w(A_\text{arch}(S, C_0), C_1), C_2), C_3)\)$ - Architect \(A_\text{arch}\): decomposes story \(S\) into ordered sections \(\{s_i\}\), identifies visual artifacts, registers them in \(V\), initializes \((B, C_1)\). - Writer \(A_w\): generates draft \(D_i = A_w(s_i, C_t)\) constrained to use visual markers and citation identifiers from the contract. - Refiner \(A_\text{ref}\): harmonizes terminology and narrative coherence across all drafts; may trigger contract updates: \((M', C_{t+1}) = A_\text{ref}(\{D_i\}, C_t)\). - Renderer \(A_\text{rend}\): produces deterministic LaTeX, resolves all visual references, validates label uniqueness and cross-references against the contract.

Generate–evaluate–adapt loop: Evaluation agents \(\{E_k\}\) assess reasoning validity, data fidelity, and visual coherence on intermediate artifacts, producing feedback signals \(r_k = E_k(X, C_t)\). The contract updates as: $\(C_{t+1} = \text{Update}(C_t, \{r_k\})\)$ The system-level optimization objective is \(R(M) = \sum_k w_k r_k\).

Key Contributions

  • Story2Proposal framework: contract-governed coordination of architect, writer, refiner, and renderer agents through a persistent shared visual contract, explicitly enforcing structural and visual obligations during generation.
  • Generate–evaluate–adapt mechanism: evaluation agents (reasoning verification, data fidelity, visual coherence) inject feedback mid-generation to dynamically update contract constraints, distinguishing this from post-hoc validation.
  • Cross-backbone empirical evaluation: consistent improvements over DirectChat (single-stage, no contract) and FARS (structured but no continuous feedback) across GPT, Claude, Gemini, and Qwen backends, assessed by 10 expert reviewers on 8 dimensions.

Results

  • vs. DirectChat (single-stage prompt-based): Story2Proposal scores 6.145 vs. 3.963 (+2.182 average expert score), with Story2Proposal outperforming DirectChat in every model configuration (GPT, Claude, Gemini, Qwen).
  • vs. FARS (structured generation baseline): Story2Proposal scores 5.705 vs. 5.197 on four benchmark research papers.
  • Evaluation used a \(2 \times 4 \times 10\) design: 2 methods Ă— 4 LLM backbones Ă— 10 expert reviewers per condition.
  • Eight expert dimensions: structural integrity, writing clarity, methodological rigor, experimental substance, citation hygiene, reproducibility, formatting stability, visual communication.
  • Qualitative analysis confirms the persistent contract preserves figure placement and narrative-artifact alignment throughout the document lifecycle.

Limitations

  • Evaluation relies exclusively on human expert scores (10 reviewers); no automated metrics are reported, making large-scale or reproducible benchmarking difficult.
  • Only two baselines compared (DirectChat and FARS); no comparison to other recent multi-agent document generation systems.
  • The manuscript itself was "primarily generated by the Story2Proposal Agent," introducing self-referential bias in both content and the implicit claim that the output is human-publication-quality.
  • Contract update mechanism (\(\text{Update}(C_t, \{r_k\})\)) is not specified algorithmically; it is unclear how conflicting feedback signals are resolved.
  • Evaluation corpus (Jericho) is a single research domain; generalizability across fields is unvalidated.

Relevance to Harnesses / Meta-Harnesses

Story2Proposal is a concrete instance of a document-generation meta-harness: the shared visual contract functions as an explicit harness state that coordinates multiple specialist agents, enforcing structural invariants across the entire pipeline rather than delegating coherence to any single model call. The generate–evaluate–adapt loop is a textbook harness pattern in which orchestration logic (contract + evaluation agents) is cleanly separated from execution agents (writer, renderer), enabling adaptive re-planning mid-run. For researchers tracking harness architectures, the contract abstraction is notable as a domain-specific state representation analogous to task graphs or scratchpads in tool-use harnesses, but with semantic provenance tracking between input story fields and generated sections. The framework also demonstrates how harness-level feedback loops can substitute for end-to-end training, treating evaluation signals as system-level rewards without any fine-tuning.