What's Missing From LLM Chatbots: A Sense of Purpose
The Gradient Kenneth Li
Large language model chatbots are evaluated primarily on static benchmarks like MMLU and HumanEval, but these metrics fail to capture performance in multi-turn interactive conversations where users have specific goals. Research shows that models like GPT-3.5-turbo and LLaMA2-chat lose instruction adherence after approximately 1.6k tokens (around 8 dialogue rounds), despite having context windows up to 100k tokens, because current training methods lack explicit goal-directed dialogue optimization. To build more effective human-AI collaboration systems, LLM training needs to incorporate purposeful dialogue frameworks that align model behavior with user intentions across extended conversations, rather than relying solely on system prompts and one-shot instruction following.
Why it matters
LLM-based chatbots’ capabilities have been advancing every month. These improvements are mostly measured by benchmarks like MMLU, HumanEval, and MATH (e.g. sonnet 3.5, gpt-4o). However, as these measures get more and more saturated, is user experience increasing in proportion to these scores? If we envision a future
