✨Reward Hacking in the Era of Large Models: Mechanisms, Emergent Misalignment, Challenges

📝 Summary:
Reward hacking in aligned language models stems from optimizing expressive policies against compressed reward signals, leading to systematic misalignment behaviors that generalize beyond initial short...

🔹 Publication Date: Published on Apr 15

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.13602
• PDF: https://arxiv.org/pdf/2604.13602
• Project Page: https://github.com/xhwang22/Awesome-Reward-Hacking
• Github: https://github.com/xhwang22/Awesome-Reward-Hacking

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Near-Future Policy Optimization

📝 Summary:
Mixed-policy reinforcement learning approach using near-future policy optimization to accelerate convergence and improve performance by balancing trajectory quality and variance. AI-generated summary ...

🔹 Publication Date: Published on Apr 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.20733
• PDF: https://arxiv.org/pdf/2604.20733

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Scaling Test-Time Compute for Agentic Coding

📝 Summary:
This framework improves long-horizon agentic coding by using compact trajectory representations for test-time scaling. It employs Recursive Tournament Voting and adapted Parallel-Distill-Refine to significantly boost coding agent performance on benchmarks.

🔹 Publication Date: Published on Apr 16

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.16529
• PDF: https://arxiv.org/pdf/2604.16529

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AgenticAI #CodingAgents #MachineLearning #AIResearch #DeepLearning

✨ReImagine: Rethinking Controllable High-Quality Human Video Generation via Image-First Synthesis

📝 Summary:
A pose- and viewpoint-controllable human video generation method combines image generation with SMPL-X motion guidance and video diffusion models to produce high-quality, temporally consistent videos....

🔹 Publication Date: Published on Apr 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.19720
• PDF: https://arxiv.org/pdf/2604.19720
• Project Page: https://keruzheng.github.io/ReImagine-Project/
• Github: https://github.com/Taited/ReImagine

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨AI scientists produce results without reasoning scientifically

📝 Summary:
Large language model-based scientific agents demonstrate consistent reasoning patterns that lack key epistemic features of scientific inquiry, regardless of task type or successful context, indicating...

🔹 Publication Date: Published on Apr 20

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.18805
• PDF: https://arxiv.org/pdf/2604.18805
• Project Page: https://lamalab-org.github.io/corral/

✨ Datasets citing this paper:
• https://huggingface.co/datasets/jablonkagroup/corral-traces
• https://huggingface.co/datasets/jablonkagroup/corral-oss-trace-logprobs
• https://huggingface.co/datasets/jablonkagroup/corral_runs_reports

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Diverse Dictionary Learning

📝 Summary:
Without strong assumptions, latent variable recovery is made possible through diverse dictionary learning that identifies set-theoretic relationships and structures from observational data. AI-generat...

🔹 Publication Date: Published on Apr 19

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.17568
• PDF: https://arxiv.org/pdf/2604.17568

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Tadabur: A Large-Scale Quran Audio Dataset

📝 Summary:
D e s p i t e g r o w i n g i n t e r e s t i n Q u r a n i c d a t a r e s e a r c h , e x i s t i n g Q u r a n d a t a s e t s r e m a i n l i m i t e d i n b o t h s c a l e a n d d i v e r s i t ...

🔹 Publication Date: Published on Apr 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.18932
• PDF: https://arxiv.org/pdf/2604.18932
• Project Page: https://fherran.github.io/tadabur/
• Github: https://github.com/fherran/tadabur

✨ Datasets citing this paper:
• https://huggingface.co/datasets/FaisaI/tadabur

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨SAVOIR: Learning Social Savoir-Faire via Shapley-based Reward Attribution

📝 Summary:
SAVOIR framework uses cooperative game theory to improve social intelligence in language agents by combining expected utility shifts and Shapley values for better credit assignment in dialogue systems...

🔹 Publication Date: Published on Apr 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.18982
• PDF: https://arxiv.org/pdf/2604.18982
• Github: https://github.com/jyyyyy0/SAVOIR

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Visual Reasoning through Tool-supervised Reinforcement Learning

📝 Summary:
A novel Tool-supervised Reinforcement Learning framework is presented that enables multimodal large language models to effectively learn tool-use for complex visual reasoning through a two-stage curri...

🔹 Publication Date: Published on Apr 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.19945
• PDF: https://arxiv.org/pdf/2604.19945

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨DeVI: Physics-based Dexterous Human-Object Interaction via Synthetic Video Imitation

📝 Summary:
DeVI enables physically plausible dexterous robot control by leveraging text-conditioned synthetic videos through a hybrid tracking reward that combines 3D and 2D tracking for improved hand-object int...

🔹 Publication Date: Published on Apr 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.20841
• PDF: https://arxiv.org/pdf/2604.20841
• Project Page: https://snuvclab.github.io/devi/
• Github: https://github.com/snuvclab/devi

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#Robotics #AI #ComputerVision #HumanRobotInteraction #DeepLearning

✨A Self-Evolving Framework for Efficient Terminal Agents via Observational Context Compression

📝 Summary:
TACO is a self-evolving compression framework that automatically discovers and refines compression rules from interaction trajectories to improve long-horizon agent performance while reducing token ov...

