The Impact of GenAI on the Future of Requirements Engineering

• Earlier shifts paired rigorous verification and validation with plan-driven development, then minimized requirements through Agile user stories and developer expertise.Source: §1 · pp. 1–2
• As implementation becomes cheaper through agents, RE shifts toward operational specifications, context curation, verification, and evaluation.Source: §1 · p. 2
• The article reviews AI’s history in RE, surveys LLMs, prompt programming, and generalist SE agents, and forecasts changes to requirements practice.Source: §1 · p. 2

• The working taxonomy spans RE activities from elicitation to verification, while AI-supported research has focused mainly on elicitation, extraction, and analysis.Source: §2 · pp. 2–3, Table 1 · p. 3
• The literature review covers prominent RE and SE venues since 2020 but is neither systematic nor comprehensive.Source: §2 · p. 3
• Elicitation research is shifting from capturing statements toward question generation and dialogue, but tacit knowledge and realistic evaluation remain challenges.Source: §2 · pp. 3–4
• Extraction research spans authoritative and user-generated sources and varied classification tasks, yet extracted requirements often remain weakly linked to intent and project context.Source: §2 · p. 4
• Generated artifacts often serve as the endpoint of research; unlike extraction, generation translates concepts and may introduce assumptions absent from source stories.Source: §2 · p. 4
• Analysis research detects ambiguity, omissions, incompatibilities, and conflicts, using supplemental domain knowledge or neuro-symbolic satisfiability reasoning in some approaches.Source: §2 · pp. 4–5
• Traceability links heterogeneous artifacts and supports retrieval and explanation, but trusted links remain difficult to establish and maintain.Source: §2 · p. 5
• Integrated RE pipelines face cascading errors because imperfect AI-produced artifacts or data feed downstream tasks that are evaluated separately.Source: §2 · p. 5

3 Advances in Language Models

• Instruction tuning, chain-of-thought prompting, multimodality, reasoning models, distillation, and mixture-of-experts broaden LLM capabilities and deployment options.Source: §3 · pp. 5–6
• The timeline traces advances from GPT-3 and chain-of-thought through tool-calling and agentic loops to skills, which together enabled SE agents and spec-driven development.Source: §3 · p. 6, Figure 1 · p. 6

3.1 Pre-training Limitations

• Pre-training data gaps and cut-off dates constrain performance and require augmented generation, while under-represented tacit and long-tail knowledge raises risks of poor answers and misplaced trust.Source: §3.1 · pp. 6–7

3.2 Symbolic and Statistical Reasoning

• LLMs struggle with logical reasoning, while chain-of-thought gains are task-limited and costly; neuro-symbolic methods can translate requirements into logic for model checking and verification.Source: §3.2 · pp. 7–8

3.3 Context Management

• Context augmentation adds relevant information to address cut-offs, but larger context windows can worsen inaccuracy when needed information is lost in the middle.Source: §3.3 · p. 8

4 Requirements Re-visited

• Requirements are also inputs to coding agents in spec-driven development, while issue reports and feature requests guide post-deployment changes.Source: §4 · p. 8

4.1 Prompt Programming

• Prompt programming treats LLM input as a natural-language program, with requirements expressed as constraints and iteratively revised to improve output.Source: §4.1 · pp. 8–9

4.2 Advances in Agentic AI

• Agentic AI combines LLM advances in a harness of sensors, actuators, and control loops.Source: §4.2 · p. 9

4.2.1 Generalist Agents

• Generalist agents use thought-action-observe loops and tools to act on software tasks; skills package reusable procedures, and multi-agent systems divide generation and synthesis roles.Source: §4.2.1 · p. 9

4.2.2 Long-Horizon Tasks

• Long-horizon tasks risk compounding errors, and live benchmarks assess completion and tool use; LLM-as-a-Judge evaluations can be biased and only slightly better than random guessing.Source: §4.2.2 · pp. 9–10

4.2.3 Software Engineering Agents

• Software engineering agents use repository tools to resolve issues, but hand-crafted pipelines can outperform synthetic data and agentic frameworks.Source: §4.2.3 · p. 10

4.3 Spec-Driven Development

• Spec-driven development gives specs increasing authority over code, but evaluation gaps, executable-specification difficulty, and code-spec drift remain challenges.Workflows range from specs guiding initial implementation to specs maintained alongside code or serving as the source from which code is generated.Source: §4.3 · pp. 10–11

