Personal LLM Agents: Insights and Survey about the Capability,...

Since the advent of personal computing devices, intelligent personal assistants (IPAs) have been one of the key technologies that researchers and engineers have focused on, aiming to help users...

https://arxiv.org/abs/2401.05459

Introduction

• we focus primarily on the aspects related to “personal” parts within Personal LLM Agents, encompassing the analysis and utilization of users’ personal data, the use of personal resources, deployment on personal devices, and the provision of personalized services

• Some approaches have attempted to automatically learn to support tasks through supervised learning or reinforcement learning. However, these methods also rely on a substantial amount of manual demonstrations and/or the definition of reward functions.

The main content and contributions of this paper can be summarized as follows:

1. We summarize the status quo of existing IPAs in both industry and academia, while analyzing their primary limitations and future trends in the LLM era.

2. We collect insights from senior domain experts in the area of LLM and personal agents, proposing a generic system architecture and a definition of intelligence levels for personal LLM agents.

3. We review the literature on three important technical aspects of personal LLM agents, including foundamental capabilities, efficiency, and security & privacy.

A Brief History of Intelligent Personal Assistants

Timeline View of the Intelligent Personal Assistants History

• Speech Recognition System → Voice-based Software → Virtual Personal Assistant →LLM-based Chatbot

Technical View of the Intelligent Personal Assistants History

• Template-based Programming：
Given a user command, the agent first map the command to the most relevant template, then follow the predefined steps to complete the task.

• Supervised Learning Methods
how to learn a representation of software GUI and how to train the interaction model

• Reinforcement Learning Methods
During the interaction, the agent gets feedback of rewards that indicate the progress of task completion, and it gradually learns how to automate the tasks by maximizing the reward payoff.To train RL-based task automation agents, a reward function that indicates the progress towards task completion is required.

• Early Adoption of Foundation Models
Let LLMs use tools autonomously to accomplish complex tasks.
Some commercial products have attempted to integrate LLM with IPA.
Many issues related to efficiency, security and privacy have not been adequately addressed yet.

Personal LLM Agents:Definition & Insights

Key Components

The combination of these key components is analogous to an operating system.

Main components of Personal LLM Agents	traditional operating systems
Foundation Model	kernel
Local recourse	driver programs
User context and memory	program contexts and system logs
Skills	software application

 

Intelligence Levels of Personal LLM Agents

Opinions on Common Problems

what features are desired for an ideal IPA?

what are the most urgent technical challenges?

• Intelligent: 1️⃣Multimodal Support 、2️⃣Context Understanding and Context-aware Actions 、3️⃣Enhancing Domain-specific Abilities of Lightweight LLM

• Performance: 1️⃣Effective LLM Compression or Compact Architecture2️⃣Practical Local-Remote Collaborative Architecture

• Security & Privacy: 1️⃣Data Security and Privacy Protection 2️⃣Inference Accuracy and Harmlessness

• Personalization & Storage:efficient data storage solutions to manage and leverage userrelated data

• Traditional OS Support:LLM-friendly interfaces and support (API) of traditional operating systems like Android

Fundamental Capabilities

Task Automation Methods

1. code-based Task Automation

• Slot-filling method : slot-value pairs

• Program synthesis method
• fine-tune LLMs to use specific APIs
• utilize the chain reasoning and in-context learning ability of LLMs→show descriptions and demonstrations of the tools (e.g. APIs, other DNNs, etc.) in context

2. UI-based Task Automation

• text-based GUI representation

• Multimodal representation

Autonomous Agent Frameworks

Evaluation

• Metrics and Benchmarks

Context Sensing

• Hardware-based:acquisition through various sensors

• Software-based:a form of software-base sensing(typing habits)

Sensing sorces:

• Hardware Sensor,Software Sensor,Combination of Multiple Sensors

Sensing Targets:

• Sensing the Environment
• Scene sensing:more tangible environmental factors(locations and places)
• Occasion perception:deeper environmental information(religious and cultural backgrounds…)

• Sensing the User
• Short-term user sensing
• Long-term user sensing

Memorizing

Obtaining Memory

Managing and Utilizing Memory

• Raw Data Management and Processing：selecting, filtering, transforming to other formats, etc

• Memory-augmented LLM Inference:
• Short-term memory : the form of symbolic variables during the current decision cycle.This includes perceptual inputs, active knowledge and other core information from memory or the previous.
• Long-term memory : experiences from earlier decision cycles. This includes history event flows, game trajectories from previous episodes, interaction information.

• Agent Self-evolution:
• Learning Skills: skills as code or APIs through the strategic use of prompts.Agents could acquire novel skills by linking skills within a foundational skill set.
• Finetuning LLM:
reasons:
1. LLMs were not specifically designed for agent-specific use cases.
2. Limited device makes it difficult to acquire new skills through prior knowledge and in-context learning abilities.
3. New knowledge and tools change the task schemas,demanding adaptation of LLMs.
Fine-tuned smaller LLMs could outperform prompted larger LLMs with reduced inference and expense in specific case.

Effiency

• Inference(bottleneck):a lot of both computation and memory resources

• Customization:different context tokens or tune the LLM with domain-specific data

• Memory manipulation:require access to longer contexts or external memories,handle and managemeny.

