LLM Chatbots: The Great Void of Purpose in Conversation

What happened?

Large language models (LLMs) have achieved impressive advances on benchmarks like MMLU, HumanEval, and MATH. For instance, GPT-4o and Claude 3.5 Sonnet have scored close to 90% on MMLU, and on HumanEval they exceed an 80% success rate. However, as an article in The Gradient points out, these improvements do not necessarily translate into a better user experience. The reason: benchmarks measure single-turn capabilities, while real human interaction is multi-turn and purpose-oriented. In fact, the saturation of these benchmarks suggests that a new evaluation paradigm is needed. The Gradient article notes that most current tests are non-interactive, ignoring the collaborative nature of human communication.

Why is it important?

Purposeful dialogue refers to multi-round conversations focused on a goal: from being a travel agent to a virtual therapist. In travel planning, for example, conveying all preferences in a single message is costly; instead, iterative exchange allows negotiation and refinement. As Terry Winograd said:

“All use of language can be thought of as a way of activating procedures in the listener.”

Each utterance is a deliberate action to alter the other's model of the world. In human-AI collaboration, this is essential. Negotiation theory supports that iterative bargaining yields better results than an all-or-nothing offer. Moreover, in areas like customer service, purposeful dialogue allows solving complex problems without overwhelming the user with initial questions. A Gartner study suggests that by 2025, 80% of customer service interactions will be managed by AI, but without goal-oriented dialogue, user frustration will increase.

Consequences for the future

The lack of purpose limits critical applications like code generation. Benchmarks like SWE-bench show that solving GitHub issues requires bidirectional communication: the AI must ask, confirm requirements, and request help. Without iterative dialogue, full automation is unfeasible. In fact, on SWE-bench, current models solve less than 20% of problems without human intervention. Additionally, turn-taking allows building long-term memory and user profiles, like a personal assistant that learns preferences and summarizes news. Companies like Google and Microsoft are already investing in conversational assistants with persistent memory, but they still lack a true sense of purpose. The next frontier is not just accuracy, but the ability to maintain goal-oriented conversations. This also has ethical implications: dialogue without purpose can lead to misunderstandings or users attributing wrong intentions to the system.

What readers should know

• Current benchmarks are insufficient to measure the quality of human-AI interaction. Interactive metrics are needed, such as the number of turns to complete a task or user satisfaction in multi-turn dialogues.
• Purposeful dialogue is fundamental for applications like virtual assistants, customer service, and pair programming. For example, GitHub Copilot already offers context-aware suggestions but cannot hold a conversation to refine requirements.
• The next frontier for chatbots is not just accuracy, but the ability to maintain goal-oriented conversations. Research like Anthropic's on constitutional models aims to align dialogue with human intentions, but a sense of purpose is still missing.

In summary, the true potential of LLMs will not be realized until they integrate a sense of purpose into dialogue, enabling genuine collaboration rather than just one-way responses. The industry must rethink how it evaluates and designs these systems, prioritizing iterative interaction and understanding long-term goals.