• The section organizes future RE research around software process and the experience of requirements.Source: §5 · p. 11
• The software-process dimension concerns how AI changes the way software is constructed.Source: §5 · p. 11
• The requirements-experience dimension concerns how requirements are collected, analyzed, and realized in software.Source: §5 · p. 11

5.1.1 From Validation to Verification

• Validated requirements can become executable or checkable specifications that test whether generated code, tests, and documentation preserve intended behavior.Behavioral specifications encode pre-conditions and post-conditions; auto-formalization and neuro-symbolic methods may extend verification.Source: §5.1.1 · pp. 11–12

5.1.2 Singular to Long-Horizon Tasks

• Automating isolated RE tasks can fail downstream when the next task is vague or under-specified, leaving a singleton output unfit for purpose.A feature-removal workflow, for example, also needs dependency checks, code changes, and updated requirements documentation.Source: §5.1.2 · p. 12
• Long-horizon benchmarks should test real requirements elicitation, analysis, implementation planning, trace preservation, and satisfiability rather than isolated artifact fluency.The proposed arenas and gyms use real data, measurable objectives, and difficult tasks, with baseline accuracy around 10%.Source: §5.1.2 · p. 12

5.1.3 Innovation in Evaluation

• LLM-based RE needs evaluation beyond overlap metrics and LLM judges, which can reward hallucinated text or produce biased labels.BLEU and ROUGE have known limits, and LLM judges can perform only slightly better than random guessing in certain settings.Source: §5.1.3 · p. 13
• Fluency-dominated human evaluation can mistake plausible requirements or trace explanations for artifacts that preserve intent or satisfy specifications.Fluent text can attract trust and higher factuality ratings regardless of its factual content.Source: §5.1.3 · p. 13
• Requirements evaluation may shift from text-quality judgments toward measuring whether generated artifacts are useful and fit for purpose in long-horizon tasks.Examples of outcome measures include pass@k for issue resolution and exact-match, string-subset, and fuzzy-match metrics for web tasks.Source: §5.1.3 · pp. 13–14

5.1.4 Requirements Evolution & Management

• Spec-driven development treats change requests as proposed specification changes, followed by updates to affected scenarios, constraints, plans, tests, and code.This is especially important in brownfield projects, where existing behavior must be preserved while functionality changes.Source: §5.1.4 · p. 14
• Agent-facing context files must remain minimal, current, and task-relevant, with change-impact methods to detect conflicts, propagate accepted changes, and preserve traceability.Irrelevant context can increase task failure, and vanilla RAG may be insufficient for context management over long-horizon tasks.Source: §5.1.4 · p. 14

5.2 Matters of Experiencing Requirements

• The section frames these shifts as less implementation authorship, easier access to behavioral previews, and more responsive ways for stakeholders to express requirements.The anticipated changes span developers, stakeholders, and users.Source: §5.2 · pp. 14–15

5.2.1 Reductions in Authorship Experience

• As SE agents take on implementation, developers may encounter requirements later in testing and lose some of the technical understanding gained through authorship.The paper argues that developers need transparency into agent design choices so they can review, supervise, and document them.Source: §5.2.1 · p. 15

5.2.2 Previewing Behavioral Experiences

• Lower-cost code generation can make benchmarks, prototypes, and simulations easier to create, helping stakeholders preview and compare behavioral outcomes before requirements are fully explored.Benchmarks assess design choices, prototypes explore integrations, and simulations test scalability and edge cases.Source: §5.2.2 · pp. 15–16

5.2.3 Responsive and Adaptive Specification

• SE agents could let end-users express unmet needs during the moment of experience, including ephemeral needs that must be addressed within minutes or hours.Supporting real-time refactoring requires software designed for modifiability, evolution, and potentially self-adaptivity.Source: §5.2.3 · p. 16

6 Conclusion

• Agentic planning loops combine code generation with observation, feedback, and verification, creating opportunities to reintroduce historically heavyweight RE methods.The paper notes that developers may see RE-related changes across the pipeline.Source: §6 · p. 16
• In software processes, the paper forecasts a shift from validation toward verification, particularly for evaluating automation over long-horizon tasks.Source: §6 · p. 16
• In requirements experience, developers may focus less on code authorship and more on design alternatives and their behavioral effects, while stakeholders gain access to evolving requirements and implementations.Source: §6 · p. 16