Search papers, labs, and topics across Lattice.
Lossfunk
6
0
9
VLMs exhibit a staggering drop in performance when forced to rely solely on image content, with open models plummeting to just 1-9% accuracy on the most challenging questions.
Strong coding agents leverage metaprogramming to excel in unfamiliar languages, while weaker agents struggle without this adaptive strategy.
Neural nets can implicitly learn to identify discontinuities in physical simulations just by predicting future states, unlocking massive speedups and a fallback mechanism for increased accuracy.
LLMs that ace standard coding benchmarks spectacularly fail at esoteric languages, revealing a reliance on memorization rather than true reasoning.
Coding agents can pinpoint performance bottlenecks in real-world inference codebases, but struggle to generate functional optimization patches, revealing a critical gap between identifying problems and implementing solutions.
LLMs can achieve 85% grammatical accuracy in a language with minimal training data, like Tulu, using only structured prompting and clever constraints.
