Answers are grounded in this paper and cite their sources.
Analysis quantifies economic costs of token inflation in AI, suggesting improved training metrics for efficiency.
We aggregate published measurements of RLHF-induced response length inflation across the literature and compute the first industry-scale estimate of its economic cost. Alignment training systematically inflates output length: sentences triple after SFT, DPO doubles response length within the first 10% of training, and on one benchmark 98% of PPO reward improvement is attributable to length alone. Verbosity compensation rates range from 13.6% to 74.2% across 14 models, and output tokens cost 4-8x more than input tokens across all frontier providers. Combining published verbosity rates, real-world token volumes, and current API pricing, we estimate the annual verbosity premium at 500M to 1.8B, with a central estimate of 1.2B (approximately 14% of total industry inference spend). We survey 12 training-side mitigations and show that all target response length rather than information density. A 500-token response with 50 atomic facts is efficient; the same length with 10 facts restated five ways is waste. Length penalties cannot distinguish these cases. Drawing on rate-distortion theory and evidence that factual precision degrades with response length, we argue the correct optimization target is information density (supported facts per token) and present two concrete density-aware reward formulations.
No takes yet. Share an insight, caveat, or question.
John N. Kearney (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: