Research map / Large language models / Agents, code and math
Abductive Reasoning with Probabilistic Commonsense
Joseph Cotnareanu, Chiara Roverato, Han Zhou, Didier Chételat, Yingxue Zhang, Mark Coates
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 |
|---|---|---|---|
| use score-minimizing sampling (model-count score) instead of text-based informativeness beam search | If-Beam · FOLIO (Llama 3 8B) | accuracy: 74% -> 82% | Improved |
| same swap on QUAIL | If-Beam · QUAIL (Llama 3 8B) | accuracy: 74% -> 80% | Improved |
| PACS vs text-based heuristic search on CosmosQA | If-Beam · CosmosQA (Llama 3 8B) | accuracy: 82% -> 82% | No clear effect |
| PACS vs commonsense-clause sampling ARGOS | ARGOS · FOLIO (Llama 3 8B) | accuracy: 78% -> 82% | Improved |
| PACS vs self-consistency with 20 samples | SC-20 · FOLIO (Llama 3.3 70B) | accuracy: 77% -> 88% | Improved |
| minimize model-count score to choose steps (ablation) | PACS · FOLIO, QUAIL, CosmosQA (Llama 3 8B) | average path length: 6.7 -> 4.6 (FOLIO) | Improved |
| aggregate over diverse sampled commonsense paths | PACS · FOLIO (Llama 3 8B) | % problems never sampled correct: 16% -> 3% (vs If-Beam) | Improved |
| use PACS with small non-thinking model instead of SC with thinking model | PACS Llama 3.3 70B · FOLIO | accuracy: 82% -> 88% (vs SC GPT-OSS-120B) | Improved |
| PACS with Llama 3 8B vs thinking 120B model | PACS Llama 3 8B · FOLIO | accuracy: 82% vs 82% | No clear effect |
| LLM-generated steps without guidance | COT · FOLIO | score progression | Hurt |
Closest papers by meaning
- RULEBREAKERS: Challenging LLMs at the Crossroads between Formal Logic and Human-like Reasoning (ICML 2025 · similarity 0.88)
- Does Reasoning Emerge? Examining the Probabilities of Causation in Large Language Models (NeurIPS 2024 · similarity 0.87)
- Boule or Baguette? A Study on Task Topology, Length Generalization, and the Benefit of Reasoning Trace (NeurIPS 2026 · similarity 0.87)
- Learning to Reason via Mixture-of-Thought for Logical Reasoning (ICLR 2026 · similarity 0.87)
- The Generalization and Robustness of Transformer-Based Language Models on Commonsense Reasoning (AAAI 2024 · similarity 0.87)
- From Passive to Active Reasoning: Can Large Language Models Ask the Right Questions under Incomplete Information? (ICML 2025 · similarity 0.87)
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.
