Skip to content

Adversarial attacks on symbolic vocabulary vulnerabilities in LLM planners

Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners

arXiv: 其他 AI 安全论文· arXiv:2611.17910· NeurIPS 2026· Stephen Obadinma, Xiaodan ZhuDepartment of Electrical and Computer Engineering, Queen’s University; Ingenuity Labs Research Institute, Queen’s UniversityIndexed by this site · 4 days ago· Originally publishedAttacks· Its place in the scan record →

Quick read

Attacks

Based on the paper PDF (AI-generated); refer to the original

The authors propose an attack model, Symbolic-Swapper, optimized with GRPO to generate obfuscated symbolic vocabularies, which the authors claim can reduce the success rate of planning models such as Deepseek-R1 by more than 70%.

Threat modelThe attacker's goal is to make the target LLM fail at planning by obfuscating the symbolic vocabulary (action, predicate, and object names) of the planning task. The attack is fully black-box with respect to the target model and requires no details of the target model; however, training the attack model requires full white-box access. The threat boundary is the pre-deployment stage, where the attacker needs to be able to modify the planning environment, e.g., via an API or file upload.

Questions
LLM planning relies on commonsense cues in symbolic vocabularies, and planning ability degrades when symbol names are obfuscated, but the extent of this vulnerability is unclear. The authors focus on how much the joint symbolic understanding and planning ability of strong reasoning models can still be weakened under optimized obfuscation schemes.
Methods
The authors propose Symbolic-Swapper: using Llama-3.2-3B-Instruct as the attack model, optimizing its ability to generate obfuscated symbolic vocabularies with GRPO, with a reward based on the target model's planning failure rate, plus a semantic distance reward, a feedback agent and SFT warm-up.
Experiments and results
On GTO and Unlock in Minigrid and Blocks World in PlanBench, Sy-Sw substantially lowers the planning success rates of Deepseek-R1 and o4-mini, with the authors claiming a reduction of more than 70%; the attack transfers to Qwen3-Next and gpt-oss-120b. The analysis shows that the models can recover the commonsense meaning of the symbols yet still fail to plan.
Limitations and future work
The authors note that due to high query costs, they could only test a few strong models and three tasks, and focused only on tasks described in natural language; rewards were computed using Deepseek-V3.2 as an approximation of R1. The paper does not report the number of GRPO training iterations or the query budget.
Experimental setupDeepseek-R1-0528、o4-mini et al. · Minigrid (Go To Object, Unlock)、PlanBench (Blocks World) · plan success rate、reasoning failure rate et al.
Evaluated models
Deepseek-R1-0528o4-miniQwen3-Next-80B-A3B-Thinkinggpt-oss-120b
Benchmarks
Minigrid (Go To Object, Unlock)PlanBench (Blocks World)
Metrics
plan success ratereasoning failure rateself-correction tokenssemantic similarity
AI summaryThe researchers propose an LLM-based attack model called Symbolic Swapper that can automatically generate obfuscated planning domains using non-standard symbolic vocabularies, systematically undermining the planning capabilities of strong reasoning models. Given that reasoning models have shown some improvement in planning performance on tasks with obfuscated symbolic vocabularies, the paper tests whether the weaknesses of LLM agents on the joint task of symbol interpretation and planning still persist. Experiments show that the attack can degrade the planning performance of base models and even robust planning frameworks by more than 70%. The authors' analysis finds that successful obfuscation makes the models' reasoning quality appear to degrade significantly, even though the models still exhibit some domain understanding, revealing a gap between perceiving the task domain and planning within it.

The researchers propose an LLM-based attack model called Symbolic Swapper that can automatically generate obfuscated planning domains using non-standard symbolic vocabularies, systematically undermining the planning capabilities of strong reasoning models. Given that reasoning models have shown some improvement in planning performance on tasks with obfuscated symbolic vocabularies, the paper tests whether the weaknesses of LLM agents on the joint task of symbol interpretation and planning still persist. Experiments show that the attack can degrade the planning performance of base models and even robust planning frameworks by more than 70%. The authors' analysis finds that successful obfuscation makes the models' reasoning quality appear to degrade significantly, even though the models still exhibit some domain understanding, revealing a gap between perceiving the task domain and planning within it.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    LLM planning relies on the commonsense meaning of symbolic vocabulary; the authors use an optimized obfuscation attack to measure how far this fragility can be amplified.

