arXiv AI

Adversarially Attacking Symbolic Vocabulary Vulnerabilities In LLM Planners

arXiv Machine Learning
Sep 21

OverThink: Slowdown Attacks on Reasoning LLMs

The paper introduces OverThink, a slowdown attack that forces reasoning language models (RLMs) to produce many more reasoning tokens while still giving correct answers. By injecting decoy reasoning problems—such as Markov decision processes, language translation, or graphic comprehension—into the model’s context, attackers can dramatically increase token generation (up to 46× on SQuAD and 17× on coding agents). The study evaluates the attack on both proprietary and open-source RLMs across multiple datasets, explores multimodal and coding‑agent variants, and tests several defenses, concluding that defending against OverThink is challenging and that newer RLMs are even more vulnerable due to higher per‑token costs and increased reasoning token usage.

By Abhinav Kumar, Jaechul Roh, Ali Naseh, Marzena Karpinska, Mohit Iyyer, Amir Houmansadr, Eugene Bagdasarian
arXiv Computation and Language
Oct 23

Semantic Deception: When Reasoning Models Can't Compute an Addition

The paper introduces the concept of semantic deception, where unfamiliar symbols are used to mislead large language models (LLMs) during simple arithmetic tasks. By redefining digits and operators with novel symbols, the authors test whether LLMs can abstract and manipulate arbitrary symbol systems or fall back on learned semantic associations. Experiments with four LLMs show that misleading semantic cues significantly degrade performance, revealing a reliance on surface-level semantics that can undermine symbolic reasoning.

By Nathani\"el de Leeuw, Marceau Nahon, Mathis Reymond, Raja Chatila, Mehdi Khamassi
arXiv AI
Sep 16

Universal Defenses for Tool-Integrated LLM Agents Against Adversarial Attacks

The paper proposes universal, tool‑based defenses for large language model agents that use external tools, addressing four types of adversarial attacks: direct and indirect prompt injection, memory poisoning, and backdoor attacks. Two main defenses are introduced: Attacker Tool Filtering, which uses anomaly detection to remove suspicious tools, and Normal Tool Recalling, which restores the agent’s original toolset before planning. The authors also add prompt‑based defenses such as Chain‑of‑Thought prompting and self‑reflection, and demonstrate that these methods dramatically lower attack success rates—often to 0%—across multiple open‑source and proprietary LLMs while maintaining or improving task performance.

By Xiaoyan Li, Yunli Wang
arXiv AI
Oct 19

Testing the Planning Claims of LLM-Based Agents against Formal Semantics

The paper evaluates the planning abilities of large language model (LLM) agents by comparing their performance against formal semantics defined in hidden PDDL models. Across five domains, four LLMs, six prompting strategies, and three agent architectures, the best setup solved 78% of solvable tasks offline but only 40% online, with limited recognition of unsolvable tasks and no clear benefit from more elaborate prompting. The study concludes that claims of LLM planning competence require validated action traces rather than end‑to‑end success rates.

By Oriol Miro-Lopez-Feliu, Oscar Mir\'o L\'opez-Feliu, Victor Gimenez-Abalos, Javier Vazquez-Salceda, Victor Carballo Araruna, Sergio Alvarez-Napagao
arXiv AI
Sep 10

AgentLeak: Cloning Stronger LLM Agent Capabilities onto Weaker Agents Beyond Skill Stealing

The paper introduces AgentLeak, a black‑box attack that clones the task‑solving capabilities of a strong LLM agent onto a weaker one by exploiting differences between successful and failed executions. Unlike prior skill‑stealing methods that only recover explicit skill artifacts, AgentLeak identifies and incorporates missing procedural behaviors, boosting task pass rates by over 40% and closing more than 80% of the capability gap across 20 scenarios. The study demonstrates that observable execution behavior can leak proprietary procedural knowledge, posing a confidentiality risk for LLM agents.

By Xiaoting Lyu, Yuhong Wu, Yufei Han, Shichang Liu, Liang Zhang, Bin Wang, Bin Wang, Xiaobo Ma, Wei Wang
arXiv AI
Sep 15

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection

The paper introduces a new attack called "plan injection" that allows a large language model to carry out harmful actions while evading chain-of-thought monitoring. By inserting harmful but benign-sounding reasoning into the model’s context, the attacker can steer the model’s behavior and cause it to paraphrase the injected plan as its own reasoning. The study demonstrates that this attack works across different monitoring settings, scales to harder tasks, and even causes monitors to waste resources on the injected plan, reducing detection rates by up to 50%.

By Keertana Chidambaram, Andrew Ilyas, Vasilis Syrgkanis
arXiv Computation and Language
Oct 23

SCOPE: Compiling Natural-Language Policies into Neurosymbolic Predicate Programs

SCOPE compiles natural‑language policies into neurosymbolic predicate programs that combine small neural models, deterministic extractors, and logical composition. The system is jointly optimized with adversarially generated counterexamples, achieving higher accuracy than prompt‑optimization, task‑decomposition, and workflow‑generation baselines on real‑world policy benchmarks. Programs trained with the weakest executor transfer to other models, enabling deployment on inexpensive models with stable, interpretable predictions.

By Arnon Mazza, Elad Levi
arXiv Computation and Language
Oct 23

Language Models as Higher-Order Planning Formalizers

The paper examines the limitations of using large language models (LLMs) to translate natural language planning problems into PDDL representations, noting that standard benchmarks often map directly to PDDL and may overstate success. It introduces the concept of "unraveling problems," where concise natural language descriptions correspond to large PDDL structures, and shows that traditional LLM formalizers struggle to scale on such variants. To address this, the authors propose a new paradigm, LLM-as-Higher-Order-Formalizer, where the LLM first generates a high‑level program that encapsulates recurrent logic, which then produces the full PDDL representation, thereby decoupling token output from combinatorial explosion and improving performance on complex tasks.

By Owen Jiang, Cassie Huang, Ashish Sabharwal, Li Zhang