GRACE uses typed semantic graphs to improve agentic context reliability under distribution shift

Researchers propose Graph-Regularized Agentic Context Evolution (GRACE) to address the difficulty of verifying mutable system instructions in long-horizon LLM agents. GRACE maintains persistent instructions as a typed semantic graph, validating updates within local neighborhoods before reconstructing them as incremental textual edits.

• Evaluated on a fixed telecom agent harness derived from $\tau^2$-bench under a controlled distribution-shift protocol.
• Improved strict reliability (pass^3) from 0.091 to 0.673±0.136 across five replications, exceeding the Gemini 3.1 Pro zero-shot reference of 0.242 and the flat-text baseline of 0.191±0.051.
• Identifies structural substrates for local verification and consolidation mechanisms for usable accumulated content as key requirements for reliable evolution.

The approach demonstrates that maintaining context as a typed graph enables effective verification and usability during long-term operational updates.