This paper investigates a critical but often overlooked vulnerability in Large Language Model (LLM) planners: their over-reliance on the "commonsense" meaning of variable names. While a traditional symbolic planner doesn't care if you call an action pick_up or x7_alpha, LLMs struggle significantly when the symbolic vocabulary is obfuscated.
TL;DR
The authors introduce Symbolic-Swapper, an adversarial framework that uses reinforcement learning to generate deceptive variable names that systematically break LLM planning. Their attacks reduce the success rates of state-of-the-art models like DeepSeek-R1 and o4-mini by over 70%, revealing that even "reasoning" models rely heavily on linguistic shortcuts rather than robust symbolic logic.
What Problem Does This Paper Address?
As we move toward "Agentic AI," we are increasingly tasking LLMs with sequential decision-making in complex environments. However, unlike classical symbolic planners (like SMT solvers), which are invariant to variable renaming, LLMs bring massive "semantic priors" to the table. They "know" what pick_up means before they even read the task rules.
The authors argue that this is a double-edged sword. While it helps LLMs solve tasks quickly, it masks a fundamental lack of robustness. If an adversary (or even just a non-standard software API) uses arbitrary or counter-intuitive names for actions and objects, the LLM’s planning ability collapses. This paper seeks to quantify this fragility and understand why it happens.
Methods
The core contribution is Symbolic-Swapper, an adversarial attack model designed to find the "weakest points" in an LLM's symbolic interpretation.
- The Attacker: A lightweight LLM (Llama-3.2-3B) is trained to generate a mapping from original symbols (e.g.,
stack) to deceptive ones (e.g.,Aeeoakr). - Optimization via GRPO: The authors use Group Relative Policy Optimization (GRPO)—the same RL algorithm used to train DeepSeek-R1—to fine-tune the attacker. The reward function prioritizes:
- Target Failure: Maximizing the planning failure rate of the victim LLM.
- Semantic Distance: Ensuring the new names are semantically far from the originals.
- Formatting: Penalizing the attacker for invalid outputs.
- Feedback Loop: An auxiliary "Feedback Agent" analyzes the victim's failed reasoning traces and provides natural language tips to the attacker on how to be more deceptive.
Datasets
The authors evaluated their approach on three popular planning benchmarks, presented in natural language:
- Minigrid (Go To Object): Navigating a grid to interact with a specific item.
- Minigrid (Unlock): A multi-step task requiring the agent to find a key, navigate to a door, and unlock it.
- PlanBench (Blocks World): The classic AI challenge of stacking blocks in a specific order.
Key Insights and Ideas
- The Perception-Reasoning Gap: In a fascinating experiment, the authors asked the LLMs to "translate" the obfuscated names back to commonsense terms. The models were surprisingly good at this, often correctly identifying that a nonsense string meant
pick_up. However, they still failed to plan correctly using those strings. This suggests that the models can perceive the domain's structure but cannot maintain that mapping during active reasoning. - Chain-of-Thought Collapse: When under attack, the "reasoning quality" of models like DeepSeek-R1 degrades. Instead of logical steps, their internal traces become vague, repetitive, or poorly justified.
- Generalization: Attacks optimized against one model (e.g., DeepSeek) successfully transferred to others (e.g., OpenAI’s o4-mini), indicating that these symbolic vulnerabilities are common across different architectures.
Key Results
The results demonstrate a staggering drop in performance across all tested frameworks, including zero-shot prompting, ReACT, and even LLMFP (a framework that uses formal solvers).
- DeepSeek-R1: On the "Unlock" task, success dropped from 83% (None) to just 14% under the Symbolic-Swapper attack.
- o4-mini: On "Blocks World," performance plummeted from 92% to 40%.
- Framework Vulnerability: Even robust planning methods like ReACT saw performance cuts of over 50% in several scenarios.
As shown in Figure 1, the optimized Symbolic-Swapper is significantly more effective at breaking models than a simple random word generator.
Limitations and Future Work
The authors acknowledge several constraints:
- Computational Cost: Running RL on planning tasks is extremely token-intensive and expensive.
- Model Scale: The study focused on large, high-performing models; smaller models were already too weak at planning to be interesting targets for "stress testing."
- Natural Language Focus: The tasks were presented in natural language. Future work could explore if strictly formalized languages like PDDL offer more inherent protection (though current trends suggest LLMs struggle even more with formal syntax).
The ultimate takeaway is a call to action for the research community: to build agents that are representation-invariant, capable of reasoning through the logic of a task regardless of what the variables are called.
Summary updated