Lean Research Inception Framework
- Lean Research Inception (LRI) is a structured, agile-inspired approach for formulating and assessing research problems with practical significance in software engineering.
- It employs a five-phase workflow centered on a Problem Vision board to iteratively refine problem definitions and guide Go/Pivot/Abort decisions.
- LRI uses a seven-point semantic differential scale to evaluate research problems on being valuable, feasible, and applicable, driving actionable insights.
Lean Research Inception (LRI) is a structured, agile-inspired approach for formulating and initially assessing software engineering research problems in relation to industrial practice. It was introduced to address a recurrent diagnosis in software engineering research: limited practical relevance often begins at problem formulation, where researchers may rely on oversimplified views of practice, weak industry connections, or poorly identified problems. In its initial formulation, LRI combines a collaborative problem-definition artifact, a five-phase workflow, and a seven-point semantic differential assessment across three criteria—valuable, feasible, and applicable—to support an early Go/Pivot/Abort decision before substantial resources are committed to solution development or evaluation (Pereira et al., 15 Jun 2025).
1. Origins, purpose, and methodological position
LRI was proposed as a practical framework to formulate and assess software engineering research problems for industrial relevance. Its stated goal is to deliberately bridge academic and industrial perspectives early, so that research problems are valuable to industry, feasible to investigate with available resources, and applicable to produce usable outcomes. The framework is intended for software engineering researchers planning or conducting applied research, practitioner partners when available, and research leaders and teams within empirical software engineering and design science traditions who want an early, lightweight test of practical relevance (Pereira et al., 15 Jun 2025).
The framework is positioned at the very beginning of a project, which it names “inception.” This placement is substantive rather than procedural: LRI treats misalignment as something that typically begins at problem definition, not merely at solution transfer. It therefore complements the technology transfer model of Gorschek et al. by concentrating on the “Problem Formulation” stage and by adding an explicit collaborative pathway and an initial relevance assessment (Pereira et al., 15 Jun 2025).
Its methodological influences are explicitly lean and agile. The framework is described as grounded in agile principles and methodologies such as Design Thinking, Lean Startup, and Lean Inception, with emphasis on early co-creation, iteration, and validation of practical relevance with practitioners (Pereira et al., 20 Mar 2026). A related empirical background comes from research on Lean internal startups in large companies, where authorisation, top management support, cross-functional teams, MVPs, Build-Measure-Learn loops, and actionable metrics are central to lean experimentation under corporate governance (Edison et al., 2018). That literature does not define LRI itself, but it clarifies the organizational logic behind LRI’s early-stage, low-commitment, feedback-oriented structure.
2. Five-phase structure and the Problem Vision artifact
LRI comprises five sequential phases. The central artifact in the framework is the “Problem Vision” board, used to externalize and iteratively refine the research problem. The sequence is stable across the foundational and educational papers, although the names of some board fields are slightly adapted in the later educational presentation (Pereira et al., 15 Jun 2025).
| Phase | Primary function |
|---|---|
| 1. Problem Vision Outline | Sketch the initial problem |
| 2. Problem Vision Alignment | Co-refine the problem with practitioners |
| 3. Research Problem Formulation | Consolidate the refined problem |
| 4. Research Problem Assessment | Rate practical relevance |
| 5. Go/Pivot/Abort Decision | Decide whether to continue, refine, or stop |
In the original framework, Phase 1 uses a “Problem Vision” board with seven attributes: problem outline, context, implications/impacts, practitioners, evidence, objective, and research questions. These attributes are intended to cover what the problem is, where it occurs, why it matters, who is involved, what evidence grounds it, what objective is sought, and which research questions will guide the investigation. Phase 2 brings practitioners into a collaborative workshop to refine these attributes and improve shared understanding. Phase 3 documents the resulting research problem that will then be assessed in the next phase (Pereira et al., 15 Jun 2025).
The educational deployment reformulates the same artifact in slightly more pedagogical language: practical problem, context, implications/impacts, stakeholders, evidence, objective, and research questions. In that setting, the board functions as a cognitive artifact for reasoning, clarity, contextualization, and communication, and is supported by a public visual template, LMS materials, videos, examples, and synchronous guidance sessions (Pereira et al., 20 Mar 2026).
Phase 5 makes the framework explicitly decision-oriented. If perceived relevance is high, the team proceeds; if relevance is low but fixable, it pivots and revisits alignment; if relevance is low and irreparable, it aborts to avoid wasted effort. This decision logic is one of the features that most clearly distinguishes LRI from more open-ended forms of early problem discussion: the framework is not only descriptive, but also selective.
