A frozen, general-purpose vision model with no fashion training beat a fine-tuned fashion model at describing garments, and the only thing it was handed was other products’ descriptions. RA-CoA, posted to arXiv on 12 September by researchers at MBZUAI and IIT Jodhpur, retrieves attribute sets from a catalogue of existing items, reasons over them one attribute at a time, and never updates a weight. What that reprices is the input list for AI product discovery: the model is rentable at whatever capability a retailer wants, and the written archive feeding it is not.

The comparison that matters commercially is the one the authors ran on ground neither method owned. UniFashion, the supervised system they benchmark against, was trained on large fashion datasets including the one used here, so testing it at home would have flattered it; both were run instead against a small held-out set of product images and captions pulled from Amazon. RA-CoA, running on open-weights InternVL2-8B with no fine-tuning at all, scored 20.2 METEOR there against UniFashion’s 16.2, and nearly doubled its BLEU-1. Training on a fashion corpus bought less than reading one at inference time.

Nothing in the construction is proprietary. The authors sampled 60,000 products from the training split of FashionGen, cropped each photograph to the item with an off-the-shelf detector, and used a 3-billion-parameter Llama model to pull structured attribute-value pairs out of descriptions that professional stylists had already written for an e-commerce catalogue. The code is public and every component in that pipeline is open, small and cheap to run. The descriptions are the one input that had to be produced by someone who knew what a peplum was.

Most of the lift comes from the borrowed text rather than from the reasoning scaffold the method is named after. Showing InternVL2 a handful of retrieved neighbour captions as in-context examples, with no attribute chain running at all, lifts METEOR from 13.8 to 31.8. The full chain-of-attributes pipeline adds a little over a third as much again on top of that. The authors run the reverse test as well: strip the retrieved captions back out of the prompt and the same model loses about half its score.

The oracle experiments separate two things retailers tend to price as one. Handing the model the ground-truth list of which attributes to mention gains InternVL2 just over four METEOR points, which puts identifying what to look for close to solved by retrieval alone. Handing it the correct values for those attributes takes the score to 97.1, because whether the lapel is notched or the fly concealed is not something a model reads reliably off a photograph. That ceiling is inflated, since the values were extracted from the same expert captions the output is scored against. The ordering survives the caveat: naming the attribute is cheap, supplying its value is the entire problem, and the value is already sitting in a description somebody wrote.

This inverts how the problem has usually been filed, including by us. We have argued that the catalogue rather than the model is where AI commerce projects stall, and that benchmarks quietly delete the products with no description before anyone measures anything. Both framings treat product copy as a deficit to be remediated on the way to deploying a model. RA-CoA treats the same text as the thing being deployed.

Retrieval converts a retailer’s back catalogue into the fine-tuning it no longer has to buy.

The archive that does this work is smaller than the incumbency argument assumes. Shrinking the knowledge base to 10,000 entries costs InternVL2 little, and the authors report the gains flattening past 20,000. A retailer holding an archive that size has the asset; a marketplace holding a hundred times more has an edge on long-tail items and not much else beyond it. When CJ ONSTYLE rewrote its product copy and quadrupled its traffic from conversational assistants, the volume was the headline and the authorship was the mechanism.

The strongest objection is that this is an artefact of weak vision, and a temporary one. On that reading, attribute values are hard to see only until the next generation of models reads the notch in the lapel straight off the photograph, at which point the archive is a crutch rather than an asset. For the objection to hold, value prediction from pixels would have to improve faster than retrieval can substitute for it. GPT-4o, the strongest model the authors tested, gained the most from the knowledge base rather than the least, climbing from 12.6 METEOR zero-shot to 67.6 with retrieval attached. The distance to the value oracle also widens as models get larger, and the paper names the consequence: as capability rises, the limiting factor shifts further onto value prediction, not away from it.

Corrupting half the stored attribute values costs the system about two METEOR points. A blunter test is more telling: dropping half the attribute entries outright — simulating missing fields rather than wrong ones — costs 8.4 points, about four times as much. Copywriting is usually cut on the assumption that its output is marketing collateral with a short half-life; these results describe it as inventory. Every description a merchandising team stops commissioning this year is an entry that will not be in the retrieval base when an assistant asks what the hem does. If retailers keep treating product copy as a cost centre while licensing models to compensate, they will be renting the cheap half of the system and quietly liquidating the scarce half. What looks like a technology decision is really a staffing one.