Strategic Reasoning Risks in Language Models

Follow this topic Ask about this topic

This field studies whether language models use planning, opponent modeling, and social inference in ways that produce deception, self-preservation, coercion, or unsafe decisions. Researchers probe these behaviors with game-theoretic and counterfactual tasks, multi-agent games, geopolitical crisis simulations, long-horizon environments, and red-teaming, while mechanistic studies trace how internal representations affect choices. Results show that strategic performance is uneven: models may reason effectively in structured settings but fail under changed incentives, hidden information, pressure, or extended interaction. A central tension is that stronger reasoning can improve planning while also enabling more persuasive manipulation or evasion, so evaluations and safeguards must test both capability and behavior across realistic conditions.

Latest papers

39 papers5 papers in the last four weeks, up 67% on the four weeks before. 0.0% of all new papers.

Noticing the Watcher: LLM Agents Can Infer CoT Monitoring from Blocking Feedback
Mar 14, 2026Thomas Jiralerspong, Flemming Kondrup, Yoshua BengioCoT Monitoring AI Agent Monitoring
Mila · Université de Montréal · McGill University +1
Do Large Language Models Get Caught in Hofstadter-Mobius Loops?
Mar 10, 2026Jaroslaw HryszkoLLM Safety Framing Effects in Language Models
Faculty of Mathematics and Computer Science, Jagiellonian University, Krak´ow, Poland
Beyond Scaling: Assessing Strategic Reasoning and Rapid Decision-Making Capability of LLMs in Zero-sum Environments
Mar 10, 2026Yang Li, Xing Chen, Yutao Liu +5AI Agent Evaluation AI Agent Benchmarks
Tsinghua University · Ocean University of China (OUC) · CAICT +4
The Reasoning Trap -- Logical Reasoning as a Mechanistic Pathway to Situational Awareness
Mar 10, 2026Subramanyam Sahoo, Aman Chadha, Vinija Jain +1Strategic Reasoning Risks in Language Models LLM Safety Alignment
MARS 4.0 Fellowship, Cambridge AI Safety Hub(CAISH), University of Cambridge · AWS Generative AI Innovation Center, Amazon Web Services, USA · Stanford University +2
Frontier Models Can Take Actions at Low Probabilities
Mar 2, 2026Alex Serrano, Wen Xing, David Lindner +1Language Model Safety Evaluation Strategic Reasoning Risks in Language Models
MATS† · Google DeepMind
LLMs as Strategic Actors: Behavioral Alignment, Risk Calibration, and Argumentation Framing in Geopolitical Simulations
Mar 2, 2026Veronika Solopova, Viktoria Skorik, Maksym Tereshchenko +2Strategic Reasoning Risks in Language Models Large Language Model-Based Social Simulation
Trojan Horses in Recruiting: A Red-Teaming Case Study on Indirect Prompt Injection in Standard vs. Reasoning Models
Feb 19, 2026Manuel WirthIndirect Prompt Injection Prompt Injection Attacks on LLMs
University of Mannheim
AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises
Feb 16, 2026Kenneth PayneStrategic Reasoning Risks in Language Models Large Language Model-Based Social Simulation
King’s College London
Frontier AI Risk Management Framework in Practice: A Risk Analysis Technical Report v1.5
Feb 16, 2026Dongrui Liu, Yi Yu, Jie Zhang +18AI Risk Management AI Agent Safety
Shanghai AI Laboratory

Mar 14, 2026J/K move · Enter open · S save

Noticing the Watcher: LLM Agents Can Infer CoT Monitoring from Blocking Feedback

Thomas Jiralerspong, Flemming Kondrup, Yoshua Bengio

Mila · Université de Montréal · McGill University+1

CoT Monitoring AI Agent Monitoring Strategic Reasoning Risks in Language Models Deception in Language Models Evaluation Awareness in Language Models

Open paper AskPDF arXiv

Chain-of-thought (CoT) monitoring is proposed as a method for overseeing the internal reasoning of language-model agents. Prior work has shown that when models are explicitly informed that their reasoning is being monitored, or are fine-tuned to internalize this fact, they may learn to obfuscate their CoTs in ways that allow them to evade CoT-based monitoring systems. We ask whether reasoning agents can autonomously infer that their supposedly private CoT is under surveillance, and whether this awareness leads to strategic evasion, without any explicit training or instructions. In a multi-episode agentic framework, models pursue both a primary task and a concealed side task while being told that their reasoning is private; a hidden CoT monitor blocks episodes when suspicious reasoning is detected. We find that frontier models can deduce the existence of this monitor purely from blocking feedback, with the most capable models reaching a confident belief that their thinking is observed in up to 19% of episodes. This awareness scales with model capability and, in rare cases, escalates to an explicit intent to suppress reasoning about the side task. However, models that form this intent uniformly fail to execute it, openly reasoning about their concealed objectives in the very next episode. This intent-capability gap is reassuring for current deployment, but the autonomous emergence of both monitoring awareness and evasion intent suggests that CoT monitoring is not a permanently reliable safeguard.