What LLM Chatbots Lack: A Sense of Purpose

LLM chatbots are constantly improving, but their progress is measured by benchmarks like MMLU, HumanEval, and MATH, which test one-shot responses. However, real-world tasks such as travel planning or automating GitHub issues require multi-turn dialogue with requirement refinement. Researchers have shown that even advanced models like LLaMA2-chat-70B and GPT-3.5-turbo-16k lose adherence to the system prompt after about 1.6 thousand tokens (approximately 16 rounds). This is confirmed by an experiment where two LLM agents communicated with each other, and then it was checked how well each followed its instruction. The problem is that training via Reinforcement Learning from Human Feedback (RLHF) does not account for multi-step planning—it optimizes one-shot responses rather than long-term goals. The author proposes a "goal-directed dialogue" approach, where the chatbot acts as an assistant that gradually clarifies the user's intentions. Examples of such systems include "solution-oriented dialogue" and the Sotopia platform, which model collaborative decision-making. To create true assistants, we need not just to generate text but to design dialogues as goal-directed games, where each utterance changes the interlocutor's world model.