ARXIV Long Term Memory - The Foundation of AI Self-Evolution.pdf

ARXIV Long Term Memory - The Foundation of AI Self-Evolution.pdf

@aaaal

3 Followers

1 week ago 442

Tech #ai #producthunt #memory

ARXIV Long Term Memory - The Foundation of AI Self-Evolution.pdf

@aaaal1 week ago

Long Term Memory : The Foundation of AI
Self-Evolution
Xun JIANGµθ Feng LIθ∗ Han ZHAOθ∗ Jiaying W…

1/56

Figure 1: The overview of LTM and AI Self-Evolution.
1 Introduction
Artificial Intelligence (AI) …

2/56

• Phase 1: Cognitive Accumulation in the Physical World. Data accumulation is the first
and vital …

3/56

how to effectively integrate long-tail individual data into the model’s memory mechanisms, hoping t…

4/56

• Data Framework for LTM. To implement long-term memory, we developed a data collection, analysis,…

5/56

2.1 Definition of AI Self-Evolution
Definition: AI self-evolution refers to the process by which A…

6/56

significant breakthroughs in large-scale multi-agent collaboration, it is crucial to rely on agent
…

7/56

2.2.3 Self-Correction and Evolution Mechanism
To achieve AI self-evolution, models must possess a …

8/56

knowledge. Realizing this potential may be the key to overcoming the current limitations of LLMs,
…

9/56

isolation; they contribute their insights to the broader LTM collective, creating a shared knowledg…

10/56

outcomes: either the model fails to retain individual data, or it overfits and forgets previouslys…

11/56

Figure 3: The difference between human learning and growth versus artificial intelligence training
…

12/56

3.5 The Importance of LTM for AI Systems
3.5.1 LTM in Single Models
LTM is crucial for diversity …

13/56

Figure 4: Comparison of LLM Memory and Long-Term Memory
memory, enabling the model to recognize pa…

14/56

deficiency makes the model fully simulate human cognitive processes, resulting in outputs
that may…

15/56

data privacy and security issues are paramount, and it is essential to protect user data during col…

16/56

agent-based interaction frameworks collectively forms a robust methodology for LTM in personalized
…

17/56

These approaches transform raw data into different forms of long-term memory through various
techn…

18/56

increasingly large context windows (ranging from 4K to 1M tokens or more), can hold significantly
…

19/56

Besides, post-retrieval mechanisms such as re-ranking and self-reflection, as seen in models
like …

20/56

output (Y ) pairs. LTM data can also be reorganized in this format and be encoded into LLMs. For
L…

21/56

window to incorporate all necessary information during prompt construction. However, decoder-based
…

22/56

Figure 5: How personal data can be utilised through textual and parametric memory retrieval
specia…

23/56

process [131] can be integrated into model generation, guiding the model through complex reasoning
…

24/56

1. Medical Records Accumulation. In this work, the doctor agent first accumulates LTM
from success…

25/56

6.1.1 Real-world LTM Data Collection
Developing robust data collection and labeling protocols is c…

26/56

6.1.2 Synthetic LTM Data Acquisition
Doctor Agent Patient Agent
Patient Experience
Knowledge Gra…

27/56

Dataset Professionalism Communication (i) Communication (ii) Fluency (i) Fluency (ii) Similarity Sa…

28/56

into several explicit reasoning processes, including: 1) Determining whether the current
paragraph…

29/56

6.2 Practice on LTM Data Utilization
6.2.1 LTM Utilization with SFT and RAG
We used the private L…

30/56

Table 5: Experimental results of MedAgent-Zero and LTM-enhanced MedAgent-Zero in the MedQA
subset …

31/56

Figure 7: The Tertiary Memory Structure. The utterance of the diagnosis conversation will be
store…

32/56

We conducted series of experiment to assess the impact of these memory types on depression diagnosi…

33/56

learning steps. This innovative approach allows the model to continue learning even during test tim…

34/56

Figure 9: Left: PPL on test tokens after inference on Chinese books. Right: PPL on test
tokens aft…

35/56

Future work could focus on developing more sophisticated instantiations of f, particularly in the
…

36/56

and data. By offering a standardized memory abstraction mechanism, the Omne framework allows
devel…

37/56

With these built-in components, Omne Assistant allows developers to focus on implementing custom
f…

38/56

Figure 12: GAIA leaderboard in validation set
Figure 13: GAIA leaderboard in test set
7 Our Futur…

39/56

7.2 How to Design New Model Architectures for LTM?
One perspective is that as models progress, the…

40/56

How can long-term memory guide pruning? Can agents identify inefficient or unfeasible paths
based …

41/56

Nevertheless, in complex task scenarios, on one hand, collecting a large amount of high-quality
gr…

42/56

Looking forward, model personalization, inspired by human diversity, could drive a new level
of in…

43/56

References
[1] Min Xu, Jeanne M. David, and Suk Hi Kim. The fourth industrial revolution: Opportun…

44/56

[18] E. Talvitie. Self-correcting models for model-based reinforcement learning. In Proceedings of
…

45/56

[35] Kun Zhao and Ginestra Bianconi. Social interactions model and adaptability of human behavior.
…

46/56

[53] R. Wang, W. Zhou, and M. Sachan. Let’s synthesize step by step: Iterative dataset synthesis w…

47/56

[70] Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian
Min, Beic…

48/56

Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Franco,
Aparajita Sara…

49/56

[74] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee …

50/56

[89] Hao Li, Chenghao Yang, An Zhang, Yang Deng, Xiang Wang, and Tat-Seng Chua. Hello again!
llm-p…

51/56

[105] Dawei Zhu, Nan Yang, Liang Wang, Yifan Song, Wenhao Wu, Furu Wei, and Sujian Li. Pose:
Effic…

52/56

[119] Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Le Hou, Kevin Clark,
Ste…

53/56

[132] Yanyang Li, Shuo Liang, Michael R. Lyu, and Liwei Wang. Making long-context language
models …

54/56

Niklas Muennighoff, Fares Obeid, Atsushi Saito, Guangyu Song, Haoqin Tu, Stanisław Wo´zniak, Ruich…

55/56

A RTG prompt
Figure A shows the prompt we use to perform RTG.
{contexts}
{question}
Answer the …

56/56

@aaaal

Drag Me

Comments 0 Info

ARXIV Long Term Memory - The Foundation of AI Self-Evolution.pdf

@aaaal

Tech

#ai #producthunt #memory

This website uses cookies to improve your experience. By using this site, you consent to the use of cookies.