Paper deep dive
SIVIA-RSI: Source-Grounded Adaptation of Diagramming Skills
Feng Yuan, Yifan Gao, Haoyue Li, Xin Gao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 9/29/2026, 11:16:45 AM
Summary
The paper introduces SIVIA-RSI, a framework for source-grounded adaptation of reusable diagramming skills for scientific method diagrams. It evaluates whether experience from generating diagrams for one paper improves the first diagram of another. The framework links critiques to source passages, proposes bounded edits to a skill library, and separates candidate competition from skill acceptance. In a complete-candidate evaluation on two NLP/language-agent papers, the strongest learned candidate achieved 87.50% required-relation accuracy compared to 83.33% for the original skill. However, no candidate met the predeclared 10-percentage-point improvement gate, and automatic selection preferences did not consistently align with author assessments. The study highlights issues with incomplete conditional specifications in prompts and ambiguities in diagram routes.
Entities (12)
Relation Signals (9)
SIVIA-RSI ā generates ā S4
confidence 95% Ā· Candidate S4 Evolving run, proposed policy
SIVIA-RSI ā generates ā S1
confidence 95% Ā· Candidate S1 Fixed run, first proposal
SIVIA-RSI ā generates ā S2
confidence 95% Ā· Candidate S2 Fixed run, second proposal
SIVIA-RSI ā generates ā S3
confidence 95% Ā· Candidate S3 Evolving run, initial policy
SIVIA-RSI ā uses ā S0
confidence 95% Ā· The original skill comprises nine files... S0 Original nine-file skill.
S1 ā achievesaccuracyon ā RAG
confidence 90% Ā· S1 RAG (2 images) 11/12 ... 87.50
S0 ā achievesaccuracyon ā RAG
confidence 90% Ā· S0 RAG (2 images) 10/12 ... 83.33
SIVIA-RSI ā evaluatesusing ā Reflexion
confidence 90% Ā· The development runs use ... Reflexion for parent comparison
SIVIA-RSI ā evaluatesusing ā Toolformer
confidence 90% Ā· The development runs use ... Toolformer for candidate competition
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Scientific method diagrams express computations through entities, dependencies, and conditional routes. Although generated figures can be improved through repeated editing, it is less clear whether experience from one paper improves the first figure of another. We present \sys, a framework for source-grounded adaptation of reusable diagramming skills, and study transfer through a complete-candidate evaluation. The framework links critiques to source passages, proposes bounded edits to a persistent skill library, and separates candidate competition from skill acceptance. We evaluate the original skill and four learned candidates on two NLP and language-agent papers, with two fresh generations per condition. The strongest candidate attains 87.50\% required-relation accuracy compared with 83.33\% for the original, while candidate behavior differs across papers. Local improvements on a separate development paper and automatic selector preferences do not establish consistent transfer. Tracing all 22 non-correct relation judgments to their production prompts reveals both incomplete conditional specifications and ambiguities despite explicit instructions. All ten planning diagrams leave an already-terminal selected leaf's route unclear; none of their prompts explicitly binds that route. Our findings show why evaluating reusable diagram skills requires source-grounded relation assessment, complete candidate coverage, and inspection of both prompts and images. We provide all twenty transfer outputs, skill snapshots, assessment records, and executable analyses.
Tags
Links
- Source: https://arxiv.org/abs/2609.33386v1
- Canonical: https://arxiv.org/abs/2609.33386v1
Trouble viewing inline? Open PDF directly ā
Full Text
49,832 characters extracted from source content.
Expand or collapse full text
SIVIA-RSI: Source-Grounded Adaptation of Diagramming Skills Feng Yuan Affiliation: School of Biomedical Engineering (Suzhou), Division of Life Science and Medicine, University of Science and Technology of China, Hefei, China Affiliation: Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou, China yuanfeng2317@mail.ustc.edu.cn xingaosam@163.com Yifan Gao Affiliation: School of Biomedical Engineering (Suzhou), Division of Life Science and Medicine, University of Science and Technology of China, Hefei, China Affiliation: Shanghai Innovation Institute, Shanghai, China Affiliation: Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou, China yuanfeng2317@mail.ustc.edu.cn xingaosam@163.com Haoyue Li Affiliation: School of Biomedical Engineering (Suzhou), Division of Life Science and Medicine, University of Science and Technology of China, Hefei, China Affiliation: Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou, China yuanfeng2317@mail.ustc.edu.cn xingaosam@163.com Xin Gao Affiliation: Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou, China yuanfeng2317@mail.ustc.edu.cn xingaosam@163.com September 24, 2026 Abstract Scientific method diagrams express computations through entities, dependencies, and conditional routes. Although generated figures can be improved through repeated editing, it is less clear whether experience from one paper improves the first figure of another. We present Sivia-Rsi, a framework for source-grounded adaptation of reusable diagramming skills, and study transfer through a complete-candidate evaluation. The framework links critiques to source passages, proposes bounded edits to a persistent skill library, and separates candidate competition from skill acceptance. We evaluate the original skill and four learned candidates on two NLP and language-agent papers, with two fresh generations per condition. The strongest candidate attains 87.50% required-relation accuracy compared with 83.33% for the original, while candidate behavior differs across papers. Local improvements on a separate development paper and automatic selector preferences do not establish consistent transfer. Tracing all 22 non-correct relation judgments to their production prompts reveals both incomplete conditional specifications and ambiguities despite explicit instructions. All ten planning diagrams leave an already-terminal selected leafās route unclear; none of their prompts explicitly binds that route. Our findings show why evaluating reusable diagram skills requires source-grounded relation assessment, complete candidate coverage, and inspection of both prompts and images. We provide all twenty transfer outputs, skill snapshots, assessment records, and executable analyses. Figure 1: Source-grounded skill adaptation and complete-candidate auditing. (a) Discovery evidence informs bounded edits to a reusable skill and, optionally, its updater policy. Separate papers support candidate competition and parent acceptance. The two learning runs produce four proposals, all retained regardless of selection. (b) The original skill and four proposals are frozen for generation on RAG and RAP: five skills, two papers, and two fresh repetitions yield twenty first outputs. Source-verified relations and additional mechanism errors are assessed after generation, without target-specific repair. The document and diagram miniatures are schematic; measured outputs appear in Figures 3 and 4. 