3. Assessment logic: valuable, feasible, and applicable
The assessment core of LRI is a seven-point semantic differential scale, with one bipolar item for each of three criteria: “Worthless – Valuable,” “Infeasible – Feasible,” and “Inapplicable – Applicable.” The scale is explicitly inspired by agile principles and Lean Startup’s MVP concept, and is meant to support early, collective judgments before larger investments are made (Pereira et al., 15 Jun 2025).
The criterion valuable is defined as whether solving the problem can generate meaningful impact for industrial practice. Its intended interpretation emphasizes stakeholder salience, tangible benefits, and real-world outcomes. Qualitative feedback proposed refinements that more clearly include business value, ROI and cost, measurable benefits, originality, and problem importance. The criterion therefore extends beyond purely technical merit.
The criterion feasible is defined as whether the problem can realistically be investigated given available resources. It is explicitly framed from the research perspective, including time, budget, expertise, data access, tooling, infrastructure, and execution risks. Participants in the initial evaluation noted that feasibility can be difficult to judge during initial problem definition and requested a clearer statement that the vantage point is the research side rather than the industry side.
The criterion applicable is defined as whether the research problem leads to practical and usable outcomes in industry. It is framed from the industry perspective, centering on adoptability, integration, usability, scalability, generalizability, and barriers to use. Qualitative feedback suggested that applicability requirements or thresholds might be made more explicit and that intention to use, including adoption-oriented constructs, could sharpen the construct (Pereira et al., 15 Jun 2025).
A persistent discussion in the framework concerns the distinction between feasible and applicable. LRI stresses that feasibility refers to the research team’s ability to conduct the study, whereas applicability concerns whether industry can realistically adopt or use the outcomes. Some participants observed correlation between the two dimensions, but the framework maintains the distinction to avoid conflating research-side constraints with industry-side adoption conditions.
The paper does not prescribe a formal aggregation formula. Ratings are used to support a consolidated Go/Pivot/Abort decision. A non-canonical pseudo-formula included in the summary normalizes , , and from the seven-point scale to and combines them as , with . This suggests one way to operationalize the assessment numerically, but it is explicitly not part of the canonical framework definition (Pereira et al., 15 Jun 2025).
4. Empirical evaluation and educational deployment
The initial evaluation of LRI was conducted retroactively in an ISERN 2024 workshop in Barcelona with 42 senior software engineering researchers experienced in industry–academia collaboration. Participants applied LRI to a published research problem from Cabral et al. (2024), discussed and refined the formulation, individually rated the three criteria, and then evaluated the criteria’s importance and completeness. The reported results showed 83.3% agreement for valuable, 76.2% agreement for feasible, and 73.8% agreement for applicable. More specifically, valuable received 21.4% Slightly Agree and 61.9% Agree; feasible received 16.7% Slightly Agree and 59.5% Agree; applicable received 14.3% Slightly Agree and 59.5% Agree. A pilot with 12 master’s and PhD students at ExACTa (PUC-Rio) was used to refine workshop materials, and the study combined descriptive frequencies with qualitative analysis and mixed-methods triangulation (Pereira et al., 15 Jun 2025).
The qualitative findings were as important as the percentages. Participants suggested that “Worthless” might be too negative as an anchor, requested clearer definitions, emphasized the distinction between research-side feasibility and industry-side applicability, and argued that valuable should more explicitly address business value, ROI, costs, measurable benefits, originality, and problem importance. The general observation was that early-stage judgments are necessarily fuzzy, but that a simple MVP-inspired scale is suitable for initial collective assessment before detailed metrics exist.
A later case study transferred LRI to software engineering education. In a distance-learning Capstone Project course at the University of Vassouras, 60 final-year software engineering students and 7 faculty advisors used Phases 1–3 of LRI over approximately three weeks. Students reported benefits in reasoning (60%), clarity and definition (61.7%), contextualization (60%), communication (50%), ease of use (60%), and intention to adopt the board in future projects (58.3%). Advisors observed clearer and more structured problems (57.1%), easier communication (57.1%), strong ease of use for students (71.4%), and a high intention to recommend the board in other software engineering projects (85.7%) (Pereira et al., 20 Mar 2026).
