PaperlayerView exact revision on arXiv

Source-linked AI summary

Hello Again! LLM-powered Personalized Agent for Long-term Dialogue

Hao Li, Chenghao Yang, An Zhang, Yang Deng, Xiang Wang, Tat-Seng Chua

arXiv:2406.05925v2cs.CLcs.AI

TL;DR

Most dialogue systems focus on brief, single-session exchanges, while long-term dialogue calls for both event memory and persona consistency. LD-Agent combines dedicated memory, persona, and response modules, and the authors report state-of-the-art results on two benchmarks plus improvements across models and tasks.

  • Problem

    Most dialogue systems focus on brief single-session interactions, while long-term open-domain dialogue requires jointly maintaining event memory and persona consistency.

    Source: §1 · p. 1

  • Method

    LD-Agent is a model-agnostic framework with separate event-memory, persona-extraction, and response-generation modules that combine retrieved memories with user and agent personas.

    Source: §1 · p. 2, §3.4 · p. 5

  • Results

    LD-Agent achieves state-of-the-art performance on MSC and Conversation Chronicles and shows substantial improvements across evaluated models and multiparty dialogue tasks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The authors conclude that LD-Agent's modular decomposition supports adaptability and transferability for long-term dialogue tasks.

    Source: §5 · p. 9

    Evaluation is confined to synthetic long-term dialogue datasets, leaving validation on authentic real-world dialogue data for future work.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Open-domain dialogue systems have seen remarkable advancements with the development of large language models (LLMs). Nonetheless, most existing dialogue systems predominantly focus on brief single-session interactions, neglecting the real-world demands for long-term companionship and personalized interactions with chatbots. Crucial to addressing this real-world need are event summary and persona management, which enable reasoning for appropriate long-term dialogue responses. Recent progress in the human-like cognitive and reasoning capabilities of LLMs suggests that LLM-based agents could significantly enhance automated perception, decision-making, and problem-solving. In response to this potential, we introduce a model-agnostic framework, the Long-term Dialogue Agent (LD-Agent), which incorporates three independently tunable modules dedicated to event perception, persona extraction, and response generation. For the event memory module, long and short-term memory banks are employed to separately focus on historical and ongoing sessions, while a topic-based retrieval mechanism is introduced to enhance the accuracy of memory retrieval. Furthermore, the persona module conducts dynamic persona modeling for both users and agents. The integration of retrieved memories and extracted personas is subsequently fed into the generator to induce appropriate responses. The effectiveness, generality, and cross-domain capabilities of LD-Agent are empirically demonstrated across various illustrative benchmarks, models, and tasks. The code is released at https://github.com/leolee99/LD-Agent.

1 Introduction

LD-Agent is a model-agnostic framework for long-term open-domain dialogue that integrates event memories, user and agent personas, and response generation. Experiments report strong benchmark performance, model and task generality, and competitive cross-domain results.

1 Introduction

  • Open-domain dialogue systems aim to support long-term personalized interaction and companionship, beyond the 2–15-turn sessions common in prior studies.The paper motivates applying LLM-powered agents to open-domain dialogue, an application it describes as unexplored among agent work in other fields.Source: §1 · pp. 1–2
  • Long-term dialogue requires retaining event histories while consistently reflecting and updating user and agent personas, unlike brief single-session systems.The introduction identifies maintaining long-term event memory and persona consistency together as the central challenge; existing work often treats them separately and depends on specific architectures, limiting adaptation and zero-shot generalization.Source: §1 · p. 1
  • LD-Agent combines long- and short-term event memory banks, tunable event summarization, disentangled user-agent persona extraction, and response generation.Retrieved memories and continuously updated personas guide response generation; the framework is modular and can adapt to dialogue tasks through module retraining.Source: §1 · p. 2
  • LD-Agent achieves state-of-the-art performance on MSC and Conversation Chronicles, significantly outperforming existing methods.The evaluation also reports generality across models and tasks and competitive performance when transferring between the two datasets.Source: §1 · p. 2
  • Generality is evaluated across online and offline LLM and non-LLM models, and multiparty dialogue tasks, where LD-Agent demonstrates substantial improvements.The introduction frames broad generality, cross-domain adaptability, and dynamic refinement of event and persona information as requirements for LLM-powered long-term dialogue agents.Source: §1 · p. 2
  • In both cross-domain directions between MSC and Conversation Chronicles, LD-Agent performs competitively, nearly matching in-domain training results.The settings tune on MSC and test on CC, then tune on CC and test on MSC.Source: §1 · p. 2

