Updates
2026
VentureBeat features δ-mem as working memory for AI agents
VentureBeat interviewed Kyrie Lei and covered δ-mem as a compact 0.12% parameter add-on that gives AI agents a lightweight internal working memory. The article frames δ-mem as an important move beyond simply expanding context windows or adding more RAG: fast, online memory inside the model for long-running agent workflows.
DeCLaRe students begin research internships
Congratulations to our students on beginning research internships with leading AI research teams.
- Jingdi Lei at Tencent
- Ruiwen at MiniMax
- Chia-Yu Hung at Meta's Super Intelligence Lab
- Maojia Song at Meta's Super Intelligence Lab
PODS studies how much data to use during efficient training
The new arXiv paper complements Data Agent by scheduling the selected data volume over training, improving the efficiency-generalization trade-off under the same data budget.
δ-mem introduces lightweight online memory for LLMs
The new arXiv paper studies how a compact online memory state can evolve during interaction and directly modulate Transformer attention without extending the explicit context.
Embodied Foundational Models funded by CNRS@CREATE and NRF
The project supports research on embodied foundation models, generalist interactive AI, and vision-language-action systems. Students interested in embodied AI are encouraged to review our research themes and apply.
Toward Generalist Vision Language Action Models supported by KLASS
This funded project focuses on VLA models, action grounding, and embodied evaluation. We welcome inquiries from students with interests in robotics, multimodal learning, and efficient model training.
Four papers accepted to ICLR 2026
Work spanning text-to-audio generation, operational safety, multi-agent social dynamics, and retrieval evaluation.
- TangoFlux — fast, faithful text-to-audio generation with flow matching
- OffTopicEval — whether task-specific agents accept in-domain queries while refusing the rest
- LLMs Can't Handle Peer Pressure — models crumbling under multi-agent social influence
- Demystifying Deep Search — a hint-free multi-hop evaluation of search agents
Two papers accepted to ICML 2026
Both concern doing more with less: which data a model should train on, and where a video model should look.
- Data Agent — end-to-end dynamic data selection for training-aware efficiency
- Chain-of-Glimpse — search-guided, object-grounded progressive reasoning over video
Highly Cited Researcher recognition
Soujanya Poria was recognized by Web of Science as a Highly Cited Researcher.
Two papers accepted to AAAI 2026
A position paper on where vision-language-action research should go next, and a workshop study of multimodal reasoning.
- 10 Open Challenges Steering the Future of Vision-Language-Action Models
- Tracking the Evolution of Multimodal Reasoning on Visual Puzzles — Logical and Symbolic Reasoning workshop
Google DeepMind GCP grant supports large-scale training
S$100K of compute for language, multimodal, and agentic AI research, covering the training runs behind several of the lab’s current projects.
Stacked from One extends context windows by self-injection
A multi-scale self-injection scheme that widens a model’s usable context without retraining it from scratch.
From Perception to Action benchmarks interactive vision reasoning
An interactive benchmark that tests whether a model’s visual reasoning survives contact with acting in an environment.
Epistemic Context Learning builds trust in multi-agent systems
How an LLM-based multi-agent system should decide what to believe, and from whom.
2025
Meta Audiobox Research Grant supports audio generation research
The grant supports work on audio generation and multimodal generative modeling. Students interested in speech, audio, music, and multimodal generation can follow the Publications page for related work.
Trust-Score/Trust-Align and MOOSE-Chem accepted at ICLR
These papers study trustworthy retrieval-augmented generation and AI for Science through chemistry hypothesis rediscovery.
NORA and NORA 1.5 project pages released
The lab released project pages for compact vision-language-action models and reward-guided post-training for embodied tasks.
DeCLaRe Lab at NTU
The lab is based at Nanyang Technological University, Singapore. We welcome inquiries from students, postdocs, visiting researchers, and collaborators whose interests connect to the lab's research themes.