Jun 3, 2025ML Times

• Covert Web-to-App Tracking via Localhost on Android
  Meta and Yandex have developed a covert tracking method on Android that allows their apps to listen on local ports, enabling them to link web browsing data to user identities without consent. This method exploits the Android OS's permission model, allowing JavaScript from websites to communicate with native apps via localhost sockets, effectively bypassing privacy protections.

• Vision Language Models Are Biased
  Vision Language Models (VLMs) achieve 100% accuracy on familiar images but plummet to ~17% on modified versions, revealing a reliance on memorized knowledge rather than genuine visual analysis, which is a critical flaw in their design.

• Vision Language Models are Biased
  Vision Language Models (VLMs) exhibit significant biases, scoring an average of 17.05% accuracy in counting tasks, revealing their inability to adapt to simple changes in visual stimuli, such as recognizing alterations in logos.

• [D] Is overfitting still relevant in the era double descent?
  Double descent suggests that increasing model capacity can lower testing error, challenging traditional views on overfitting and model complexity in machine learning.

• [R] Soft Thinking: Unlocking the Reasoning Potential of LLMs in Continuous Concept Space
  Soft Thinking introduces a novel method that enables human-like reasoning in LLMs by utilizing continuous concept space, allowing for smoother transitions and richer representations beyond discrete token limitations.

• Yoshua Bengio Launches LawZero: A New Nonprofit Advancing Safe-by-Design AI
  Yoshua Bengio has launched LawZero, a nonprofit focused on developing safe-by-design AI systems to mitigate risks associated with current AI technologies, such as deception and goal misalignment.

• Deep learning gets the glory, deep fact checking gets ignored
  Deep learning models, despite their glamour, can produce significant errors; a recent study revealed that a Transformer model made hundreds of incorrect predictions about enzyme functions, undermining its credibility.

• [R] System Prompt Learning: A Third Paradigm for LLM Learning Beyond Pretraining and Fine-tuning
  System Prompt Learning (SPL) introduces a novel paradigm for LLMs, enabling them to learn explicit problem-solving strategies from experience, resulting in a 4% to 8.6% improvement on challenging mathematical reasoning benchmarks like Arena Hard and AIME24.

• Advanced audio dialog and generation with Gemini 2.5
  Gemini 2.5 introduces advanced AI-powered audio dialog and generation, enabling real-time, natural conversations with features like tone control, multilingual support, and context awareness, enhancing user interaction significantly.

• 🤗SmolVLA: Efficient Vision-Language-Action Model trained on Lerobot Community Data
  SmolVLA is a compact, open-source Vision-Language-Action model that excels in robotics, outperforming larger models like ACT on various tasks while being trainable on consumer hardware.

• 🤗NO GPU left behind: Unlocking Efficiency with Co-located vLLM in TRL
  Co-locating vLLM with TRL enhances GPU efficiency by allowing both training and inference to share the same GPUs, significantly reducing idle time and hardware costs.

• Bring Receipts: New NVIDIA AI Blueprint Detects Fraudulent Credit Card Transactions With Precision
  The NVIDIA AI Blueprint enhances fraud detection in credit card transactions by utilizing accelerated data processing and advanced algorithms, significantly improving accuracy and reducing false positives compared to traditional methods.