2 Related Work

Prior work extends dialogue through external guidance and dynamically maintained internal knowledge; this study focuses on historical events and personas. The paper positions its dialogue system among simulation-oriented agents, using memory retrieval and character analysis.

Long-term Dialogues

  • Earlier open-domain dialogue studies typically covered only about 2–15 turns in one session, while later work explored external and internal knowledge for longer conversations (Li et al., 2017; Rashkin et al., 2019).External guidance includes commonsense, medical, and psychological knowledge; internal knowledge includes historical events and personas.Source: §2 · pp. 2–3
  • Historical events are summarized into memory banks and personas maintained dynamically to support cross-session coherence and character consistency; this study integrates both forms of internal knowledge for personalized dialogue.The cited approaches cover historical events (Xu et al., 2022a; Zhang et al., 2023b, 2022; Jang et al., 2023) and personas (Gu et al., 2019; Xu et al., 2022b; Cao et al., 2022).Source: §2 · p. 3

LLM-based Autonomous Agents

  • LLM agents include task-oriented systems for predefined tasks and simulation-oriented systems that emulate human behavior; this paper takes the simulation-oriented approach for open-domain dialogue.It uses memory retrieval and character analysis modules (Shao et al., 2023; Zhou et al., 2023; Wang et al., 2023d).Source: §2 · p. 3

3 Method

LD-Agent structures long-term multi-session dialogue around event memory, dynamic user and agent personas, and response generation, combining current-session context with selected historical information.

3.1 Task Definition

  • The task generates responses from the current session together with selected information extracted from historical sessions.Unlike single-session systems, it integrates current and long-term historical conversational cues.Source: §3.1 · pp. 3–4

3.2 Event Perception

  • Event perception separates memory for historical events from memory for ongoing-session context.The event module has distinct long-term and short-term memory components.Source: §3.2 · p. 4, Figure 2 · p. 5

3.2.1 Long-term Memory

  • Long-term memory encodes timed event summaries, instruction-tunes summarization, and retrieves memories using semantic relevance, topic overlap, and recency decay.A semantic-score threshold of 0.5 filters retrieval; if no memory qualifies, the module returns “No relevant memory.”Source: §3.2.1 · p. 4

3.2.2 Short-term Memory

  • Short-term memory caches timestamped utterances and, after a 600-second gap, summarizes the cache into long-term event records before clearing it.The new utterance is then added to the refreshed cache.Source: §3.2.2 · pp. 4–5

3.3 Dynamic Personas Extraction

  • The persona module maintains user and agent persona banks, extracting traits dynamically through LoRA instruction tuning or a tuning-free prompted approach.It returns “No Trait” for utterances without personality traits; the tuning-free user-persona strategy uses Chain-of-Thought reasoning.Source: §3.3 · p. 5

3.4 Response Generation

  • The response generator combines retrieved memories, short-term context, and both personas; it is tuned on simulated five-session conversations built from MSC and CC.Original MSC and CC responses serve as ground-truth sentences for generator tuning.Source: §3.4 · p. 5

4 Experiments

The experiments assess LD-Agent on long-term dialogue tasks and examine its generality and practicality across models, domains, and tasks.

