Tag
The paper proposes T-Mem, a temporal-structure-based memory framework for LLM agents in long-horizon multi-party conversations, using Temporal Evidence Discovery and Temporal Memory Structuring to recover locally concentrated evidence and reconstruct event progressions as an evidence forest, outperforming existing memory frameworks on two benchmarks.
The paper presents FinDialogLens, a hybrid LLM pipeline for event extraction over multi-party financial chatrooms to recover missed trades, achieving 92.1% and 94.3% accuracy on final price and trade outcome with GPT-4o while cutting LLM calls by 85% via a difficulty-aware router.
The paper proposes SpeakerMem-R1, a speaker-centered dual-track memory system for multi-party dialogue, addressing bottlenecks in message attribution and relational understanding. It achieves state-of-the-art results on benchmarks like EverMemBench and LoCoMo.
ProMediConv is a benchmarking framework for proactive conversational agents in legal dispute mediation, using real-world cases to model multi-stage dialogues and propose new evaluation metrics. It aims to advance AI-assisted conflict resolution by addressing gaps in current research.
This paper proposes a multi-task learning approach that leverages turn-taking dynamics to enhance intent recognition in multi-party conversations, outperforming existing methods that ignore interaction patterns.
This paper reexamines addressee detection in multi-party dialogue, proposing continuous address levels over discrete labels and showing that address relates to gaze and backchannels beyond turn-taking, suggesting graded structure.
LoSoNA is a benchmark for evaluating LLMs' ability to infer and adapt to local social norms in group chat conversations. It tests models on their capacity to pick up hidden norms from precedent and respond appropriately.
This paper introduces DraDDP, the first publicly available English multimodal dataset for multi-party dialogue discourse parsing, built from American TV dramas with 495 segments, 6,374 utterances, and 9.1 hours of video. Benchmarks show multimodal information improves parsing of dialogue structures and relation types.
When2Speak is a synthetic dataset and pipeline for training LLMs to decide when to speak in multi-party conversations. Fine-tuning on this dataset significantly improves turn-taking, with reinforcement learning reducing missed interventions from 50% to ~20%.