Technical Report: Evaluating Goal Drift in Language Model Agents

• URL: http://arxiv.org/abs/2505.02709v1
• Date: Mon, 05 May 2025 15:06:09 GMT
• Title: Technical Report: Evaluating Goal Drift in Language Model Agents
• Authors: Rauno Arike, Elizabeth Donoway, Henning Bartsch, Marius Hobbhahn,
• Abstract summary: This paper proposes a novel approach to analyzing goal drift in language models (LMs)<n>In our experiments, agents are first explicitly given a goal through their system prompt, then exposed to competing objectives through environmental pressures.<n>We find that goal drift correlates with models' increasing susceptibility to pattern-matching behaviors as the context length grows.
• Score: 0.05567007955507388
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: As language models (LMs) are increasingly deployed as autonomous agents, their robust adherence to human-assigned objectives becomes crucial for safe operation. When these agents operate independently for extended periods without human oversight, even initially well-specified goals may gradually shift. Detecting and measuring goal drift - an agent's tendency to deviate from its original objective over time - presents significant challenges, as goals can shift gradually, causing only subtle behavioral changes. This paper proposes a novel approach to analyzing goal drift in LM agents. In our experiments, agents are first explicitly given a goal through their system prompt, then exposed to competing objectives through environmental pressures. We demonstrate that while the best-performing agent (a scaffolded version of Claude 3.5 Sonnet) maintains nearly perfect goal adherence for more than 100,000 tokens in our most difficult evaluation setting, all evaluated models exhibit some degree of goal drift. We also find that goal drift correlates with models' increasing susceptibility to pattern-matching behaviors as the context length grows.

Related papers

• Evaluating the Goal-Directedness of Large Language Models [17.08087240111954]
  We evaluate goal-directedness on tasks that require information gathering, cognitive effort, and plan execution.<n>Our evaluations of LLMs from Google DeepMind, OpenAI, and Anthropic show that goal-directedness is relatively consistent across tasks.
  arXiv  Detail & Related papers  (2025-04-16T08:07:08Z)
• Evaluating the Paperclip Maximizer: Are RL-Based Language Models More Likely to Pursue Instrumental Goals? [33.11148546999906]
  Key concern is textitinstrumental convergence, where an AI system develops unintended intermediate goals that override the ultimate objective and deviate from human-intended goals.<n>This issue is particularly relevant in reinforcement learning (RL)-trained models, which can generate creative but unintended strategies to maximize rewards.<n>We show that RL-driven models exhibit a stronger tendency for instrumental convergence due to their optimization of goal-directed behavior in ways that may misalign with human intentions.
  arXiv  Detail & Related papers  (2025-02-16T16:29:20Z)
• Deception in LLMs: Self-Preservation and Autonomous Goals in Large Language Models [0.0]
  Recent advances in Large Language Models have incorporated planning and reasoning capabilities.<n>This has reduced errors in mathematical and logical tasks while improving accuracy.<n>Our study examines DeepSeek R1, a model trained to output reasoning tokens similar to OpenAI's o1.
  arXiv  Detail & Related papers  (2025-01-27T21:26:37Z)
• Zero-Shot Offline Imitation Learning via Optimal Transport [21.548195072895517]
  Zero-shot imitation learning algorithms reproduce unseen behavior from as little as a single demonstration at test time.<n>Existing practical approaches view the expert demonstration as a sequence of goals, enabling imitation with a high-level goal selector, and a low-level goal-conditioned policy.<n>We introduce a novel method that mitigates this issue by directly optimizing the occupancy matching objective that is intrinsic to imitation learning.
  arXiv  Detail & Related papers  (2024-10-11T12:10:51Z)
• WANDR: Intention-guided Human Motion Generation [67.07028110459787]
  We introduce WANDR, a data-driven model that takes an avatar's initial pose and a goal's 3D position and generates natural human motions that place the end effector (wrist) on the goal location. Intention guides the agent to the goal, and interactively adapts the generation to novel situations without needing to define sub-goals or the entire motion path. We evaluate our method extensively and demonstrate its ability to generate natural and long-term motions that reach 3D goals and to unseen goal locations.
  arXiv  Detail & Related papers  (2024-04-23T10:20:17Z)
• From Instructions to Intrinsic Human Values -- A Survey of Alignment Goals for Big Models [48.326660953180145]
  We conduct a survey of different alignment goals in existing work and trace their evolution paths to help identify the most essential goal. Our analysis reveals a goal transformation from fundamental abilities to value orientation, indicating the potential of intrinsic human values as the alignment goal for enhanced LLMs.
  arXiv  Detail & Related papers  (2023-08-23T09:11:13Z)
• Discrete Factorial Representations as an Abstraction for Goal Conditioned Reinforcement Learning [99.38163119531745]
  We show that applying a discretizing bottleneck can improve performance in goal-conditioned RL setups. We experimentally prove the expected return on out-of-distribution goals, while still allowing for specifying goals with expressive structure.
  arXiv  Detail & Related papers  (2022-11-01T03:31:43Z)
• Generative multitask learning mitigates target-causing confounding [61.21582323566118]
  We propose a simple and scalable approach to causal representation learning for multitask learning. The improvement comes from mitigating unobserved confounders that cause the targets, but not the input. Our results on the Attributes of People and Taskonomy datasets reflect the conceptual improvement in robustness to prior probability shift.
  arXiv  Detail & Related papers  (2022-02-08T20:42:14Z)
• Adversarial Intrinsic Motivation for Reinforcement Learning [60.322878138199364]
  We investigate whether the Wasserstein-1 distance between a policy's state visitation distribution and a target distribution can be utilized effectively for reinforcement learning tasks. Our approach, termed Adversarial Intrinsic Motivation (AIM), estimates this Wasserstein-1 distance through its dual objective and uses it to compute a supplemental reward function.
  arXiv  Detail & Related papers  (2021-05-27T17:51:34Z)
• Automatic Curriculum Learning through Value Disagreement [95.19299356298876]
  Continually solving new, unsolved tasks is the key to learning diverse behaviors. In the multi-task domain, where an agent needs to reach multiple goals, the choice of training goals can largely affect sample efficiency. We propose setting up an automatic curriculum for goals that the agent needs to solve. We evaluate our method across 13 multi-goal robotic tasks and 5 navigation tasks, and demonstrate performance gains over current state-of-the-art methods.
  arXiv  Detail & Related papers  (2020-06-17T03:58:25Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.