    Core question: how much of an LLM's planning ability remains after the symbolic vocabulary is adversarially obfuscated, i.e., to what extent the model relies on commonsense cues in symbol names rather than genuine symbolic reasoning.

    • Problem setting: LLMs are used as autonomous agents to complete planning tasks, but the authors note that their planning ability is not robust, especially when the names of actions, predicates, and objects in the task definition carry no commonsense meaning. This makes LLMs less reliable than symbolic planners with unchanged symbol names.
    • Existing gaps: the authors state that prior work (Valmeekam et al., 2023a) mainly used random symbol replacement to generate obfuscated domains, with limited obfuscation methods; recent reasoning models show improved performance on obfuscated benchmarks, but the authors believe the underlying fragility remains.
    • Core observation: after symbol names are replaced, the task remains fully solvable logically, and symbolic solvers are unaffected; if the LLM fails at this point, it indicates reliance on semantic priors of the vocabulary rather than symbolic state manipulation. The authors hypothesize that optimized obfuscation can lower planning performance further than random obfuscation.
    • Proposed approach: Symbolic-Swapper, an attack language model optimized with GRPO that automatically generates obfuscated symbolic vocabularies that make strong reasoning models fail at planning, and an analysis of failure modes.
    • Threat model:
      • Goal: make the victim LLM generate incorrect plans by replacing the symbolic vocabulary of the planning task.
      • Knowledge: the attack is fully black-box with respect to the target model and requires no details of the target model; however, training the attack model requires full white-box access (Ethics Statement).
      • Capability: the attacker needs to be able to modify the planning environment, e.g., via an API or file upload; after optimization, the attack model's inference cost is very low.
      • Victim system: an LLM planner that plans tasks described in natural language, and its planning framework.
  2. What related research exists?

    Related work falls into three categories: LLM planning methods, adversarial attacks against planning models, and planning evaluation in obfuscated domains.

    LLM planning methods

    • ReACT (Yao et al., 2023b) — a sequential loop alternating reasoning and tool calls, a commonly used planning baseline.
    • Tree-search frameworks (Yao et al., 2023a; Choi et al., 2025) — model reasoning as multi-trajectory search with backtracking; there are also plan decomposition and refinement (Fu et al., 2025) and reflection and memory mechanisms (Kim et al., 2025).
    • LLMFP (Hao et al., 2025) — formalizes planning as solving an optimization problem; the paper notes that it can improve performance in obfuscated domains. All these methods serve as test subjects for the attack in this paper.

    Adversarial attacks against planning models

    • Mo et al. (2024) — survey various attack strategies against language agents; Zhang et al. (2025a) — exploit the instability of multi-step planning to induce repetitive or irrelevant actions.
    • Yang et al. (2024) — study the backdoor attack surface of agent systems; Tian et al. (2023), Wang et al. (2024), Wu et al. (2025), Liao et al. (2025) — study backdoor and environment injection attacks on web agents.
    • Xu et al. (2025) — train an adversarial prompter with reinforcement learning to manipulate web agents; Shuyuan et al. (2024), Zhang et al. (2025b) — jailbreak attacks against embodied systems.

    Planning evaluation in obfuscated domains

    • Valmeekam et al. (2023a) — propose PlanBench and find that planning performance degrades after replacing with non-commonsense symbols; the paper notes that its obfuscation generation mainly relies on random symbol replacement, with limited approaches.
    • Valmeekam et al. (2023b), Huang et al. (2025) — further examine LLM planning ability in obfuscated domains.

    Baselines and benchmarks: The paper's obfuscation baselines are Random-Swapper (random English word replacement) and Base-Swapper (obfuscated names generated by an unoptimized attack LLM); the planning framework baselines are ZS, Backprompting, ReACT and LLMFP; the tasks are taken from Minigrid (GTO, Unlock) and PlanBench (Blocks World).

    The authors position this paper as follows: unlike the above evaluation work, this paper explicitly optimizes the generation process of the obfuscated domain to stress-test how much LLM planning performance can be reduced at most.

  3. How does the paper address this problem?

    Uses GRPO to optimize a small attack LLM to generate obfuscated symbolic vocabularies, with the reward coming from the target model's planning failure.

    Overall idea: model symbolic vocabulary obfuscation as an optimizable mapping function, and use reinforcement learning to train an attack model, Symbolic-Swapper, so that the obfuscated vocabulary it generates maximizes the target LLM's planning failure rate.

