Best Practice Critic Optimization
Paper ā¢ 2608.23566 ā¢ Published ā¢ 18
AI & ML interests
None defined yet.
Recent Activity
View all activity
zhouxiangxinĀ
authored a
paper 21 days ago
Jiaqi-hkustĀ
posted an update about 1 month ago
Post
2406
š§ Remember-R1: Our fix for MLLMs forgetting the image during long reasoning
We noticed a frustrating problem: when multimodal models reason over long chains, they gradually stop looking at the imageāand start hallucinating based on their own text.
So we built RememberāR1, a simple RL framework that directly supervises visual attention on the original reasoning trajectoryāno inference overhead, no proxy tasks.
We use three complementary rewards: coverage, persistence, and focus. They encourage the model to keep attending to relevant visual evidence even in later reasoning steps.
Results across 7 benchmarks and 2 model sizes: better reasoning, andāmore importantlyāvisual attention decays much more slowly during generation.
No extra cost at inference, just cleaner supervision where it counts.
š Paper: https://arxiv.org/abs/2608.01314
š» Code: https://github.com/Ch921-cell/Remember-R1
Happy to answer any questions and receive feedback!
#MultimodalAI #RL #MLLM #CoT #VisualReasoning
We noticed a frustrating problem: when multimodal models reason over long chains, they gradually stop looking at the imageāand start hallucinating based on their own text.
So we built RememberāR1, a simple RL framework that directly supervises visual attention on the original reasoning trajectoryāno inference overhead, no proxy tasks.
We use three complementary rewards: coverage, persistence, and focus. They encourage the model to keep attending to relevant visual evidence even in later reasoning steps.
Results across 7 benchmarks and 2 model sizes: better reasoning, andāmore importantlyāvisual attention decays much more slowly during generation.
No extra cost at inference, just cleaner supervision where it counts.
š Paper: https://arxiv.org/abs/2608.01314
š» Code: https://github.com/Ch921-cell/Remember-R1
Happy to answer any questions and receive feedback!
#MultimodalAI #RL #MLLM #CoT #VisualReasoning
Jiaqi-hkustĀ
posted an update about 2 months ago
Post
3101
IQA-T1: EvidenceāBased Image Quality Assessment with MLLMs
Most MLLMs are blind to lowālevel degradationsānoise, blur, compression artifacts look the same as clean images in their internal representations. That leads to quality scores based on semantic āgut feelingā rather than real perceptual evidence.
IQA-T1 changes that. We equip the model with a toolbox of 15 perceptual tools (noise residual maps, Fourier spectra, gradient maps, etc.) that generate structured visual evidence on demand. The model learns how to use tools via supervised fineātuning on our QāTool dataset (11k evidenceāgrounded reasoning chains), and when to call them via GRPO reinforcement learning that balances accuracy, tool count, and redundancy.
The result: SOTA performance across 7 benchmarks (avg PLCC 0.795), using only 2.34 tools per image on average. Every predicted score is now interpretable and backed by hard visual evidence.
All code, weights, dataset, and demo are open. Check them out and give it a spin!
š arxiv.org/abs/2607.12375v1
š» github.com/zibuyu-02/IQA-T1
š¤ model/data: huggingface.co/zibuyu-02/IQA-T1
š® demo: huggingface.co/spaces/Jiaqi-hkust/IQA-T1
Most MLLMs are blind to lowālevel degradationsānoise, blur, compression artifacts look the same as clean images in their internal representations. That leads to quality scores based on semantic āgut feelingā rather than real perceptual evidence.
IQA-T1 changes that. We equip the model with a toolbox of 15 perceptual tools (noise residual maps, Fourier spectra, gradient maps, etc.) that generate structured visual evidence on demand. The model learns how to use tools via supervised fineātuning on our QāTool dataset (11k evidenceāgrounded reasoning chains), and when to call them via GRPO reinforcement learning that balances accuracy, tool count, and redundancy.
The result: SOTA performance across 7 benchmarks (avg PLCC 0.795), using only 2.34 tools per image on average. Every predicted score is now interpretable and backed by hard visual evidence.