4.1 Evaluation Settings

  • Experiments use MSC and CC, each with five sessions and about 50 turns per sample, and assess response quality, persona classification, and human-rated coherence, fluency, and engagingness.Baselines include ChatGPT, ChatGLM, BlenderBot, BART, and HAHT; the study also evaluates transfer to Ubuntu IRC.Source: §4.1 · p. 6

4.2 Evaluation Pipeline

  • Automatic evaluation initializes from session one, scores agent utterances, and feeds preceding ground-truth dialogue into LD-Agent so memories and personas can be extracted during simulation.For MSC, annotated personas are also evaluated as an alternative to extracted personas.Source: §4.2 · p. 6

4.3 Results of Multi-Session Dialogue

  • LD-Agent improves all tested models across sessions and metrics on MSC and CC; BlenderBot with LD-Agent surpasses HAHT on BLEU-2 in sessions 2–5 on both datasets.The reported comparison highlights gains in both zero-shot and tuned settings.Source: §4.3 · pp. 6–7, Table 1 · p. 7

4.4 Ablation Studies

  • All tested modules improve long-term dialogue, with event memory contributing most and yielding more stable performance across sessions than either persona module.The ablation compares a current-session ChatGLM baseline with versions adding event memory, agent personas, or user personas.Source: §4.4 · p. 7

4.5 Persona Extraction Analysis

  • The tuned persona extractor outperforms zero-shot ChatGLM with Chain-of-Thought on all reported extraction metrics and performs better for response generation in most sessions, with the gap widening over sessions.Extraction is assessed using BLEU-2/3, ROUGE-L, and accuracy.Source: §4.5 · pp. 7–8

4.6 Human Evaluation

  • Topic-based retrieval significantly outperforms direct semantic retrieval on human-annotated accuracy and recall, while human evaluations also demonstrate LD-Agent’s advantages in long-term dialogue coherence, fluency, and engagingness.Eight students assessed memory retrieval and response generation; the response-generation comparison used ChatGLM with and without LD-Agent across multiple chat sessions.Source: §4.6 · p. 8, Figure 3 · p. 8

4.7 Generality Analysis

  • Models tuned on MSC or CC perform well on the other dataset with only a slight decrease versus same-dataset tuning, and cross-domain tuned models outperform zero-shot models by a significant margin.The evaluation crosses the two datasets, which differ in collection method; results on CC are shown in Table 4, with full results in Appendix Table 8.Source: Table 4 · p. 8, §4.7 · p. 9
  • On Ubuntu IRC, BART with LD-Agent achieves state-of-the-art performance on most metrics, surpassing the BART-based HeterMPCBART and supporting transfer to multiparty dialogue.The comparison also includes GPT-2, GSN, and BART tuned without a prompt.Source: §4.7 · p. 9, Table 5 · p. 9

5 Conclusion

LD-Agent is a modular framework for coherent, personalized dialogue over extended periods, integrating historical events and user-agent personas. Experiments demonstrate its capability across multiple benchmarks, models, and tasks.

5 Conclusion

  • LD-Agent combines historical events and user-agent personas to support coherent, consistent long-term conversation.The framework targets companionship and personalized interactions over extended periods, including significant gaps between sessions.Source: §5 · p. 9
  • Its three learnable modules significantly enhance the framework’s adaptability and transferability.Source: §5 · p. 9
  • Extensive experiments demonstrate LD-Agent’s strong capability in long-term dialogue tasks across multiple benchmarks, models, and tasks.Source: §5 · p. 9

Limitations … A.3 Baselines

The paper identifies synthetic-data and basic-module-design limitations, then details its evaluation settings, datasets, metrics, and baseline models in the appendix.