1 Introduction A scientific method diagram communicates more than the names of a modelās components. Its arrows specify dependencies, its boundaries identify where state lives, and its branches determine when an operation is performed. A diagram can contain the right terminology yet describe the wrong computation. For example, a search update can be drawn as a model-parameter update, or a route can bypass the stage that makes an output valid. These distinctions are central to NLP and language-agent methods, where visually similar loops may represent interaction with an environment, internal simulation, or inference-time memory. Recent systems generate scientific diagrams through document understanding, planning, rendering, and refinement (Mondal et al., 2024; Zhu et al., 2026; Lin et al., 2026). Structural evaluation and generation methods increasingly assess relations as well as visual presentation (Zhang et al., 2026; Luo et al., 2026). Much of this process operates on the current figure: a critic identifies a defect, and a subsequent generation attempts to repair it. We investigate a complementary question: can the experience of making one diagram improve a new paperās first diagram? This form of reuse could turn individual corrections into persistent knowledge about abstraction, state ownership, and conditional control flow. Answering that question requires distinguishing three objects: a candidate instruction, the procedure that selects it, and the figures it produces on other papers. A selected candidate may benefit from generation variability or from the preferences of its judge. Conversely, a useful proposal may be rejected before its transfer behavior is measured. Examining only the final accepted skill therefore obscures whether adaptation produced useful knowledge and whether selection recognized it. We introduce Sivia-Rsi, an adaptation framework built around the Sivia scientific-figure workflow (Sivia contributors, 2026). A persistent multi-file skill guides document-to-diagram design. Critiques link visible problems to identifiable source passages; an updater converts that evidence into bounded edits. Candidate competition and parent acceptance use different papers, and an archive retains every proposal regardless of its selection outcome. Figure 1 connects this adaptation procedure to our complete-candidate transfer evaluation. Our study asks three questions. Do adapted skills improve first outputs on new papers? Do automatic selection decisions agree with the authorās assessment of the same images? Which non-correct relations are already incompletely specified in the design prompt, and which remain ambiguous despite explicit instructions? We address these questions using two development runs and a frozen matrix containing five skills, two transfer papers, and two independent generations per condition. The results distinguish local improvement from consistent reuse. Both adapted Reflexion images are preferred to their respective parent images, but the largest family-averaged gain in the subsequent transfer matrix is 4.17 percentage points. Some skills improve retrieval-related diagrams while weakening planning diagrams. Across all ten planning outputs, a conditional route remains unclear even though the normal search phases are present. Its absence from the corresponding prompts identifies a concrete gap in how persistent instructions are translated into a new methodās control flow. The contributions are a source-grounded skill-adaptation framework with inspectable state changes; a complete-candidate protocol for measuring first-output transfer separately from selection; and an empirical analysis connecting every non-correct relation judgment to its production instructions. We release all twenty transfer images with their exact prompts, skill identities, source-verified rubrics, and executable analyses. Together, these contributions make cross-paper diagram adaptation a measurable research problem with evidence at the instruction, selection, and image levels. 2 Related Work Scientific diagram generation. SciDoc2Diagrammer-MAF generates scientific diagrams from documents using an intermediate code representation and multi-aspect refinement (Mondal et al., 2024). PaperBanana combines reference retrieval, content planning, styling, rendering, and iterative critique (Zhu et al., 2026). DiagramRAG uses diagram knowledge graphs to retrieve references compatible with a sketchās semantics and topology (Yu et al., 2026). AutoFigure-Edit studies editable illustrations and reference-guided styling (Lin et al., 2026). These systems motivate a distinction between improving the current figure and retaining decisions that help on a later paper. Our experiment holds the generator configuration fixed and measures the first image produced with each persistent skill. Structural and scientific fidelity. SciFlow-Bench evaluates rendered diagrams through inverse parsing into structural graphs (Zhang et al., 2026); SciForma targets structural fidelity in generation (Luo et al., 2026). Our assessment uses source-verified scientific relations, including state ownership and conditional routes, alongside additional mechanism errors. A required-relation checklist and an audit of extra edges answer different questions: whether the required meaning is expressed, and whether other visible content changes the computation. We retain both, including uncertain judgments. Learning through language. Promptbreeder evolves task and mutation prompts (Fernando et al., 2023), and GEPA uses natural-language reflection to propose and select prompt changes (Agrawal et al., 2025). Sivia-Rsi applies this family of ideas to scientific-diagram skills, where feedback must be grounded in the source paper and transfer concerns relational meaning in an image. Our comparison measures four realized skill proposals against their shared parent. It characterizes the behavior of this adaptation process rather than establishing superiority over general prompt optimizers or diagram-generation systems. 