Paper Circle
Learning Reactive Synthesis from Model Checking Feedback
How your CV stacks up
Upload your CV to see how well it fits this job role
?%
Current Multimodal Large Language Models (MLLMs) Limitations
Current Multimodal Large Language Models (MLLMs) may struggle with tasks requiring deep logical reasoning about video content, primarily stemming from the feed-forward processing nature, which limits their ability for self-correction and iterative refinement.
Proposed Framework: CyberV
To address these limitations, we propose a novel framework inspired by cybernetic principles, redesigning video MLLMs as adaptive systems capable of self-monitoring, self-correction, and dynamic resource allocation during inference.
CyberV Framework Components
Reasons to use Rodeo
Start with a chat, not a search bar
Grad scheme, placement, apprenticeship? Not sure what you want yet — that's fine. Your agent talks it through with you and turns "I have no idea" into a shortlist.
It searches the market for you
Every day your agent scans the market matching roles against what actually matters to you, not just keywords on a CV.
Only hits
No noise. No "maybe this fits." Just roles with a clear explanation of why they're right — and where to focus when applying.
Our approach, CyberV, introduces a cybernetic loop consisting of:
- MLLM Inference System: The core system responsible for video understanding.
- Sensor: Monitors MLLM forward processes and collects intermediate interpretations, such as attention drift.
- Controller: Determines when and how to trigger self-correction and generates feedback to guide the next round.
This test-time adaptive scaling framework enhances frozen MLLMs without requiring training or additional components.
Performance Improvements
Experiments demonstrate significant improvements on complex reasoning benchmarks:
Get help with your application
Your very own career expert that helps elevate your application to the next level.
- CyberV boosts Qwen2.5-VL-7B by 8.3% and InternVL3-8B by 5.5% on VideoMMMU, surpassing the competitive proprietary model GPT-4o.
- When applied to Qwen2.5-VL-72B, it yields a 10.0% improvement, achieving performance even comparable to human experts.
- On other reasoning-focused benchmarks, our method shows consistent gains of 4.6% on the multiple-choice question section of MMVU and 2.4% on MMR-V, highlighting its robustness in enhancing logical reasoning for video understanding.
Future Research
The code will be released to support further research.
“It took my CV and asked me questions relevant to understanding what kind of jobs to suggest for me. Suggestions were almost perfect. Jobs were exactly what I’ve been looking for.”
Jessica, London
Skills