DeCLaRe Lab · NTU Singapore
Updates
In the news
- VentureBeat Coverage A 0.12% parameter add-on gives AI agents the working memory RAG can't
- Clarivate Recognition Highly Cited Researcher
- The Business Times Opinion Building the next chapter of scientific competitiveness
- MarkTechPost Coverage NVIDIA and SUTD Singapore introduce TangoFlux and CRPO
- MIT Technology Review Recognition Innovators Under 35, Asia Pacific
2026
ScrambleToolBench: adapting to hidden tool changes
A terminal benchmark for agents that must infer undocumented tool behaviour and recover when the tool-to-function mapping changes.
Σ-Mem: online reliability memory for multi-agent systems
Σ-Mem updates a symmetric record of pairwise reliability from post-decision feedback without retraining the agents.
IDEAgent: quality-diversity search for research ideas
IDEAgent searches for a set of ideas that meet quality thresholds without repeating one another, and tracks each candidate through repair and refinement.
RQ-Bench: testing LLM judgements of scientific novelty
RQ-Bench derives research questions from published papers and authors' accounts, then compares LLM novelty scores with expert judgements.
GRAIL: token-level advantages for RL reasoning
GRAIL uses gradient-activation saliency to assign token-level advantage weights in reinforcement learning with verifiable rewards.
Two papers accepted to ICML 2026
The papers study adaptive data selection during training and search-guided reasoning over video.
- Data Agent: end-to-end dynamic data selection for training-aware efficiency
- Chain-of-Glimpse: search-guided, object-grounded progressive reasoning over video
δ-mem featured in VentureBeat
The article covers δ-mem's compact online state and its low-rank interface to a frozen language model.
Research internships for DeCLaRe students
Four students started internships at Tencent, MiniMax and Meta.
PODS: scheduling data volume during training
PODS alternates low- and high-data phases under a fixed cumulative budget and can be used with existing data-selection methods.
Stacked from One: longer context through self-injection
A multi-scale self-injection method for extending usable context without retraining the base model from scratch.
From Perception to Action: interactive visual reasoning
A benchmark for spatial and visual reasoning during interaction with simulated physical environments.
Four papers accepted to ICLR 2026
The papers cover text-to-audio generation, task boundaries, multi-agent social influence 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: susceptibility to multi-agent social influence
- Demystifying Deep Search: a hint-free multi-hop evaluation of search agents
Epistemic Context Learning: evidence and peer reliability
Studies how LLM agents update beliefs about one another from evidence gathered during multi-agent interaction.
Embodied Foundational Models funded by CNRS@CREATE and NRF
A 2026–2029 project on generalist vision-language-action models for embodied AI.
Toward Generalist Vision Language Action Models funded by KLASS
A 2026–2028 project on action grounding and evaluation across robot platforms.
Highly Cited Researcher recognition
Soujanya Poria was recognized by Web of Science as a Highly Cited Researcher.
Google DeepMind GCP grant
S$100K in cloud compute for language, multimodal and agent research.
2025
Two papers accepted to AAAI 2026
A position paper on future directions in vision-language-action research and a study of multimodal reasoning on visual puzzles.
- 10 Open Challenges Steering the Future of Vision-Language-Action Models
- Tracking the Evolution of Multimodal Reasoning on Visual Puzzles
NORA-1.5: preference post-training for vision-language-action models
NORA-1.5 adds a flow-matching action expert and preference pairs scored by a world model and trajectory deviation.
10 Open Challenges Steering the Future of Vision-Language-Action Models
A position paper on spatial reasoning, world dynamics, post-training and cross-embodiment action generalisation for VLA models.
Demystifying Deep Search: hint-free multi-hop evaluation
WebDetective tests whether web agents can find multi-hop reasoning chains without hints embedded in the question.
OffTopicEval: out-of-domain refusal in LLM agents
Tests whether task-specific agents accept requests within scope and refuse requests outside it across 20 open-weight models.
Training Vision-Language PRMs for Test-Time Scaling in Multimodal Reasoning
Studies dataset synthesis, perception-level supervision and test-time scaling for vision-language process reward models.
LLMs Can't Handle Peer Pressure
KAIROS tests how rapport, peer actions and model confidence affect consensus in multi-agent LLM systems.
DeCLaRe Lab moves to Nanyang Technological University
DeCLaRe moved from SUTD to NTU Singapore in August 2025.
Meta Audiobox Research Grant
Research funding for audio generation and multimodal generative modelling.
NORA and NORA-1.5 released
The lab released code and model weights for both vision-language-action models.
Trust-Score/Trust-Align and MOOSE-Chem accepted at ICLR
Papers studying trustworthy retrieval-augmented generation and chemistry hypothesis discovery.