Research

What we work on. The full paper list is on the Publications page and on Google Scholar.

We work toward AI systems that stay reliable and efficient for diverse people, languages, and real-world settings — increasingly settings where many agents act alongside each other and alongside us. About as much of the work goes into understanding why today's systems fall short as into building ones that don't:

  • Multilingual & Multicultural LLMs — understanding how multilinguality works inside a model, and building ones whose competence holds up outside English, in text and in speech
  • Collective Intelligence — what populations of AI agents do when they act among each other and with people, e.g., which norms emerge, and whether diversity survives the group
  • Efficiency & Mechanism — what capability costs, and where inside a model it actually lives
  • Evaluation => Improvement — a static benchmark says little about a system that acts; we work on evaluation that is dynamic and automatic, and on turning what it finds back into a signal that improves the model

Multilingual & Multicultural LLMs

Two overviews: an AAAI 2026 talk on the research agenda, and a podcast conversation (2025) on open models.

Open models and benchmarks from this line:

Collective Intelligence

  • Value Diversity — treating value pluralism as a property of the collective, not of one model/agent arXiv 2026
  • What Should Agents Say? — action-state communication for efficient multi-agent systems arXiv 2026
  • MoltNet — emerging social behavior of AI agents in an agent-native network COLM 2026
  • MASim — multilingual agent-based simulation for social science COLM 2026

Efficiency & Mechanism

  • PEAR — rewarding reasoning by phase entropy, so models think only where it pays ICLR 2026
  • Difficulty–Diversity Filtering — selecting the fine-tuning data that is actually worth training on ICLR 2026
  • Parameter Specialization — how knowledge gets localized inside the weights NeurIPS 2025
  • Custom Pruning — turning a general LLM into a compact domain/language/task expert ACL Findings 2025
  • Safety Neurons — locating the small set of neurons that carry safety behavior ICLR 2025

Evaluation => Improvement

  • DR-Arena — fully automated evaluation for deep research agents ACL 2026
  • Judging Bias — what breaks when a reasoning model is the one doing the grading COLM 2025
  • Auto-Arena — models evaluating models through peer battles, without human votes ACL 2025

Research Grants

Awards & Recognition