Search papers, labs, and topics across Lattice.
Affiliation:
6
0
8
This method introduces Instance-level Difficulty-Aware Control (IDAC), a reward-shaping mechanism that leverages pre-computed reference statistics (accuracy and token usage) to regulate reasoning depth and enables stable critic-free optimization without learned reward models or online reference-model queries.
LLMs frequently recognize boundary violations yet still generate out-of-character responses, revealing a critical compliance failure in character role-play.
GRAFT reveals that optimized workflows can evolve in real-time, adapting to execution feedback without the computational burden of complete re-optimization.
TPOUR achieves a 15.21% improvement in retrieval accuracy on implicit queries while being 72.7x smaller than its larger counterpart.
