SAGE: Symbolic Action-Gating and Editing for LLM Task Planners

简体中文

L1 core-mechanism diagnostic. Runs a zero-token symbolic gate and local suffix editing, not AI2-THOR.

Paper information

Field	Details
Paper	SAGE: Symbolic Action-Gating and Editing for LLM Task Planners
Organization	No first-author affiliation is listed on the paper's first page.
First public release	2026-09-28 (arXiv v1)
Authors' code	mtbui2010/sage is public.
Adapter / method	sage-planner
Local reproduction code	src/auto_research/agent_research/latest_20260930_closure.py

Original paper

Motivation and method

A symbolic precondition gate blocks unsafe actions before execution and returns a typed reason. On failure, rewrite only the suffix for the relevant subgoal and retain the completed prefix.

flowchart LR
  I[Public input and current state] --> M[sage-planner mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or state update]

Key figure from the paper

Figure 2: Overall workflow, key stages, and data flow. From the original paper; copyright remains with the authors. Click the image for the source.

Core computation

The reference kernel retains the paper's decisive gates, masks, recursion, credit weights, or promotion conditions. Tests also check gradient isolation, type boundaries, and determinism.

Paper results

Benchmarks, speed, and training curves belong to the paper. The local three-seed mini-suite checks mechanisms and invariants, not paper-scale results or online gains.

Local reproduction

See the three-seed diagnostic, marked diagnostic_only=true and ineligible for formal capability ranking.

Reproduction boundaries

The zero-token symbolic gate and local suffix editor run. AI2-THOR was not run.