Research map / Large language models / AI and society
When Models Know More Than They Can Explain: Quantifying Knowledge Transfer in Human-AI Collaboration
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.
Introduces KITE, a two-phase human-AI study (N=118) measuring how well LLM explanations transfer knowledge enabling humans to solve coding and math problems independently.
Problem
Whether improvements in LLM reasoning translate into better knowledge transfer, i.e., models communicating reasoning that humans can understand, apply, and learn from, which is under-evaluated.
Approach
A conceptual framework (shared, AI-exclusive, human-exclusive knowledge regions with projections) plus a two-phase protocol: humans ideate with an LLM without solution content, then solve the same problem alone with no model access; success measured against model solo performance and Bradley-Terry human preference win rates.
Key results
- Collaborative outcomes correlate with solo model performance (r=0.84 code, r=0.69 math), but slope below unity, with outliers such as Claude-3.7-Sonnet (+25.0% code, +60.9% math) and Gemini-2.5-Pro (-10.0% code).
- Human preference correlates with solution success overall (r=0.86) but weakly in math (r=0.14, non-significant).
- Covariates such as user expertise, LLM familiarity, and collaboration-tool experience had no significant effect on solve rates (p>0.19).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Collaborate with GPT-4.1 rather than use it solo | GPT-4.1 · Code | accuracy: -3.8 | Hurt |
| Collaborate with GPT-4o rather than use it solo | GPT-4o · Code | accuracy: +23.3 | Helped |
| Collaborate with o1 rather than use it solo | o1 · Code | accuracy: +0.0 | No clear effect |
| Collaborate with GPT-4.5 rather than use it solo | GPT-4.5 · Code | accuracy: +0.6 | Helped |
| Collaborate with DeepSeek-V3 rather than use it solo | DeepSeek-V3 · Code | accuracy: +18.6 | Helped |
| Collaborate with LLaMA-4-Maverick rather than use it solo | LLaMA-4-Maverick · Code | accuracy: +7.6 | Helped |
| Collaborate with Claude-3.7-Sonnet rather than use it solo | Claude-3.7-Sonnet · Code | accuracy: +25.0 | Helped |
| Collaborate with Gemini-2.5-Pro rather than use it solo | Gemini-2.5-Pro · Code | accuracy: -10.0 | Hurt |
| Collaborate with GPT-4.1 rather than use it solo | GPT-4.1 · Math | accuracy: +28.1 | Helped |
| Collaborate with GPT-4o rather than use it solo | GPT-4o · Math | accuracy: +48.4 | Helped |
| Collaborate with o1 rather than use it solo | o1 · Math | accuracy: +2.5 | Helped |
| Collaborate with GPT-4.5 rather than use it solo | GPT-4.5 · Math | accuracy: +27.5 | Helped |
Surprising findings
- Higher-performing models such as Gemini-2.5-Pro did not yield proportionally better collaborative outcomes, indicating diminishing returns in knowledge transfer.
- In math, Llama-4-Maverick won 25.9% preference in MHT settings despite modest solo performance, while Gemini-2.5-Pro scored only 4.4% in that setting.
Limitations
- Sample of 118 participants skewed toward STEM students limits generalizability.
- Assumes some projection of model reasoning could enable the human to solve each task, which is unverifiable.
- Incentives may inflate measured collaboration effectiveness relative to real-world users.
Experimental setup
Datasets: LiveCodeBench AMC AIME LeetCode problem ratings
Benchmarks: LiveCodeBench AMC/AIME math problems
Metrics: solve rate (human-AI collaboration vs model solo) Bradley-Terry win rate Pearson correlation
Baselines: GPT-4.1 GPT-4o o1 GPT-4.5 DeepSeek-V3 Llama-4-Maverick Claude-3.7-Sonnet Gemini-2.5-Pro
Builds on: Clio Bradley-Terry model Elo rating LiveCodeBench
Closest papers by meaning
- Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.88)
- Language Model Goal Selection Differs from Humans' in a Self-Directed Learning Task (NeurIPS 2026 · similarity 0.88)
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.87)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.87)
- Does GPT Really Get It? A Hierarchical Scale to Quantify Human and AI’s Understanding of Algorithms (AAAI 2025 · similarity 0.87)
- Do as We Do, Not as You Think: the Conformity of Large Language Models (ICLR 2025 · similarity 0.87)
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.