Research map / Large language models / Agents, code and math
Not All Neuro-Symbolic Concepts Are Created Equal: Analysis and Mitigation of Reasoning Shortcuts
Emanuele Marconato, Stefano Teso, Antonio Vergari, Andrea Passerini
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Force disentanglement (DIS) (ablation) | DPL · XOR | % models with RS: 100\% -> 0\% | Improved |
| Force disentanglement (DIS) (ablation) | DPL · MNIST-Addition | % models with RS: 96.7\% -> 0\% | Improved |
| Force disentanglement (DIS) (ablation) | SL · MNIST-Addition | % models with RS: 82.9\% -> 0\% | Improved |
| Add reconstruction (R) alone (ablation) | LTN · MNIST-EvenOdd | F1 labels (Y): 98.1 \pm 0.2 -> 76.3 \pm 1.1 | Hurt |
| Add reconstruction (R) alone (ablation) | DPL · MNIST-EvenOdd | F1 labels (Y): 85.1 \pm 4.6 -> 79.8 \pm 1.0 | Hurt |
| Add entropy loss (H) alone (ablation) | DPL · MNIST-EvenOdd | F1 labels (Y): 85.1 \pm 4.6 -> 98.1 \pm 0.1 | Improved |
| Add entropy loss (H) alone (ablation) | DPL · MNIST-EvenOdd | F1 concepts (C): 0.1 \pm 0.1 -> 0.1 \pm 0.1 | No clear effect |
| Add concept supervision (C) alone (ablation) | SL · MNIST-EvenOdd | F1 concepts (C): 0.2 \pm 0.1 -> 21.5 \pm 6.2 | Improved |
| Combine H and C (ablation) | LTN · MNIST-EvenOdd | F1 concepts (C): 0.3 \pm 0.1 -> 98.6 \pm 0.1 | Improved |
| Combine R, H and C (ablation) | DPL · MNIST-EvenOdd | F1 concepts (C): 0.1 \pm 0.1 -> 96.2 \pm 0.2 | Improved |
| Multi-task learning (MultiOp) vs single task (ablation) | DPL · MNIST-AddMul | F1 concepts (C): 0.0 \pm 0.0 -> 99.8 \pm 0.1 | Improved |
| Add entropy loss (H) to DPL (ablation) | DPL · BDD-OIA | F1-mean concepts (C): 39.4 \pm 6.2 -> 48.1 \pm 0.3 | Improved |
Closest papers by meaning
- Analysis for Abductive Learning and Neural-Symbolic Reasoning Shortcuts (ICML 2024 · similarity 0.92)
- Right for the Right Reasons: Avoiding Reasoning Shortcuts via Prototypical Neurosymbolic AI (NeurIPS 2025 · similarity 0.90)
- Shortcuts and Identifiability in Concept-based Models from a Neuro-Symbolic Lens (NeurIPS 2025 · similarity 0.90)
- A learnability analysis on neuro-symbolic learning (NeurIPS 2025 · similarity 0.90)
- A Neuro-Symbolic Benchmark Suite for Concept Quality and Reasoning Shortcuts (NeurIPS 2024 · similarity 0.90)
- Position: The Road to Generalizable Neuro-Symbolic Learning Should be Paved with Foundation Models (NeurIPS 2026 · similarity 0.90)
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.
