Giuseppe Carenini

My long-term vision is human–AI collaboration on data-intensive tasks, in which people and AI agents jointly analyze complex multimedia data to answer hard questions and discover knowledge, communicating their goals and intentions through both natural language and interactive visualizations.

Toward this vision, my research in AI/NLP and HCI follows two complementary threads. The first is discourse and intent: processing long documents, document collections, multimodal documents, and conversations, through tasks such as multi-document summarization, topic modeling, and intent understanding. The second combines NLP with information visualization, for text analytics, interpretability, and decision support.

Today much of this work focuses on benchmarking and improving LLMs, vision-language models, and the AI agents and multi-agent systems built on them: how they understand discourse and communicative intentions, intentionally generate text and visualizations, reason about charts and dashboards, and help people analyze data and make decisions. For agents, we examine how they plan and execute complex workflows, and when multiple agents actually outperform a single one.

I received my PhD in Intelligent Systems from the University of Pittsburgh and joined UBC in 2000. Along the way my group has worked on (multi-doc) summarization of conversations and opinions, discourse parsing, visual text analytics, user-adaptive visualization, and interactive visualization for preferential choice. I am also an Amazon Scholar.

Recent news

  • MCompassRAG and SproutRAG logos
    Four papers at EMNLP 2026:
    • CLARUS: hypergraph modeling for unsupervised discharge summarization (main)
    • MCompassRAG: topic metadata as a compass for paragraph-level retrieval (Findings)
    • SproutRAG: attention-guided tree search for long-document RAG (Findings)
    • ReVision: dropping redundant visual tokens in computer-use agents (Findings)
  • NeurIPS
    2026
    The Illusion of Multi-Agent Advantage, showing that automatically designed multi-agent systems underperform a single agent with self-consistency, at NeurIPS 2026.
  • ICML
    2026
    Neural topic modeling with semantically-grounded soft labels from language models, at ICML 2026.
  • ACL
    2026
    Infusing Theory of Mind into socially intelligent LLM agents, in Findings of ACL 2026.
  • Unit circle plot generated to solve a VAMPS math problem
    Three new benchmarks released: VAMPS for visual-assisted math problem solving, IntentGrasp for intent understanding, and UnpredictaBench for distributional randomness in LLMs.
  • LREC
    AVI 2026
    Meta-prompting follow-ups for unsupervised dialogue evaluation at LREC 2026, and human gaze supervision for VLM robustness to misleading visualizations at AVI 2026.
  • Example from BeDiscovER: the discourse particle just in two sentences
    BeDiscovER, a benchmark of discourse understanding for reasoning language models, and MIMIC, on multi-party dialogue augmentation, at EACL 2026.
  • ICLR
    2026
    DRBench, a benchmark for AI agents doing deep research in enterprise settings, at ICLR 2026.
  • Human gaze heat map over a bar chart, from ChartGaze
    ChartGaze and CEMTM at EMNLP 2025, plus two Findings papers on multimodal question answering and multi-document summarization.
  • REALM
    @ ACL 2025
    AgentAda received an oral paper award (top 6%) at the REALM workshop at ACL 2025.
  • A multi-party dialogue and its discourse structure graph
    Best Paper Award at SIGDIAL 2024 for dialogue discourse parsing as generation with LLMs.

On X

Feed not loading? See my posts on X

Research

My research focuses on benchmarking and improving LLMs, vision-language models, and the AI agents and multi-agent systems built on them: how they understand discourse and communicative intentions, intentionally generate text and visualizations, reason about charts and dashboards, and help people analyze data and make decisions. For agents, we examine how they plan and execute complex workflows, and when multiple agents actually outperform a single one.

Some of the Questions we are asking now

  • What discourse structures do LLMs capture, and how can we inject more of them?
  • How can LLMs recognize communicative goals and intentions, and act on them when generating text and visualizations?
  • How can vision-language models attend to the right parts of a chart or dashboard?
  • How can LLMs and AI agents generate useful, trustworthy insights from a given dataset?
  • How can AI agents plan and carry out complex analytical and research workflows, following procedures and interacting with visualizations and dashboards as people do?
  • When do multi-agent systems actually outperform a single well-designed agent, and why?
  • How should people and AI agents divide the work, and communicate, when analyzing data together?
  • How do vision-language-action models generalize?
Left: a heat map of human eye fixations. Right: a bar chart of power transistor sales with a gaze heat map over the bars for 2008 and 2009.
ChartGazeVision-language models often look at the wrong part of a chart. We use human eye-tracking data to refine their attention, improving chart question answering and interpretability. EMNLP 2025 paper
Two plots generated to solve a trigonometry problem: points on the unit circle forming a square, and the zeros of the equation.
VAMPSA bilingual benchmark for visual-assisted math problem solving, where multimodal LLMs build their own visualizations as tools and reason over them. arXiv preprint
A six-turn multi-party dialogue next to four graphs showing its discourse structure, with sub-dialogues highlighted in different colours.
Discourse structure of conversationsMulti-party dialogues can be represented as directed acyclic graphs over turns, and decomposed into sub-dialogues. We extract these structures with open-source LLMs. SIGDIAL 2024 paper
Simulated robot arm placing a red block on a magenta block on a wooden table.
Language for robotsLexical generalization in vision-language-action models: can a robot that learned to "stack blocks" follow the same task described step by step?
Benchmarks from the past yearMuch of our recent work builds benchmarks that probe what models can and cannot do.

BeDiscovER (discourse), ORQA (operations research; code), SOP-Bench (industrial procedures; code), DRBench (enterprise deep research; code), VAMPS (visual math), IntentGrasp (intent understanding), UnpredictaBench (randomness in LLMs), and work in progress on interactive visualization world models.

Selected recent publications

For the full list, see Google Scholar or DBLP.

Book

Giuseppe Carenini, Gabriel Murray, Raymond T. Ng. Methods for Mining and Summarizing Text Conversations. Synthesis Lectures on Data Management, Morgan & Claypool, 2011.

Earlier systems and resources

CODRA discourse parser, ConVis and MultiConVis for exploring online conversations, ValueCharts for preferential choice, the BC3 conversational corpus, and eye-tracking studies toward user-adaptive visualization, which received best paper awards at UMAP 2014 and in ACM TiiS.

My Group

PhD students

MSc students

Postdoctoral fellows

Collaborators

I'm part of the UBC NLP group and work closely with colleagues across computer science, medicine, and industry.

Teaching

  • CPSC 322: Introduction to Artificial Intelligenceredesigned in 2025–26, with connections to LLM-based generative AI
  • CPSC 436N: Natural Language Processingdesigned the course in 2021 with Vered Shwartz and Varada Kolhatkar
  • CPSC 503: Computational Linguisticsgraduate course mixing lectures, paper discussion, and a project
  • Graduate course on NLP and interactive visualizationLLMs, VLMs, and human–AI collaboration for data-intensive problems
  • Master of Data Sciencedirector since 2020, co-director from 2016

Contact

Email
carenini@cs.ubc.ca
Office
CICSR #105, Department of Computer Science
2366 Main Mall, Vancouver, BC V6T 1Z4, Canada
Profiles
Google Scholar   DBLP   LinkedIn