Efficient Inference

Model Compression

• Quantization:reduce the model size by using fewer bits to represent the model parameters, and also reduces computations with system-level support for quantized kernels.

post-training quantization(PTQ) : quantize after train(available and flexible) quantization-aware training(QAT) : train after quantization

• weight only quantization(WOQ):integer quantization(INT4 and INT8) on weights only, while preserving activations in float formats (FP16 and FP32).
• INT8 quantization for both weights and activations: the activations including KV pairs are more difficult to quantize because of outliers.
• low-bit floating point quantization(FP4 and FP8): higher computational performance

• Pruning: removing less important connections in the network

Structured pruning: removes weights in regular patterns, such as a rectangle block Unstructured pruning: eazier to maintain model accuracy but less hardware-friendly

• Knowledge Distillation(KD): using a well-performing teacher model (a large number of parameters and high precision) to guide the training of a lightweight student model(fewer parameters and lower precision)

White-box: require teacher model’s parameters Black-box: don’t require

KD is also adopted in QAT and pruning techniques to enhance the training performance.

• Low-rank Factorization: Low-rank Factorization can be combined with quantization and pruning.

Inference Acceleration

• KV Cache：storing (i.e., “caching”) and incrementally updating the Key-Value (KV) pairs in each token’s generation.

• Context Compression:
1. Co-quantization of weights and activations, including KV cache.
2. compress the context at the prefill stage based on different importance of tokens(one-shot and cannot prune the KV cache).
3. a learnable mechanism to continuously determine and drop uninformative tokens.
4. reduce computations of less important tokens instead of directly removing them.

• Kernel Optimization:
Small-batch or single-batch inference is especially important for edge scenarios.
The complexity of attention scales quadratically with the sequence length, while that of the FFN scales linearly.
• efficient attention kernels including FlashAttention and Flashdecoding++
• reduce the computational complexity of attention from the algorithm aspect and achieve linear complexity for self-attention in the prefill phase.
• reduce dequantization overhead.

• Speculative Decoding:
An effective approach in small-batch inference to improve the latency.
Speculative decoding mitigates this challenge by “guessing” several subsequent tokens through a lightweight “draft model”, and then validating the draft tokens in batches using the large “oracle model”.

Energy Optimization

software perspective: model compression,KV cache,efficient attention kernels hardware perspective: utilize efficient processors including NPUs and TPUs/FPGA-based.

Efficient Customization

Context Loading Efficiency:

• A straightforward way is to prune some redundant tokens or shorten the context length.

• Another way is to reduce the bandwidth consumption during context data transmission.

• different input prompts may have overlapping text segments,pre-computing,storing and reuse these segments.

Fine-tuning Efficiency:

• Parameter-efficient fine-tuning (PEFT)

• Efficient Optimizer Design(skip)

• Training Data Curation
a small amount of high-quality data can lead to significantly reduced training cost and achieve capabilities comparable to large-scale datasets and models.

Efficient Memory Manipulation

Search Efficiency

• Typical Indexing Algorithms:partitioning methods include randomization→data structures

• Hardware-aware Index Optimization：the utilization of disk-based indexes or the co-design of hardware and algorithms

• Search Mechanism Design：
Multiple similarity criteria can be employed to evaluate vector similarity.Rule-based or estimated-cost-based methods configured offline are often employed to determine the optimal search plan.
Combine vector search with metadata filters.

• Search Process Execution:
Several hardware acceleration methods,multi-threading,multi-core parallelism,GPU,distributed clusters.

Workflow Optimization

• Pipelining
RaLMSpec: enable a local cache for speculative retrieval,cache prefetching, optimal speculation stride scheduler, and asynchronous verification.
PipeRAG:adjust the vector search space depending on the latency expectation of the upcoming token.

• Caching
Select reason: the temporal and spatial locality of retrieved documents.use knowledge tree to organize in the GPU and host memory hierarchy.
the embedding and generative model are equivalent→Query Caching or Query-Doc Caching and save computation overhead

Security and Privacy

Confidentiality

Local Processing

Secure Remote Processing

• Homomorphic encryption(HE):employ encryption to encode in the client→server conducts inference on the ciphertext→decryption
Challenge:certain operations in the LLMs, such as max, min, and softmax, cannot be accurately performed using HE;inference speed is slow.

Data Masking

• hide or replace sensitive content

• embedding-based data anonymization approaches(outperform HE but more risky)

Information Flow Control

Intergrity

Adversarial Attacks

• adversarial defense, abnormal input detection, input preprocessing, output security verification, and more.

• adversarial training through parameter-efficient fine-tuning.

Backdoor Attacks

• through data poisoning , inserting maliciously modified samples into the model’s training data, enabling the model to learn deliberate hidden decision logic

• modify the model input during the test time

• modify the prompts, essentially fine-tunes the model’s parameters and thus alters its decision logic

• distill benign knowledge from poisoned pre-trained encoders and transfer it to a new encoder

Prompt Injection Attacks

• ensure the transparency and security of the LLM’s prompts.

• distinguish the third-party from the system’s inherent promptsRe.

Reliability

Problems

• Hallucination:incorrect answers,coherent and fluent but ultimately erroneous.

• Unrecognized Operation:agent is required to execute actions,which have higher requirements for the format and executability of their outputs.

• Sequential Reliability:LLMs are initially pre-trained on sequential data,while problems in the real world may not be fully addressed sequentially.

Improvement

• Alignment:
the use of pre-training and finetuning,incorporating human values and intentions into their training.
Or reinforcement learning techniques.

• Self-Reflection:
leverage the model’s self-reflection and check the consistency between the responses.
Or enable multiple large model agents to engage in mutual discussion and verification.

• Retrieval Augmentation.

Inspection

• Verification:Constrained Generation.

• Explanation:incorporate user opinions or human assistance.

• Intermediate Feature Analysis:harness hierarchical information across different layers of the model.

Conclusion and Outlook