Large Language Models (LLMs) are increasingly deployed in business-critical domains such as finance, education, healthcare, and customer support, where users expect consistent and reliable recommendations. Yet LLMs often exhibit variability when prompts are phrased with minor differences, even when semantically equivalent. Such inconsistency undermines trust, complicates compliance, and disrupts user experience. While personalization is desirable in certain contexts, many enterprise scenarios-such as HR onboarding, customer support, or policy disclosure-require invariant information delivery regardless of phrasing or prior conversational history. Existing approaches, including retrieval-augmented generation (RAG) and temperature tuning, improve factuality or reduce stochasticity but cannot guarantee stability across equivalent prompts. In this paper, we propose a reinforcement learning framework based on Group Relative Policy Optimization (GRPO) to directly optimize for consistency. Unlike prior applications of GRPO, which have been limited to reasoning and code generation, we adapt GRPO to enforce stability of information content across groups of semantically equivalent prompts. We introduce entropy-based helpfulness and stability rewards, treating prompt variants as groups and resetting conversational context to isolate phrasing effects. Experiments on investment and job recommendation tasks show that our GRPO-trained model reduces variability more effectively than fine-tuning or decoding-based baselines. To our knowledge, this is a novel application of GRPO for aligning LLMs toward information consistency, reframing variability not as an acceptable feature of generative diversity but as a correctable flaw in enterprise deployments.
Large Language Models (LLMs) such as Llama-3 are increasingly deployed in domains requiring decision support and recommendations. A critical requirement in these deployments is that the AI provides consistent and reliable outputs regardless of how a user phrases a prompt. Consistency is not merely a technical consideration-it underpins trust, usability, and compliance in business applications. In practice, organizations depend on stable outputs to ensure operational reliability, regulatory adherence, brand integrity, and user satisfaction. At the same time, consistent behavior is also essential to safeguard fairness and prevent systemic harms.
It is important to acknowledge that consistency is not always universally desirable. In some settings, such as personalized learning platforms or health coaching, users benefit when the system tailors its responses to their unique profiles and histories. Here, variation in outputs can reflect meaningful personalization rather than unreliability. However, there are many business-critical scenarios where consistency must prevail regardless of how well the LLM knows the user or their prior interaction context. For example, the use of LLMs for answering organizational policy-related questions, personal financial planning questions, and educational planning questions are common applications. In such applications, providing a consistent answer is important in building the user trust on LLM based answers. In these cases, personalization should not alter the essential information content of the response. For example, in human resources onboarding, new employees should always receive the same explanation of company policies; in customer support, answers about product warranty coverage should remain unchanged no matter how the question is phrased; and in compliance-driven settings, financial disclosures or insurance terms must be delivered identically to every user. These are cases where personalization may add conversational nuance but should never compromise the consistency of core information.
Solutions such as RAG [1] have been proposed as a way to support consistency by grounding answers in external knowledge. While effective in many contexts, RAG does not fully eliminate inherent inconsistencies in LLM behavior. For instance, consider two applicants preparing for the same job interview and querying an LLM-powered assistant: one asks, “What are the most common data structures I should review?” while another asks, “Which data structures should I prepare for in coding interviews?” Even when both queries are backed by the same retrieved documents, the LLM may generate divergent lists or emphasize different topics. Such inconsistencies create confusion and reduce trust in the system, especially in high-stakes contexts like interview preparation. Literature has often embraced this variability as an acceptable property of generative models, framing it as diversity in responses [2] and proposing workarounds such as RAG or temperature tuning to control stochasticity. However, this acceptance does not resolve the fundamental issue. Simply lowering temperature or relying on retrieval does not guarantee that semantically equivalent prompts will produce consistent outputs. For many business and educational applications, overlooking this inconsistency is untenable.
Across industries such as finance, education, healthcare, and customer services, unpredictable or inconsistent responses can have significant consequences and may not help build the required user trust in using such systems. A bank providing different disclosures depending on phrasing, risks compliance failures; a chatbot that answers the same question differently for two customers undermines confidence in customer support; and in educational or hiring contexts, demographic attributes such as gender introducing unintended inconsistencies result in ethical issues. In particular, job recommendation scenarios highlight both the operational and ethical risks of inconsistency and demonstrate the need for robust methods to enforce stable and equitable outputs.
While contextual methods such as RAG can help improve consistency by grounding model responses in relevant documentation, their applicability depends on whether contextual retrieval is available at query time. In enterprise deployments, RAG can reduce hallucinations and tie answers to authoritative sources, thereby increasing factual consistency. However, many user interactions occur without such context-when individuals query general-purpose assistants directly, without attached documents or retrieval layers. In such settings, the model must still produce internally consistent responses across semantically equivalent prompts, regardless of who is asking the question. In this paper, we focus on this latter class of scenarios-direct, context-free user interactions-and leave the extension to contextual or retrieval-grounded querying as a direction for future work.
In addi
This content is AI-processed based on open access ArXiv data.
