Cutting-edge AI breakthrough eliminates prompt fatigue, boosts efficiency

AI agents today struggle with efficiently mastering multiple tasks due to their heavy reliance on prompts. The traditional method of embedding task-specific hints within prompts results in long and complex instructions that AI must process at every step. As the number of tasks increases, these prompts grow larger, making AI agents slower and less efficient.

The challenge of enabling artificial intelligence (AI) agents to master multiple complex tasks efficiently has long been a limitation in AI research. Current large language model (LLM)-based AI agents depend on prompts to store and retrieve knowledge, much like a person relying on written notes to compensate for memory loss. While this approach allows AI to function across various scenarios, it creates a dependency on ever-expanding prompts that increase computational costs and degrade performance.

A new study proposes a groundbreaking alternative - Memento No More (MNM) - which allows AI agents to internalize knowledge instead of relying on prompts. This novel method helps AI improve through iterative human feedback, enabling it to refine its abilities over multiple training rounds. The study, titled "Memento No More: Coaching AI Agents to Master Multiple Tasks via Hints Internalization", was conducted by Minttu Alakuijala, Ya Gao, Georgy Ananov, Samuel Kaski, Pekka Marttinen, Alexander Ilin, and Harri Valpola from Aalto University and System 2 AI. It is currently under review, and a preprint version is available on arXiv. 

Overcoming the limitations of prompt-based AI agents

AI agents today struggle with efficiently mastering multiple tasks due to their heavy reliance on prompts. The traditional method of embedding task-specific hints within prompts results in long and complex instructions that AI must process at every step. As the number of tasks increases, these prompts grow larger, making AI agents slower and less efficient. This challenge is comparable to a person with anterograde amnesia who depends on written notes to function but becomes overwhelmed when the notes become too extensive to manage effectively.

The MNM framework addresses this issue by internalizing hints directly into the AI model’s neural weights, eliminating the need for ever-expanding prompts. Instead of processing long prompts for every task, the AI undergoes iterative training, absorbing human-provided hints into its core learning structure. This process is achieved through context distillation, which gradually refines the AI’s decision-making abilities while improving its efficiency. As a result, AI agents trained with MNM become less dependent on external instructions and are better equipped to handle complex tasks autonomously.

How MNM trains AI agents more effectively

MNM employs an iterative hint internalization process, which enhances AI learning through structured training rounds. In the first phase, the AI agent is introduced to a set of training tasks with minimal guidelines, including basic tool descriptions and formatting instructions. Human experts monitor the AI’s behavior, identify errors, and provide corrective hints. Unlike conventional methods where hints remain in the prompt, MNM integrates these corrections directly into the AI’s neural network using context distillation. This is facilitated by a LoRA adapter, which allows the model to internalize feedback without requiring explicit prompt modifications.

With each subsequent training round, the AI undergoes performance evaluations where human experts review its mistakes, refine the corrective hints, and retrain the model. This iterative approach ensures that the AI agent continuously improves its ability to complete tasks without external hints. Over multiple rounds, the AI progressively eliminates the need for manual intervention and becomes increasingly proficient at solving complex problems. By the final stage, the AI model operates independently, demonstrating high accuracy and efficiency while using significantly fewer computational resources than traditional prompt-based systems.

Benchmark Results: Outperforming leading AI models

The MNM framework was tested against ToolQA, a benchmark designed to evaluate AI performance in tasks involving information retrieval, tool use, and question answering. To measure its effectiveness, the researchers trained a Llama-3-based AI agent using MNM and compared it against leading AI models such as GPT-4o and DeepSeek-V3. The results showed a remarkable improvement in AI efficiency and accuracy.

After just three training rounds, the MNM-trained agent achieved an impressive 97.9% success rate, significantly outperforming GPT-4o (92.8%) and DeepSeek-V3 (87.5%). One of the key advantages of MNM was its ability to complete tasks using only 7–9% of the token count required by other models, drastically reducing computational overhead. This optimization led to 4x faster inference speeds and lower operational costs. Moreover, the MNM-trained agent demonstrated exceptional generalization capabilities, maintaining high accuracy even when presented with new, previously unseen tasks. These findings highlight the potential of MNM to create more efficient, scalable, and cost-effective AI systems by shifting from prompt-heavy learning to internalized knowledge retention.

The future of AI learning: Scalable, efficient, and adaptive

The success of MNM presents a transformative shift in AI training methodologies. By removing the reliance on expanding prompts, this approach significantly reduces computational costs and enhances inference efficiency, making AI more scalable for real-world applications. The ability to refine decision-making through structured human feedback ensures that AI agents can continuously improve without suffering from information overload. Industries such as automation, research, education, and professional services stand to benefit immensely, as AI systems trained with MNM can handle diverse and complex tasks with greater autonomy.

Additionally, integrating structured feedback into AI training fosters greater trust and reliability in AI-driven decision-making. The MNM method enables AI to become more adaptable, allowing it to evolve and improve without requiring frequent manual intervention. As researchers continue to refine this approach, its applications are expected to expand across various fields, marking a major advancement in the development of intelligent AI agents. This study paves the way for a new era in AI learning, where internalized knowledge replaces extensive prompts, setting the foundation for more efficient and intelligent AI systems.