Research map / Probabilistic and scientific modeling / Molecules and proteins
Evolving Minds: Logic-Informed Inference from Temporal Action Patterns
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 a logic-informed temporal point process trained with amortized variational EM, using a discrete-time renewal process posterior over latent mental events, to infer intentions and predict human actions.
Problem
Inferring latent mental states (intentions, desires) from irregularly timed human action sequences and predicting future actions, especially with limited data.
Approach
Integrates first-order temporal logic rules as priors into a TPP intensity function over actions and mental events, uses a discrete-time renewal process with neural hazards (cross-attention over actions) as amortized posterior, alternates E-step (encoder) and M-step (logic weights and rules via column generation), and adds a backtracking mechanism for action prediction.
Key results
- On Syn Data-1, the method reaches 41.72% ER and 2.32 MAE, versus best baseline 43.47% ER and 2.42 MAE (PromptTPP).
- On Hand-Me-That, the method reaches 75.28% ER and 1.26 MAE, versus best baseline 75.88% ER (HYPRO) and 1.68 MAE (PromptTPP).
- On Car-Following, the method reaches 32.72% ER and 1.80 MAE, with HYPRO at 32.86% ER and 2.03 MAE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Ours instead of the strongest baseline | Ours · Syn Data-1 | ER%: 43.47 ± 0.90 -> 41.72 ± 1.45 | Helped |
| use Ours instead of the strongest baseline | Ours · Syn Data-2 | MAE: 2.60 ± 0.05 -> 2.52 ± 0.15 | Helped |
| use Ours instead of the strongest baseline | Ours · Hand-Me-That | MAE: 1.68 ± 0.08 -> 1.26 ± 0.07 | Helped |
| use Ours instead of the strongest baseline | Ours · Car-Follow | ER%: 32.75 ± 1.90 -> 32.72 ± 2.90 | Helped |
| use Ours instead of the strongest baseline | Ours · MultiTHUMOS | ER%: 58.25 ± 2.42 -> 57.20 ± 2.32 | Helped |
| use Ours instead of the strongest baseline | Ours · EPIC-Kitchen | MAE: 33.21 ± 1.78 -> 32.19 ± 2.24 | Helped |
| add prior knowledge (ablation) | Ours without prior knowledge · Syn Data-1 | ER%: 41.72 -> 45.64 | Helped |
| add prior knowledge (ablation) | Ours without prior knowledge · Syn Data-2 | ER%: 46.85 -> 48.37 | Helped |
| add backtracking (ablation) | Ours without backtracking · Syn Data-1 | ER%: 45.64 -> 41.72 | Helped |
| add backtracking (ablation) | Ours without backtracking · Syn Data-2 | ER%: 48.37 -> 46.85 | Helped |
Surprising findings
- On Car-Following, the proposed method's ER (32.72%) is only marginally better than HYPRO (32.86%) and STLR (32.75%), the gap being within standard deviations.
Limitations
- Relies on prior rule templates for real-world datasets that may be partially incorrect
- Rule-based decoder trades expressiveness for interpretability; black-box encoder remains
Experimental setup
Datasets: Syn Data-1 Syn Data-2 Hand-Me-That Car-Following MultiTHUMOS EPIC-Kitchen-100
Benchmarks: Syn Data-1 Syn Data-2 Hand-Me-That Car-Following MultiTHUMOS EPIC-Kitchen-100
Metrics: ER% MAE
Baselines: RMTPP THP PromptTPP HYPRO TELLER CLNN STLR AVAE GNTPP VEPP STVAE
Builds on: neural Hawkes process TELLER amortized variational inference variational autoencoder
Closest papers by meaning
- Overcoming Multi-step Complexity in Multimodal Theory-of-Mind Reasoning: A Scalable Bayesian Planner (ICML 2025 · similarity 0.88)
- Mental-Models for Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
- AutoToM: Scaling Model-based Mental Inference via Automated Agent Modeling (NeurIPS 2025 · similarity 0.87)
- MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (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)
- Reflection-Bench: Evaluating Epistemic Agency in Large Language Models (ICML 2025 · 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.