✨Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation

📝 Summary:
Researchers used Chinese LLMs censored on political topics as a natural testbed for honesty elicitation and lie detection. They found prompt modifications and fine-tuning increased truthful responses, while self-classification was effective for detection. No method fully eliminated falsehoods.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05494
• PDF: https://arxiv.org/pdf/2603.05494
• Github: https://github.com/cywinski/chinese_auditing

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

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

#LLMs #Censorship #LieDetection #AISafety #NLP

✨IF-RewardBench: Benchmarking Judge Models for Instruction-Following Evaluation

📝 Summary:
IF-RewardBench is a new meta-evaluation benchmark for instruction-following. It employs a preference graph for listwise evaluation to assess judge models ability to rank responses. This reveals current judge model deficiencies and shows stronger correlation with downstream task performance.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04738
• PDF: https://arxiv.org/pdf/2603.04738
• Project Page: https://github.com/thu-coai/IF-RewardBench
• Github: https://github.com/thu-coai/IF-RewardBench

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

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

#InstructionFollowing #LLMEvaluation #AIBenchmarks #JudgeModels #AIResearch

✨τ-Knowledge: Evaluating Conversational Agents over Unstructured Knowledge

📝 Summary:
τ-Knowledge extends τ-Bench to evaluate conversational agents in fintech customer support, integrating external knowledge with tool use. Its τ-Banking domain involves navigating 700 documents and executing tool-mediated updates. Frontier models achieve only ~25.5% pass, struggling with document r...

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04370
• PDF: https://arxiv.org/pdf/2603.04370

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

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

#ConversationalAI #Fintech #LLMEvaluation #KnowledgeIntegration #ToolUse

✨Operator Learning Using Weak Supervision from Walk-on-Spheres

📝 Summary:
WoS-NO trains neural PDE solvers using Monte Carlo weak supervision from Walk-on-Spheres, avoiding expensive data and higher-order derivatives. This method improves accuracy, speeds up training, and reduces memory compared to traditional physics-informed approaches.

🔹 Publication Date: Published on Mar 1

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.01193
• PDF: https://arxiv.org/pdf/2603.01193
• Github: https://github.com/neuraloperator/WoS-NO

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

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

#OperatorLearning #WeakSupervision #NeuralPDE #MonteCarlo #SciML

✨Physics Informed Viscous Value Representations

📝 Summary:
This work introduces a physics-informed regularization for offline GCRL, based on the Hamilton-Jacobi-Bellman equation's viscosity solution. Using Monte Carlo estimation, it improves value estimation and geometric consistency for complex navigation and manipulation tasks.

🔹 Publication Date: Published on Feb 26

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2602.23280
• PDF: https://arxiv.org/pdf/2602.23280

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

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

#ReinforcementLearning #PhysicsInformed #OfflineRL #MachineLearning #Robotics

✨U6G XL-MIMO Radiomap Prediction: Multi-Config Dataset and Beam Map Approach

📝 Summary:
This paper improves XL-MIMO radiomap prediction for 6G by creating a large dataset and benchmark framework. A novel physics-informed beam map feature enhances generalization to unseen array configurations and environments. This method significantly reduces prediction error by decoupling array rad...

🔹 Publication Date: Published on Mar 6

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.06401
• PDF: https://arxiv.org/pdf/2603.06401
• Project Page: https://lxj321.github.io/MulticonfigRadiomapDataset/
• Github: https://github.com/Lxj321/MulticonfigRadiomapDataset

✨ Datasets citing this paper:
• https://huggingface.co/datasets/lxj321/Multi-config-Radiomap-Dataset

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

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

#6G #MIMO #WirelessCommunication #MachineLearning #Radiomap

✨DreamCAD: Scaling Multi-modal CAD Generation using Differentiable Parametric Surfaces

📝 Summary:
DreamCAD is a multi-modal generative framework that creates editable BReps from point-level supervision using parametric patches and differentiable tessellation, achieving superior geometric fidelity ...

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.05607
• PDF: https://arxiv.org/pdf/2603.05607
• Project Page: https://sadilkhan.github.io/dreamcad2026/

✨ Datasets citing this paper:
• https://huggingface.co/datasets/SadilKhan/CADCap-1M

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨nabla-Reasoner: LLM Reasoning via Test-Time Gradient Descent in Latent Space

📝 Summary:
nabla-Reasoner improves LLM reasoning by integrating differentiable optimization directly into the decoding loop. It leverages gradient signals from the LLM and a reward model to refine textual representations, achieving over 20% accuracy improvement while reducing model calls.

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.04948
• PDF: https://arxiv.org/pdf/2603.04948
• Github: https://github.com/VITA-Group/Nabla-Reasoner

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Generalizable Knowledge Distillation from Vision Foundation Models for Semantic Segmentation

📝 Summary:
Generalizable Knowledge Distillation GKD improves out-of-domain generalization for semantic segmentation. GKD decouples representation learning from task learning, using query-based soft distillation to transfer knowledge from vision foundation models. It consistently outperforms other methods, a...

