✨ARM-Thinker: Reinforcing Multimodal Generative Reward Models with Agentic Tool Use and Visual Reasoning

📝 Summary:
ARM-Thinker is an agentic reward model that uses external tools like image cropping and document retrieval to verify judgments in multimodal reasoning tasks. This significantly improves accuracy, interpretability, and visual grounding compared to existing reward models, achieving substantial perf...

🔹 Publication Date: Published on Dec 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.05111
• PDF: https://arxiv.org/pdf/2512.05111
• Project Page: https://github.com/InternLM/ARM-Thinker
• Github: https://github.com/open-compass/VLMEvalKit/pull/1334

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#MultimodalAI #AgenticAI #RewardModels #VisualReasoning #AIResearch

✨Multimodal RewardBench 2: Evaluating Omni Reward Models for Interleaved Text and Image

📝 Summary:
MMRB2 is a new benchmark for multimodal reward models, evaluating them on interleaved image and text tasks using 4,000 expert-annotated preferences. It shows top models like Gemini 3 Pro achieve 75-80% accuracy, still below human performance, highlighting areas for improvement in these models.

🔹 Publication Date: Published on Dec 18

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.16899
• PDF: https://arxiv.org/pdf/2512.16899
• Github: https://github.com/facebookresearch/MMRB2/tree/main

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#MultimodalAI #RewardModels #AIbenchmark #MachineLearning #AIResearch

✨SWE-RM: Execution-free Feedback For Software Engineering Agents

📝 Summary:
This paper introduces SWE-RM, a robust, execution-free reward model for software engineering agents. It overcomes limitations of execution-based feedback, improving coding agent performance in both test-time scaling and reinforcement learning. SWE-RM achieves new state-of-the-art results for open...

🔹 Publication Date: Published on Dec 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2512.21919
• PDF: https://arxiv.org/pdf/2512.21919

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#SoftwareEngineering #AI #ReinforcementLearning #CodingAgents #RewardModels

✨Real-Time Aligned Reward Model beyond Semantics

📝 Summary:
RLHF faces reward overoptimization from reward model misalignment. R2M introduces a new framework that uses real-time policy feedback to dynamically adapt the reward model. This improves alignment by responding to continuous policy distribution shifts beyond just semantics.

🔹 Publication Date: Published on Jan 30

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.22664
• PDF: https://arxiv.org/pdf/2601.22664

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#ReinforcementLearning #AI #MachineLearning #RewardModels #AIAlignment

✨RM -RF: Reward Model for Run-Free Unit Test Evaluation

📝 Summary:
RM-RF is a lightweight reward model predicting unit test outcomes directly from source code, skipping compile and run. It forecasts test suite success, coverage, and mutation kill rate, offering faster, cheaper evaluation for AI generated tests. This enables scalable feedback for test generation.

🔹 Publication Date: Published on Jan 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2601.13097
• PDF: https://arxiv.org/pdf/2601.13097
• Github: https://github.com/trndcenter/RM-RF-unit-tests

==================================

For more data science resources:
✓ https://xn--r1a.website/DataScienceT

#RewardModels #UnitTesting #AIGeneratedTests #SoftwareEngineering #MachineLearning