Research map / Large language models / Language, safety and interpretability
Turning large language models into cognitive models
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.
Finetunes a linear readout on LLaMA-65B embeddings of text-described psychology experiments (CENTaUR), showing it outperforms domain-specific cognitive models in decision-making and generalizes to a held-out task.
Problem
Large language models often show unhuman-like behavior, so the paper asks whether they can be turned into accurate models of human cognition and decision-making.
Approach
Extract hidden embeddings from a frozen LLaMA-65B on text prompts describing psychological choice tasks, then fit a regularized logistic regression (linear layer) on top to predict human choices, yielding CENTaUR; extend with random effects for individual differences and joint multi-task training for hold-out generalization.
Key results
- On choices13k, CENTaUR reached NLL 48002.3 versus 49448.1 for BEAST and 96248.5 for untuned LLaMA.
- On the horizon task, CENTaUR reached NLL 25968.6 versus 29042.5 for the hybrid model; with random effects NLL improved to 23929.5, and CENTaUR best-fit 52 of 60 participants.
- Trained jointly on choices13k and horizon data, CENTaUR predicted held-out experiential-symbolic choices (NLL 4521.1 vs 6307.9 for LLaMA and 5977.7 random) and reproduced the S-option overvaluation pattern.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune a linear layer on LLaMA embeddings | LLaMA 65B · choices13k | negative log-likelihood: 96248.5 -> 48002.3 | Helped |
| Use CENTaUR instead of BEAST | choices13k · choices13k | negative log-likelihood: 49448.1 -> 48002.3 | Helped |
| Fine-tune a linear layer on LLaMA embeddings | LLaMA 65B · horizon task | negative log-likelihood: 46211.4 -> 25968.6 | Helped |
| Use CENTaUR instead of the hybrid model | horizon task · horizon task | negative log-likelihood: 29042.5 -> 25968.6 | Helped |
| Fine-tune a linear layer on LLaMA embeddings | LLaMA 65B · choices13k | regret: M = 1.85, SE =0.01 -> 1.35 (SE =0.01) | Helped |
| Fine-tune a linear layer on LLaMA embeddings | LLaMA 65B · horizon task | regret: M = 7.21, SE =0.02 -> M = 2.38, SE =0.01 | Helped |
| Add random effects for each participant and embedding dimension (ablation) | CENTaUR · horizon task, participant-level | negative log-likelihood: 25968.6 -> 23929.5 | Helped |
| Fine-tune on choices13k and horizon task | LLaMA 65B · experiential-symbolic task (hold-out) | negative log-likelihood: 6307.9 -> 4521.1 | Helped |
| Fine-tune on choices13k and horizon task | LLaMA 65B · experiential-symbolic task (hold-out) | negative log-likelihood: 5977.7 -> 4521.1 | Helped |
| Fine-tune on psychological decision-task data | LLaMA 65B · horizon task, equal-information and unequal-information conditions | human-like exploratory choice patterns | Helped |
| Fine-tune on choices13k and horizon task | LLaMA 65B · experiential-symbolic task (hold-out) | description-option preference | Helped |
Surprising findings
- Untuned LLaMA-65B performs near chance on the decision tasks, yet a linear readout on its embeddings beats domain-specific cognitive models.
- A model finetuned on two tasks captures human behavior on a third, unseen task without seeing its data.
Limitations
- Only the final linear layer is finetuned; the LLM itself is not updated, and the approach is evaluated on few tasks.
- Interpretability of what the embeddings encode about human cognition is limited by the authors' own account.
- Hold-out evaluation uses a single unseen task, so generalization breadth is preliminary.
Experimental setup
Datasets: choices13k horizon task experiential-symbolic task
Benchmarks: choices13k horizon task experiential-symbolic task
Metrics: negative log-likelihood regret choice curves indifference points
Baselines: LLaMA-65B without finetuning BEAST hybrid model (horizon task) random guessing
Builds on: LLaMA BEAST PyTorch
Closest papers by meaning
- Generating Computational Cognitive models using Large Language Models (NeurIPS 2025 · similarity 0.93)
- Language Models Trained to do Arithmetic Predict Human Risky and Intertemporal Choice (ICLR 2025 · similarity 0.92)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.91)
- Using Artificial Populations to Study Psychological Phenomena in Neural Models (AAAI 2024 · similarity 0.90)
- Simulating Human Memory with Language Models (NeurIPS 2026 · similarity 0.89)
- Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.89)
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.