Model
IGPO
A process-supervised reinforcement learning method that attempts to assign a reward to each step of an agent. It is related work on credit assignment for search agents.
Timeline (last 90 days)
Oct 19
Process-supervision methods (such as GiGPO and IGPO) face two failure modes: surface anchoring (grouping by observation text or tool name while ignoring semantic equivalence) and internal model signals (using hidden states or attention patterns as progress proxies, which are fragile when the context is long, noisy and heterogeneous).
Unverified50%