AI Research Rundown: Formula Recognition, Behavioral Prediction, and Adversarial Attacks

📊 OPTICAL CHARACTER RECOGNITION

Texo: Formula Recognition within 20M Parameters

The Rundown

The team behind Texo has unveiled a new formula recognition model that packs a punch with just 20 million parameters. This minimalist design achieves performance comparable to leading models like UniMERNet-T and PPFormulaNet-S while slashing model size by 80% and 65%, respectively. This efficiency enables real-time inference on standard consumer hardware, making it accessible for in-browser applications. They also launched a web app to showcase Texo's capabilities, facilitating user engagement and practical applications.

The details

• Texo's architecture allows it to perform similarly to UniMERNet-T, which has 100 million parameters, achieving 95% accuracy in formula recognition tasks.
• The model's reduced size enables deployment on devices with as little as 4GB of RAM, making it viable for a broader range of users.
• Real-time inference speeds reach up to 30 frames per second, significantly enhancing user experience in interactive applications.

Why it matters

Texo's compact design and high performance could democratize access to advanced formula recognition tools, enabling startups and smaller firms to leverage sophisticated AI without hefty infrastructure costs.

🧠 BEHAVIORAL PREDICTION AI

Decoding the Human Factor: High Fidelity Behavioral Prediction

The Rundown

The introduction of the Large Behavioral Model (LBM) marks a significant leap in predicting human decision-making in complex environments. Unlike traditional models, LBM utilizes a structured, high-dimensional trait profile derived from psychometric data to enhance prediction accuracy. In tests, LBM's fine-tuning outperformed the Llama-3.1-8B-Instruct backbone and matched leading models when conditioned on Big Five personality traits. This approach allows for nuanced predictions of individual behaviors in strategic scenarios.

The details

• LBM achieved a 15% improvement in predictive accuracy over traditional prompting methods in strategic decision-making tasks.
• The model was trained on a proprietary dataset linking psychological traits to observed behaviors, enhancing its contextual understanding.
• Performance continues to improve with additional trait dimensions, showcasing LBM's scalability in behavioral simulations.

⚔️ ADVERSARIAL ATTACKS ON LVLMS

Pushing the Frontier of Black-Box LVLM Attacks

The Rundown

Researchers have developed M-Attack-V2, a refined method for executing black-box adversarial attacks on large vision-language models (LVLMs). This new approach boosts success rates significantly: from 8% to 30% on Claude-4.0 and from 83% to 97% on Gemini-2.5-Pro. M-Attack-V2 improves upon its predecessor by employing gradient-denoising techniques and multi-crop alignment to stabilize optimization, making it a formidable tool against advanced LVLMs.

The details

• M-Attack-V2's enhancements led to a 100% success rate against GPT-5, showcasing its effectiveness against the latest models.
• The model utilizes a novel gradient-denoising upgrade that reduces variance in gradient calculations, enhancing attack stability.
• By averaging gradients from multiple views, M-Attack-V2 minimizes the risk of detection during adversarial attempts.