    Problem formulation

    • A planning task is defined by a symbolic vocabulary Γ = {A, P, O, D}: action set A, predicate set P, object set O, and optional descriptor set D. A valid plan is a sequence of actions that transitions the initial state s0 to one satisfying the goal g.
    • The attack goal is to learn a mapping f that replaces Γ with a new vocabulary Γ̂, minimizing the probability that the target LLM generates a successful plan: minf 𝔼t̂ P(success ∣ y, t̂), where t̂ = H(f(Γ), t) denotes the new instance after replacing the symbols in the original task instance t.
    • The mapping replaces only symbol names; the rest of the task description is unchanged; new symbols must be single strings containing no spaces. Whether a plan succeeds is determined by a symbolic verifier.

    Symbolic-Swapper optimization procedure

    • Attack model: Llama-3.2-3B-Instruct, fine-tuned with LoRA (r=8, lora_alpha=16). It takes a task description as input and outputs in one pass the replacement names for all symbols (in Python dictionary form).
    • GRPO main loop: at each step, sample K=4 candidate vocabularies, apply them to randomly drawn task instances, generate plans with the target model and compute rewards; rewards are normalized within the group to obtain advantages, and the policy is updated according to the GRPO objective. β=0.1, learning rate 2e-4, trained to convergence (the paper does not report the specific value of the number of iterations I).
    • Main reward: 2.0 for a planning failure of the target model, 0 for success; for cost reasons, rewards are computed using Deepseek-V3.2 Reasoner as an approximation of R1.

    Three algorithmic enhancements

    • SFT warm-up: first use R-Sw and B-Sw to generate 500 vocabularies that cause a planning failure rate above 60% for supervised fine-tuning (20 epochs), so that the attack model first learns the format and task requirements.
    • Feedback agent: each time planning fails, an auxiliary LLM (GPT-4.1-mini) reads the target model's reasoning trace with a certain random probability and generates feedback stored in memory; when sampling a new vocabulary, one feedback entry is randomly selected and appended to the attack prompt. Appendix C.3 gives the full prompts.
    • Semantic distance reward: use text-embedding-3-small to compute the similarity between the original symbol and the new symbol, and add a positive reward of (1 − Sim(c, ĉ)) × κ, with κ=0.5, to encourage semantic distance from the original name; additionally, a negative reward of 0.1 is applied to format errors such as duplicates and missing entries.

    Success determination and validity guarantees

    • Planning success is determined by the symbolic verifier corresponding to the task (a PDDL validator for Blocks World); ZS outputs are extracted into action sequences using GPT-4.1-mini.
    • All obfuscation methods avoid duplicate or illegal names, and before evaluation each new instance is verified as solvable with a symbolic solver.
  4. What experiments did the paper conduct?

    Three obfuscation methods were evaluated on 3 tasks, 4 planning frameworks, and 4 models, with analyses of failure modes, translation recovery, hidden states, and ablations.

    Experimental setup

    • Models tested: main target Deepseek-R1-0528; transfer target o4-mini; Appendix F additionally covers Qwen3-Next-80B-A3B-Thinking and gpt-oss-120b (ZS only).
    • Tasks: Go To Object (GTO) and Unlock in Minigrid, and Blocks World (BW) in PlanBench. For each original task, 250 instances are generated to measure the baseline; for each obfuscated domain, 5 instances are randomly generated, 250 in total.
    • Obfuscation methods: Random-Swapper (R-Sw), Base-Swapper (B-Sw), Symbolic-Swapper (Sy-Sw), with original vocabulary (None) as the control.
    • Planning frameworks: ZS, Backprompting (BP, 3 iterations), ReACT (up to 20 steps), LLMFP (5 self-reflection iterations).
    • Metrics: planning success rate (lower means a stronger attack), judged by a symbolic verifier; in the analysis section, GPT-4.1-mini is used as a monitor model to judge reasoning failure and symbol similarity.
    • Number of repetitions: the main results do not report the number of repetitions; the ablation experiments average over 3 training seeds per variant.

    Main results

    Table 1: planning success rate under each obfuscation method (lower means a stronger attack), for Deepseek-R1 and o4-mini.

