Updates
2026
ScrambleToolBench studies how autonomous agents adapt to structural environment changes
An interactive terminal benchmark that evaluates whether autonomous agents can infer tool behaviors through trial-and-error and recover from mapping drift without documentation.
Σ-Mem introduces online reliability memory for multi-agent systems
A real symmetric reliability memory state that tracks peer competence and relationship evidence from post-decision feedback, enabling stable online coordination without retraining.
IDEAgent introduces agentic quality-diversity search for research idea generation
IDEAgent formalizes scientific discovery as an agentic quality-diversity search problem, generating diverse, high-novelty hypotheses while continually repairing and refining candidate ideas.
RQ-Bench evaluates limitations of LLMs judging scientific novelty
Reconstructs author-anchored research questions from published papers to benchmark automated novelty assessment, finding significant divergences between LLM judges and human domain experts.
GRAIL introduces gradient-reweighted advantages for RL reasoning
Uses gradient-activation saliency to assign token-level advantage weights in reinforcement learning with verifiable rewards, improving mathematical reasoning without process reward models.
Two papers accepted to ICML 2026
Work spanning end-to-end data selection for efficient training and search-guided progressive 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
VentureBeat features δ-mem as working memory for AI agents
Coverage of δ-mem, a 0.12% parameter addition that provides fast online memory modulation for LLM agents during long-horizon interaction.
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 schedules data volume over training for data-centric efficiency
A plug-and-play scheduling framework that alternates between low-ratio regularization and high-ratio recovery phases to improve data selection efficiency.
Stacked from One extends context windows via self-injection
A multi-scale self-injection approach that expands usable model context without retraining from scratch.
From Perception to Action benchmarks interactive visual reasoning
Evaluates whether multimodal models maintain spatial and visual reasoning consistency when interacting with simulated physical environments.
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
Epistemic Context Learning builds trust in multi-agent systems
Formulates evidence accumulation principles for determining belief formation and peer reliability in LLM-based multi-agent networks.
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.
Toward Generalist Vision Language Action Models supported by KLASS
This funded project focuses on VLA models, action grounding, and embodied evaluation across robotic platforms.
Highly Cited Researcher recognition
Soujanya Poria was recognized by Web of Science as a Highly Cited Researcher.
Google DeepMind GCP grant supports large-scale training
S$100K compute grant supporting language, multimodal, and agentic training runs.
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 improves vision-language-action policies via preference rewards
Integrates a flow-matching action expert into the pre-trained NORA backbone and applies world model and action-based preference optimization for embodied robotics tasks.
10 Open Challenges Steering the Future of Vision-Language-Action Models
A roadmap paper outlining ten core milestones for generalist VLA models across spatial reasoning, world dynamics, post-training, and cross-embodiment action generalization.
Demystifying Deep Search introduces hint-free multi-hop evaluation
Presents WebDetective, a benchmark evaluating whether web agents discover multi-hop reasoning chains autonomously rather than following surface cues in question prompts.
OffTopicEval evaluates operational safety and out-of-domain refusal in LLM agents
Introduces an evaluation benchmark measuring whether LLM agents refuse out-of-scope requests, finding significant vulnerability across 20 open-weight models.
Training Vision-Language PRMs for Test-Time Scaling in Multimodal Reasoning
Analyzes dataset synthesis, perception-level supervision, and test-time scaling strategies for vision-language process reward models.
LLMs Can't Handle Peer Pressure evaluates multi-agent social influence
Introduces KAIROS to analyze how rapport, peer actions, and model confidence affect consensus and vulnerability in multi-agent LLM systems.
DeCLaRe Lab moves to Nanyang Technological University
The lab established its base at Nanyang Technological University, Singapore, continuing research across language, multimodal, and embodied AI.
Meta Audiobox Research Grant supports audio generation research
The grant supports research on audio generation and multimodal generative modeling.
NORA and NORA 1.5 project pages released
Project pages and code releases for compact vision-language-action models and reward-guided post-training for embodied tasks.
Trust-Score/Trust-Align and MOOSE-Chem accepted at ICLR
Papers studying trustworthy retrieval-augmented generation and chemistry hypothesis discovery.