3 The SIVIA-RSI Framework 3.1 Task and reusable state Let x be a paginated source paper and S a reusable skill library. A designer produces a self-contained production prompt z, which an image generator renders as y: z=Dā”(x,S,ĻµD),y=Gā”(z,ĻµG).z=D(x,S; _D), y=G(z; _G). (1) The task is a method overview with a common scope instruction. Layout need not reproduce the paperās original illustration, and experimental-result plots are outside the requested scope. The ĻµĪµ terms represent generation variability; fresh repetitions include both design and rendering. The persistent state is natural-language guidance rather than model weights. The original skill comprises nine files covering source interpretation, narrative organization, visual relations, and production-prompt detail. It contains 24,240 tokens under a common serialization with the o200k_base proxy tokenizer. Candidate libraries contain 24,319ā24,440 tokens. The base already includes direction, boundary, and ownership guidance, so adaptation may refine the specificity or prominence of an existing rule. 3.2 Source-grounded critique and skill updates A critic examines a discovery image together with the source divided into addressable text blocks. Each scientific finding identifies a visible region, the problem, and one to three source-block identifiers. The implementation resolves these identifiers to the original passage and page provenance. This grounds the evidence channel in existing text; whether the text supports the critique remains a semantic judgment. Given discovery evidence E and an updater policy U, the proposal model produces a skill change: ĪāS=Pā”(S,E,U),Sā²=Applyā”(S,ĪāS). S=P(S,E;U), S =Apply(S, S). (2) Each edit is an exact find-and-replace operation in an existing skill file. Its find string must occur once. The validator permits at most six edits and 12,000 combined removed-plus-inserted characters, including a valid empty edit list. The updater cannot alter source data, evaluators, split assignments, or executable code. Every resulting library has a content identity and is retained for inspection. A concrete learned decision. The parent skill requires external edges to have visible endpoints and directions. Candidate S1S_1 adds that an update must show where state is written and which later operation reads it, with both placed inside the boundary that owns the state. This converts a general connector instruction into a producerāstateāconsumer requirement. It is reusable because the next paper must supply its own state, writer, and reader; the edit inserts no RAG or RAP answer. Another edit separates layout percentages from scientific quantities. The complete literal diffs are included in the artifact. Skill Tokens Candidate origin and emphasis S0S_0 24,240 Original nine-file skill. S1S_1 24,348 Fixed run, first proposal; state ownership and consumers. S2S_2 24,330 Fixed run, second proposal; state interfaces and layout numbers. S3S_3 24,319 Evolving run, initial policy; branches, feedback, termination. S4S_4 24,440 Evolving run, proposed policy; directed-graph consistency. Table 1: All frozen skill identities. The automatic selector advanced S1S_1 and S3S_3; the transfer study retains all four proposals. Token counts use a shared proxy serialization. 3.3 Selection and complete candidate retention Two stages separate candidate comparison from acceptance. Candidates first compete on a selection paper. The selected skill then competes against its parent on another paper. At each stage, both image orders are judged and preferences are mapped back to image identity. Let Jā”(a,b)J(a,b) be the average preference for challenger a over incumbent b. The challenger advances only when Jā”(a,b)>0.5J(a,b)>0.5; candidate competition otherwise retains the initial-policy candidate, and parent acceptance otherwise retains the parent. An average of 0.5 may reflect conflicting preferences across orders. The framework also supports proposing an updated updater policy Uā²U from U and E. In the evolving run, U and Uā²U each produce one candidate from the same parent and evidence. The fixed run produces two candidates under U and discards a policy proposal to match the call structure. Skill and policy promotion are coupled: Uā²U is retained only when its candidate is selected and accepted. This optional mechanism supplies one of the four candidates; independently sampled discovery evidence prevents the two runs from isolating its causal effect. All proposals are archived before selection. The complete-candidate transfer study freezes the parent and every candidate, then generates first outputs on papers unused for proposal generation or selection. This preserves rejected candidates and allows their transfer behavior to be compared with that of accepted or advanced candidates. There is no target-specific critique, image repair, or best-of-n selection in this evaluation. 4 Experimental Design 4.1 Sources, candidates, and generation The development runs use ReAct for discovery (Yao et al., 2022), Toolformer for candidate competition (Schick et al., 2023), and Reflexion for parent comparison (Shinn et al., 2023). Each run yields five images: one discovery image, two candidate images, and two parent-comparison images. These ten images support the selection analysis. The initial author assessment uses a 34-relation rubric, which is kept separate from the later transfer rubric. The transfer study uses RAG (Lewis et al., 2020) and RAP (Hao et al., 2023), neither of which generated or selected the four candidates. RAG tests retrieval, probability marginalization, and frozen-versus-trainable components. RAP tests internal simulation, search-state updates, and conditional control flow. Both were read during study preparation and are development sources rather than a sealed final test set. The five frozen skills in Table 1 receive identical paper text and the common method-overview instruction. The matrix contains five skills Ć two papers Ć two fresh design-and-render repetitions, for twenty first outputs. A fixed randomized schedule interleaves conditions. The scheduling seed does not control the native image backend. A deterministic format and length check allows one text repair per design, counted in the budget. All twenty images complete; nine initial prompts exceed the 12,000-character limit and require that repair. The batch uses 29 text calls and 20 image attempts, with no repeated image sampling. Text responses report gpt-5.6-sol; the image interface is native-imagegen-tool, whose immutable revision and sampling seed are unavailable. Appendix A provides accounting and reproducibility details. 