Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Estimating cognitive biases with attention-aware inverse planning
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.
Formalizes attention-aware inverse planning by extending value-guided construal with attentional bias heuristics, and infers those biases from simulated trajectories of PPO agents in tabular DrivingWorld and in GPUDrive scenes from the Waymo Open Motion Dataset.
Problem
Estimating a person's attentional/cognitive biases from their goal-directed behavior, since standard inverse RL assumes unlimited attention and optimal behavior, which misspecifies attention-limited decision-makers.
Approach
Extends value-guided construal with a parameterized heuristic bias over construal selection, then performs maximum-likelihood inference over bias weights, amortizing construed policies by masking states passed to a pre-trained generalist PPO policy.
Key results
- In tabular DrivingWorld, ice and cone biases are recovered robustly while parked-car bias is harder to recover, across 1000 simulated agents.
- IRL fit to attention-limited trajectories is noisier, sometimes fails to reach the goal, and misses the ice-visiting pattern of the original policy, as shown by negative log-likelihood comparison.
- In GPUDrive, 215 simulated agents with 80 trajectories (about 12 minutes of driving) per agent allowed recovery of the three heuristic weights, with convergence around 12 minutes of data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use attention-aware inverse planning instead of IRL that assumes unlimited attention | DrivingWorld attention-limited behavior · DrivingWorld | negative log-likelihood | Helped |
| Use 80 trajectories (about 12 minutes of driving data) for inference | Attention-aware inverse planning · GPUDrive Waymo scenarios | squared error between true and estimated heuristic weights | Helped |
| Infer heuristic weights from behavior | Attention-aware inverse planning · DrivingWorld | heuristic-weight recovery | Mixed |
| Infer heuristic weights from behavior using masked states and a pre-trained generalist policy | Attention-aware inverse planning · GPUDrive Waymo scenarios | heuristic-weight recovery | Helped |
Surprising findings
- Standard IRL can fail to capture attention-limited behavior even with a reward function that is a sensible fit, because the mis-specified unlimited-attention model cannot represent the behavior.
- More consequential biases (e.g., ignoring parked cars) are less identifiable because they are outweighed by value-guided construal.
Limitations
- Assumes a single fixed construal per episode, without dynamic construal switching.
- Validation uses synthetic trajectories generated by the model rather than real human behavior.
- Assumes a known reward model and a pre-specified set of heuristic features.
Experimental setup
Datasets: Waymo Open Motion Dataset DrivingWorld synthetic tabular scenarios
Benchmarks: DrivingWorld tabular MDP GPUDrive with Waymo Open Motion Dataset scenes
Metrics: R^2 between true and estimated heuristic weights negative log-likelihood mean squared error of lambda estimates
Baselines: inverse reinforcement learning
Builds on: value-guided construal inverse reinforcement learning GPUDrive PPO
Closest papers by meaning
- Inverse Attention Agents for Multi-Agent Systems (ICLR 2025 · similarity 0.87)
- Attention Trajectories as a Diagnostic Axis for Deep Reinforcement Learning (NeurIPS 2026 · similarity 0.86)
- $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.85)
- More Than Meets the Eye? Uncovering the Reasoning-Planning Disconnect in Training Vision-Language Driving Models (NeurIPS 2026 · similarity 0.84)
- Continuously Learning, Adapting, and Improving: A Dual-Process Approach to Autonomous Driving (NeurIPS 2024 · similarity 0.84)
- Massively Scalable Inverse Reinforcement Learning in Google Maps (ICLR 2024 · 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.