LLM routing based on token count is flawed, mirroring human biases

A recent study suggests that routing Large Language Models (LLMs) based solely on token count is an ineffective proxy for complexity, potentially sending difficult tasks to less capable models. This approach mirrors human biases where processing effort, rather than content, influences judgment of intelligence and credibility. The research highlights that stripped-down language, while shorter, is not simpler and can be misinterpreted by both machines and humans, leading to inaccurate assessments of capability and potentially negative interactions. AI

IMPACT Misinterpreting LLM complexity based on token count can lead to inefficient resource allocation and suboptimal performance.

RANK_REASON The item is an opinion piece discussing research findings and their implications, rather than a direct announcement or release.

Read on dev.to — LLM tag →

• arXiv:2609.17542
• arXiv:2609.38788
• Cognition 2026
• doi:10.1002/acp.1178
• doi:10.1016/j.cognition.2026.106697
• doi:10.1073/pnas.2415254122
• Oppenheimer
• PNAS 2025
• Vashiel

• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →