Adaptive Consistency Graph for Long-Horizon Agents
ACG keeps long agent runs tied to the original requirement by rebuilding the context around each decision.
The paper describes a persistent graph that records execution evidence and provenance as an agent works through dependent tool calls. For each step, it creates a temporary requirement-centered view that fits within a bounded context budget. In the matched evaluation, ACG raised GPT-5.6-luna’s average success rate from 44.5% with ReAct to 50.2%, with its largest reported gain on BrowseComp-Plus. HF Daily Papers' note
The paper describes a persistent graph that records execution evidence and provenance as an agent works through dependent tool calls. For each step, it creates a temporary requirement-centered view that fits within a bounded context budget. In the matched evaluation, ACG raised GPT-5.6-luna’s average success rate from 44.5% with ReAct to 50.2%, with its largest reported gain on BrowseComp-Plus. HF Daily Papers' note
score 5