Research map / Large language models / AI and society
Language Model Goal Selection Differs from Humans' in a Self-Directed Learning Task
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Tests whether LLMs (GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5, Qwen3 32B, Centaur) select goals like humans in a text-adapted self-directed alchemy learning task, finding substantial divergence from human goal selection and performance.
Problem
Whether LLMs can validly replace humans in self-directed goal selection, which is assumed in agentic, social-science, and interestingness-driven systems but has not been empirically tested.
Approach
Adapts a human cognitive-science alchemy task with self-chosen goals into a text-based multi-turn format and compares LLM goal selection, learning, and testing against 175 human participants, with CoT, persona steering, semantic manipulations, and a computer-use replication as controls.
Key results
- Gemini 2.5 Pro showed near-perfect in-distribution test accuracy (0.99) versus humans' 0.64, while GPT-5 and Qwen3 32B learned at super-human levels but dropped at testing (in-distribution 0.22 and 0.19).
- Most models repeated the same goal almost always (0.93-0.94 for GPT-5, Claude, Qwen3) versus humans' 0.54, and goal entropy was far lower than humans' 1.67.
- All models were significantly farther from their closest human than humans were from each other in feature space (all Mann-Whitney U tests p < 0.002); CoT and persona steering had limited effects.
Surprising findings
- Claude Sonnet 4.5 showed near-zero learning accuracy (0.07) and never cycled goals, despite general capability.
- Persona steering and chain-of-thought gave only limited improvements in human-likeness, and the divergence persisted even when GPT-5.4 interacted through the actual visual web interface.
Limitations
- LLMs had full interaction history in context, unlike memory-limited humans, which may explain the absence of systematic hypothesis testing.
- Only five models were tested, and newer or other providers' models may behave differently.
- Controlled setting; generalization to real-world applications remains open.
Experimental setup
Datasets: human alchemy task data (175 participants, re-analyzed) fishing net task human data
Benchmarks: alchemy goal-selection task skill acquisition variant non-words variant fishing net goal commitment task
Metrics: task choice accuracy probability of choosing two-action goals goal repetition probability goal selection entropy preferred goal position goal cycles hypothesis testing count indifference point nearest-neighbor distance
Baselines: Gemini 2.5 Pro GPT-5 Claude Sonnet 4.5 Qwen3 32B Centaur human participants
Builds on: Centaur CogBench
Closest papers by meaning
- Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.88)
- When Models Know More Than They Can Explain: Quantifying Knowledge Transfer in Human-AI Collaboration (NeurIPS 2025 · similarity 0.88)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.88)
- Eliciting Human Preferences with Language Models (ICLR 2025 · similarity 0.88)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.88)
- Latent Learning Progress Drives Autonomous Goal Selection in Human Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.