Research map / Large language models / AI and society

Cognitive Bias and Reassignment: Who Can Contribute High Quality LLM Data

AAAI 2025 · Large language models › AI and society · approach: Critiques of LLM claims and judging · 0 citations (OpenAlex)

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.

Analyzes user profiles and cognitive bias (Dunning-Kruger-like overestimation) on an LLM data co-creation platform, and proposes GDBM, a multi-task LoRA fine-tuning of small LLMs to reassign users to tasks.

Problem

Crowdsourced LLM data collection suffers from mismatches between contributors' self-assessed and actual abilities, causing wasted effort, low-quality submissions, and reduced engagement; existing approaches focus on data rather than contributors.

Approach

Builds an observable user cohort from platform data to expose over/underestimation zones, then trains small language models with three instruction-formatted matching tasks (user-task matching, task-to-user and user-to-task assignment) to reallocate users to tasks they are competent in.

Key results

• GDBM (Qwen2-7B) reaches 0.611 overall accuracy versus 0.450 for the base Qwen2-7B, a 35.78% improvement (as reported).
• GDBM (Llama3.1-8B) reaches 0.583 overall accuracy versus 0.416 for base Llama3.1-8B; EART gains reported as 16.62% (Llama) and 20.68% (Qwen).
• Doctoral users made up 81.75% of participants in 'Weak Correlation Reasoning' but passed only 42.92% versus 50.77% overall average pass rate.

Measured effects reported in the paper

Surprising findings

• Users with high school education often produced higher-quality data than those with bachelor's or advanced degrees.
• GPT-4-turbo judged users competent in 84.62% of point-wise cases, showing permissiveness and not outperforming fine-tuned SLMs.

Limitations

• Evaluated only offline on platform data; reallocation not tested in a live environment.
• Small dataset (about 3,700 instruction entries) and limited user sample (371 users after cleaning).
• Ground truth built from observed pass rates, which may reflect platform biases.

Experimental setup

Datasets: Platform single-submission data (46,224 valid entries, 371 users, March-April 2024)

Benchmarks: Platform user-task matching datasets (D_match, D_t-pair, D_u-pair, D_u-multi)

Metrics: accuracy Average Reward per Task (ART) Expected Average Reward per Task (EART)

Baselines: BERT-base GTE-base GPT-4o-mini GPT-4-turbo Llama3.1-8B-Instruct Qwen2-7B-Instruct

Builds on: Dunning-Kruger effect LoRA Pareto improvement

Closest papers by meaning

• Do Large Language Models Perform the Way People Expect? Measuring the Human Generalization Function (ICML 2024 · similarity 0.87)
• Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.87)
• Fixing It in Post: A Comparative Study of LLM Post-Training Data Quality and Model Performance (NeurIPS 2025 · similarity 0.87)
• Correlated Errors in Large Language Models (ICML 2025 · similarity 0.86)
• DATE-LM: Benchmarking Data Attribution Evaluation for Large Language Models (NeurIPS 2025 · similarity 0.86)
• Position: Rethinking LLM Bias Probing Using Lessons from the Social Sciences (ICML 2025 · similarity 0.86)