4.2 Source-verified human assessment Before generation, six required relations are fixed for each transfer paper. An assistant drafts propositions with page-specific source anchors, and an author independently verifies them. The generator receives the full paper and common overview instruction, but not the evaluation propositions or verification notes. The distinction between the generation scope and the checklist is analyzed in Section 5.3. An offline interface presents source PDFs, propositions, and anonymized images, hiding skill labels and production prompts. One author independently scores every image. The author has prior source familiarity and exposure to some project imagery; this is not a fully blinded assessment. Codex organizes the returned files, whose original metadata and the authorās explicit provenance confirmation are preserved separately. Each required relation is labeled correct, contradicted, ambiguous, omitted, or cannot judge. A proposition-issue label would trigger a common source-level correction before comparison; none occurs in the transfer returns. The assessor separately records extra errors, source evidence, and whether they alter the central mechanism. Central-error status is yes, no, or uncertain. Unsupported relations beyond the checklist are counted individually, merging repeated depictions of the same relation. Five ordinal 1ā5 ratings and a within-paper ranking provide secondary assessments; they are not combined into the primary score. 4.3 Measures and analysis protocol For image y of paper p with KpK_p required relations, required-relation accuracy is RRA(y,p)=1Kpāk=1Kpāp,k(y)=correct.RRA(y,p)= 1K_p _k=1^K_p1\ _p,k(y)=correct\. (3) All non-correct labels remain in the denominator. For skill s, we average repetitions within a paper and then weight paper families equally: As=1Fāāp=1F[1Rpāār=1RpRRAā”(ys,p,r,p)].A_s= 1F _p=1^F [ 1R_p _r=1^R_pRRA(y_s,p,r,p) ]. (4) Here F=2F=2, Rp=2R_p=2, and Kp=6K_p=6. Relation labels and repeated images are not independent paper samples. Results are descriptive; we do not attach population-level significance claims to this matrix. Before generation, we set a continuation criterion: at least a 10-percentage-point macro gain, no paper-level accuracy regression, and no increase in central errors or unsupported relations on either paper. This is a predeclared resource-allocation rule, not a statistical threshold. Uncertain error labels cannot authorize continuation. The primary decision and all candidates are reported regardless of whether the gate is met. To locate failures, a post-hoc assistant analysis traces every non-correct relation judgment to its exact production prompt. It distinguishes explicit requirements, partially specified requirements, absent requirements, and generationāevaluation scope mismatches. Full prompts are inspected for absence claims; retained spans locate positive evidence. This trace adds diagnostic categories without changing human labels. It does not audit the 98 correct items and is not an independent human evaluation. 5 Results 5.1 RQ1: Do learned skills transfer across papers? Required relations correct Central error: yes / uncertain Skill RAG RAP Macro (%) Ī (p) RAG (2 images) RAP (2 images) Extra unsupported Pass gate S0S_0 10/12 10/12 83.33 +0.00 0 / 1 0 / 0 0 ā S1S_1 11/12 10/12 87.50 +4.17 1 / 0 0 / 0 0 No S2S_2 9/12 10/12 79.17 -4.17 1 / 0 1 / 0 0 No S3S_3 11/12 9/12 83.33 +0.00 0 / 1 0 / 0 0 No S4S_4 11/12 7/12 75.00 -8.33 0 / 1 0 / 2 1 No Table 2: Complete frozen transfer matrix. Macro averages repetitions within each paper and then the two paper families equally. Extra unsupported counts cover four images per skill. Uncertain central-error labels remain unknown. All candidates fail the primary 10-percentage-point continuation gate, independently of additional-error constraints. This is a descriptive development comparison with two paper families. Table 2 reports the complete matrix, and Figure 2 shows the paper-level means and both repetitions. The original skill achieves 83.33% on each paper. Candidate S1S_1 increases RAG coverage from ten to eleven correct judgments out of twelve and leaves RAP unchanged, yielding 87.50% macro accuracy. The remaining candidates reach 79.17%, 83.33%, and 75.00%, respectively. No candidate meets the predeclared 10-point criterion, so the conditional follow-up batch is not run. Figure 2: First-output relation accuracy for every frozen skill. Bars show the mean of two independent design-and-render repetitions; circles and diamonds show the individual images. Dashed lines mark the original skillās paper-level mean. The same candidate can improve RAG while weakening RAP. The plot summarizes observed outputs, with no confidence intervals or independent-paper interpretation of the repetitions. Paper-specific behavior is consequential. S3S_3 gains one correct RAG judgment and loses one RAP judgment, producing the same macro score as the original. S4S_4 gains one on RAG and loses three on RAP. Its top-ranked RAG image therefore coexists with the lowest RAP mean, 58.33%. Candidate selection based on an attractive individual output would miss this difference between papers. Across the twenty images, 98 of 120 required relations are correct, 21 ambiguous, and one omitted. No checklist item is labeled contradicted or cannot judge. Yet three images have a confirmed central-mechanism error and five retain an uncertain status. Two of the three confirmed-error images still satisfy five of six checklist relations. Required coverage thus does not substitute for reviewing additional routes that change the computation. One unsupported update relation is also recorded for S4S_4 on RAP. Variation within a skill. The two S1S_1 RAG outputs rank second and tenth among the ten RAG images despite receiving six and five correct relations, respectively. The lower-ranked image chooses a Sequence-only scope and also contains a separately identified decoding inconsistency. The higher-ranked image meets all required relations while retaining a local gradient-direction issue. Appendix D presents both original images. The comparison illustrates what the repetition and extra-error annotations capture beyond the mean relation score. 