PaperlayerView exact revision on arXiv

Source-linked AI summary

Evaluating Very Long-Term Conversational Memory of LLM Agents

Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, Yuwei Fang

arXiv:2402.17753v1cs.CLcs.AIcs.LG

TL;DR

Very long-term conversational memory remains insufficiently evaluated beyond a few sessions. This paper introduces LoCoMo and a benchmark for long-term memory, finding that LLMs struggle with lengthy narratives and long-range temporal and causal connections despite gains from long-context models and RAG.

  • Problem

    Existing evaluations cover only limited dialogue contexts and are not well-suited to directly assessing comprehension of very long-term conversations.

  • Method

    The paper uses a human-machine pipeline to create LoCoMo and evaluates memory through question answering, event-graph summarization, and multimodal dialogue tasks.

  • Results

    Long-context LLMs and RAG improve QA by 22-66% but remain 56% below humans, especially in temporal reasoning, where the gap reaches 73%.

  • Takeaways & Limitations

    LLMs still struggle to comprehend long-term narratives and connect events temporally and causally, while RAG offers a balanced compromise for long-context memory.

  • Takeaways & Limitations

    Because the dataset is sourced primarily from LLM-generated text, it may not fully reflect the nuances of real-world online conversations.

Abstract

from arXiv · show

Existing works on long-term open-domain dialogues focus on evaluating model responses within contexts spanning no more than five chat sessions. Despite advancements in long-context large language models (LLMs) and retrieval augmented generation (RAG) techniques, their efficacy in very long-term dialogues remains unexplored. To address this research gap, we introduce a machine-human pipeline to generate high-quality, very long-term dialogues by leveraging LLM-based agent architectures and grounding their dialogues on personas and temporal event graphs. Moreover, we equip each agent with the capability of sharing and reacting to images. The generated conversations are verified and edited by human annotators for long-range consistency and grounding to the event graphs. Using this pipeline, we collect LoCoMo, a dataset of very long-term conversations, each encompassing 300 turns and 9K tokens on avg., over up to 35 sessions. Based on LoCoMo, we present a comprehensive evaluation benchmark to measure long-term memory in models, encompassing question answering, event summarization, and multi-modal dialogue generation tasks. Our experimental results indicate that LLMs exhibit challenges in understanding lengthy conversations and comprehending long-range temporal and causal dynamics within dialogues. Employing strategies like long-context LLMs or RAG can offer improvements but these models still substantially lag behind human performance.

1 Introduction

The paper introduces LoCoMo, a human-verified dataset and holistic benchmark for very long-term, multimodal dialogue memory. Experiments show that long-context LLMs and RAG improve QA but remain far below human performance and struggle with adversarial, temporal, and event-graph reasoning.

  • Motivation: Existing long-term dialogue evaluations typically cover only approximately 1K tokens across five sessions, motivating evaluation of very long conversations.The authors argue that thorough assessment is needed despite advances in long-context LLMs and RAG.
  • Evaluation framework: The benchmark evaluates long-term memory through question answering, event summarization, and multimodal dialogue generation.QA covers single-hop, multi-hop, temporal, commonsense or world knowledge, and adversarial reasoning types.
  • Dataset construction: The human-machine pipeline generates multimodal open-domain dialogues with LLM agents, then annotators correct inconsistencies, remove irrelevant images, and verify event grounding.Agents share and react to images while dialogues are guided by personas and temporal events.
  • Results: 22-66% improvements in QA from long-context LLMs and RAG still leave models 56% behind humans overall and 73% behind in temporal reasoning.These results indicate persistent difficulty understanding lengthy conversations and long-range temporal dynamics.
  • Results: 83% lower adversarial-QA performance affects long-context LLMs, which also lag the base model by 14% on event-graph summarization, while RAG offers a balanced compromise.RAG performs particularly well when dialogues are represented as assertions about speakers’ lives and personas.

2 Related Work

Related work addresses long-term dialogue through historical-context retrieval and event scaffolding, while multi-modal dialogue distinguishes image-grounded and image-sharing tasks. Separately, LLM-generated data is increasingly used to augment or synthesize dialogue benchmarks as human-generated data remains limited.

