Research map / Large language models / Agents, code and math
Interpretable Concept-Based Memory 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 Concept-based Memory Reasoner (CMR), a concept bottleneck model whose task predictor selects a logic rule from a learnable rulebook via neural attention and evaluates it symbolically, enabling global interpretability and formal verification.
Problem
Existing concept-based models have task predictors that are either weak, non-differentiable, or black-box, so their global decision process cannot be inspected or formally verified before deployment.
Approach
Model each task prediction as a neural selection over a jointly learned memory (rulebook) of conjunctive logic rules over concepts, followed by symbolic evaluation of the selected rule; the rulebook is a global, verifiable disjunctive theory, and a regularizer pushes rules toward prototypes of concept activations.
Key results
- CMR reaches 97.25% task accuracy on MNIST+ and 85.14% on CEBaB, comparable to the black-box DNN (99.19% on C-MNIST, 88.67% on CEBaB) and above CBM baselines on most datasets.
- On MNIST+* (incomplete concepts), CMR gets 94.65% task accuracy versus 92.94% for the best CBM competitor (CEM).
- CMR recovers all ground-truth rules in all runs on MNIST+ and C-MNIST, and rule interventions during training correct the learned rulebook.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CMR instead of the strongest CBM baseline | MNIST+ · MNIST+ | task accuracy: 97.41 {\scriptstyle \pm 0.55} -> 97.25 {\scriptstyle \pm 0.24} | No clear effect |
| Use CMR instead of the strongest CBM baseline | MNIST+* · MNIST+* | task accuracy: 92.94 {\scriptstyle \pm 1.15} -> 94.65 {\scriptstyle \pm 1.99} | Helped |
| Use CMR instead of the strongest CBM baseline | C-MNIST · C-MNIST | task accuracy: 99.44 {\scriptstyle \pm 0.03} -> 99.12 {\scriptstyle \pm 0.04} | Hurt |
| Use CMR instead of the strongest CBM baseline | CelebA · CelebA | task accuracy: 65.44 {\scriptstyle \pm 0.25} -> 63.17 {\scriptstyle \pm 1.13} | Hurt |
| Use CMR instead of the deep neural network | CUB · CUB | AUC: 64.07 {\scriptstyle \pm 0.33} -> 60.07 {\scriptstyle \pm 1.70} | Hurt |
| Use CMR instead of the deep neural network | CEBaB · CEBaB | task accuracy: 88.67 {\scriptstyle \pm 0.19} -> 85.14 {\scriptstyle \pm 0.43} | Hurt |
| Replace predicted concepts with ground-truth concepts | CMR · MNIST+ | task accuracy: 97.25 {\scriptstyle \pm 0.24} -> 100.0 {\scriptstyle \pm 0.00} | Helped |
| Allow CMR to learn additional rules alongside injected decision-tree rules (ablation) | CMR on CEBaB · CEBaB (first task, first 6 concepts) | accuracy | Helped |
Surprising findings
- On CUB, CMR's task AUC (60.07) is below the CBM+linear/CBM+MLP results (62.97 and 55.83 are mixed; the CBM+linear entry is 55.83 and CBM+MLP 51.83 in the table), showing no uniform dominance over concept-based baselines.
- Some baseline rows in the task accuracy table are degenerate (e.g., 0.00 on MNIST+ and CEBaB 45.15 ± 29.93), suggesting unstable training for some competitors.
Limitations
- Focuses on positive-only explanations; negative-reasoning explanations not explored.
- Verification demonstrated only on simple properties in MNIST+ and CelebA, not safety-critical domains.
- Performance depends on hyperparameters such as rulebook size, beta, and selector re-initialization frequency, chosen by validation search.
Experimental setup
Datasets: MNIST+ MNIST+* C-MNIST CelebA CUB CEBaB
Benchmarks: MNIST+ MNIST+* C-MNIST CelebA CUB CEBaB
Metrics: task accuracy subset accuracy AUC-ROC concept accuracy task accuracy after concept intervention
Baselines: CBM+linear CBM+MLP CBM+DT CBM+XGBoost CEM DCR Deep neural network without concepts
Builds on: Concept Bottleneck Models Concept Embedding Models Deep Concept Reasoner ResNet BERT
Closest papers by meaning
- Causally Reliable Concept Bottleneck Models (NeurIPS 2025 · similarity 0.89)
- Learning to Intervene on Concept Bottlenecks (ICML 2024 · similarity 0.88)
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.87)
- Predictive Concept Decoders: Training Scalable End-to-End Interpretability Assistants (NeurIPS 2026 · similarity 0.86)
- Language Guided Concept Bottleneck Models for Interpretable Continual Learning (CVPR 2025 · similarity 0.86)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · 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.