From hallucinations to {hardware}: Classes from a real-world pc imaginative and prescient challenge gone sideways

Pc imaginative and prescient initiatives not often go precisely as deliberate, and this one was no exception. The thought was easy: Construct a mannequin that would have a look at a photograph of a laptop computer and establish any bodily harm — issues like cracked screens, lacking keys or damaged hinges. It appeared like an easy use case for picture fashions and massive language mannequins (LLMs), but it surely shortly changed into one thing extra sophisticated.

Alongside the way in which, we bumped into points with hallucinations, unreliable outputs and pictures that weren’t even laptops. To unravel these, we ended up making use of an agentic framework in an atypical approach — not for process automation, however to enhance the mannequin’s efficiency.

On this submit, we are going to stroll by way of what we tried, what didn’t work and the way a mixture of approaches finally helped us construct one thing dependable.

The place we began: Monolithic prompting

Our preliminary strategy was pretty commonplace for a multimodal mannequin. We used a single, massive immediate to move a picture into an image-capable LLM and requested it to establish seen harm. This monolithic prompting technique is easy to implement and works decently for clear, well-defined duties. However real-world knowledge not often performs alongside.

We bumped into three main points early on:

• Hallucinations: The mannequin would typically invent harm that didn’t exist or mislabel what it was seeing.
• Junk picture detection: It had no dependable approach to flag pictures that weren’t even laptops, like photos of desks, partitions or folks sometimes slipped by way of and obtained nonsensical harm experiences.
• Inconsistent accuracy: The mix of those issues made the mannequin too unreliable for operational use.

This was the purpose when it grew to become clear we would wish to iterate.

First repair: Mixing picture resolutions

One factor we observed was how a lot picture high quality affected the mannequin’s output. Customers uploaded all types of pictures starting from sharp and high-resolution to blurry. This led us to seek advice from analysis highlighting how picture decision impacts deep studying fashions.

We educated and examined the mannequin utilizing a mixture of high-and low-resolution pictures. The thought was to make the mannequin extra resilient to the big selection of picture qualities it might encounter in apply. This helped enhance consistency, however the core problems with hallucination and junk picture dealing with endured.

The multimodal detour: Textual content-only LLM goes multimodal

Inspired by latest experiments in combining picture captioning with text-only LLMs — just like the approach coated in The Batch, the place captions are generated from pictures after which interpreted by a language mannequin, we determined to provide it a strive.

Right here’s the way it works:

• The LLM begins by producing a number of doable captions for a picture. 
• One other mannequin, known as a multimodal embedding mannequin, checks how effectively every caption suits the picture. On this case, we used SigLIP to attain the similarity between the picture and the textual content.
• The system retains the highest few captions based mostly on these scores.
• The LLM makes use of these prime captions to jot down new ones, making an attempt to get nearer to what the picture truly exhibits.
• It repeats this course of till the captions cease enhancing, or it hits a set restrict.

Whereas intelligent in idea, this strategy launched new issues for our use case:

• Persistent hallucinations: The captions themselves typically included imaginary harm, which the LLM then confidently reported.
• Incomplete protection: Even with a number of captions, some points had been missed fully.
• Elevated complexity, little profit: The added steps made the system extra sophisticated with out reliably outperforming the earlier setup.

It was an attention-grabbing experiment, however in the end not an answer.

A inventive use of agentic frameworks

This was the turning level. Whereas agentic frameworks are normally used for orchestrating process flows (assume brokers coordinating calendar invitations or customer support actions), we puzzled if breaking down the picture interpretation process into smaller, specialised brokers may assist.

We constructed an agentic framework structured like this:

• Orchestrator agent: It checked the picture and recognized which laptop computer parts had been seen (display, keyboard, chassis, ports).
• Part brokers: Devoted brokers inspected every part for particular harm sorts; for instance, one for cracked screens, one other for lacking keys.
• Junk detection agent: A separate agent flagged whether or not the picture was even a laptop computer within the first place.

This modular, task-driven strategy produced way more exact and explainable outcomes. Hallucinations dropped dramatically, junk pictures had been reliably flagged and every agent’s process was easy and targeted sufficient to regulate high quality effectively.

The blind spots: Commerce-offs of an agentic strategy

As efficient as this was, it was not excellent. Two principal limitations confirmed up:

• Elevated latency: Operating a number of sequential brokers added to the full inference time.
• Protection gaps: Brokers may solely detect points they had been explicitly programmed to search for. If a picture confirmed one thing sudden that no agent was tasked with figuring out, it might go unnoticed.

We would have liked a approach to steadiness precision with protection.

The hybrid resolution: Combining agentic and monolithic approaches

To bridge the gaps, we created a hybrid system:

1. The agentic framework ran first, dealing with exact detection of recognized harm sorts and junk pictures. We restricted the variety of brokers to probably the most important ones to enhance latency.
2. Then, a monolithic picture LLM immediate scanned the picture for anything the brokers might need missed.
3. Lastly, we fine-tuned the mannequin utilizing a curated set of pictures for high-priority use circumstances, like regularly reported harm eventualities, to additional enhance accuracy and reliability.

This mixture gave us the precision and explainability of the agentic setup, the broad protection of monolithic prompting and the arrogance enhance of focused fine-tuning.

What we realized

A number of issues grew to become clear by the point we wrapped up this challenge:

• Agentic frameworks are extra versatile than they get credit score for: Whereas they’re normally related to workflow administration, we discovered they might meaningfully enhance mannequin efficiency when utilized in a structured, modular approach.
• Mixing completely different approaches beats counting on only one: The mix of exact, agent-based detection alongside the broad protection of LLMs, plus a little bit of fine-tuning the place it mattered most, gave us way more dependable outcomes than any single technique by itself.
• Visible fashions are susceptible to hallucinations: Even the extra superior setups can leap to conclusions or see issues that aren’t there. It takes a considerate system design to maintain these errors in examine.
• Picture high quality selection makes a distinction: Coaching and testing with each clear, high-resolution pictures and on a regular basis, lower-quality ones helped the mannequin keep resilient when confronted with unpredictable, real-world photographs.
• You want a approach to catch junk pictures: A devoted examine for junk or unrelated photos was one of many easiest adjustments we made, and it had an outsized influence on general system reliability.

Closing ideas

What began as a easy thought, utilizing an LLM immediate to detect bodily harm in laptop computer pictures, shortly changed into a a lot deeper experiment in combining completely different AI methods to deal with unpredictable, real-world issues. Alongside the way in which, we realized that a number of the most helpful instruments had been ones not initially designed for the sort of work.

Agentic frameworks, typically seen as workflow utilities, proved surprisingly efficient when repurposed for duties like structured harm detection and picture filtering. With a little bit of creativity, they helped us construct a system that was not simply extra correct, however simpler to grasp and handle in apply.

Shruti Tiwari is an AI product supervisor at Dell Applied sciences.

Vadiraj Kulkarni is an information scientist at Dell Applied sciences.