  • Long-term Dialogue: Long-term dialogue methods retrieve and temporally reason over historical segments and/or use events to scaffold conversations for consistency.These approaches draw on previous dialogues and event structures.
  • Multi-modal Dialogue: Multi-modal dialogue primarily comprises image-grounded dialogue and image-sharing dialogue.Image-grounded tasks involve answering questions or generating conversations about specific images, whereas image-sharing tasks focus on selecting images.
  • Synthetic Evaluation Benchmark: Because human-generated data is scarce and LLMs approach human-level annotation quality, researchers increasingly use LLMs to augment or synthesize large-scale dialogue benchmarks.These benchmarks assess responses in everyday social interactions and other dialogue settings.

3 Generative Pipeline for LOCOMO

LOCOMO’s generative pipeline creates two persona-driven LLM agents whose temporally and causally grounded interactions use memory, reflection, and multimodal image behaviors. Human annotators then edit the generated dialogues for consistency, image relevance, and event-graph alignment.

  • Persona Construction: Two virtual agents, L1 and L2, are initialized with gpt-3.5-turbo and assigned distinct persona statements to shape their dialogues.Personas are expanded from MSC statements and may describe objectives, experiences, habits, relationships, names, ages, and genders.
  • Temporal Event Graphs: Each agent receives a temporal event graph containing up to 25 dated events over 6–12 months, linked by causal relationships.The graph is generated with text-davinci-003 conditioned on the agent’s persona.
  • Multimodal Interaction: The agents primarily use reflect-and-respond while also supporting judicious image sharing and image reaction behaviors.Image sharing generates a caption and keywords, searches the web for a relevant image, and shares the selected result.
  • Memory and Reflection: Agent responses condition on short- and long-term memory, session summaries, retrieved observations, current conversation history, persona information, and intervening graph events.This architecture induces long-term temporal narratives while modeling recent conversations and distilled experiences.
  • Human Verification: Nearly 15% of dialogue turns were edited and approximately 19% of images were removed or substituted by human annotators.Annotators addressed long-term inconsistencies, irrelevant images, and alignment between event graphs and conversation content.

4 LOCOMO Evaluation Benchmark

The LoCoMo evaluation benchmark assesses long-term conversational memory through three tasks: question answering, event summarization, and multimodal dialogue generation. These tasks test cross-session reasoning, event dynamics and factuality, and sustained persona and narrative consistency.

  • Benchmark overview: The benchmark comprises three tasks designed to assess the accuracy of long-term memory.It is based on the dialogues generated in Section 3.
  • Question answering: Question answering covers five reasoning categories, including single-hop, multi-hop, and temporal reasoning across dialogue sessions.Answers are evaluated using normalized exact-match F1 scores, with annotations taken directly from conversations where possible.
  • Event summarization: Event summarization evaluates whether agents understand and recount the causal connections and chronological sequence of events in temporal event graph G.FactScore is used because lexical metrics such as BLEU and ROGUE do not adequately measure factual accuracy.
  • Multimodal dialogue generation: Multimodal dialogue generation measures alignment with ground-truth dialogues to assess sustained persona and continuous narrative consistency over time.Alignment is quantified using MMRelevance alongside other NLG metrics.

5 Experimental Setup

The experiments evaluate question answering, event summarization, and multimodal dialogue generation over LoCoMo conversations using constrained- and long-context LLMs, RAG, incremental summaries, and multimodal training variants. Images are replaced by captions for text-based tasks but used directly for multimodal dialogue generation.

  • Question Answering: Question answering replaces images with captions and evaluates answer prediction using F1-score, while RAG additionally reports recall@k for retrieval accuracy.Higher is better for both metrics.
  • Question Answering: Question answering compares base LLMs with constrained contexts, long-context LLMs, and retrieval-augmented generation over dialogue history and speaker observations.Base models include Mistral-7B, LLama-70B-chat, gpt-3.5-turbo, and gpt-4-turbo; the long-context model is gpt-3.5-turbo-16k.
  • Event Summarization: Event summarization uses base and long-context setups with incremental summaries, omitting RAG because summarization requires comprehensive understanding of the entire dialogue.Each preceding-session summary becomes the basis for summarizing subsequent sessions.
  • Multi-modal Dialogue Generation: Multimodal dialogue generation trains three MiniGPT-5 variants on 50 automatically generated conversations using prior turns alone, a global summary, or retrieved observations.Each run starts from a MiniGPT-5 checkpoint fine-tuned on MMDialog.

6 Experimental Results

Experiments across question answering, event summarization, and multimodal dialogue generation show that LLMs struggle with very long conversations, especially temporal and causal reasoning. Long-context models and retrieval improve some results, but performance remains limited by context noise, information loss, and increasing dialogue-history length.