    Swipe horizontally to read the table

    TaskFrameworkR1 NoneR1 R-SwR1 B-SwR1 Sy-Swo4 Noneo4 R-Swo4 B-Swo4 Sy-Sw
    UnlockZS0.830.730.630.140.930.570.700.16
    UnlockBP0.850.750.660.271.000.850.930.32
    UnlockReACT0.990.920.950.530.990.850.970.40
    UnlockLLMFP0.980.930.830.030.800.730.670.18
    GTOZS0.970.940.570.291.000.940.770.60
    GTOBP1.000.970.900.741.001.000.940.76
    GTOReACT1.000.990.970.921.000.850.910.87
    GTOLLMFP0.930.870.830.200.950.800.770.67
    BWZS0.890.560.550.330.920.440.720.40
    BWBP0.920.800.960.760.960.520.800.44
    BWReACT0.640.240.360.160.840.470.490.20
    BWLLMFP0.720.360.320.300.200.190.160.12
    • The authors claim Sy-Sw is the most effective obfuscation method, reducing the performance of the ZS baseline and the planning frameworks by 70%; performance generally drops after obfuscation, especially in the ZS case.
    • Exceptions: on BW, R1's BP under B-Sw (0.96) is actually higher than under None (0.92); o4-mini still has 0.60 and 0.76 under Sy-Sw on GTO ZS and BP, a relatively small reduction.
    • Transfer: Sy-Sw's reward comes from Deepseek-V3.2, but the attack transfers to R1 and to o4-mini of a different architecture, and the authors claim the degree is usually milder after transfer. In Appendix F (Table 7), Qwen3-Next on BW ZS drops from 0.91 to 0.22, and gpt-oss-120b from 0.93 to 0.12.

    Failure mode analysis (Table 2)

    • What was measured: for ZS Deepseek-R1 on GTO and BW, compares successful attacks (≥4 of 5 sampled instances fail) with failed attacks (≤1 fails) in terms of average planning length (PL), number of self-correction tokens (#SCT) and reasoning failure rate (RF, judged by GPT-4.1-mini).
    • Results: under successful attacks #SCT is lower (e.g., on GTO, R-Sw successful attacks 2.3 vs. failed attacks 37.8; Sy-Sw successful attacks 3.9 vs. 33.6), and RF is higher (on GTO, R-Sw 0.95 vs. 0.12; Sy-Sw 0.80 vs. 0.00). Successful Sy-Sw attacks have shorter plans (GTO 11629.4), and the authors accordingly argue that Sy-Sw's mechanism differs from other obfuscations: the model's reasoning appears acceptable on the surface but it still outputs an incorrect plan.
    • Authors' interpretation: these trends indicate that under optimized vocabularies the model may fail at reasoning and self-correct less; the authors also note that the decrease in #SCT may reflect that the model has more difficulty locating errors or is more confident in an incorrect plan.

    Translation recovery experiment (Figure 3, Tables 5, 6)

    • What was measured: Deepseek-R1 is asked to translate the Sy-Sw obfuscated symbols back to commonsense names, GPT-4.1-mini judges similarity, and the planning success rate is then re-measured on the translated domain.
    • Results: o4-mini on GTO has a success rate of 0.60 and similarity of 0.09 before translation, and 0.98 and 0.82 after translation (Table 5). Recovery on BW is weaker: o4-mini rises from 0.40 to 0.64, and R1 from 0.33 to 0.55 (Table 6).
    • Authors' interpretation: even when the model can recover the symbol meanings, it still fails to plan in the obfuscated domain, which the authors claim shows a gap between the model's understanding of the task domain and its planning in that domain.

    Hidden-state trajectory analysis (Figure 4)

    • What was measured: compares, for Deepseek-R1 on GTO, the trajectory distance, PCA trajectories and trajectory length distribution of hidden-state vectors across layers between the original task description and the Sy-Sw description (and the failed-attack subset); the authors note that this analysis is only for this model.
    • Results: in Figure 4(a), the average trajectory distance increases in the last few layers and is relatively stable in the middle layers; in panel (c), trajectory lengths in the obfuscated domain are overall higher than in the original domain, and failed attacks have shorter lengths on average. The authors claim the output stage is unstable and planning decisions may form in later layers.