All code, weights, dataset, and demo are open. Check them out and give it a spin!
š arxiv.org/abs/2607.12375v1
š» github.com/zibuyu-02/IQA-T1
š¤ model/data: huggingface.co/zibuyu-02/IQA-T1
š® demo: huggingface.co/spaces/Jiaqi-hkust/IQA-T1
FlippyDoraĀ
submitted a
paper to Daily Papers about 2 months ago
zhouxiangxinĀ
authored 11
papers about 2 months ago
Stale but Stable: Staleness-Adaptive Trust Regions for Stabilizing Asynchronous Reinforcement Learning
Paper ā¢ 2607.18722 ā¢ Published ā¢ 37
MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators
Paper ā¢ 2607.15273 ā¢ Published ā¢ 17
TempAct: Advancing Temporal Plausibility in Autoregressive Video Generation via Planner-Executor RL
Paper ā¢ 2606.28016 ā¢ Published ā¢ 1
Exploring the Design Space of Reward Backpropagation for Flow Matching
Paper ā¢ 2606.11075 ā¢ Published ā¢ 12
Reinforcing Few-step Generators via Reward-Tilted Distribution Matching
Paper ā¢ 2605.26108 ā¢ Published ā¢ 7
Fine-tuning Flow Matching Generative Models with Intermediate Feedback
Paper ā¢ 2510.18072 ā¢ Published
Modeling All-Atom Glycan Structures via Hierarchical Message Passing and Multi-Scale Pre-training
Paper ā¢ 2506.01376 ā¢ Published
OS Agents: A Survey on MLLM-based Agents for General Computing Devices Use
Paper ā¢ 2508.04482 ā¢ Published ā¢ 10
DecompOpt: Controllable and Decomposed Diffusion Models for Structure-based Molecular Optimization
Paper ā¢ 2403.13829 ā¢ Published
Global Sparse Momentum SGD for Pruning Very Deep Neural Networks
Paper ā¢ 1909.12778 ā¢ Published ā¢ 1
GSLB: The Graph Structure Learning Benchmark
Paper ā¢ 2310.05174 ā¢ Published
FlippyDoraĀ
updated a
dataset 2 months ago
FlippyDoraĀ
published a
dataset 2 months ago
Jiaqi-hkustĀ
authored 2
papers 3 months ago
Jiaqi-hkustĀ
posted an update 3 months ago
Post
3993
š Introducing Robust-U1: Teaching MLLMs to Self-Recover Corrupted Visual Content
Multimodal Large Language Models (MLLMs) have achieved impressive visual understanding, yet they remain highly brittle under real-world corruptionsānoise, blur, compression artifacts, adverse weather.
Standard MLLMs suffer dramatic performance drops, and existing robustness solutions come with fundamental limits: blackābox feature alignment lacks interpretability, while whiteābox text reasoning cannot restore the lost pixelālevel visual details. This raises a crucial question:
š§ Can MLLMs recover corrupted visual content by themselves?
If the answer is yes, we can move beyond merely ācompensatingā for corruption and instead build a more intrinsic, generalizable form of resilience. Robust-U1 is our answer to that question.
š” Paper: https://arxiv.org/abs/2606.08063
š Code: github.com/jqtangust/Robust-U1
š Demo: Jiaqi-hkust/Robust-U1
Multimodal Large Language Models (MLLMs) have achieved impressive visual understanding, yet they remain highly brittle under real-world corruptionsānoise, blur, compression artifacts, adverse weather.
Standard MLLMs suffer dramatic performance drops, and existing robustness solutions come with fundamental limits: blackābox feature alignment lacks interpretability, while whiteābox text reasoning cannot restore the lost pixelālevel visual details. This raises a crucial question:
š§ Can MLLMs recover corrupted visual content by themselves?
If the answer is yes, we can move beyond merely ācompensatingā for corruption and instead build a more intrinsic, generalizable form of resilience. Robust-U1 is our answer to that question.
š” Paper: https://arxiv.org/abs/2606.08063
š Code: github.com/jqtangust/Robust-U1
š Demo: Jiaqi-hkust/Robust-U1
- 1 reply