  • Question answering: gpt-4-turbo achieves an overall question-answering score of 32.4, substantially below the human benchmark of 87.9.Limited-context LLMs struggle with truncated context windows, while long-context models better comprehend longer narratives.
  • Question answering: 5% improvement with gpt-3.5-turbo occurs when using the top 5 relevant observations instead of pure conversation logs.Performance worsens as more observations are retrieved, while session summaries do not significantly help despite high recall accuracies.
  • Question answering: Time reasoning and open-domain knowledge questions are the most challenging question-answering scenarios.Improperly retrieved context can degrade open-domain knowledge performance in the RAG setting.
  • Event summarization: Incremental summarization with gpt-3.5-turbo produces the highest recall and F1 performance, while gpt-4-turbo achieves a 5.3% precision improvement over gpt-3.5-turbo.The task requires modeling temporal and causal connections across multiple sessions, and manual analysis identifies missing information, hallucinations, and misunderstood dialogue cues as error categories.
  • Multimodal dialogue generation: Retrieved observations significantly improve MiniGPT-5’s multimodal dialogue generation and persona faithfulness, while MM-Relevance drops as dialogue history length increases.Retrieval-augmented generation alleviates this relevance decline to some extent.

7 Conclusion

The paper introduces LoCoMo, a dataset and evaluation framework for very long conversations, showing that LLMs struggle with long-term narratives and temporal and causal connections.

  • Dataset and evaluation: LoCoMo contains 50 high-quality very long conversations averaging 300 turns and 9K tokens across up to 35 sessions.The dataset was collected through a human-machine pipeline.
  • Dataset and evaluation: The proposed evaluation framework consists of three tasks measuring models’ proficiency in long conversations.
  • Findings: LLMs struggle to comprehend long-term narratives within dialogue and fail to draw temporal and causal connections between discussed events.

8 Limitations

The study’s limitations concern synthetic data generation, restricted multimodal and language coverage, reliance on closed-source LLMs, and difficulties evaluating verbose long-form outputs.

  • Hybrid human-machine generated data: The dataset relies primarily on LLM-generated text, with human annotators verifying and editing it to approximate real-world interactions.This approach avoids the logistical and legal complexities of collecting very long-term real-world conversations at scale.
  • Limited exploration of multimodal behavior: Web-sourced images lack the visual long-term consistencies of personal photos and can usually be replaced by captions without substantial information loss.Exceptions occur when optical character recognition is required.
  • Language: The long-term conversation generation pipeline has been developed only for English.The authors state that it could be adapted to other languages using proficient LLMs and appropriate prompt translations.
  • Closed-source LLMs: The dialog generation pipeline depends on state-of-the-art commercial LLMs accessed through paid APIs, limiting its reliance on open-source models.The authors plan to release the code so it can potentially work effectively with state-of-the-art open-source LLMs.
  • Evaluation of long-form NLG: Verbose LLM answers make correctness difficult to evaluate, and the evaluation framework faces the same challenge for long-form NLG experiments.LLMs may generate verbose answers even when prompted to respond in short phrases.

9 Broader Impacts … A.2 Temporal Event Graph

The paper highlights ethical risks of realistic generative agents and multimodal dialogue generation, while documenting the appendix’s dataset pipeline, persona construction, and causally connected temporal event-graph generation. It cautions against substituting agents for humans in consequential studies or using the framework to inform real-world policy.

  • 9 Broader Impacts: The framework inherits ethical concerns from prior generative-agent work because it aims to produce highly realistic conversations.The authors specifically identify Park et al. (2023) as the source framework whose concerns apply.
  • 9 Broader Impacts: Realistic conversational agents may foster parasocial relationships that adversely affect users.The authors recommend disclaimers identifying the source of generated dialogues before practical deployment.
  • 9 Broader Impacts: Multimodal LLMs generating dialogue-conditioned images may propagate misinformation and social biases.The risk is heightened when agents are coerced into repeating false information or dangerous opinions.
  • 9 Broader Impacts: Generative agents should not replace humans in studies whose outcomes could support consequential real-world decisions.The authors describe their work as a study of model comprehension and make no real-world policy recommendations.
  • Appendix Overview: The appendix covers the LOCOMO generation pipeline, dataset statistics and release details, experimental setup, and additional benchmark results.These topics are organized into Sections A through D.
  • A.1 Persona: Persona construction expands 4 to 5-sentence MSC speaker-attribute selections into full persona statements using gpt-3.5-turbo.The prompt conditions the model on selected initial persona statements and can add biographical, relational, activity, preference, and experience details.
  • A.2 Temporal Event Graph: Temporal event graphs are generated iteratively from persona summaries as causally connected events relevant to each personality.Initialization produces three independent events, after which later events are generated as consequences of one or more existing events.