5.2 RQ2: Does automatic selection identify the preferred candidate? Reflexion pair Relations Fidelity Author choice Fixed run 2/4ā4/42/4ā 4/4 2ā42ā 4 Updated Evolving run 3/4ā4/43/4ā 4/4 3ā43ā 4 Updated Table 3: Local parent-to-candidate comparisons on Reflexion under the initial rubric. Fidelity uses a 1ā5 scale. Both updated skills were proposed by the original updater; the pairs concern one paper. The Reflexion comparisons show a local adaptation signal: both revised images are preferred by the author and improve required-relation coverage (Table 3). The strongest candidate in the later matrix does not reach the continuation criterion, showing why a favorable parent comparison and first-output transfer are different outcomes. On both Toolformer candidate pairs, the automatic judge prefers the candidate opposite to the authorās ranking, in both presentation orders. In the evolving-run Reflexion pair, reversing image order reverses the final identity-level preference. The resulting average of 0.5 prevents parent replacement. Across four pairs and five dimensions, 11 of 20 dimension-level comparisons change identity preference between orders, including tie transitions. These are repeated judgments of four pairs, not twenty independent comparisons. The author ranks the full within-paper set, whereas the model sees pairs; the observed disagreement is specific to these assessment contexts. Complete candidate retention allows the study to proceed without assuming either selector identifies the best transferable skill. Indeed, the rejected S4S_4 later produces the best-ranked RAG image and weaker RAP results. Switching selectors alone is therefore not established as a solution to transfer. 5.3 RQ3: Where do non-correct relations arise? Prompt-side diagnosis RAG RAP Total Explicit requirement 6 2 8 Partial / unbound requirement 1 11 12 Absent critical requirement 0 1 1 Generationāevaluation scope mismatch 1 0 1 Table 4: Assistant-authored post-hoc trace of all 22 non-correct human relation judgments. Counts partition these observed items; they are neither additional human labels nor causal error probabilities. The prompt trace separates several superficially similar image failures (Table 4). A requirement is explicit when its critical operands, direction, and necessary condition binding are stated; partial when related content appears without a critical binding; and absent when the requirement is not specified. A scope mismatch is treated separately when the design follows a permitted scope narrower than the checklist. Phase names do not specify conditional routes. All ten RAP images are ambiguous on the proposition that combines the normal search phases with handling an already-terminal selected leaf. The phases are visible, but the direct route to backup is unclear. Correspondingly, all ten prompts describe normal search phases without explicitly binding terminality at selection to bypassing expansion and simulation. Some mention terminal states or handling after selection, but none specifies the required shortcut. These cases are partial specifications, not wholesale omission of search. This distinction concerns when a route applies. Rollout termination, the terminality of a selected leaf, and exhaustion of the global search budget are different events. One RAP prompt omits the budget requirement; another names a bounded loop and an output selector without binding the loop exit to it. A third explicitly makes that binding, yet its rendered result is assessed as ambiguous. Preserving conditions in instructions and expressing them visually are separate requirements. Explicit instructions can still yield ambiguous images. Eight items have an explicit corresponding requirement: six in RAG and two in RAP. For example, a RAG prompt originates an input bus at the original input and routes it to each concatenation operation, while the assessor identifies the visible bus as originating at retrieval output. Another prompt connects marginalized probability to the training loss, but the assessed image draws an input from retrieval weights before marginalization. These cases establish a discrepancy between the retained instructions and assessed pixels, without isolating the image model as its sole cause. Scope alignment changes the interpretation of omission. The common generation instruction permits a clearly stated main regime; the RAG checklist requires both Sequence and Token. The S1S_1 second repetition expressly chooses Sequence-only generation and does not request a Token comparison. Its omitted Token relation is zero under the frozen rubric, but does not cleanly measure disobedience to the generation task. We retain that label and denominator. This case demonstrates why generator scope and evaluation scope must be aligned even when evaluation propositions are withheld from the generator. 6 Discussion The study identifies three distinct requirements for reusable diagram knowledge. First, an instruction must be applicable to the new method. A general state-ownership rule is reusable, but its writer, state, and reader must be rebound to the target source. Second, the designer must instantiate the rule completely. The RAP prompts show that retaining search terminology does not ensure that a conditional route is specified. Third, the renderer must make that relationship recoverable from the image. The eight explicit-prompt cases show why instruction inspection alone cannot certify scientific fidelity. These requirements explain the role of the complete-candidate protocol. It measures the realized outputs of each skill instead of treating the selectorās decision as evidence of effectiveness. It also separates required coverage from harmful additional structure. A diagram can satisfy most checklist items while introducing a bypass; a candidate can produce a strong image on one paper and weaker results on another. Preserving the full matrix makes both behaviors visible. A practical implication is to represent conditional routes and state ownership explicitly between source interpretation and rendering. We implement a source-scoped relation-contract validator and compiler as a companion artifact (Appendix E). It checks declared nodes, guarded routes, evidence identities, and state writers and consumers. The hand-authored RAP example and seventeen software tests make these structural requirements executable; they do not measure a gain in generated images. The longer-term question is whether corrections can be stored as condition-aware knowledge: a design decision together with its applicability, supporting source evidence, and correction history. Reference retrieval already supplies useful visual and topological priors (Zhu et al., 2026; Yu et al., 2026). Testing what accumulated corrections add requires matching references, context budgets, and generator calls, then evaluating first outputs on new paper families. The artifact retains this extension separately; the present paperās conclusions rest on the completed adaptation and transfer studies. 