Reasoning Graphs: Deterministic Agent Accuracy through Evidence-Centric Chain-of-Thought Feedback

Use this page to decide whether the paper is strong enough to influence an eval design. It summarizes the abstract plus available structured metadata. If the signal is thin, use it as background context and compare it against stronger hub pages before making protocol choices.

Abstract

Language model agents reason from scratch on every query: each time an agent retrieves evidence and deliberates, the chain of thought is discarded and the next similar query starts with no prior insight. This produces lower accuracy and high variance, as the same type of query can succeed or fail unpredictably. We introduce reasoning graphs, a graph structure that persists an agent's per-evidence chain of thought as structured edges connected to the evidence items they evaluate. Unlike prior memory mechanisms that store distilled strategies as flat records indexed by query similarity or appended by recency, reasoning graphs enable evidence-centric feedback: given a new candidate set, the system traverses all incoming evaluation edges for each evidence item across all prior runs, surfacing how that specific item has been judged before. This backward traversal from evidence inward is a structurally different capability from query-similarity retrieval, because the feedback is tied to the specific evidence the agent is currently examining, not to the query. We further introduce retrieval graphs, a complementary structure that feeds a pipeline planner to tighten the candidate funnel over successive runs. Together, both graphs form a self-improving feedback loop: accuracy rises and variance collapses over successive runs, with every decision fully traceable through the graph. This improvement requires no retraining; the base model remains frozen and all gains come from context engineering via graph traversal. We formalize the graph structure, traversal algorithms, and feedback mechanisms, and describe a sequential cluster evaluation protocol for measuring accuracy convergence and variance collapse on multi-hop question answering benchmarks.