Research map / Large language models / Agents, code and math
A Neuro-Symbolic Benchmark Suite for Concept Quality and Reasoning Shortcuts
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 rsbench, a benchmark suite of learning-and-reasoning tasks prone to reasoning shortcuts, with data generators, concept-quality metrics, and a model-counting algorithm (countrss) to formally verify RS presence, and evaluates DPL, LTN, CBM, NN and CLIP on it.
Problem
Reasoning shortcuts in neuro-symbolic learning let models reach high label accuracy while learning concepts with unintended semantics, but suitable benchmarks with known RSs and concept supervision are scarce.
Approach
A configurable benchmark suite of RS-prone tasks (arithmetic, logic, high-stakes) with OOD and continual-learning data generators, concept-quality metrics (concept F1, concept collapse), and a model-counting algorithm (countrss) that formally counts optimal reasoning shortcuts a priori.
Key results
- On MNAdd-EvenOdd, DPL reaches FY 0.94 but FC only 0.06, indicating concepts with poor semantics despite high label accuracy.
- On SDD-OIA, all models show a visible OOD mean-F1 drop; NN is most affected (~46% difference) while CLIP drops only ~9%.
- Under the countrss analysis, the exhaustive XOR task admits 24 RSs versus 6 for AND, growing to 192 for XOR with a single training example.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evaluate models on the changed-knowledge OOD scenario | SDD-OIA models · SDD-OIA OOD | mean label F1: DPL drops by 18%; LTN by 10%; CBM by ~15%; NN average 46% difference; CLIP only 9% drop | Hurt |
| Train with a single example instead of an exhaustive training set | MNLogic XOR task · MNLogic XOR | number of RSs: 24 -> 192 | Hurt |
| Train with a single example instead of an exhaustive training set | MNLogic AND task · MNLogic AND | number of RSs: 6 -> 48 and 336 | Mixed |
| Provide concept supervision to CBM | CBM · MNAdd-EvenOdd | concept macro F1 | Helped |
| Use TCAV to extract concepts from NN and CLIP embeddings | NN and CLIP · MNAdd-EvenOdd, Kand-Logic, SDD-OIA | concept collapse | Mixed |
Surprising findings
- High label F1 coexists with very low concept F1 across all models, e.g. CBM reaches only 0.60 label F1 on SDD-OIA while NN and CLIP reach high label F1 with poor concepts.
- Lower collapse for NN and CLIP partly stems from TCAV noise rather than genuinely better concepts.
Limitations
- TCAV-based concept extraction for black-box NN and CLIP may be unreliable and bias collapse estimates.
- countrss relies on assumptions A1 (invertibility) and A2 (determinism) and works only for tasks satisfying them; BDD-OIA is excluded.
- Experiments are limited to five architectures on three tasks, with small training sets.
Experimental setup
Datasets: MNAdd-EvenOdd MNAdd-Half Kand-Logic CLE4EVR BDD-OIA MNMath MNLogic SDD-OIA MNIST CLEVR
Benchmarks: MNAdd-EvenOdd Kand-Logic SDD-OIA MNMath MNLogic
Metrics: macro F1 label (FY) macro F1 concept (FC) concept collapse mean F1 for SDD-OIA OOD mean F1 label accuracy concept accuracy
Baselines: DPL LTN CBM NN CLIP
Builds on: DeepProbLog Logic Tensor Networks concept bottleneck models TCAV MNIST addition
Closest papers by meaning
- Shortcuts and Identifiability in Concept-based Models from a Neuro-Symbolic Lens (NeurIPS 2025 · similarity 0.90)
- Right for the Right Reasons: Avoiding Reasoning Shortcuts via Prototypical Neurosymbolic AI (NeurIPS 2025 · similarity 0.88)
- Analysis for Abductive Learning and Neural-Symbolic Reasoning Shortcuts (ICML 2024 · similarity 0.88)
- Deciphering Raw Data in Neuro-Symbolic Learning with Provable Guarantees (AAAI 2024 · similarity 0.86)
- Concise and Logically Consistent Conformal Sets for Neuro-Symbolic Concept-Based Models (NeurIPS 2026 · similarity 0.86)
- RBench: Graduate-level Multi-disciplinary Benchmarks for LLM & MLLM Complex Reasoning Evaluation (ICML 2025 · similarity 0.85)
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.