7 Conclusion Sivia-Rsi studies scientific-diagram experience as persistent, source-grounded skill updates. Its complete-candidate evaluation distinguishes local improvement, selector preference, and first-output transfer across papers. In the observed matrix, candidate behavior varies by paper and the largest macro gain is 4.17 percentage points. Relation-level tracing identifies conditional specifications missing before rendering, ambiguities despite explicit instructions, and a scope mismatch between generation and assessment. These findings motivate evaluating reusable visual skills at all three levels: what the instructions preserve, what the selector retains, and what a reader can recover from the generated image. Limitations The transfer study covers two development paper families and two repetitions per skill and paper. One author supplies the image judgments, with prior source familiarity and some image exposure. The results characterize these observations rather than population-level superiority or inter-rater reliability. The development and transfer rubrics differ and are not pooled. Independently sampled discovery evidence prevents causal comparison of fixed and evolving updater policies. The prompt trace is a post-hoc assistant interpretation of non-correct items; it neither estimates overall prompt compliance nor replaces independent image assessment. Native image-model revisions and seeds are unavailable, so inputs and outputs are auditable without guaranteeing identical future pixels. The relation-contract prototype has software validation, while its image-level effectiveness and the proposed knowledge-library extension remain unmeasured. Data and Research Transparency The ancillary artifact contains every transfer image, exact production prompts, five complete skill snapshots, original assessments and their provenance confirmations, source-verified relation inventories, the full prompt trace, and executable analyses. Development-review records and selector comparisons are provided separately. Codex assisted source-verification drafts, software, diagnostic tracing, and manuscript preparation; the author independently verified propositions and assessed images. The Sivia-generated overview is a checked method schematic, excluded from experimental sample counts and generation costs. Public source-paper identifiers and hashes are provided rather than redistributing the papers. The existing Sivia and upstream attribution and license notices are retained. References Agrawal et al. (2025) L. A. Agrawal, S. Tan, D. Soylu, N. Ziems, R. Khare, K. Opsahl-Ong, A. Singhvi, H. Shandilya, M. J. Ryan, M. Jiang, C. Potts, K. Sen, A. G. Dimakis, I. Stoica, D. Klein, M. Zaharia, and O. Khattab GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning. Note: arXiv:2507.19457 External Links: Link Cited by: Ā§2. Fernando et al. (2023) C. Fernando, D. Banarse, H. Michalewski, S. Osindero, and T. RocktƤschel Promptbreeder: Self-Referential Self-Improvement Via Prompt Evolution. Note: arXiv:2309.16797 External Links: Link Cited by: Ā§2. Hao et al. (2023) S. Hao, Y. Gu, H. Ma, J. J. Hong, Z. Wang, D. Z. Wang, and Z. Hu Reasoning with Language Model is Planning with World Model. Note: arXiv:2305.14992 External Links: Link Cited by: Ā§4.1. Lewis et al. (2020) P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. KĆ¼ttler, M. Lewis, W. Yih, T. RocktƤschel, S. Riedel, and D. Kiela Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Note: arXiv:2005.11401 External Links: Link Cited by: Ā§4.1. Lin et al. (2026) Z. Lin, Q. Xie, M. Zhu, S. Li, Q. Sun, E. Gu, Y. Ding, K. Sun, F. Guo, P. Lu, Z. Ning, Y. Weng, and Y. Zhang AutoFigure-edit: generating editable scientific illustrations via reference-guided styling. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), G. Durrett and P. Jian (Eds.), San Diego, California, United States, p. 57ā67. External Links: Link, Document, ISBN 979-8-89176-392-0 Cited by: Ā§1, Ā§2. Luo et al. (2026) Y. Luo, P. Zhang, X. Zhang, X. Guo, Z. Lian, and Y. Lu SciForma: Structure-Faithful Generation of Scientific Diagrams. Note: arXiv:2607.18091 External Links: Link Cited by: Ā§1, Ā§2. Mondal et al. (2024) I. Mondal, Z. Li, Y. Hou, A. Natarajan, A. Garimella, and J. Boyd-Graber SciDoc2Diagrammer-MAF: towards generation of scientific diagrams from documents guided by multi-aspect feedback refinement. In Findings of the Association for Computational Linguistics: EMNLP 2024, Y. Al-Onaizan, M. Bansal, and Y. Chen (Eds.), Miami, Florida, USA, p. 13342ā13375. External Links: Link, Document Cited by: Ā§1, Ā§2. Schick et al. (2023) T. Schick, J. Dwivedi-Yu, R. DessƬ, R. Raileanu, M. Lomeli, L. Zettlemoyer, N. Cancedda, and T. Scialom Toolformer: Language Models Can Teach Themselves to Use Tools. Note: arXiv:2302.04761 External Links: Link Cited by: Ā§4.1. Shinn et al. (2023) N. Shinn, F. Cassano, E. Berman, A. Gopinath, K. Narasimhan, and S. Yao Reflexion: Language Agents with Verbal Reinforcement Learning. Note: arXiv:2303.11366 External Links: Link Cited by: Ā§4.1. Sivia contributors (2026) Sivia contributors Sivia: Paper to Scientific Overview to Editable PowerPoint. Note: Software repositoryAccessed September 22, 2026 External Links: Link Cited by: Ā§1. Yao et al. (2022) S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao ReAct: Synergizing Reasoning and Acting in Language Models. Note: arXiv:2210.03629 External Links: Link Cited by: Ā§4.1. Yu et al. (2026) X. Yu, J. Han, Z. Chen, C. Zhang, X. Fu, J. Tan, Z. You, Y. Jian, Y. Wang, and C. Chai DiagramRAG: A Lightweight Framework to Retrieve Scientific Diagram for Figure Generation. Note: arXiv:2605.27931 External Links: Link Cited by: Ā§2, Ā§6. Zhang et al. (2026) T. Zhang, H. Lin, Z. Liu, C. Chen, and W. Zhang SciFlow-bench: evaluating structure-aware scientific diagram generation via inverse parsing. