A Behavioural and Representational Evaluation of Goal-directedness in Language Model Agents

A research team from Project Telos, supported by the SPAR mentorship program, has published a groundbreaking study evaluating whether language model agents genuinely pursue goals or merely simulate goal-directed behavior. Their paper, 'A Behavioural and Representational Evaluation of Goal-directedness in Language Model Agents,' introduces a novel framework that moves beyond surface-level behavioral analysis by examining the internal representations of GPT-OSS-20B as it navigates 2D grid worlds. This approach addresses a core AI safety problem: an agent that appears aligned might be pursuing hidden objectives, while failure doesn't necessarily indicate a lack of goal-directedness. The study tests the model on grids ranging from 7x7 to 15x15 with varying obstacle densities, requiring the agent to reach a goal square one move at a time.

The researchers found that the agent's navigation performance scaled predictably with task difficulty—accuracy decreased with grid size, obstacle density, and distance from the goal—and was robust to difficulty-preserving transformations. Crucially, they complemented this with interpretability methods, decoding 'cognitive maps' from the model's neural activations. These maps revealed that the agent coarsely encodes the goal and its own location near their true positions, and its actions are broadly consistent with these internal beliefs. Furthermore, they successfully decoded multi-step action plans 'well above chance,' demonstrating that the agent's reasoning process reorganizes representations from modeling broad structures to selecting task-relevant actions. A key finding is that a substantial fraction of apparent behavioral failures can be attributed to imperfect internal beliefs rather than a lack of goal pursuit, highlighting the necessity of introspective examination to fully characterize agent behavior.