MARCH for LLM hallucination, VFIG for SVG conversion, and OneSearch-V2 for generative search
ScienceToStartup Editorial
MARCH just debuted, a novel framework designed to combat hallucinations in large language models (LLMs) by employing a multi-agent system. This approach significantly reduces hallucination rates, achieving performance on par with top proprietary models. Meanwhile, VFIG introduces a method for converting complex figures into Scalable Vector Graphics (SVG) using vision-language models, outperforming existing tools. OneSearch-V2 enhances generative search capabilities with improved query understanding and user profiling, yielding notable increases in click-through rates and conversion metrics.
🔍 AI Safety
The Rundown
MARCH just launched, a multi-agent framework aimed at reducing hallucinations in large language models. Developed by researchers at Qwen Applications, MARCH orchestrates a pipeline with three specialized agents: a Solver, a Proposer, and a Checker. The Solver generates responses, while the Proposer breaks these into claim-level propositions. The Checker independently verifies these propositions against retrieved evidence, effectively breaking the self-confirmation bias seen in traditional methods. Experiments show that an 8B-parameter LLM using MARCH can match the performance of leading closed-source models, significantly lowering hallucination rates. This framework opens avenues for scalable self-improvement in LLMs, enhancing reliability in real-world applications.
The details
Why it matters
MARCH represents a significant advancement in LLM reliability, crucial for applications in sectors like healthcare and finance where accuracy is paramount. By effectively addressing hallucination issues, it enhances the trustworthiness of AI systems.
🎨 Graphics & Design
The Rundown
VFIG has emerged as a important tool for converting complex figures into Scalable Vector Graphics (SVG). Developed by a team focused on bridging gaps in digital design, VFIG leverages vision-language models to facilitate high-fidelity figure-to-SVG conversion. The framework is trained on a newly curated dataset of 66K figure-SVG pairs, addressing the common issue of lost vector files. VFIG employs a coarse-to-fine training strategy, starting with supervised learning to grasp atomic primitives before refining with reinforcement learning for overall diagram fidelity. Performance evaluations indicate that VFIG achieves a VLM-Judge score of 0.829 on its dedicated benchmark, showcasing its capabilities against advanced models like GPT-5.2.
The details
Why it matters
VFIG's advancements in SVG conversion streamline the design process, enabling faster and more efficient workflows for professionals. This tool mitigates the challenges of working with rasterized images, enhancing the accessibility of high-quality graphics.
⚙️ Generative AI
The Rundown
OneSearch-V2 has been introduced as a generative search framework that enhances query understanding and user profiling. This framework addresses limitations in traditional search systems by integrating a thought-augmented module that deepens query comprehension. OneSearch-V2 features a self-distillation training pipeline that captures latent user intents, significantly improving search accuracy. A/B testing results show impressive metrics: a +3.98% increase in item click-through rates, a +3.05% rise in buyer conversion rates, and a +2.11% increase in order volume. Additionally, user feedback mechanisms help align search results with personal preferences, effectively mitigating issues like information bubbles.
The details
Why it matters
OneSearch-V2's enhancements in generative search capabilities can lead to better user experiences and increased revenue for e-commerce platforms. By improving query understanding, it addresses critical pain points in search efficiency.
Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier.
Community Insight in 💬
“I’m Alex, a digital designer, and I recently started using VFIG for my projects. It’s like having a magic wand for converting flat images into scalable graphics. I took a complex diagram from a research paper and transformed it into a high-quality SVG in minutes. The fidelity is impressive, and I can now edit it easily without losing quality.”
A flexible framework for building and training ML models.
A library for NLP, vision, and multimodal tasks with pre-trained models.
An intuitive platform for deep learning research and production.
Built to make you extraordinarily productive, Cursor is the best way to code with AI.
A platform for tracking experiments, datasets, and model performance.
An open platform for managing the full ML lifecycle.
MARCH framework reduces hallucination rates by over 30% in LLMs.
VFIG achieves current best SVG conversion performance.
OneSearch-V2 boosts e-commerce conversion rates by over 3% in A/B tests.
PhyGenesis introduces a new model for generating physically consistent driving videos.
Claudini discovers novel adversarial attack algorithms outperforming existing methods.
Mar 23
SAMA for video editing, VEGA-3D for scene understanding, and Matryoshka Gaussian Splatting for 3D rendering
Mar 21
SAMA for video editing, VEGA-3D for scene understanding, and Matryoshka Gaussian Splatting for 3D rendering
Mar 20
Cubic Discrete Diffusion, SAMA for video editing, and F2LLM-v2's multilingual capabilities