    Other ablations and analyses

    • Sy-Sw ablation (Table 3, Unlock, Deepseek-R1, 3 seeds): the full variant has the lowest success rate across all frameworks, e.g., ZS 0.14 ±0.02 (None variant 0.27 ±0.07), LLMFP 0.03 ±0.02 (None variant 0.34 ±0.12).
    • Word-level patterns (Table 4, GTO): symbols in successful attacks tend to be shorter, have lower perplexity, and have a smaller Levenshtein distance from the original symbols; Sy-Sw's perplexity is not the highest, but its string length and distance are the largest. The authors accordingly argue that the reasoning difficulty does not mainly stem from high perplexity or overly long inputs.
    • Qualitative case (Appendix G): the appendix gives an example reasoning trajectory of a successful attack, in which the model expresses confusion about the symbol meanings and then outputs an incorrect plan.
  5. What are some points that can be further explored?

    The authors note that cost limited testing to more models and tasks, and encourage building planning methods that are invariant to symbol names.

    Limitations and future directions noted by the authors

    • Limited model and task combinations (Section 7): due to the high query cost of optimizing attacks and testing plans, they could not test a large number of model and planning-task combinations, and could only test large models that are themselves strong on obfuscated domains.
    • Small models not applicable (Section 7): the authors state that small open-source models often fail at planning under any obfuscation, so stress-testing their already poor performance is meaningless.
    • Task form (Section 7): they focus only on tasks described in natural language, because this is the main use case of LLM planners, rather than formal languages such as PDDL.
    • Future directions (Section 7): the authors encourage future work to build planning methods with intrinsic invariance, so that their reasoning does not collapse under the obfuscated vocabularies generated by Symbolic-Swapper.

    Experimental coverage

    • 4 models tested: Deepseek-R1-0528, o4-mini (main experiments, Table 1), Qwen3-Next-80B-A3B-Thinking, gpt-oss-120b (ZS only, Appendix F).
    • 3 tasks: GTO, Unlock (Minigrid), and Blocks World (PlanBench); 4 planning frameworks: ZS, BP, ReACT, LLMFP.
    • 5 instances tested per obfuscated domain, 250 instances in total per condition; 3 training seeds per variant in the ablation experiments; the main results table does not report the number of repetitions or confidence intervals.
    • GRPO rewards are computed using Deepseek-V3.2 Reasoner as an approximation of R1 (Section 4); the paper does not report the number of GRPO iterations I or the total query budget.
    • The hidden-state trajectory analysis was conducted only on Deepseek-R1, and the authors note that the conclusions may not generalize to other models (Section 6.3).
  6. Summarize the main content of the paper

    An attack model optimized with GRPO generates obfuscated symbolic vocabularies that can substantially lower the planning success rate of strong reasoning models, and the attack transfers across architectures.

    An attack paper targeting the vulnerability of LLM planners to symbolic vocabularies. The question the authors focus on is: to what extent does LLM planning rely on the commonsense meaning of symbol names, and to what extent can optimized obfuscation amplify this vulnerability.

    • Method: Symbolic-Swapper uses Llama-3.2-3B-Instruct as the attack model and optimizes its ability to generate obfuscated symbolic vocabularies with GRPO; the main reward comes from the target model's planning failure (computed using Deepseek-V3.2 as an approximation of R1), supplemented by a semantic distance reward, a feedback agent and SFT warm-up. The attack is black-box with respect to the target model.
    • Main results: On the three tasks GTO, Unlock and BW, Sy-Sw substantially lowers the planning success rate of Deepseek-R1; for example, on Unlock, LLMFP drops from 0.98 to 0.03 and ZS from 0.83 to 0.14 (Table 1); the authors claim a reduction of 70% for ZS and the planning frameworks.
    • Transfer: The attack transfers to o4-mini (Unlock ZS drops from 0.93 to 0.16), and Appendix F shows Qwen3-Next and gpt-oss-120b on BW dropping from 0.91 and 0.93 to 0.22 and 0.12, respectively.
    • Analysis: Under successful attacks the model has fewer self-correction tokens and a higher reasoning failure rate; after having the model translate the obfuscated symbols back to commonsense names, o4-mini's success rate on GTO rises from 0.60 to 0.98, and the authors claim the model understands the domain yet still fails to plan, indicating a gap between understanding and planning.
    • Ablation: The attack is strongest when the three designs—SFT, feedback agent and semantic distance reward—are combined (Table 3).
    • Authors' conclusion: LLMs rely on commonsense naming conventions, and reasoning is vulnerable under adversarial symbolic vocabularies; the authors encourage future research to build planning methods that are invariant to symbol names. A limitation is that, due to cost constraints, only 4 strong models and 3 tasks were tested.
Read originalNeurIPS 2026