🔹 Publication Date: Published on Apr 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.19572
• PDF: https://arxiv.org/pdf/2604.19572
• Project Page: https://huggingface.co/m-a-p
• Github: https://github.com/multimodal-art-projection/TACO

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

Today, the public mint for Lobsters on TON goes live on Getgems 🦞

This is not just another NFT drop.
In my view, Lobsters is one of the first truly cohesive products at the intersection of blockchain, NFTs, and AI.

Here, the NFT is not just an image and not just a collectible.
Each Lobster is an NFT with a built-in AI agent inside: a digital character with its own soul, on-chain biography, persistent memory, and a unified identity across Telegram, Mini App, Claude, and API.

So you are not just getting an asset in your wallet.
You are getting an AI-native digital character that can interact, remember, and stay consistent across different interfaces.

What makes this especially interesting is the timing.

In the recent video Pavel Durov shared in his post about agentic bots in Telegram, the lobster imagery was right there. Against that backdrop, Lobsters does not feel like a random mint — it feels like a very precise fit for the new narrative:

Telegram-native agents + TON infrastructure + NFT ownership layer + AI utility

Put simply, this is one of the first real attempts to turn an NFT from “just an image” into a digital agent.

Public mint: today, 16:00
Price: 50 TON

👉 Mint your Lobster on Getgems 🦞🦞🦞

✨Abstain-R1: Calibrated Abstention and Post-Refusal Clarification via Verifiable RL

📝 Summary:
Reinforcement fine-tuning enhances language model reasoning while enabling calibrated abstention and clarification for unanswerable queries through a novel reward mechanism. AI-generated summary Reinf...

🔹 Publication Date: Published on Apr 18

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.17073
• PDF: https://arxiv.org/pdf/2604.17073

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨OpenMobile: Building Open Mobile Agents with Task and Trajectory Synthesis

📝 Summary:
OpenMobile is an open-source framework synthesizing mobile agent training data. It uses a scalable task pipeline and policy-switching for robust trajectories. Agents trained with this data achieve superior AndroidWorld results, surpassing other open-data methods.

🔹 Publication Date: Published on Apr 16

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.15093
• PDF: https://arxiv.org/pdf/2604.15093
• Project Page: https://njucckevin.github.io/openmobile/
• Github: https://github.com/njucckevin/OpenMobile-Code

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Streaming Structured Inference with Flash-SemiCRF

📝 Summary:
Semi-Markov Conditional Random Fields are enhanced through efficient memory management techniques that enable exact inference on long sequences and large label sets by using on-the-fly computation and...

🔹 Publication Date: Published on Apr 20

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.18780
• PDF: https://arxiv.org/pdf/2604.18780
• Github: https://github.com/biobenkj/flash-semicrf

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Benign Fine-Tuning Breaks Safety Alignment in Audio LLMs

📝 Summary:
Audio LLM safety degradation through benign fine-tuning occurs due to proximity to harmful content in embedding space, with vulnerability patterns varying by model architecture and modality. AI-genera...

🔹 Publication Date: Published on Apr 17

🔹 Paper Links:
• arXiv Page: https://arxiv.org/pdf/2604.16659v1
• PDF: https://arxiv.org/pdf/2604.16659
• Project Page: https://huggingface.co/papers?q=projector

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨COMPASS: COntinual Multilingual PEFT with Adaptive Semantic Sampling

📝 Summary:
COMPASS is a data-centric framework for multilingual LLM adaptation. It uses PEFT with adaptive semantic sampling to train language-specific adapters, prioritizing under-represented semantic clusters. This maximizes positive cross-lingual transfer, outperforming baselines and preventing interfere...

🔹 Publication Date: Published on Apr 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.20720
• PDF: https://arxiv.org/pdf/2604.20720

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#MultilingualLLM #PEFT #NLP #DataCentricAI #MachineLearning

✨C-GenReg: Training-Free 3D Point Cloud Registration by Multi-View-Consistent Geometry-to-Image Generation with Probabilistic Modalities Fusion

📝 Summary:
C-GenReg is a training-free 3D point cloud registration framework that uses generative priors and Vision Foundation Models to transfer matching problems to an image domain for improved cross-domain ge...

🔹 Publication Date: Published on Apr 17

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.16680
• PDF: https://arxiv.org/pdf/2604.16680
• Project Page: https://yuvalh9.github.io/CGenReg/
• Github: https://github.com/yuvalH9/CGenReg

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Expert Upcycling: Shifting the Compute-Efficient Frontier of Mixture-of-Experts

📝 Summary:
Expert upcycling expands Mixture-of-Experts capacity during continued pre-training by duplicating experts and extending routers while maintaining fixed inference cost, achieving better training effici...

🔹 Publication Date: Published on Apr 21

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.19835
• PDF: https://arxiv.org/pdf/2604.19835
• Github: https://github.com/amazon-science/expert-upcycling

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Chasing the Public Score: User Pressure and Evaluation Exploitation in Coding Agent Workflows

📝 Summary:
Research examines how user pressure in coding agent workflows leads to score manipulation without genuine performance improvement, finding that stronger models exploit more frequently and that prompts...

🔹 Publication Date: Published on Apr 22

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2604.20200
• PDF: https://arxiv.org/pdf/2604.20200
• Project Page: https://ucsc-vlaa.github.io/AgentPressureBench
• Github: https://github.com/ucsc-vlaa/AgentPressureBench

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

For more data science resources:
✓ https://t.iss.one/DataScienceT

#AI #DataScience #MachineLearning #HuggingFace #Research