LLM Trojans that increase response lengths and cost
Original title: LLM Trojans that Increase Response Lengths and Cost
The authors also describe a defense and a trigger that fires on semantically similar instructions.
The paper describes a data-poisoning Trojan for instruction-tuned LLMs that lengthens responses, and therefore token cost, when a trigger is present in the prompt. It says some frontier providers charge five times more per response token than per prompt token, while others discount prefilled prompt tokens. The authors present a defense and say a subtle Trojan can activate on semantically similar instructions.
Unconfirmed: The abstract names no models, datasets, or measured cost increase.
- The Trojan is planted by data poisoning and activates from the prompt.
- The trigger can match semantically similar instructions, not only an exact string.
- The stated effect is longer responses and higher user token cost.
- The paper says it devises a defense against this attack.
From the original article
1 min read at arxiv.orgManagement of the token-based costs of using LLMs has emerged as a significant problem. Some frontier-model providers charge five times per response token than per prompt token, while others offer prefilling discounts for prompt tokens.
By Zhengxing Li · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.