The Stubborn Prolixity of LLMs: Why Expert Prompts Fail
Coverage of lessw-blog
A recent LessWrong discussion highlights the persistent difficulty of forcing Large Language Models to be concise, even when explicit 'expert' personas and strict constraints are applied.
In a recent post, lessw-blog discusses a growing friction point for advanced users of Large Language Models (LLMs): the inability to effectively curb model verbosity. As LLMs transition from experimental novelties to essential development tools, the efficiency of the interaction has become a critical metric. The author argues that despite utilizing advanced personalization features and explicit system instructions, current models remain excessively "prolix," burying answers in unrequested context.
The central tension described in the post is the failure of "expert" personas. The author details their attempts to configure personalization settings with specific constraints: stating a Computer Science background, claiming 30+ years of programming experience, and explicitly demanding brevity without background explanations. Theoretically, these instructions should signal the model to bypass introductory concepts. In practice, the author reports that models continue to generate "walls of text," often explaining basic principles that the user already understands.
This issue suggests a fundamental rigidity in current model alignment. Most consumer-facing models undergo extensive Reinforcement Learning from Human Feedback (RLHF) to prioritize safety and helpfulness. However, this tuning appears to over-index on comprehensive explanations, making it difficult for the model to adapt to high-context, low-verbosity requests. The post notes that even API-level system prompts often fail to override this behavior. The author humorously notes that when asked for advice on how to fix this, the model itself suggested the user had "done all you can," implying an architectural or fine-tuning limitation.
Among the major platforms, the author identifies Anthropic's Claude as "probably the least annoying" regarding this specific issue, though it still struggles. This discussion is significant for developers and power users because it highlights a lack of granular control. If models cannot reliably adhere to length and tone constraints, it complicates their utility in agentic workflows where precision and token efficiency are paramount.
For those interested in the nuances of prompt engineering and the limits of current model steerability, the full discussion offers a relatable and technical look at the problem.
Read the full post on LessWrong
Key Takeaways
- Personalization settings and 'expert' personas (e.g., claiming a CS degree) frequently fail to prevent LLMs from explaining basic concepts.
- Current RLHF tuning likely prioritizes comprehensive 'helpfulness' over strict adherence to brevity constraints.
- The author notes that Claude appears slightly better at managing verbosity than its competitors, though the issue remains prevalent across the board.
- The inability to control output length impacts the efficiency of using LLMs for professional coding and analysis workflows.