Research map / Multi-agent systems and games / LLM agents

FaSTA*: Fast-Slow Toolpath Agent with Subroutine Mining for Efficient Multi-turn Image Editing

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: GUI agent memory and trajectories · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FaSTA*, a neurosymbolic multi-turn image editing agent that mines reusable tool-call subroutines from past toolpaths via LLM inductive reasoning and uses fast subroutine-based planning with A* search as a fallback, cutting execution cost about 49% versus CoSTA*.

Problem

Multi-turn image editing requires long toolpaths over many AI tools, and A* search over tool subgraphs (as in CoSTA*) is computationally expensive, with no reuse of past exploration.

Approach

LLMs inductively mine frequently successful tool-call subroutines with activation rules from explored toolpaths and store them in a rule table; planning tries these subroutines first (fast plan) and falls back to per-subtask A* search only when no subroutine applies or a VLM quality check fails (slow plan).

Key results

• FaSTA* reduces average execution cost by over 49.3% relative to CoSTA* (about 29.5s vs 46.8s) with about 3.2% quality drop (0.91 vs 0.94 overall accuracy on the CoSTA* benchmark).
• On a Complex-Edit subset, average cost falls from 78.27s to 55.12s while accuracy goes from 0.89 to 0.87.
• Fast-slow planning falls back to A* on only 9% of subtasks; fast-plan-only quality drops to 0.84 versus 0.91 for FaSTA*.

Surprising findings

• Using a smaller Qwen 2.5VL 3B for subroutine selection picked wrong subroutines about 40% of the time, raising total cost by about 20% while quality stayed consistent.
• Subroutine verification reduced low-level fallback from 28% to 9% of subtasks.

Limitations

• Cold start: performance on entirely new tasks is suboptimal until enough experiences are collected.
• Performance depends on the capabilities of the underlying LLM; smaller LLMs make more selection errors.
• Quality on the CoSTA* benchmark is slightly lower than CoSTA* (about 3.2% drop).

Experimental setup

Datasets: CoSTA* benchmark (121 image-prompt pairs, 550 manipulations) Complex-Edit (subset) MagicBrush (qualitative)

Benchmarks: CoSTA* benchmark Complex-Edit subset MagicBrush (qualitative)

Metrics: human-evaluated task accuracy execution time (s) per image

Baselines: CoSTA* VisProg CLOVA GenArtist InstructPix2Pix MagicBrush

Builds on: CoSTA* in-context reinforcement learning VisProg CLOVA

Closest papers by meaning

• JarvisArt: Liberating Human Artistic Creativity via an Intelligent Photo Retouching Agent (NeurIPS 2025 · similarity 0.86)
• RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.86)
• Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing (NeurIPS 2026 · similarity 0.85)
• AdaReasoner: Dynamic Tool Orchestration for Iterative Visual Reasoning (ICLR 2026 · similarity 0.85)
• Meta-CoT: Enhancing Granularity and Generalization in Image Editing (CVPR 2026 · similarity 0.84)
• EdiVal-Agent: An Object-Centric Framework for Automated, Fine-Grained Evaluation of Multi-Turn Editing (ICLR 2026 · similarity 0.84)