marktechpost

Weak-for-Strong (W4S): A Novel Reinforcement Learning Algorithm that Trains a weak Meta Agent to Design Agentic Workflows with Stronger LLMs

W4S operates in turns. The state contains task instructions, the current workflow program, and feedback from prior executions. An action has 2 components, an analysis of what to change, and new Python ...
marktechpost

QeRL: NVFP4-Quantized Reinforcement Learning (RL) Brings 32B LLM Training to a Single H100—While Improving Exploration

What QeRL changes in the Reinforcement Learning (RL) loop? Most RLHF/GRPO/DAPO pipelines spend the bulk of wall-clock time in rollouts (token generation). QeRL shifts the policy’s weight path to NVFP4 ...
marktechpost

Baidu’s PaddlePaddle Team Releases PaddleOCR-VL (0.9B): a NaViT-style + ERNIE-4.5-0.3B VLM Targeting End-to-End Multilingual Document Parsing

https://ernie.baidu.com/blog/publication/PaddleOCR-VL_Technical_Report.pdf ...
marktechpost

Microsoft AI Debuts MAI-Image-1: An In-House Text-to-Image Model that Enters LMArena’s Top-10

Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data ...
marktechpost

Andrej Karpathy Releases ‘nanochat’: A Minimal, End-to-End ChatGPT-Style Pipeline You Can Train in ~4 Hours for ~$100

Andrej Karpathy has open-sourced nanochat, a compact, dependency-light codebase that implements a full ChatGPT-style stack—from tokenizer training to web UI inference—aimed at reproducible, hackable ...
marktechpost

Alibaba’s Qwen AI Releases Compact Dense Qwen3-VL 4B/8B (Instruct & Thinking) With FP8 Checkpoints

Do you actually need a giant VLM when dense Qwen3-VL 4B/8B (Instruct/Thinking) with FP8 runs in low VRAM yet retains 256K→1M context and the full capability surface? Alibaba’s Qwen team has expanded ...
marktechpost

Agentic Context Engineering (ACE): Self-Improving LLMs via Evolving Contexts, Not Fine-Tuning

ACE positions “context engineering” as a first-class alternative to parameter updates. Instead of compressing instructions into short prompts, ACE accumulates and organizes domain-specific tactics ...
marktechpost

Google Introduces Speech-to-Retrieval (S2R) Approach that Maps a Spoken Query Directly to an Embedding and Retrieves Information without First Converting Speech to Text

In the traditional cascade modeling approach, automatic speech recognition (ASR) first produces a single text string, which is then passed to retrieval. Small transcription errors can change query ...
marktechpost

Sentient AI Releases ROMA: An Open-Source and AGI Focused Meta-Agent Framework for Building AI Agents with Hierarchical Task Execution

ROMA provides a setup.sh quick start with Docker Setup (Recommended) or Native Setup, plus flags for E2B sandbox integration (--e2b, --test-e2b). The stack lists Backend: Python 3.12+ with ...
marktechpost

How to Evaluate Your RAG Pipeline with Synthetic Data?

Evaluating LLM applications, particularly those using RAG (Retrieval-Augmented Generation), is crucial but often neglected. Without proper evaluation, it’s almost impossible to confirm if your ...
marktechpost

How to Evaluate Voice Agents in 2025: Beyond Automatic Speech Recognition (ASR) and Word Error Rate (WER) to Task Success, Barge-In, and Hallucination-Under-Noise

Optimizing only for Automatic Speech Recognition (ASR) and Word Error Rate (WER) is insufficient for modern, interactive voice agents. Robust evaluation must measure ...
marktechpost

This AI Research Proposes an AI Agent Immune System for Adaptive Cybersecurity: 3.4× Faster Containment with <10% Overhead

What does “Profile → Reason → Neutralize” mean at the primitive level? Profile. Agents are deployed as sidecars/daemonsets alongside microservices and API gateways. They build behavioral fingerprints ...