Today’s multimodal AI systems can often produce answers that sound convincing but aren’t always based on what they actually see or observe. This can lead to mistakes that are hard to predict and risky in real-world situations. To address this, a new framework called Argos focuses on teaching AI to generate answers grounded in visual and temporal evidence, using automated checks instead of relying solely on human labels.

Why Grounding Matters in AI

Modern AI models are great at recognizing images, generating language, and navigating environments. However, they sometimes make errors like grasping a blocked object or describing something that isn’t there. These mistakes happen because models are often trained to produce plausible responses rather than ones based on real data from their environment. This gap between what they say and what they actually observe can be dangerous, especially as AI becomes more involved in physical tasks and decision-making.

Training AI to be more reliable involves ensuring that their answers are rooted in the actual information they receive. It’s not enough for an answer to seem correct; it needs to be correct for the right reasons, based on verifiable evidence. This is where Argos comes in, offering a way to verify what the AI is referencing and how it’s reasoning about its observations.

How Argos Enhances AI Training

Argos is a verification system layered on top of existing multimodal AI models. It works by checking whether the objects and events the model mentions actually exist in its input data, like images or videos. It also evaluates whether the model’s reasoning aligns with what it observes. To do this, Argos uses a pool of larger, more capable teacher models and rule-based checks to verify these aspects.

Instead of just rewarding the AI for giving the right answer, Argos rewards it when its reasoning is grounded in real evidence. This approach helps create higher-quality training data and guides the model to improve its ability to reason accurately. As a result, models trained with Argos show better spatial understanding, fewer visual hallucinations, and more stable learning behaviors, all while requiring less training data.

This method has shown promising results in tasks like robotics and real-world applications, where reliable perception and reasoning are crucial. By focusing on verification, Argos helps AI systems become safer and more dependable in complex environments.

Overall, integrating verification tools like Argos into AI training pushes the field toward models that not only produce correct answers but also understand the reasons behind them. This progress is essential for making AI systems safer and more trustworthy as they take on increasingly complex roles in our daily lives.

Inspired by

• https://www.microsoft.com/en-us/research/blog/multimodal-reinforcement-learning-with-agentic-verifier-for-ai-agents/

Sources

• aka.ms