Limitations

  • LD-Agent is evaluated only on synthetic long-dialogue datasets, leaving validation on authentic real-world long-term conversations for future work.The authors attribute this constraint to the difficulty of gathering authentic long-term dialogue data.Source: Limitations · p. 9
  • The modules use basic methods, leaving memory summarization and retrieval plus persona extraction and persona-based retrieval as future design directions.The paper frames LD-Agent as a modular framework whose implementations could be made more sophisticated.Source: Limitations · p. 9

Appendix

A.1 Datasets

  • MSC and CC each contain five sessions of about 50 turns per sample, with session intervals; both use 4,000 training, 500 validation, and 501 test conversations.MSC uses PersonaChat for its initial session and crowd-worker dialogues afterward, with persona records; CC is ChatGPT-compiled from event graphs and relationships, with intervals spanning years.Source: §A.1 · p. 12
  • DialogSum supplies labeled summaries and topics, MSC personas support extraction, generated memory-and-persona examples support response training, and Ubuntu IRC tests multi-party transfer.DialogSum has 13,460 dialogues (12,460 train, 500 validation, 1,500 test); the persona set has 26,605 train and 17,660 validation samples, response generation has 34,907 train and 11,851 validation samples, and Ubuntu IRC has 311,725 train, 5,000 validation, and 5,000 test dialogues.Source: §A.1 · pp. 12–13

A.2 Metrics

  • Automatic evaluation uses METEOR, accuracy, and Distinct-1/2/3; human evaluation measures retriever accuracy and recall and assesses response coherence, fluency, and engagingness.METEOR complements BLEU in multi-party tasks.Source: §A.2 · p. 13

A.3 Baselines

  • The baseline suite spans HAHT, BlenderBot, ChatGLM3, ChatGPT, and BART across online and offline, tuned and zero-shot models.BlenderBot is used for a parameter-scale-fair comparison with HAHT; the ChatGPT API model is gpt-3.5-turbo-1106, and ChatGLM3 has 6B parameters.Source: §A.3 · p. 13

A.4 Implementation Details

The event summarizer, persona extractor, and response generator share a base model and use LoRA across configurations. Training, evaluation, and key model and retrieval settings are specified.

A.4 Implementation Details

  • All modules use the same base model with LoRA, and training and evaluation run on one NVIDIA A100 GPU.ChatGLM3-6B uses Adam (Kingma and Ba, 2015) at 5e-5, batch size 4, and 3 epochs; BlenderBot uses a 2e-5 learning rate, batch size 4, and 5 epochs. The interval threshold β is 600 seconds, time-decay temperature τ is 1e+7, and semantic cosine-similarity threshold γ is 0.5.Source: §A.4 · p. 13

B Qualitative Analysis · B.1 Persona Ablation · B.2 Memory Ablation

Because practical open-domain dialogue lacks a single gold reference answer, the section qualitatively examines how persona and memory modules affect LD-Agent’s responses. The examples show persona-guided proactive engagement and memory-supported conversational continuity.

B Qualitative Analysis

  • Practical open-domain dialogue lacks a single “gold” reference answer, motivating additional qualitative analysis of LD-Agent.The analysis supplements the paper’s quantitative experiments to further assess the agent.Source: B Qualitative Analysis · p. 14

B.1 Persona Ablation

  • The persona ablation compares responses to the same user query with and without the initial personas.The authors use this comparison to examine the persona module’s specific impact on the dialogue process.Source: §B.1 · p. 14
  • With initial personas, LD-Agent proactively asks about the new company, as illustrated in Figure 5.The response aligns with the “enthusiastic” and “proactive” personas.Source: §B.1 · p. 14, Figure 5 · p. 14
  • The Figure 5 response aligns its proactive inquiry about the new company with the “enthusiastic” and “proactive” personas.This illustrates the impact of using initial personas in the response.Source: §B.1 · p. 14, Figure 5 · p. 14

B.2 Memory Ablation

  • The memory ablation manually initializes a historical session, then compares new-session responses with and without the memory module.This setup tests the module’s impact on the dialogue process.Source: §B.2 · p. 14
  • With the memory module, the Agent recalls movies recommended in the historical session, as shown in Figure 6.The comparison is between responses in a new session with and without the module.Source: §B.2 · p. 14, Figure 6 · p. 14
  • The recalled recommendations enhance conversation continuity in the new session.The section attributes this effect to the memory module’s recall of movies recommended during the historical session.Source: §B.2 · p. 14

B.3 Event Summarizer Analysis

The trained event summarizer improves on zero-shot ChatGLM across metrics on in-domain and out-of-domain datasets, and produces more concise summaries that preserve important conversational information.

