Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Affiliation:
3
0
6
LLM agents can learn complex tasks more efficiently by adaptively injecting skills only when their utility is confirmed through outcome-aware likelihood comparisons on existing data, bypassing expensive environment rollouts.
Forget outcome-based filtering: TopoCurate uses interaction topology to surface informative tool-use trajectories and tasks, boosting SFT and RL performance by up to 6.9%.
Forget expensive MoE training from scratch: ExpertWeaver unlocks inherent MoE structure within dense LLMs using GLU activation patterns, offering a training-free conversion.