In the educational context, advisors acted as representatives of professional practice. The study therefore did not validate the full industrial deployment of LRI, and it deliberately excluded Phases 4 and 5 to keep the focus on problem formulation. Even so, it established that the board can serve as a practical instrument for making assumptions explicit, challenging vague problem statements, and improving alignment among stakeholders in a supervised learning setting.
5. AI-agent-supported LRI
A later vision paper proposes integrating AI agents into each LRI phase. In that formulation, LRI remains the underlying framework, but AI agents are introduced as facilitators and knowledge mediators that can support pre-filling problem attributes, aligning stakeholder perspectives, refining research questions, simulating multiperspective assessments, and supporting decision making (Pereira et al., 14 Dec 2025).
The five agent-supported functions are:
- Pre-filling problem attributes: agents synthesize scientific literature, white papers, industry reports, and repositories to draft entries for the seven Problem Vision attributes.
- Aligning stakeholder perspectives: agents generate role-specific views, examples, boundary conditions, and disagreement points for workshops.
- Refining research questions: agents examine objectives and research questions for internal consistency, novelty, and methodological fit.
- Simulating multiperspective assessments: agents produce value, feasibility, and applicability ratings from multiple stakeholder standpoints, with justifications.
- Supporting decision making: agents analyze consolidated assessments, risks, failure patterns, and economic considerations to recommend Go/Pivot/Abort actions.
The descriptive evaluation scenario is grounded in a published study on code maintainability in machine learning and uses the seed problem that “data scientists produce ML code that is difficult to maintain.” Within that scenario, agent support is reported to reduce cognitive effort in Phase 1, improve mutual understanding in Phase 2, strengthen research-question precision in Phase 3, and better prepare decision arguments in Phase 5. The paper explicitly describes the outcome as a scenario-based and descriptive evaluation rather than a formal experiment, and it calls for empirical validation before strong effectiveness claims can be made (Pereira et al., 14 Dec 2025).
The same paper references Google’s AI Co-Scientist as a suitable multi-agent, asynchronous architecture for implementing these functions. This positions AI-supported LRI not as an autonomous replacement for collaborative judgment, but as a way to enrich collaborative discussions and enhance critical reflection on value, feasibility, and applicability.
6. Limitations, debates, and broader reinterpretations
LRI is explicitly presented as ongoing research. The foundational paper identifies limitations in internal validity, external validity, construct validity, and conclusion validity, including the restricted sample of senior software engineering researchers, the descriptive nature of the statistics, and the absence of broader industry stakeholder participation in the initial evaluation. Future work is said to require more industry-driven case studies, clearer terminology, better examples, and a more explicit treatment of business value, ROI, originality, and the feasible/applicable distinction (Pereira et al., 15 Jun 2025).
The educational paper introduces additional boundary conditions. Its setting is a single undergraduate software engineering program in Brazil, delivered at distance across three campuses; it uses advisor-mediated alignment rather than direct practitioner participation, relies on self-reported perceptions gathered by Likert-scale questionnaires, and does not apply the full assessment and decision phases. The results therefore support use in problem-formulation pedagogy, but not a general claim that LRI’s full industrial decision logic has been validated in educational settings (Pereira et al., 20 Mar 2026).
A separate development is terminological extension. In EconCSLib, “Lean Research Inception” is used to denote “the process of initiating, structuring, and scaling Lean-based formalization of research,” supported by a reusable Lean 4 library, an author-facing paper-centric workflow, and human–AI–Lean tooling for formalizing Economics and Computation papers (Garg, 11 Jun 2026). A related paper presents EconCSLib as an early Lean 4 library for computational economics with over 40,000 lines of Lean code and more than 1,300 theorems/lemmas, and describes this infrastructure as operationalizing LRI for AI-assisted formalization and machine-checked open problems (Bei et al., 15 Jun 2026). This suggests that the phrase “Lean Research Inception” is beginning to function as a broader label for lean, early-stage structuring of research activity, even though its original and most developed meaning remains the software engineering framework for practice-aligned problem formulation.
Across these strands, the stable core of LRI is early alignment under uncertainty. What changes across papers is the object being aligned: in software engineering, it is the research problem and its practical relevance; in education, it is the student’s problem formulation and supervisory dialogue; in the AI-agent vision, it is the distribution of cognitive work across humans and agents; and in EconCSLib, it is the first mile of Lean-based formalization. The canonical formulation, however, remains the five-phase software engineering framework centered on a Problem Vision board and the triad of valuable, feasible, and applicable (Pereira et al., 15 Jun 2025).
