Overview
Spark-to-Paper is a system proposed by Zhuoyang Qian, Biao Wu, Yiran Wang et al. in their August 2026 paper *Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill*. It decomposes the full research workflow into 13 composable skills executed inside an existing coding assistant, with no dedicated agent platform or orchestration service required.
A representative scenario: a user enters a research topic (e.g., *knowledge degradation in small models under long-context training*). The pipeline then runs through 13 skills covering literature retrieval, experiment design, execution, evidence comparison, claim revision, chart generation, and manuscript assembly.
Core Design Principles
1. Separation of Model Judgment and Deterministic Operations
- Model handles judgment: literature relevance, hypothesis support, writing clarity, revision decisions.
- Code handles deterministic tasks: running experiments, verifying citations exist, validating chart editability, detecting fabrication.
- Citation validity: 99.5%
- Chart editability: 96.4%
- Fabrication detection: 14% (single draft) → 92% (full integrity stack), a 6.5× improvement
- Adversarial self-review precision: 74%
- 11.9M tokens
- $8.1 at current API pricing
- 3.2 hours end-to-end
- 99.5% citation validity is not because the model is smarter—it is because the structure enforces citation verification.
- 92% fabrication detection is not because the model is more honest—it is because deterministic checks catch what the model invents.
- Title: Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill
- Authors: Zhuoyang Qian, Biao Wu, Yiran Wang, et al.
- arXiv: https://arxiv.org/abs/2608.11924
This split yields a system that is simultaneously flexible (judgment, writing, revision via LLM) and verifiable (checks, validations, executions via deterministic code).
2. Evidence Before Claim
The most distinctive design choice. Traditional workflows let results shape the narrative, which encourages selective reporting when evidence contradicts a hypothesis. Spark-to-Paper instead requires the system to write down the evidence required to support or refute a hypothesis before seeing any results. That evidence requirement is frozen, experiments are then run, and the system matches outcomes against the frozen requirements.
If results do not support the original hypothesis, the system revises the claim, not the narrative. Claims must remain consistent with experimental evidence; "the results were weak so let's rephrase to sound stronger" is structurally blocked.
This implements the scientific method—hypothesis first, evidence-driven, falsifiable conclusions—at the code level.
3. Bounded Self-Refutation Loop
A known failure mode: experiments repeatedly contradict the research goal. Human researchers either abandon the topic or weaken the hypothesis until it is trivially supported. The former wastes effort; the latter is misconduct.
Spark-to-Paper implements a bounded recovery: the system may revise experimental methods or hypothesis wording a limited number of times, but once a threshold is exceeded it halts and reports *"current evidence does not support the original research goal"*. It neither loops forever nor manipulates data to make results "look good".
4. Editable Vector Charts
Instead of exporting PNGs and editing them in presentation tools, Spark-to-Paper generates charts as programmatic, editable vector graphics (matplotlib-style for result plots; reconstructed code for methodology diagrams). Researchers can edit the chart code directly without rerunning the full pipeline.
Reported Numbers
Evaluated on 8 controlled research topics:
Cost per paper:
Why the Positioning Matters
Earlier systems such as AI Scientist and GPT-Researcher already demonstrated AI-generated papers. Spark-to-Paper's contribution is structural rather than functional:
1. No standalone platform — the 13 skills install into an existing coding assistant, lowering deployment cost. 2. Composable skills — each can be swapped or upgraded independently (e.g., replace the literature retrieval skill without touching the rest). 3. Separation of judgment and verification — the jump from 14% to 92% fabrication detection comes from deterministic checks that catch what the model would otherwise fabricate.
Relation to the ACE Workflow
The earlier ACE (Advanced Context Engineering) RPI workflow—Research, Plan, Implement—applies the same underlying principle at a coarser 3-step granularity: each stage is independent, stages connect via deterministic interfaces, and context compression per stage prevents contamination.
Spark-to-Paper generalizes this to 13 steps. ACE's per-stage context compression corresponds to per-skill context isolation; ACE's "research first" corresponds to Spark-to-Paper's "evidence before claim". The shared pattern is: decompose long agent workflows into independently verifiable steps with explicit input/output contracts.
A Broader Principle
Spark-to-Paper illustrates a recurring theme: reliability is a structural property, not a model property.
This echoes the same lesson seen in colibrì (1,300 lines of C running a 744B-parameter model): division of labor between what models do well and what code does well outperforms a single monolithic approach. Across papers, a converging consensus is forming: the robustness of agent systems is determined by their architecture, not by the underlying model's capability.