Research map / Large language models / Agents, code and math
Learning Invariants through Soft Unification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Add strong context-attention supervision to UMN (ablation) | UMN · bAbI 1k | mean task error rate (%): 18.8 -> 5.1 | Helped |
| Add strong context-attention supervision to UMN (ablation) | UMN · Logical reasoning 2k | mean task error rate (%): 37.7 -> 27.4 | Helped |
| Use 3 invariants instead of 1 (ablation) | UMN · bAbI 1k, weak supervision | mean task error rate (%): 18.8 -> 19.0 | Hurt |
| Use 3 invariants instead of 1 (ablation) | UMN · Logical reasoning 2k, strong supervision | mean task error rate (%): 27.4 -> 29.0 | Hurt |
| Train with 50 examples per task instead of 1k (ablation) | UMN, strong supervision, 3 invariants · bAbI | mean task error rate (%): 6.6 -> 28.7 | Hurt |
| Train with 100 examples per task instead of 2k (ablation) | UMN, strong supervision, 3 invariants · Logical reasoning | mean task error rate (%): 29.0 -> 47.1 | Hurt |
| Restrict logic predicates to arity 1 (ablation) | UMN, strong supervision, 1 invariant · Logical reasoning 2k | mean task error rate (%): 28.9 -> 14.3 | Helped |
| Increase training data from 2k to 20k (ablation) | UMN, strong supervision, 1 invariant, arity 2 · Logical reasoning | mean task error rate (%): 12.2 -> 2.4 | Helped |
| Use soft unification rather than the plain model | UMLP, UCNN, and URNN · Synthetic sequence and grid tasks, and sentiment analysis | test accuracy | Helped |
| Use soft unification rather than the plain model | URNN · Sentiment analysis | test accuracy | Mixed |
| Use a learning rate of 0.01 (ablation) | UMLP, UCNN, and URNN with soft unification · Synthetic sequence and grid tasks, and sentiment analysis | training performance | Hurt |
| Use additional invariants (ablation) | UMLP, UCNN, and URNN · Synthetic tasks and sentiment analysis | performance | No clear effect |
Closest papers by meaning
- Evaluating alignment between humans and neural network representations in image-based learning tasks (NeurIPS 2024 · similarity 0.85)
- Right for the Right Reasons: Avoiding Reasoning Shortcuts via Prototypical Neurosymbolic AI (NeurIPS 2025 · similarity 0.85)
- From Noise to Diversity: Random Embedding Injection in LLM Reasoning (NeurIPS 2026 · similarity 0.84)
- Selection, Reflection and Self-Refinement: Revisit Reasoning Tasks via a Causal Lens (ICLR 2026 · similarity 0.84)
- Deciphering Raw Data in Neuro-Symbolic Learning with Provable Guarantees (AAAI 2024 · similarity 0.84)
- Soft Thinking: Unlocking the Reasoning Potential of LLMs in Continuous Concept Space (NeurIPS 2025 · 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.