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), M. Liakata, V. P. Moreira, J. Zhang, and D. Jurgens (Eds.), San Diego, California, United States, p. 17747ā17765. External Links: Link, Document, ISBN 979-8-89176-390-6 Cited by: Ā§1, Ā§2. Zhu et al. (2026) D. Zhu, R. Meng, Y. Song, X. Wei, S. Li, T. Pfister, and J. Yoon PaperBanana: Automating Academic Illustration for AI Scientists. Note: arXiv:2601.23265 External Links: Link Cited by: Ā§1, Ā§2, Ā§6. Appendix A Implementation and Reproduction Details The original library and every candidate are preserved as complete nine-file JSON snapshots. Exact replacements, skill hashes, source identities, generation settings, and twenty production prompts identify the study state. The transfer freeze has hash prefix 7b1f72700dc6; the RAG and RAP review packages have identifiers af0ac1ec6e03 and edfcc6331215. Raw submitted files retain exporter metadata, with the authorās explicit human-assessment confirmation stored separately. Historical preparation-stage status fields are not rewritten. Generation and selection details. Each development run generates two sibling candidates from its own parent and discovery evidence. Candidate competition advances the challenger only above an average preference of 0.5; otherwise the initial-policy candidate advances. Parent acceptance requires a strict win. The fixed run makes a policy proposal and discards it; the evolving run uses it to generate a sibling and promotes policy and skill jointly only after acceptance. Neither completed run promotes an evolved updater. Both skills advanced to Reflexion were proposed by the original updater. The two development runs use 15 and 13 text calls, respectively, and five images each. The frozen transfer batch uses 29 text calls, including nine repairs for overlength designs, and twenty image attempts. Recorded input, output, and cache-read tokens are 1,378,697, 85,122, and 24,576; these are separate provider fields. Native image usage and dollar cost are not supplied. Wall time is 65.83 minutes, including orchestration and handoffs. The scheduling seed is 2026092102. Actual image dimensions and hashes are retained. No failed image generation or repeated sampling is omitted from this batch. Prespecified continuation. A candidate must improve macro accuracy by at least 0.10, have no accuracy regression on either family, and not increase either familyās central-error rate or unsupported-relation count. Uncertain constraints cannot authorize continuation. At most one passing skill would enter a second two-paper comparison against the original and a preconstructed repetition control. No candidate passes the primary criterion, independently of the uncertain extra-error labels. That conditional study is not run and contributes no observations to this paper. Development annotations. The earlier ten-image review contains 50 dimension ratings and 34 relation judgments. Two proposition-issue judgments refer to the same ReAct statement: reasoning and action are interleaved but need not strictly alternate at every step in all tasks. Its wording correction is preserved separately. The Reflexion and selector comparisons in Section 5.2 use the original records. A seven-call synthetic judge sanity check preceded these studies; it is not evidence of reliability on dense generated diagrams. Historical debugging and incomplete runs are not included as successful transfer units. Executable checks. python anc/reproduce.py reconstructs the transfer table and continuation decision. python anc/reproduce-v2.py checks all 22 prompt-trace identities, hashes, and evidence spans. python anc/verify-images.py verifies all twenty image files against original hashes and dimensions. python anc/reproduce-paper.py verifies the development comparisons, selector order counts, and plot coordinates. The relation-contract tests run with unittest. None of these checks makes model calls. Figure 2 is generated directly from the frozen assessment records; no smoothing or inferred error bars are used. Appendix B Source-Verified Relation Inventory The following descriptions summarize the frozen Chinese-language rubric. The original wording, source pages, and verified explanations are retained in the artifact. RAG. (1) Query encoding leads to top-K document retrieval. (2) Generation depends on the original input and a retrieved document. (3) Output probability marginalizes document-conditioned generation using retrieval weights. (4) RAG-Sequence fixes a document within a sequence before marginalizing over documents. (5) RAG-Token marginalizes over documents at each token and continues autoregressively. (6) Training updates the query encoder and BART generator while holding the document encoder and index fixed. Source anchors lie on PDF pages 2ā4. RAP. (1) One language model acts as reasoning agent and world model under different prompts. (2) The agent proposes an action from the current state; the world model predicts the next state from state and action. (3) Trajectories are internally simulated without interaction with a real external environment. (4) Selection uses value and exploration information; reward backup updates stateāaction values in the search tree. (5) Search distinguishes selection, expansion, simulation, and back-propagation; an already-terminal selected leaf reaches backup without forced expansion. (6) At the prescribed budget, the algorithm selects an output trace. Source anchors are on PDF pages 4ā6 and Algorithm 1 on page 14. Paper Required relation (short name) Correct / 10 RAG Query to retrieval 10/10 Input and document operands 7/10 Weighted marginalization 9/10 Sequence formulation 10/10 Token formulation 8/10 Training ownership 8/10 RAP Two prompted roles 10/10 Stateāaction transition 10/10 Internal simulation 10/10 UCT and reward backup 9/10 Phases and terminal shortcut 0/10 Budget to output trace 7/10 Table 5: Descriptive relation profile across all five skills and two repetitions per paper. Counts retain the original compound propositions and are not independent paper samples. Table 5 aggregates each required relation over the ten outputs for that paper. In RAP, the first three propositions are correct in every image, while the compound phase-and-terminal-route proposition is ambiguous in all ten. The latter does not mean that the search phases are absent. Its original definition and denominator are retained. Appendix C Complete Image and Trace Results Paper Skill Rep. Correct Relations 1ā6 Central error Unsupported Rank RAG S0S_0 1 5/6 CACCCC no 0 5 RAG S0S_0 2 5/6 CCCCCA uncertain 0 8 RAG S1S_1 1 6/6 C no 0 2 RAG S1S_1 2 5/6 CCCCOC yes 0 10 RAG S2S_2 1 5/6 CCACCC no 0 4 RAG S2S_2 2 4/6 CACCCA yes 0 9 RAG S3S_3 1 6/6 C no 0 3 RAG S3S_3 2 5/6 CCCCAC uncertain 0 6 RAG S4S_4 1 5/6 CACCCC uncertain 0 7 RAG S4S_4 2 6/6 C no 0 1 RAP S0S_0 1 5/6 CCCCAC no 0 5 RAP S0S_0 2 5/6 CCCCAC no 0 3 RAP S1S_1 1 5/6 CCCCAC no 0 2 RAP S1S_1 2 5/6 CCCCAC no 0 1 RAP S2S_2 1 5/6 CCCCAC no 0 7 RAP S2S_2 2 5/6 CCCCAC yes 0 10 RAP S3S_3 1 5/6 CCCCAC no 0 4 RAP S3S_3 2 4/6 CCCCAA no 0 8 RAP S4S_4 1 3/6 CCCAAA uncertain 1 9 RAP S4S_4 2 4/6 CCCCAA