Forget-Me-Not: A Proposal For A Simple Prompting Technique To Prevent Forgetting Information In Long Prompts
3 main points
✔️ Focus on the challenge of "in-context forgetting," in which LLMs forget important information in long contexts
✔️ "Forget-Me-Not (FMN)" improves memory performance by adding a sentence that highlights important information
✔️ FMN improves accuracy by up to +43 points for multiple models, including GPT-4
When Life Gives You Samples: The Benefits of Scaling up Inference Compute for Multilingual LLMs
written by Ammar Khairi, Daniel D'souza, Ye Shen, Julia Kreutzer, Sara Hooker
(Submitted on 25 Jun 2025)
Comments: Published on arxiv.
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
The images used in this article are from the paper, the introductory slides, or were created based on them.
Summary
While LLMs have shown tremendous performance in complex inference and question answering, a phenomenon known as "in-context forgetting" has been a challenge. This is where the model gradually forgets information given in a prompt over a long context, which can negatively impact accurate response generation. This study focuses on this in-context forgetting and examines in detail under what conditions current LLMs "forget" information.
In addition, the authors proposed a simple prompt design tool called Forget-Me-Not (FMN). This tool is a mechanism to remind models of important information by means of one-sentence instructions inserted in context. Notably, it does not change the structure of the LLM and does not require any training, yet it significantly improves the model's contextual memory capacity. Experiments have reported cases in which FMN insertion improves the accuracy by up to 40 points or more, even for state-of-the-art models such as GPT-4.
Proposed Methodology
The central proposal in this study is a prompt-based auxiliary sentence insertion method called Forget-Me-Not (FMN).
FMNs are brief natural language sentences that "remind" the model once again of important information that tends to get buried in longer contexts. For example, a sentence such as "This instruction is very important: Your answer to the question should be based solely on this sentence" can be added immediately after the original instruction to achieve the desired effect.
FMN is designed to emphasize specific information in the model while minimizing the number of tokens. What is unique is that it does not make any changes to the internal structure or parameters of the model, but only adjusts the wording in the prompt to improve performance. FMN also serves to compensate for ambiguity in the model's understanding of context, since it explicitly indicates "which information is important" in natural language.
Furthermore, FMN can be used in conjunction with other contextual memory support methods such as few-shot learning (few-shot learning), and is designed as a versatile approach. As a result, it can mitigate contextual forgetting more simply and effectively than conventional methods.
Experiments
The authors designed and performed a total of 10 contextual forgetting tasks on several LLMs (GPT-3.5, GPT-4, Claude 2, Gemini Pro, etc.) to evaluate in-context forgetting and to test the effectiveness of FMN. Each task measured in detail whether the model was able to retain specific information in context and how much the insertion of FMN improved its accuracy.
Experimental results showed a significant performance improvement of up to +43 points for GPT-4 and +37 points for Claude 2. In addition, trends consistent with the model's memory characteristics were confirmed, such as "the more backward the location of the information, the greater the effect," and "the greater the distance from the question, the more effective the FMN.
Furthermore, to evaluate the impact of FMN on the overall prompt, interference with other information in the prompt and the risk of misdirection were also examined, but in the majority of cases no negative effects were found, rather a stable performance improvement was reported. In particular, it is noted that in GPT-4, FMN almost always had a positive effect.
Based on these results, we conclude that FMN is a practical method that can easily and effectively mitigate in-context forgetting.
Categories related to this article
