Search papers, labs, and topics across Lattice.
1
0
3
2
LLMs can achieve 3% higher accuracy in math reasoning while using 30% fewer tokens by explicitly rewarding efficient progress toward solutions, not just verbosity.
