Research map / Large language models / Agents, code and math
Socrates or Smartypants: Testing Logic Reasoning Capabilities of Large Language Models with Logic Programming-Based Test Oracles
Zihao Xu, Junchen Ding, Yiling Lou, Kun Zhang, Dong Gong, Yuekang Li
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 |
|---|---|---|---|
| generate fallacies with Prolog rules and LLM-filled facts instead of direct LLM generation | SmartyPat vs FallacyGen-Direct · Sentence quality, GPT-4o judge | average quality score: +38.12% on average | Improved |
| constrain generation with Prolog predicates and validation instead of unconstrained LLM-written Prolog (ablation) | SmartyPat vs FallacyGen-Prolog · Fallacy of Composition, score-3 count | number of score-3 sentences: 25 -> 60 | Improved |
| group facts per fallacy instance instead of per predicate in few-shot prompts (ablation) | Prolog knowledge generation prompt · LLM fact generation | fact coherence (qualitative) | Improved |
| add inline comments clarifying each predicate in few-shot facts (ablation) | Prolog knowledge generation prompt · LLM fact generation | LLM understanding (qualitative) | Improved |
| use reasoning models instead of non-reasoning models for fallacy detection | nine LLMs · SmartyPat-Bench detection | F1: DeepSeek V3 0.934, Grok-2 0.921 vs GPT-o3-mini 0.891 | Hurt |
| use reasoning models for fallacy categorization | nine LLMs · SmartyPat-Bench categorization | rank-weighted label score: GPT-o3-mini -8.41 vs GPT-4o -126.46 | Improved |
| evaluate on synthetic SmartyPat-Bench-Augmented instead of real-world SmartyPat-Bench | LLM fallacy categorization · categorization score | rank-weighted label score: Claude 3.7 improves by over 70% | Improved |
| evaluate on synthetic SmartyPat-Bench-Augmented instead of real-world SmartyPat-Bench | LLM fallacy detection · detection F1 | FPR/FNR/F1 | No clear effect |
| move from surface-level to context-dependent fallacy types | LLM fallacy detection · SmartyPat-Bench-Augmented | detection accuracy | Hurt |
Closest papers by meaning
- Are Language Models Efficient Reasoners? A Perspective from Logic Programming (NeurIPS 2025 · similarity 0.90)
- RULEBREAKERS: Challenging LLMs at the Crossroads between Formal Logic and Human-like Reasoning (ICML 2025 · similarity 0.89)
- Reasoning-as-Logic-Units: Scaling Test-Time Reasoning in Large Language Models Through Logic Unit Alignment (ICML 2025 · similarity 0.89)
- ORACLE: Optimizing Reasoning Abilities of Large Language Models via Constraint-Led Synthetic Data Elicitation (AAAI 2026 · similarity 0.89)
- Enhancing Reasoning Capabilities of LLMs via Principled Synthetic Logic Corpus (NeurIPS 2024 · similarity 0.88)
- When LLMs Meet Cunning Texts: A Fallacy Understanding Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.88)
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.
