much quicker.” For CUB, that meant bringing 687 forward passes down to 47. Logan believes the importance of the work lies in addressing questions that researchers in applied fields are already asking. “People are focusing on LLMs, but they're not really focused on this real niche use case for a subset of scientists and ecologists,” he points out. He hopes the work will encourage further research into improving these models rather than simply measuring their limitations. One direction he finds particularly promising is fine-tuning multimodal models with stronger domain knowledge. “We’ve drilled down closer to how you’re supposed to evaluate these models in terms of the specific task,” he says, “Now, I’d really like to see some work in terms of what we can leverage to increase the built-in knowledge of these models!” 7 DAILY WACV Sunday Logan Lawrence
RkJQdWJsaXNoZXIy NTc3NzU=