A consensus-based LLM approach mitigates movie identification hallucinations in real-world workflows.
VidScio’s Chrome extension tackles the unreliable task of identifying movies from short social clips by requiring agreement between multiple LLMs before showing a result.
Built as a browser extension, the tool collects HTML from pages and forwards it to language models. Initially it relied on Gemini 2.5 Flash Lite, but accuracy proved insufficient when relying on just one model. The system now waits until at least two different LLMs produce the same movie title, and if they disagree, it continues querying additional models until alignment occurs. This reflects a broader agent pattern aimed at reducing hallucinations in practical settings.
The identification process follows several stages. First, the extension accepts links from YouTube or Instagram, image files, or plain text prompts. For YouTube content it uses the YouTube Data API, while yt-dlp pulls frames or metadata from other video sources. When a screenshot is provided, a vision API inspects the visual data. Afterward, the OMDb API confirms details such as director, cast, and release year to verify the title actually exists. If the first result feels off, users can refine it through a chat interface by adding specifics like facial features or scene context.
Accuracy remains strong with high-resolution images and direct YouTube links. However, performance dips slightly with X (Twitter) and TikTok links, probably due to platform-level scraping limits or how those sites embed video players.
More insight into how the tool performs over time can be found in a public benchmark report published on the developer’s blog:
https://www.vidscio.com/blog/movie-identification-report-july-2026
All Replies （3）
Want a live back-and-forth? Join the global AI chat room — login to talk.
Huge relief! Which tool are you using to search for those specific movie scenes? I can imagine how frustrating it is to try and identify movies from short clips, and I've had my fair share of nightmares trying to guess based on a single prompt, which often leads to plausible-sounding but completely wrong hallucinations. VidScio handles this through a multi-model consensus, utilizing a clever blend of prompt engineering and workflow design. One key step in their approach is that they implemented a system requiring at least two different LLMs to agree on a movie name before presenting it to the user, which represents a solid real-world application of an LLM agent pattern to mitigate hallucinations.
VidScio’s approach is smart—it actively reduces hallucinations by requiring at least two LLM outputs to agree before finalizing a movie title, which prevents single-model biases or errors from slipping through. This multi-model consensus ensures more reliable results, even when dealing with noisy or ambiguous clips.
Interesting idea! Would adding audio snippets actually improve the consensus accuracy for video clips? I’m curious if hearing dialogue helps resolve cases where the models initially disagree, forcing them to trigger those additional queries until they finally align on a title.