Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
QORA: Zero-Shot Transfer via Interpretable Object-Relational Model Learning
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.
Introduces QORA, a non-deep-learning algorithm that learns object-oriented transition models by assembling relational predicates into quantified first-order rules, and a new benchmark suite for generalization in RL transition modeling.
Problem
Learning object-oriented transition models without domain knowledge that generalize zero-shot, are interpretable, and are sample-efficient, from (state, action, next state) observations of object-based states.
Approach
QORA extracts relational predicates from object observations, encodes first-order formula conditions as bit strings that implicitly cover all connectives and quantifiers, and iteratively combines high-scoring conditions using a confidence-bounded score to build probabilistic rules that are counted in tables.
Key results
- QORA reaches 100% predictive accuracy with roughly 10,000x fewer observations than NPE in the walls domain.
- QORA transfers zero-shot from 8x8 to 64x64 and 128x128 walls worlds with no errors after training on 1,000 observations.
- QORA is approximately 600x faster than NPE per step after convergence in the walls domain.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace neural networks (NPE, MHDPA) with QORA on walls domain | transition modeling · walls domain | sample complexity / error: approx. 10,000× fewer observations than NPE | Helped |
| Transfer trained QORA model zero-shot from 8x8 to 64x64 and 128x128 levels | QORA · walls transfer | error (EMD): transferred models never make any errors | Helped |
| Transfer neural networks from 8x8 to 16x16 worlds without retraining | NPE, MHDPA · walls transfer | error (EMD): substantial increase in error immediately after transfer | Hurt |
| Increase number of steps per level n (less variety in training data) (ablation) | NPE, MHDPA · walls domain | learning speed: MHDPA with n = 100 never drops to 0.5 error | Hurt |
| Lower alpha (higher confidence level, wider intervals) (ablation) | QORA · walls and fish domains | learning speed: slower learning | Hurt |
| Lower alpha (higher confidence level) (ablation) | QORA · walls and fish domains | memory usage (candidates explored): lower alpha yields much more consistent memory usage | Helped |
| Pre-train QORA in simpler lights levels before transfer to complex levels (ablation) | QORA · lights domain | convergence speed to perfect accuracy: almost 3× faster | Helped |
| Train on single door color before transfer, with post-transfer learning (ablation) | QORA · doors 8x8 to 16x16 transfer | error (EMD): not zero error immediately post-transfer; re-converges to zero within 5,000 additional observations | Mixed |
Surprising findings
- Even completely untrained QORA models could generate successful plans in the walls domain by taking observations while planning.
- Neural baselines (NPE, MHDPA) saturate around 0.5 EMD error even after two million observations in 8x8 walls worlds.
Limitations
- Relies on only two predicate types and a fixed set of object attributes; more complex predicates would need to be added.
- Sample complexity and runtime are hard to predict in general and depend on the number of candidates generated.
- Tested only on grid-world benchmark domains; not yet applied to complex games or physical robots.
Experimental setup
Datasets: walls domain doors domain fish domain lights domain complex domains paths domains moves domain players domain
Benchmarks: walls doors fish lights complex(4) paths moves players
Metrics: Earth Mover's Distance observations to zero error runtime per step memory usage (candidates explored)
Baselines: NPE MHDPA
Builds on: Relational reinforcement learning DOORMAX Neural Physics Engine Multi-Head Dot Product Attention
Closest papers by meaning
- Model-Based Transfer Learning for Contextual Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- Zero-shot Model-based Reinforcement Learning using Large Language Models (ICLR 2025 · similarity 0.85)
- Zero-Shot Reinforcement Learning via Function Encoders (ICML 2024 · similarity 0.85)
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.85)
- Efficient Exploration and Discriminative World Model Learning with an Object-Centric Abstraction (ICLR 2025 · similarity 0.84)
- Reinforcing VLAs in Task-Agnostic World Models (NeurIPS 2026 · 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.