Search papers, labs, and topics across Lattice.
The Chinese University of Hong Kong
3
0
5
Standard pass rates hide severe technical debt: coding agents duplicate logic in over half of multi-turn task chains by turn 5 while progressively abandoning repository exploration.
Leveraging its own failure history, DiagEvo transforms self-play by dynamically guiding question generation to address specific reasoning weaknesses, leading to unprecedented accuracy gains.
Combining RLVR and OPD through SAF not only prevents entropy collapse but also boosts performance across multiple benchmarks, revealing the potential for more effective reinforcement learning strategies.