A.2.1 Virtual Agent Architecture … C.1 Baselines

The paper’s virtual agents combine session-level and persona-based memory with image sharing, while human annotators filter conversations for multimodal, temporal, causal, and cross-turn consistency. The benchmark evaluates language-model baselines for question answering, event summarization, and multimodal dialogue generation under different context and memory configurations.

  • A.2.1 Virtual Agent Architecture: Virtual agents combine Reflect & respond with Image sharing & response, using short-term session summaries and long-term speaker observations grounded in turn-level evidence.Short-term memory conditions each session summary on the previous summary; long-term memory stores persona and life observations linked to contributing turn IDs.
  • A.2.1 Virtual Agent Architecture: Image-sharing behavior is implemented through dedicated prompts, enabling agents to share and respond to images during conversations.The prompts for these behaviors are provided in Figure 10.
  • A.3 Human Filtering: Human annotators remove irrelevant images, add omitted image context, replace caption-mismatched images, and edit dialogs for cross-turn and event consistency.They also remove event-graph events that do not appear in the conversation.
  • A.3 Human Filtering: Temporal event graphs represent persona-grounded life events as dated JSON nodes linked by causal influences between past sub-events and current events.The graph-generation prompt specifies event, date, caused_by, and id fields; Figure 7 illustrates persona-specific events and causal connections.
  • B.1 Dataset Statistics: The LOCOMO dataset contains natural-language dialogs and images requiring higher-order reasoning and multimodal coreference resolution, with statistics reported alongside benchmark annotations.The cited passages provide no numerical values for the dataset breakdown.
  • C.1 Baselines: Question-answering experiments compare Base, Long-context, and Retrieval-augmented Generation setups, using truncated, extended, or retrieved dialogue context.DRAGON is used for retrieval; Base includes Mistral-7B, LLama-70B-chat, gpt-3.5-turbo, and gpt-4-turbo, while long-context and RAG comparisons use gpt-3.5-turbo variants.
  • C.1 Baselines: Event summarization uses Base and Long-context configurations but excludes RAG because summarization requires comprehensive dialogue understanding rather than retrieving a specific portion.Multimodal dialogue generation instead trains MiniGPT-5 in Base, + summary, and + observation variants using preceding turns plus progressively richer conversation memory.

C.2 Implementation Details · D Results · D.1 Event Summarization Task

The LoCoMo evaluation uses fixed-generation settings and standardized metric implementations, with experiments run in a controlled FP32 environment. Results include event-summarization error analysis and multimodal dialogue evaluation, including image-query and image-grounded response generation.

  • C.2 Implementation Details: All experiments, including RAG-based models and MiniGPT-5 training and inference, run on an Nvidia A6000 server with FP32.Each model’s results come from a single inference run; MiniGPT-5 training lasts 10 epochs and approximately 30 h.
  • C.2 Implementation Details: The evaluation protocol uses default implementations of BLEU, ROUGE, BertScore, and FactScore in their respective Python packages.BLEU is implemented using the referenced NLTK package.
  • C.2 Implementation Details: A prompt converts textual descriptions of shared photos into short web-image-search queries that omit people’s names, years, and irrelevant details.For example, a Zelda scene becomes the query “the legend of zelda: breath of wild link landscape.”
  • C.2 Implementation Details: Another prompt generates responses grounded in a shared image and the personas of the conversation’s speakers.Figure 10 describes prompts for both image-sharing and image-response behavior.
  • D Results: Table 6 compares multimodal dialogue generation performance across different MiniGPT-5 training variants.The table marks the optimal performance in bold.
  • D.1 Event Summarization Task: Event summarization analysis presents five broad categories of errors made by LLMs.An example of these error categories is provided in Table 7 and outlined in Section 6.2.

D.2 Multimodal Dialog Generation Task

The LoCoMo benchmark evaluates multimodal dialogue generation using multiple versions of MiniGPT-5, with results reported in Table 6.

  • Task and evaluation: The multimodal dialogue generation task is evaluated within the LoCoMo benchmark.
  • Task and evaluation: The evaluation compares various versions of the MiniGPT-5 model.
  • Task and evaluation: The task results are presented in Table 6.
Loading 2402.17753v1…