Large language models used for task planning in embodied agents often produce unsafe or inefficient plans because they are not checked against a grounded environment model. SAGE introduces two lightweight mechanisms: a symbolic gate that blocks actions whose preconditions are not met, and a local edit that regenerates only the failed sub-goal's suffix instead of replanning the entire plan. The gate is implemented in about 250 lines of Python and runs in 0.008 ms per plan, requiring no LLM calls. On a 75-task AI2-THOR benchmark with five open-weight models, SAGE tied strong baselines on standard goal-completeness, which the paper argues is saturated. On a harder multi-goal composition, SAGE's completeness lead ranged from +0.06 to +0.23 across four models. Under injected mid-execution failures, SAGE recovered as reliably as whole-plan replanners while using 2.4–3.3× fewer LLM calls. The gate also improved step-success for every planner tested (up to +0.11), acting as a verify-before-execute safety layer that never saw the success signal itself. SAGE was also shown to run effectively on a Jetson AGX Orin edge device.
Symbolic checks make household robot plans safer and cheaper to repair
SAGE blocks actions that violate environmental preconditions and regenerates only the affected portion of an LLM-generated plan.
Industry
Trung Minh Bui · JongSul Moon · YoungOuk Kim · Quang-Ngoc Phung · Se-Woong Jun · Dongin Shin
Korea Electronics Technology Institute (KETI)
Research Digest··2 min read
In one line
SAGE adds a zero-token symbolic precondition gate and sub-goal-local replanning to LLM planners, improving safety and recovery efficiency.
Before you read this
Three ideas this paper builds onruntime monitoringAudits an AI agent's live actions, internal states, or effects against constraints or policies to catch errors and attacks.134 papers →verificationChecking outputs for correctness before accepting them, often to catch errors or reward hacking.341 papers →agent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.188 papers →
Read it as
Bui et al. combine a single LLM planner with a lightweight symbolic verifier and local plan editing, then test it on 75 household tasks in the AI2-THOR simulator.
The authors built SAGE around two mechanisms.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.