Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Neural Reasoning about Agents’ Goals, Preferences, and Actions
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 IRENE, a GNN-plus-transformer model for intuitive psychology that reaches state-of-the-art on three of five Baby Intuitions Benchmark tasks and analyses how training-task choice affects generalisation.
Problem
Common-sense reasoning about agents' goals, preferences, and actions from observed behaviour (intuitive psychology) under a violation-of-expectation paradigm, evaluated on the Baby Intuitions Benchmark, where training and evaluation tasks differ and models must generalise.
Approach
Encodes each grid-world frame as a heterogeneous spatial-relation graph processed by a GraphSAGE-based relational GNN, then uses a transformer over familiarisation trials (with a learnable CTX token) to form a context vector that conditions an MLP predicting the agent's next position.
Key results
- IRENE achieves 74.9% on Multi-Agent versus 50.3% for the previous best (Video-RNN), a 48.9% relative improvement.
- IRENE reaches 94.7% average on Efficient Action, the best among compared models, and 85.8% on Inaccessible Goal.
- IRENE scores 71.5% on Instrumental Action average, below VT's 84.7%, with weak results on Preference (48.5%) and Inconsequential Barrier (52.4%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use IRENE rather than Video-RNN | BIB evaluation · Multi-Agent | VoE accuracy: 50.3 -> 74.9 | Helped |
| use IRENE rather than VT | BIB evaluation · Instrumental Blocking Barrier | VoE accuracy: 64.2 -> 83.5 | Helped |
| use IRENE rather than BC-MLP | BIB evaluation · Efficiency Irrational Agent | VoE accuracy: 73.8 -> 85.7 | Helped |
| replace the transformer context encoder with an LSTM (ablation) | IRENE · Multi-Agent | VoE accuracy: 74.9 -> 49.7 | Hurt |
| replace GraphSAGE layers with GCN layers (ablation) | IRENE · Inaccessible Goal | VoE accuracy: 85.8 -> 58.1 | Hurt |
| use only local directional relations (ablation) | IRENE graph inputs · Multi-Agent | VoE accuracy: 74.9 -> 98.0 | Helped |
| use only local directional relations (ablation) | IRENE graph inputs · Efficiency Path Control | VoE accuracy: 98.1 -> 31.8 | Hurt |
| use only remote directional relations (ablation) | IRENE graph inputs · Multi-Agent | VoE accuracy: 74.9 -> 50.0 | Hurt |
| use mean rather than maximum prediction error for expectedness (ablation) | IRENE · Instrumental Blocking Barrier | VoE accuracy: 83.5±0.4 -> 99.4±0.0 | Helped |
| use mean rather than maximum prediction error for expectedness (ablation) | IRENE · Multi-Agent | VoE accuracy: 74.9±0.0 -> 63.6±0.5 | Hurt |
| train on MPS instead of all training tasks (ablation) | IRENE · Efficient Action Average | VoE accuracy: 94.7±0.1 -> 93.2±0.2 | No clear effect |
| train on MPS instead of all training tasks (ablation) | IRENE · Instrumental Action Average | VoE accuracy: 71.5±0.1 -> 80.4±0.1 | Helped |
Surprising findings
- Training only on a subset of training tasks generally hurts, but training on MPS (excluding the instrumental task) matches full-training average performance (75.1 vs 75.4).
- Replacing the transformer context encoder with an LSTM or GraphSAGE with GCN drops performance to chance on Multi-Agent, showing both components are needed.
Limitations
- Does not outperform existing models on all tasks, notably Preference, No Barrier, Inconsequential Barrier.
- Re-implementation from the paper alone was extremely difficult; AGENT benchmark evaluation not possible since no code or benchmark was provided.
- Results averaged over only three seeds; expectedness defined via max prediction error.
Experimental setup
Datasets: Baby Intuitions Benchmark
Benchmarks: Baby Intuitions Benchmark
Metrics: VoE accuracy mean squared error prediction error
Baselines: BC-MLP BC-RNN Video-RNN VT
Builds on: Baby Intuitions Benchmark Video Transformer (VT) GraphSAGE Relational GNN ToMnet
Closest papers by meaning
- Eliciting Reasoning in Language Models with Cognitive Tools (NeurIPS 2025 · similarity 0.83)
- MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.83)
- Reflection-Bench: Evaluating Epistemic Agency in Large Language Models (ICML 2025 · similarity 0.83)
- Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks (ICML 2026 · similarity 0.83)
- AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling (NeurIPS 2025 · similarity 0.83)
- Emergent Symbolic Mechanisms Support Abstract Reasoning in Large Language Models (ICML 2025 · similarity 0.83)
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.