TLDR: A new research paper introduces “Visual Thinking Drift,” a phenomenon where Chain-of-Thought (CoT) reasoning in video AI leads to hallucinations and errors due to models relying on internal biases rather than actual visual evidence. To address this, the paper proposes Visual Evidence Reward (VER), a reinforcement learning framework that explicitly rewards reasoning steps grounded in visual evidence. Their Video-VER model consistently outperforms other methods across various video understanding benchmarks, demonstrating that grounding reasoning in visual facts is crucial for robust video AI.
In the rapidly evolving world of artificial intelligence, enabling machines to understand and reason from dynamic visual content, like videos, is a critical challenge. While a technique called Chain-of-Thought (CoT) has significantly boosted reasoning in text-based AI, its application to video understanding has revealed some unexpected pitfalls.
A recent research paper, “When Thinking Drifts: Evidential Grounding for Robust Video Reasoning”, delves into these challenges, identifying a phenomenon the authors term “visual thinking drift.” This occurs when AI models, attempting to reason step-by-step through video content, generate verbose but often misleading internal thoughts. Instead of sticking to what’s actually visible in the video, these models tend to hallucinate visual details, amplify their internal biases, or rely on language patterns, leading them astray from the true visual evidence.
Imagine an AI watching a cooking video. If it experiences visual thinking drift, it might correctly identify a chef chopping vegetables but then incorrectly assume the next step is adding salt because of a common language pattern, even if the video clearly shows the chef reaching for sugar. This drift can cause models to “storytell” rather than engage in reasoning truly grounded in what they “see.”
The researchers explain this drift through a Bayesian perspective, suggesting that as an AI’s thought process lengthens, its reliance on actual visual evidence diminishes, and it increasingly leans on its internal language knowledge. This means that even if an AI could answer a simple question directly and correctly, forcing it to “think out loud” with CoT can sometimes make it perform worse, especially on tasks requiring quick visual perception.
To counteract this pervasive problem, the paper introduces a novel solution: Visual Evidence Reward (VER). VER is a reinforcement learning framework designed to explicitly reward AI models when their reasoning steps are verifiably grounded in visual evidence. The core idea is to encourage models to “see while thinking,” ensuring that their internal thought processes are actively and granularly tied to the perceived content of the video.
How does VER work? During training, an auxiliary AI model acts as a judge. It evaluates whether the intermediate thoughts generated by the main AI align factually with the visual inputs. If the reasoning references accurate and relevant visual details, it receives a reward. This mechanism encourages the AI to produce not just coherent, but also correct and visually-backed reasoning, stabilizing its inferences and improving overall accuracy.
The effectiveness of Video-VER, the model developed using this framework, was rigorously tested across 10 diverse video understanding benchmarks. The results were compelling: Video-VER consistently achieved top performance, often ranking first or second among strong base models and existing reasoning techniques. It showed significant accuracy gains, sometimes as much as +9.0% absolute accuracy, and an average of +4.0% across all benchmarks, compared to its base model without VER.
Qualitative examples further highlighted Video-VER’s strength. While other models often included speculative or hallucinated details in their reasoning, Video-VER maintained a clear alignment between its intermediate thought steps and the observable evidence in the video. This research underscores a crucial insight for the future of AI: in video reasoning, true intelligence comes from grounding thoughts in visual evidence, not just from generating verbose explanations.
Also Read:
- Unveiling the Disconnect: How AI Driving Models Plan Without True Reasoning
- The Evolving Landscape of AI Evaluation: From Simple Recognition to Complex Reasoning
This work paves the way for large multimodal models that not only “think before answering” but also genuinely “see while thinking,” leading to more robust and reliable AI systems for understanding our dynamic visual world.
