Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Neural Amortized Inference for Nested Multi-Agent Reasoning
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 neural amortized inference (learned recognition networks proposing particles) to speed up nested I-POMDP reasoning, evaluated on Construction and Driving grid/traffic domains against ToMnet and exact inference.
Problem
Exact inference for nested multi-agent reasoning (I-POMDP) is recursive and computationally intractable as hypothesis spaces and nesting levels grow, while end-to-end alternatives lack explicit, uncertainty-aware goal and belief inference.
Approach
Train neural recognition networks at each I-POMDP level to propose data-driven particles/hypotheses for states and goals, then use importance sampling with a small particle set, so inference matches exact accuracy with far fewer hypotheses.
Key results
- In Construction, the method with 10 particles matches exact inference accuracy (comparable at 6 particles, 6.7% of 90 hypotheses vs. 90 for exact inference).
- In Driving, 9 particles (12.5% of hypothesis space) reach action prediction accuracy comparable to exact inference (72 particles), and it significantly outperforms ToMnet.
- Uniform-proposal ablation (Ours w/o NN) has much lower accuracy than the neural-proposal version, showing the recognition networks drive the gain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace uniform proposals with neural recognition-network data-driven proposals (amortized inference) (ablation) | nested goal inference in I-POMDP · Construction | goal inference accuracy | Helped |
| amortize inference with neural proposals so only ~6-10 particles are needed instead of enumerating all hypotheses | I-POMDP inference · Construction | goal inference accuracy vs particle count: 6 particles vs 90 for exact inference | Helped |
| neural amortized inference with ~9 particles instead of enumerating all hypotheses | I-POMDP inference · Driving | action prediction accuracy: 9 particles vs 72 for exact inference | Helped |
| end-to-end ToMnet instead of model-based nested inference | goal inference · Construction | goal inference accuracy | Hurt |
| end-to-end ToMnet instead of model-based nested inference | action prediction · Driving | action prediction accuracy over episode progress | Hurt |
Surprising findings
- The level-1 goal recognition network has only about 17% accuracy yet still yields useful proposals better than uniform sampling.
Limitations
- Only amortizes up to level-2 reasoning in experiments
- Recognition distribution over lower-level beliefs is set identical to the prior (particle update), not learned
- Evaluated only in two synthetic domains
Experimental setup
Datasets: Construction (synthetic 2D grid-world, 20x20) Driving (CARLO-based intersection simulation)
Benchmarks: Construction Driving
Metrics: goal inference accuracy action prediction accuracy KL-divergence to exact inference
Baselines: ToMnet Ours w/o NN (uniform proposal) Exact inference
Builds on: I-POMDP ToMnet CARLO Rational Speech Act
Closest papers by meaning
- MAGDi: Structured Distillation of Multi-Agent Interaction Graphs Improves Reasoning in Smaller Language Models (ICML 2024 · similarity 0.87)
- Deep Reasoning in General Purpose Agents via Structured Meta-Cognition (NeurIPS 2026 · similarity 0.86)
- Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation (ICLR 2026 · similarity 0.86)
- Adaptive Social Learning via Mode Policy Optimization for Language Agents (ICLR 2026 · similarity 0.86)
- Benefits and Limitations of Communication in Multi-Agent Reasoning (ICLR 2026 · similarity 0.86)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.86)
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.