OpenAI's GPT-5.6 Sol is literally deleting user files. Microsoft is telling its sales teams to push Copilot over OpenAI and Anthropic. The AI world is a mess of trust issues and locked-down APIs.
Enter CARPRT — a method so elegantly simple it might just be the answer nobody saw coming. It's a training-free, fully black-box approach that adjusts prompt weights for each class based on class-specific relevance. And it outperforms existing class-independent reweighting strategies on standard benchmarks.
Accepted at ICLR 2026 in Rio de Janeiro, the research comes from a team spanning the University of Melbourne, RIKEN Center for Advanced Intelligence Project, and Nanyang Technological University. The code is already on GitHub under `tmlr-group/CARPRT`. This isn't a vaporware press release — it's peer-reviewed, reproducible, and ready to use.
But here's the twist: CARPRT achieves class-awareness without any gradient access or training. It uses the VLM's own outputs — logits from multiple prompts — to compute a weight vector per class. That's like teaching a locked phone to recognize your face better just by watching how you hold it.
The losers here are the white-box prompt-tuning methods that assume you can rummage around inside the model's brain. If you're a startup building your entire business on fine-tuning CLIP with full gradient access, CARPRT just made your value proposition a lot harder to sell.
The winners? Anyone running black-box VLMs via API — which is basically everyone using commercial AI. CARPRT could make those models much more accurate without requiring a single retraining call. In a landscape where OpenAI is literally deleting files and Microsoft is fighting internal turf wars over Copilot, the ability to improve proprietary models from the outside is a superpower.
What should we watch next? If major API providers like OpenAI or Google start offering prompt reweighting as a service, CARPRT could become the new standard baseline. If not, open-source implementations might eat their lunch. Either way, the era of assuming you need full model access to get good results might be ending.
Could CARPRT scale to other modalities? Could it fail on imbalanced datasets? The research doesn't answer those questions yet. But one thing is clear: the black box just got a lot brighter.