Factors That Support Grounded Responses in LLM Conversations: A Rapid Review

TL;DR

This rapid review addresses how LLMs can be aligned with conversational goals, grounded in context, and protected against hallucinations. It maps techniques across inference-time, post-training, and RL-based life-cycle stages, highlighting efficient inference-time methods that align without retraining. The review synthesizes 23 studies to compare approaches such as prompt-based alignment, decoding-time controls, contrastive fine-tuning, and reward-based optimization, and discusses their computational trade-offs and grounding capabilities. The findings offer a practical roadmap for deploying aligned LLMs in real-world conversations while identifying gaps for future research.

Abstract

Large language models (LLMs) may generate outputs that are misaligned with user intent, lack contextual grounding, or exhibit hallucinations during conversation, which compromises the reliability of LLM-based applications. This review aimed to identify and analyze techniques that align LLM responses with conversational goals, ensure grounding, and reduce hallucination and topic drift. We conducted a Rapid Review guided by the PRISMA framework and the PICO strategy to structure the search, filtering, and selection processes. The alignment strategies identified were categorized according to the LLM lifecycle phase in which they operate: inference-time, post-training, and reinforcement learning-based methods. Among these, inference-time approaches emerged as particularly efficient, aligning outputs without retraining while supporting user intent, contextual grounding, and hallucination mitigation. The reviewed techniques provided structured mechanisms for improving the quality and reliability of LLM responses across key alignment objectives.

Figures (1)

• Figure 1: Filtering steps and Database. This image was generated using PRISMA guidelines