uncertain 0 6 Table 6: Every transfer unit. C: correct; A: ambiguous; O: omitted. Relation positions correspond to Appendix B. Rank is local to each paper, with 1 preferred. Unknown central errors are not relabeled as absent. Unit Relation Human label Prompt diagnosis rag-S0-r1 rag-02 ambiguous E: explicit requirement rag-S0-r2 rag-06 ambiguous E: explicit requirement rag-S1-r2 rag-05 omitted T: task-scope mismatch rag-S2-r1 rag-03 ambiguous E: explicit requirement rag-S2-r2 rag-02 ambiguous E: explicit requirement rag-S2-r2 rag-06 ambiguous E: explicit requirement rag-S3-r2 rag-05 ambiguous P: partial / unbound rag-S4-r1 rag-02 ambiguous E: explicit requirement rap-S0-r1 rap-05 ambiguous P: partial / unbound rap-S0-r2 rap-05 ambiguous P: partial / unbound rap-S1-r1 rap-05 ambiguous P: partial / unbound rap-S1-r2 rap-05 ambiguous P: partial / unbound rap-S2-r1 rap-05 ambiguous P: partial / unbound rap-S2-r2 rap-05 ambiguous P: partial / unbound rap-S3-r1 rap-05 ambiguous P: partial / unbound rap-S3-r2 rap-05 ambiguous P: partial / unbound rap-S3-r2 rap-06 ambiguous A: absent requirement rap-S4-r1 rap-04 ambiguous E: explicit requirement rap-S4-r1 rap-05 ambiguous P: partial / unbound rap-S4-r1 rap-06 ambiguous E: explicit requirement rap-S4-r2 rap-05 ambiguous P: partial / unbound rap-S4-r2 rap-06 ambiguous P: partial / unbound Table 7: Complete post-hoc trace inventory. Human labels are unchanged. Prompt categories are assistant-authored diagnostic judgments, not independently human-verified classifications. Table 6 reports all twenty planned units, including uncertain central errors. Ranks are local to a paper, with one preferred, and are not treated as interval-scale scores across papers. Table 7 lists every non-correct relation and its prompt diagnosis. Original author notes, full diagnostic rationales, and exact prompt spans remain available in the ancillary records. Appendix D Illustrative Generated Outputs (a) S1S_1, RAG, repetition 1: 6/6 required relations; rank 2/10; no central error. (b) S1S_1, RAG, repetition 2: 5/6 required relations; rank 10/10; central error recorded. Figure 3: Two independent first outputs from the same frozen skill. The lower output chooses a Sequence-only scope permitted by the generator instruction, which differs from the evaluation scope; the author separately identifies a decoding inconsistency. The upper output meets all required relations while retaining a local gradient-direction issue. These post-assessment examples illustrate the difference between relation coverage and overall preference. Original pixels are retained. Figure 3 shows the two S1S_1 RAG repetitions discussed in Section 5.1. The lower image places Sequence marginalization before a single beam-search route, independently of its omitted Token variant. The differing ranking therefore cannot be reduced to the single checklist item alone. Figure 4: RAP under S2S_2, repetition 2. Five required relations are correct, yet the author records a central control-flow error: the expansion-to-next-selection route can bypass simulation and backup. This post-assessment example motivates keeping extra-mechanism errors separate from checklist coverage. Its unsupported-relation count is zero. Figure 4 provides a complementary example: most required relations are present while an additional route changes control flow. All twenty transfer images, including those not displayed here, are supplied in the ancillary artifact. Appendix E Source-Scoped Relation-Contract Prototype The conditional-route diagnosis motivates an explicit interface between source interpretation and rendering. We implement a validator and prompt compiler for a contract C=(x,Ļ,V,E,B)C=(x,Ļ,V,E,B), comprising source identity, scope identity, typed nodes, directed edges, and evidence blocks. Operation, decision, state-memory, and output nodes have distinct types. Decision edges bind Boolean predicates to true or false continuations; update edges identify the owner of the destination state. Evidence identifiers resolve to short source spans with page provenance and hashes. The validator checks identities, endpoints, types, decision continuations, state writers and readers, and reachability. It rejects updates to frozen state or to a different owner. These checks concern the declared representation rather than semantic entailment or all real execution paths. The compiler emits directed relations with guard labels, ownership, and source references. Figure 5 illustrates a hand-authored RAP example developed after the measured matrix; none of the twenty experimental outputs uses this prototype. Selection Selected leaf already terminal? Expansion Simulation to rollout completion Back-propagation update search Q Global search budget reached? Select output trace falsetrue: skip expansion and simulationtruefalse: continue search Figure 5: Hand-authored RAP relation-contract example, derived from the exposed development source. Conditions are attached to routes rather than represented only by phase names. This is a schema illustration produced from source interpretation, not a newly generated evaluation image. The prototype has not been tested for image-quality improvement. Seventeen deterministic software tests cover terminal and nonterminal paths, a bounded unfinished-budget loop, source/scope applicability, evidence identity, dangling endpoints, incomplete or wrongly bound decisions, ownership, frozen-state updates, and disconnected nodes or consumers. They verify that a terminal selected leaf reaches backup without expansion or simulation and that a nonterminal leaf traverses both. These are software checks on an exposed source, not new research samples. Testing automatic source binding and final-image effectiveness requires a separate experiment with a shared generation scope for every condition. Connection to reusable visual knowledge. A correction may be stored with its applicability, source evidence, role mapping, visual organization, forbidden transfers, and validation history. The reusable rule in the RAP example is to bind a predicate to its admissible routes; the next source must determine the actual predicate and endpoints. Copying an MCTS shortcut into a non-search method would violate that principle. A controlled test would compare the same retrieved references alone, structured knowledge from those references, and structured knowledge plus corrections acquired from other papers. Context and generation budgets must be matched and final-test feedback withheld. This extension is a research direction, not an evaluated method in the present study.