Research map / Probabilistic and scientific modeling / Dynamical systems
CogReact: A Reinforced Framework to Model Human Cognitive Reaction Modulated by Dynamic Intervention
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.
Proposes CogReact, a hybrid framework that combines an LSTM math agent, SVM mapping to humans, a drift-diffusion model, and a DRL agent to simulate how dynamic time-pressure stimuli modulate human response times in math, decision-making, and learning tasks.
Problem
Simulating how dynamic, fine-grained environmental stimuli (e.g., time pressure feedback) affect human cognitive response time, which prior work mostly modeled under ideal conditions with static or binary stimuli.
Approach
Integrates a closed-form drift-diffusion evidence accumulation process into a deep reinforcement learning loop, where a DRL agent takes per-frame stimuli as input and outputs biases that modulate the accumulator, with the SVM-based baseline response time as the starting estimate.
Key results
- CogReact achieves the lowest mean MAPE (0.2703) on the numeric-task/numeric-feedback setting versus the best baseline Transformer (0.3052) and LSTM (0.3059).
- The hybrid DRL agent converges in about 4.42 minutes versus 38.30 minutes for the pure DRL agent on the same hardware.
- The SVM with math-agent features reaches 0.9613 choice accuracy and 0.3652 MAPE, outperforming raw digit or string inputs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use CogReact instead of the strongest listed baseline, Transformer + 3D ResNet | Math reasoning, Type II · Math reasoning, Type II (encoded string task, video feedback) | MAPE: 0.3306 -> 0.2999 | Helped |
| use CogReact instead of the strongest listed baseline, Transformer | Math reasoning, Type IV · Math reasoning, Type IV (numeric task, numeric feedback) | MAPE: 0.3052 -> 0.2703 | Helped |
| use math-agent features rather than one-hot encoded raw math numbers (ablation) | SVM user choice and response-time prediction · Math reasoning | MAPE: 0.3813 -> 0.3652 | Helped |
| use math-agent features rather than raw numeric math numbers (ablation) | SVM user choice and response-time prediction · Math reasoning | MAPE: 0.3740 -> 0.3652 | Helped |
| increase LSTM output neurons to 256 rather than using the other tested sizes (ablation) | Math logical reasoning agent · Math-agent test set | accuracy: over 95% -> 99.93% | Helped |
| integrate DDM into DRL and segment the stimulus into frames (ablation) | Pure DRL agent · Math reasoning | MAPE | Helped |
| integrate DDM into DRL and segment the stimulus into frames (ablation) | Pure DRL agent · Math reasoning, training efficiency | convergence time: 38.30 minutes -> 4.42 minutes | Helped |
| remove the first-stage risk/LLM agent embeddings (ablation) | Hybrid DRL · Decision making and learning datasets | MAPE | Hurt |
| remove DDM from the DRL loop (ablation) | Hybrid DRL · Decision making and learning datasets | MAPE | Hurt |
| remove both DDM and DRL, retaining the encoded SVM variant (ablation) | Hybrid DRL · Decision making and learning datasets | MAPE | Hurt |
| use risk/LLM agent embeddings with SVM rather than a straightforward SVM baseline (ablation) | Encoded SVM · Decision making and learning datasets | MAPE | Hurt |
| add DDM to standalone deep-learning models without reinforcement learning (ablation) | LSTM, MLP, Transformer, and MLPv2 · Three datasets | MAPE | Mixed |
Surprising findings
- Human accuracy was unaffected by time-pressure stimuli, so the authors model only response time.
- Encoded SVM with risk/LLM embeddings performed worse than a plain SVM, although removing those embeddings from the hybrid DRL hurt performance.
Limitations
- Focuses only on response time, not choice accuracy.
- Relies on a single math task with a small participant pool (36 retained participants).
Experimental setup
Datasets: CogReact math time-pressure dataset (21,157 responses) CPC18 PeerEdu synthetic arithmetic dataset (20,414 samples)
Benchmarks: CogReact math time-pressure dataset CPC18 PeerEdu
Metrics: MAPE Pearson correlation accuracy F1-score Kolmogorov-Smirnov test permutation test
Baselines: hGRU LSTM + AlexNet LSTM + VGG-16 LSTM + ViT-B-16 MLP + 3D ResNet Decision Tree Linear Regression LSTM MLP Random Forest SVM Transformer
Builds on: drift-diffusion model Stable Baselines3 Roseboom et al. time perception network
Closest papers by meaning
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.85)
- CogDPM: Diffusion Probabilistic Models via Cognitive Predictive Coding (ICML 2024 · similarity 0.85)
- Habitizing Diffusion Planning for Efficient and Effective Decision Making (ICML 2025 · similarity 0.85)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.85)
- Act to See, See to Act: Diffusion-Driven Perception-Action Interplay for Adaptive Policies (NeurIPS 2025 · similarity 0.84)
- BrainVista: Modeling Naturalistic Brain Dynamics as Multimodal Next-Token Prediction (NeurIPS 2026 · similarity 0.84)
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.