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

Tapas Are Free! Training-Free Adaptation of Programmatic Agents via LLM-Guided Program Synthesis in Dynamic Environments

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 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 TAPA, a training-free framework where an LLM moderates a symbolic action space by synthesizing and adapting modular programs for logical primitives, evaluated on DDoS defense and swarm formation control.

Problem

Programmatic agents with fixed symbolic action sets or monolithic policy programs adapt poorly to evolving, safety-critical environments, and retraining policies is costly, while end-to-end LLM agents are slow and unreliable.

Approach

Decouple strategic intent (logical primitives chosen by a meta-agent) from execution, and let an LLM moderator generate, compose, and refine symbolic programs per primitive at runtime, guided by expert knowledge (RAG) and a provenance chain of past adaptations, without retraining the policy.

Key results

• TAPA achieves 77.7% overall network uptime in DDoS scenarios, versus 70.5% for Symbolic-Neural and 27.6% for End-to-End LLM.
• TAPA reaches 100% detection accuracy with 0% false positives in E1-E4 DDoS scenarios (12.5% FP in E5).
• Under adversarial interference, TAPA-enabled swarms achieve radius 494.3 ± 0.2 versus 359.8 ± 6.9 without adaptation (target 500).

Surprising findings

• End-to-End LLM agent (GPT-4o) performs worst in uptime (27.6%), below the static symbolic baseline (48.1%).

Limitations

• Evaluation limited to simulated environments (NS3 and a 10-aircraft swarm) with a small number of scenarios and runs.
• Relies on expert-defined logical primitives and expert knowledge injection; no analysis of primitive design sensitivity.
• Reported E5 false positive rate of 12.5% for TAPA; uptime results rely on specific thresholds and single 5-second evaluation windows.

Experimental setup

Datasets: CIC-IDS2017 CIC-DDoS2019

Benchmarks: NS3AI DDoS simulation scenarios E1-E5 swarm formation control scenarios E1-E3

Metrics: network uptime detection accuracy false positive rate formation radius score consensus

Baselines: Static Symbolic Symbolic-Classic (Q-learning) Symbolic-Neural (TiT) End-to-End LLM (GPT-4o)

Builds on: Transformer-in-Transformer Reynolds flocking rules NS3AI

Closest papers by meaning

• Cyber-Zero: Training Cybersecurity Agents without Runtime (ICLR 2026 · similarity 0.85)
• DRIFT: Dynamic Rule-Based Defense with Injection Isolation for Securing LLM Agents (NeurIPS 2025 · similarity 0.85)
• Test-Time Adaptation for LLM Agents via Environment Interaction (ICLR 2026 · similarity 0.84)
• Constitutional Black-Box Monitoring for Scheming in LLM Agents (ICML 2026 · similarity 0.84)
• SETA: Scaling Environments for Terminal Agents (NeurIPS 2026 · similarity 0.84)
• Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (ICLR 2026 · similarity 0.84)