Research map / Reinforcement learning / Visual and representation RL
Encoding Human Behavior in Information Design through Deep Learning
Guanghui Yu, Wei Tang, Saumik Narayanan, Chien-Ju Ho
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use HAIDNet policy instead of Bayesian-rational optimal policy (BR-Policy) | information policy for human receivers · MTurk phase 2 | average sender utility: 0.524 -> 0.667 | Improved |
| use TH-Model-optimal policy instead of BR-Policy | information policy · MTurk phase 2 | average sender utility: 0.524 -> 0.621 | Improved |
| replace TH-Model descriptor with data-driven neural network descriptor in policy design | HAIDNet · MTurk phase 2 | average sender utility: 0.621 -> 0.667 | Improved |
| replace Bayesian rational descriptor with TH-Model for predicting human choices | human behavior descriptor · MTurk phase 1 | test accuracy: 0.562 -> 0.735 | Improved |
| replace TH-Model with neural network descriptor | human behavior descriptor · MTurk phase 1 | test accuracy: 0.735 -> 0.770 | Improved |
| increase number of states M from 2 to 10 | HAIDNet · single Bayesian receiver, binary actions | gap to optimal (test utility): 0.7408 vs 0.7451 -> 0.8196 vs 0.8686 | Hurt |
| increase number of actions N from 2 to 10 | HAIDNet · single Bayesian receiver, binary states | gap to optimal (test utility): 0.7408 vs 0.7451 -> 0.6623 vs 0.6963 | Hurt |
| increase number of receivers K from 2 to 15 | HAIDNet · K Bayesian receivers | gap to optimal (test utility): 0.7756 vs 0.7873 -> 0.6621 vs 0.6843 | Mixed |
| use HAIDNet instead of LP for 18 receivers | HAIDNet · run-time, K=18 | time: 84280 -> 0.352 | Improved |
| increase beta_H (more rational receiver) | HAIDNet · non-Bayesian receiver | test sender utility: 0.5041 -> 0.6964 | Mixed |
| train a single HAIDNet for up to 10 receivers instead of fixed K | HAIDNet · at most 10 receivers, K=10 | test sender utility: 0.6650 vs 0.6901 | Hurt |
| train a single HAIDNet for up to 5 states/actions | HAIDNet · (M,N)=(2,3) | test sender utility: 0.6517 vs 0.7411 | Hurt |
Closest papers by meaning
- Turning large language models into cognitive models (ICLR 2024 · similarity 0.82)
- Mecha-nudges for Machines (NeurIPS 2026 · similarity 0.82)
- Eliciting Language Model Behaviors with Investigator Agents (ICML 2025 · similarity 0.82)
- Teaching Human Behavior Improves Content Understanding Abilities Of VLMs (ICLR 2025 · similarity 0.82)
- InfoRM: Mitigating Reward Hacking in RLHF via Information-Theoretic Reward Modeling (NeurIPS 2024 · similarity 0.82)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.82)
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.
