Research map / Large language models / Agents, code and math
Do Language Models Exhibit the Same Cognitive Biases in Problem Solving as Human Learners?
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 show three child-like cognitive biases (consistency, transfer vs comparison, carry effect) on synthetically generated arithmetic word problems, finding the first two but not the carry effect.
Problem
Whether LLMs can serve as faithful cognitive models of children by exhibiting the same biases children show when solving arithmetic word problems across text comprehension, solution planning, and solution execution.
Approach
Decomposes word problem solving into three steps and builds a neuro-symbolic generator (MathWorld-based logical forms, templated text, GPT-3.5 error correction) that produces minimally-different problem pairs to measure CATE of known child biases in LLMs.
Key results
- Almost all models show significant consistency bias (20 of 23 CATEs significant), with inconsistent problems lowering accuracy; e.g., LLaMA2 7B base CoT consistent-inconsistent gap of 10.4 points.
- Models are more accurate on transfer than comparison problems; GPT-3.5 Turbo direct shows a 27.6-point CATE (61.0% vs 33.4%).
- No carry effect: only one of 23 carry tests significant, with models performing similarly with and without carry.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use consistent rather than inconsistent comparison wording | LLMs solving arithmetic word problems · Consistency bias | Accuracy: 20 out of 23 statistically significant CATEs | Helped |
| Use transfer rather than comparison problem structures with identical arithmetic expressions | LLMs solving arithmetic word problems · Transfer vs comparison bias | Accuracy: With the exception of CoT-prompted pretrained-only Mistral and Mixtral, we observe statistically significant positive CA | Helped |
| Vary the arithmetic expression to include a carry | LLMs solving arithmetic word problems · Carry effect | Accuracy: we only observe one significant result out of the 23 tests | No clear effect |
| Use chain-of-thought instead of direct prompting | LLMs solving arithmetic word problems · Consistency bias | CATE: LLaMA2 7B: 4.8 -> 10.4; GPT-3.5 Turbo: 9.4 -> 1.4 | Mixed |
| Use chain-of-thought instead of direct prompting | LLMs solving arithmetic word problems · Transfer vs comparison bias | CATE: LLaMA2 7B: 8.8 -> 5.2; GPT-3.5 Turbo: 27.6 -> 4.0 | Mixed |
| Use chain-of-thought instead of direct prompting | LLMs solving arithmetic word problems · Carry effect | CATE: LLaMA2 7B: 4.8 -> -5.6; GPT-3.5 Turbo: 0.2 -> -0.4 | No clear effect |
Surprising findings
- Chain-of-thought prompting amplifies the consistency bias in most models even though it improves overall accuracy.
- No carry effect appears in any model, despite being a well-documented child bias.
Limitations
- Only English problems are studied, and carry effect may differ across languages.
- Cannot compare absolute performance with children; only presence or absence of effects.
- Biases from humans are operationalized as relative effects; no grade-level control.
Experimental setup
Datasets: MAWPS ASDiv-A SVAMP MathWorld annotations
Benchmarks: consistency bias test transfer vs comparison test carry effect test
Metrics: accuracy conditional average treatment effect p-value
Baselines: LLaMA2 7B LLaMA2 13B LLaMA2 70B Mistral 7B Mixtral 8x7B GPT-3.5 Turbo GPT-4 Turbo
Builds on: MathWorld chain-of-thought prompting zero-shot CoT Benjamini-Hochberg false discovery rate control
Closest papers by meaning
- Physics of Language Models: Part 2.1, Grade-School Math and the Hidden Reasoning Process (ICLR 2025 · similarity 0.90)
- CogMath: Assessing LLMs' Authentic Mathematical Ability from a Human Cognitive Perspective (ICML 2025 · similarity 0.89)
- Language Models Trained to do Arithmetic Predict Human Risky and Intertemporal Choice (ICLR 2025 · similarity 0.89)
- Advancing Mathematical Reasoning in Language Models: The Impact of Problem-Solving Data, Data Synthesis Methods, and Training Stages (ICLR 2025 · similarity 0.89)
- Evaluating Large Vision-and-Language Models on Children's Mathematical Olympiads (NeurIPS 2024 · similarity 0.89)
- Procedural Knowledge in Pretraining Drives Reasoning in Large Language Models (ICLR 2025 · 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.