B.3 Event Summarizer Analysis

  • The DialogSum-trained event summarizer is evaluated in-domain, out of distribution, and for its module impact.The stated aim is to extract more concise event summaries.Source: §B.3 · p. 14
  • The trained summarizer improves over zero-shot ChatGLM across all metrics on DialogSum and SAMSum, with the SAMSum results demonstrating stronger generalization.DialogSum matches the summarizer’s training domain; SAMSum is out of domain.Source: §B.3 · p. 14, Table 6 · p. 14
  • LD-Agent’s summary is more concise than zero-shot ChatGLM’s while effectively conveying the conversation’s most important information.The paper presents the comparison using the same dialogue context and says this concision suits long-term dialogue’s memory-storage needs.Source: §B.3 · pp. 14–15, Figure 7 · p. 15

B.4 Persona Extractor Analysis · B.5 Response Generation Analysis · C Additional Experimental Results

The analyses report that LD-Agent’s tuned persona extractor produces more concise and logical personas than zero-shot ChatGLM, while its generated response captures information from a history session better than original ChatGLM. Section C introduces additional experimental results.

B.4 Persona Extractor Analysis

  • Compared with zero-shot ChatGLM, LD-Agent’s tuned persona extractor produces more concise and logical personas from a real-world utterance.The paper presents this quality as making long-term dialogue processes more feasible.Source: §B.4 · p. 15

B.5 Response Generation Analysis

  • LD-Agent’s response successfully captures the historical discussion of “General Nathan Bedford Forrest,” performing better than original ChatGLM.The example is presented as an analysis of LD-Agent’s long-term dialogue generation ability.Source: §B.5 · p. 15

C Additional Experimental Results

C.1 Part of Speech Importance Analysis

The analysis supports topic-based retrieval by noun overlap: nouns carry the highest average information entropy in DialogSum, indicating they convey substantial information.

C.1 Part of Speech Importance Analysis

  • Nouns have the highest average information entropy in DialogSum, supporting topic retrieval based on noun overlap.The analysis treats less frequent words as more informative because information entropy is reflected through word frequency.Source: §C.1 · p. 15, Figure 9 · p. 15

C.2 Generation Diversity Analysis · D Prompt

On MSC, LD-Agent responses were more diverse than ChatGLM responses under Dist-1/2/3. The appendix also illustrates prompts for event summarization, persona extraction, and response generation.

C.2 Generation Diversity Analysis

  • LD-Agent responses were consistently more diverse than ChatGLM responses on MSC under Dist-1/2/3.The paper presents these results in Table 7 and interprets them as evidence of the generator’s powerful generation capability.Source: Table 7 · p. 15, §C.2 · p. 15

D Prompt

  • The prompt appendix separately illustrates prompts for the Event, Persona, and Response Modules.It presents prompts for event summaries, persona extraction, and response generation.Source: D Prompt · p. 16

D.1 Prompt of Event Summary

  • The event-summary prompt asks for the conversation’s main points in brief English sentences of no more than 20 words.It supplies the conversation context and asks the model to produce a summary.Source: §D.1 · p. 16

D.2 Prompt of Persona Extraction

  • The persona-extraction prompt requests personal traits from a speaker’s sentence, limits the extraction to 20 words, and requires NO_TRAIT when none can be extracted.The prompt includes examples of extracted traits and a NO_TRAIT response.Source: §D.2 · p. 16

D.3 Prompt of Response Generation

  • The base response prompt casts the model as the agent and asks it to continue the dialogue in English using at most 30 words.It provides recent conversation context and the user’s latest input.Source: §D.3 · p. 16
  • The agent-response prompt adds agent traits, linked memories, and user traits to the context before requesting a reply of at most 30 English words.It also frames the model as the agent continuing the dialogue after the user’s latest statement.Source: §D.3 · p. 17

Related papers

Loading 2406.05925v2…