🔹 Publication Date: Published on Mar 3

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.02554
• PDF: https://arxiv.org/pdf/2603.02554
• Github: https://github.com/Younger-hua/GKD

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨PIRA-Bench: A Transition from Reactive GUI Agents to GUI-based Proactive Intent Recommendation Agents

📝 Summary:
PIRA-Bench presents a benchmark for evaluating multimodal large language models on proactive GUI agent tasks using continuous visual inputs, while PIRF offers a memory-aware framework for handling com...

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08013
• PDF: https://arxiv.org/pdf/2603.08013
• Project Page: https://www.pira-bench.top

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨PureCC: Pure Learning for Text-to-Image Concept Customization

📝 Summary:
PureCC presents a concept customization method that preserves original model behavior through decoupled learning and adaptive guidance scaling. AI-generated summary Existing concept customization meth...

🔹 Publication Date: Published on Mar 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.07561
• PDF: https://arxiv.org/pdf/2603.07561
• Github: https://github.com/lzc-sg/PureCC

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨From Narrow to Panoramic Vision: Attention-Guided Cold-Start Reshapes Multimodal Reasoning

📝 Summary:
The study introduces a novel attention-based metric called Visual Attention Score to analyze cold-start initialization in multimodal large reasoning models, identifying a counter-intuitive phenomenon ...

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.03825
• PDF: https://arxiv.org/pdf/2603.03825
• Github: https://github.com/lrlbbzl/Qwen-AVAR

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence

📝 Summary:
Holi-Spatial presents the first fully automated, large-scale, spatially-aware multimodal dataset constructed from raw video inputs, supporting multi-level spatial supervision for 3D scene understandin...

🔹 Publication Date: Published on Mar 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.07660
• PDF: https://arxiv.org/pdf/2603.07660
• Project Page: https://visionary-laboratory.github.io/holi-spatial/
• Github: https://github.com/Visionary-Laboratory/holi-spatial

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨\$OneMillion-Bench: How Far are Language Agents from Human Experts?

📝 Summary:
A new benchmark evaluates language models on complex, real-world professional tasks requiring multi-step reasoning, evidence resolution, and domain-specific decision-making across multiple industries....

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.07980
• PDF: https://arxiv.org/pdf/2603.07980
• Github: https://github.com/humanlaya/OneMillion-Bench

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Believe Your Model: Distribution-Guided Confidence Calibration

📝 Summary:
Large reasoning models enhance prediction accuracy through test-time scaling techniques that generate multiple candidate responses, with the proposed DistriVoting method utilizing distributional prior...

🔹 Publication Date: Published on Mar 4

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.03872
• PDF: https://arxiv.org/pdf/2603.03872
• Github: https://github.com/yxizhong/SSC

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Scale Space Diffusion

📝 Summary:
Scale-space theory connects diffusion models' information hierarchy to low-pass filtering, leading to a framework that combines scale spaces with diffusion processes for efficient image processing. AI...

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08709
• PDF: https://arxiv.org/pdf/2603.08709
• Project Page: https://prateksha.github.io/projects/scale-space-diffusion/
• Github: https://github.com/prateksha/ScaleSpaceDiffusion

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨FVG-PT: Adaptive Foreground View-Guided Prompt Tuning for Vision-Language Models

📝 Summary:
Foreground attention shifts during CLIP-based prompt tuning are addressed through an adaptive module that enhances foreground view quality and mitigates generalization degradation. AI-generated summar...

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08708
• PDF: https://arxiv.org/pdf/2603.08708
• Github: https://github.com/JREion/FVG-PT

✨ Datasets citing this paper:
• https://huggingface.co/datasets/JREion/Prompt_Tuning_Datasets_with_Foreground

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Skip to the Good Part: Representation Structure & Inference-Time Layer Skipping in Diffusion vs. Autoregressive LLMs

📝 Summary:
Diffusion language models exhibit distinct representational structures compared to autoregressive models, with hierarchical abstractions and reduced bias, enabling efficient layer-skipping inference w...

🔹 Publication Date: Published on Mar 8

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.07475
• PDF: https://arxiv.org/pdf/2603.07475

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Scaling Agentic Capabilities, Not Context: Efficient Reinforcement Finetuning for Large Toolspaces

📝 Summary:
ATLAS enables small language models to effectively operate in large-scale tool environments through reinforcement fine-tuning that learns context control and execution structure, achieving performance...

🔹 Publication Date: Published on Mar 5

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.06713
• PDF: https://arxiv.org/pdf/2603.06713

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research

✨Agentic Critical Training

📝 Summary:
Agentic Critical Training (ACT) is a reinforcement learning approach that trains language model agents to autonomously reason about action quality by directly rewarding correct judgment between altern...

🔹 Publication Date: Published on Mar 9

🔹 Paper Links:
• arXiv Page: https://arxiv.org/abs/2603.08706
• PDF: https://arxiv.org/pdf/2603.08706
• Project Page: https://attention-is-all-i-need.github.io/ACT/

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

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

#AI #DataScience #MachineLearning #HuggingFace #Research