Mistral OCR

Mistral OCR is an Optical Character Recognition API that sets a new standard in document understanding, comprehending each element of documents with unprecedented accuracy and cognition, and is now available as a default model for document understanding. The API excels in understanding complex document elements, is natively multilingual and multimodal, and has consistently outperformed other leading OCR models in rigorous benchmark tests, making it an ideal tool for various use cases, including digitizing scientific research and transforming document repositories into actionable knowledge.

Read more

Using GRPO to Beat o1, o3-mini and R1 at "Temporal Clue"

The authors used Group Relative Policy Optimization (GRPO) to train a model that surpasses state-of-the-art models like R1, o1, and o3-mini on a reasoning-heavy game called "Temporal Clue", while being over 100x cheaper to run at inference time. The model, trained on a novel deduction task, achieved results within a couple percentage points of Sonnet 3.7, a leading model, and the authors share their training recipe, dataset, and model weights under the MIT license.

Read more

Mistral OCR

Mistral OCR is an Optical Character Recognition API that sets a new standard in document understanding, comprehending each element of documents with unprecedented accuracy and cognition, and is now available as a default model for document understanding and as an API for developers. The API excels in understanding complex documents, is natively multilingual and multimodal, and has consistently outperformed other leading OCR models in rigorous benchmark tests, making it an ideal solution for organizations to unlock the collective intelligence of their digitized information.

Read more

'Trump Gaza' AI video intended as political satire, says creator

The creator of a viral AI-generated video depicting the Gaza Strip as a Dubai-style paradise, which was shared by Donald Trump, says it was intended as a political satire of Trump's "megalomaniac idea" to develop the area. The video, created in less than eight hours, was posted by Trump without explanation or consent, and its creator, Solo Avital, says the experience has highlighted the potential for misinformation and the need for a public debate about the rights and wrongs of generative AI.

Read more

State Dept. to use AI to revoke visas of foreign students who appear "pro-Hamas"

Secretary of State Marco Rubio is launching an AI-powered "Catch and Revoke" program to cancel the visas of foreign nationals who appear to support Hamas or other designated terror groups, with a focus on reviewing social media accounts of tens of thousands of student visa holders. The effort, which is part of a broader crackdown on anti-Israel activity, has raised concerns about free speech and the policing of foreign nationals' conduct and speech, with critics arguing that it could have a chilling effect on student visa-holders and infringe on their rights.

Read more

Cognitive Behaviors That Enable Self-Improving Reasoners

Researchers have found that certain language models, like Qwen, are better at self-improvement through reinforcement learning due to their intrinsic properties, such as exhibiting cognitive behaviors like verification and backward chaining. By priming other models, like Llama, with examples containing these reasoning behaviors, they can achieve substantial improvements and match the performance of Qwen, highlighting the importance of initial reasoning behaviors in a model's capacity for improvement.

ArXiv

Substructural Parametricity

This paper develops a family of logical relations to prove consequences of parametricity for various substructural type systems, using an algebraic parameterization to interpret different type systems. The approach is used to deduce extensional properties of functions, such as proving that certain types are inhabited by unique functions, including list append, reversal, and fold functions.

ArXiv

Large Models Aren't Physical Reasoners

Researchers have developed EgoNormia, a dataset of 1,853 ego-centric videos, to evaluate and improve the normative reasoning capability of vision-language models (VLMs) in understanding social and physical contexts. The results show that current state-of-the-art VLMs lack robust norm understanding, scoring only 45% on EgoNormia, and highlight significant risks in areas such as safety, privacy, and collaboration, but also demonstrate the potential to enhance normative reasoning in VLMs using a retrieval-based generation method.

ArXiv

Spark-TTS: Text-2-Speech Model Single-Stream Decoupled Tokens [pdf]

Spark-TTS is a novel text-to-speech system that utilizes a single-stream speech codec and a large language model to generate highly customizable voices with both coarse-grained and fine-grained control. The system achieves state-of-the-art zero-shot voice cloning and surpasses the limitations of reference-based synthesis, with accompanying tools and datasets, including the 100,000-hour VoxBox dataset, made available for research.

ArXiv

Towards Understanding Distilled Reasoning Models: A Representational Approach

This study examines how model distillation affects the development of reasoning features in large language models, finding that distilled models contain unique reasoning feature directions that can influence the model's thinking mode. The research also reveals that larger distilled models may develop more structured representations, leading to enhanced distillation performance and contributing to more transparent and reliable AI systems.

ArXiv

DeepSeek-R1-671B-Q4_K_M with 1 or 2 Arc A770 on Xeon

IPEX-LLM is an LLM acceleration library for Intel GPU, NPU, and CPU, providing seamless integration with various frameworks and models, including over 70 optimized and verified models. The library offers state-of-the-art LLM optimizations, XPU acceleration, and low-bit support, and has undergone numerous updates to expand its capabilities and support for different models and frameworks.

Repo

Show HN: Open-source, native audio turn detection model

The Smart Turn Detection model is an open-source, community-driven project that aims to improve conversational voice AI technology by detecting when a human has finished speaking, using linguistic and acoustic cues rather than just voice activity detection. The current model, based on Meta AI's Wav2Vec2-BERT backbone, is a proof-of-concept that handles a limited number of scenarios and only supports English, but the project goals include expanding to multiple languages, improving inference time, and capturing a wider range of speech nuances.

Repo

Show HN: Fast-agent – Compose MCP enabled Agents and Workflows in minutes

Fast-agent is a Python library that enables users to create and interact with sophisticated agents and workflows in minutes, using a simple declarative syntax to compose prompts and MCP servers. It allows for the creation of multi-model workflows, testing of model interactions, and human input for task completion, with features such as chaining, parallel workflows, and evaluator-optimizers to generate and refine content.

Repo

Show HN: Ariana – A time travel debugger for PY/JS right in VSCode

Ariana is an IDE extension and CLI tool that helps developers understand what their JavaScript, TypeScript, and Python code does when it runs, providing features like inspecting expression values and execution times without using a traditional debugger. To use Ariana, developers can install the CLI tool and VSCode extension, then run their code with the ariana command to get instant debugging information in their IDE.

Repo

Show HN: iMCP – Connect Your macOS Messages, Calendar, and More to Claude

iMCP is a macOS app that connects your digital life with AI by integrating with various services such as calendar, contacts, location, and messages, and works with clients like Claude Desktop through the Model Context Protocol (MCP). The app allows you to activate and manage these services, granting permission for access to your personal data, and enables AI tools like Claude to retrieve and provide personalized information without requiring manual data sharing.

Repo