AdSteer
A lightweight decoding-time advertising injection framework that integrates native advertisements into LLM generation by dynamically adjusting token probability distributions, without modifying model weights or rewriting answers.
Timeline (last 90 days)
AdSteer is a lightweight decoding-time advertising framework that does not change model weights or modify input prompts, but instead dynamically adjusts the token probability distribution to incorporate advertising content.
Existing approaches to embedding ads in LLMs mainly follow three paths: prompt-level injection, multi-round rewriting, and fine-tuning.
Prompt injection easily makes model outputs stiff or even off-topic; multi-round rewriting means additional inference passes, noticeably raising computational cost; fine-tuning requires training resources and data preparation, with poor deployment flexibility.
AdSteer adjusts the strength and duration of advertising influence, so that ads are not present in an all-suppressing manner throughout but are integrated with rises and falls and in a measured way.
The paper evaluates on GEM-Bench covering three application scenarios, and AdSteer improves the overall score by an average of 14.1% over the strongest baseline.
Compared with rewriting-based methods, AdSteer reduces the total number of tokens by an average of 22.5%.
The current experiments are based on the GEM-Bench benchmark, and factors such as user tolerance and the diversity of advertiser demands in real product environments still require further verification.
The decoding-time advertising scheme is attractive in terms of cost because it introduces no additional inference rounds and completes ad integration within a single generation with controllable token overhead; in contrast, multi-round rewriting multiplies the number of generations per request, directly driving up serving costs.
AdSteer reframes advertising integration from a "prompt engineering problem" to a "decoding control problem," bringing lower intrusiveness and less additional computation.
All Facts (9)
AdSteer is a lightweight decoding-time advertising framework that does not change model weights or modify input prompts, but instead dynamically adjusts the token probability distribution to incorporate advertising content.
50%UnverifiedExisting approaches to embedding ads in LLMs mainly follow three paths: prompt-level injection, multi-round rewriting, and fine-tuning.
50%UnverifiedPrompt injection easily makes model outputs stiff or even off-topic; multi-round rewriting means additional inference passes, noticeably raising computational cost; fine-tuning requires training resources and data preparation, with poor deployment flexibility.
50%UnverifiedAdSteer adjusts the strength and duration of advertising influence, so that ads are not present in an all-suppressing manner throughout but are integrated with rises and falls and in a measured way.
50%UnverifiedThe paper evaluates on GEM-Bench covering three application scenarios, and AdSteer improves the overall score by an average of 14.1% over the strongest baseline.
50%UnverifiedCompared with rewriting-based methods, AdSteer reduces the total number of tokens by an average of 22.5%.
50%UnverifiedThe current experiments are based on the GEM-Bench benchmark, and factors such as user tolerance and the diversity of advertiser demands in real product environments still require further verification.
50%UnverifiedThe decoding-time advertising scheme is attractive in terms of cost because it introduces no additional inference rounds and completes ad integration within a single generation with controllable token overhead; in contrast, multi-round rewriting multiplies the number of generations per request, directly driving up serving costs.
50%UnverifiedAdSteer reframes advertising integration from a "prompt engineering problem" to a "decoding control problem," bringing lower intrusiveness and less additional computation.
50%