Search papers, labs, and topics across Lattice.
Information Sciences Institute, University of Southern California
5
0
8
LLMs struggle with affordance reasoning, lagging behind humans by 20 points, but a new method can boost their performance significantly.
VLMs that ace standard chart QA benchmarks can still fail spectacularly when presented with subtly altered, counterfactual versions of the same charts.
Real-world images can actually *harm* a vision-language model's ability to understand abstract concepts, as models latch onto irrelevant visual details.
MLLMs can aggressively prune visual tokens without sacrificing performance by adapting token reduction strategies to specific classes and prompts.
LLM-derived abstractions can significantly boost analogical reasoning in narratives, outperforming end-to-end LLMs and offering a more interpretable, modular approach.
