Research map / Large language models / Retrieval and knowledge
Reverse Multi-Choice Dialogue Commonsense Inference with Graph-of-Thought
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.
Proposes ReX-GoT, a three-step reverse-exclusion graph-of-thought prompting framework (option exclusion, error analysis, combine information) for multi-choice dialogue commonsense QA on CICERO datasets.
Problem
Dialogue Commonsense Multi-choice QA (DC-MCQ) is hard when multiple options may be correct, due to option saturation and clue labyrinth challenges; forward reasoning and vanilla CoT underperform on multi-answer questions.
Approach
Mimics human option elimination by first prompting the model to identify unreasonable options and reasons, then analyzing remaining options in a graph-of-thought structure, and selecting the final answer path by voting, combining backward exclusion with forward reasoning.
Key results
- Flan-T5-11B + ReX-GoT reaches 93.87 F1 / 78.46 EM on CICERO_v2 in the fine-tuned setting, above the best TEAM baseline (91.53 F1).
- In zero-shot setting on CICERO_v2, Flan-T5-11B ReX-GoT gets 69.35 F1 vs 51.68 for TEAM (11B), a 17.67-point gain.
- GPT3.5-based ReX-GoT reaches 91.12 F1 on CICERO_v2 zero-shot, reported as a 39.44% increase over the baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Combine Information step (ablation) | ReX-GoT · CICERO | F1: 88.39(-1.13) -> 89.52 | Helped |
| add the Combine Information step (ablation) | ReX-GoT · CICERO-Multi | F1: 77.71(-0.8) -> 78.51 | Helped |
| add the reverse exclusion step (ablation) | ReX-GoT · CICERO | F1: 88.02(-1.50) -> 89.52 | Helped |
| add the reverse exclusion step (ablation) | ReX-GoT · CICERO-Multi | F1: 77.54(-0.97) -> 78.51 | Helped |
| add Graph-of-Thought (ablation) | ReX-GoT · CICERO | F1: 87.25(-2.27) -> 89.52 | Helped |
| add Graph-of-Thought (ablation) | ReX-GoT · CICERO-Multi | F1: 76.33(-2.18) -> 78.51 | Helped |
| use ReX-GoT rather than CoT prompting | Flan-T5 (11B) · CICERO-Multi | EM: 47.26 -> 53.08 | Helped |
| use ReX-GoT rather than CoT prompting | Flan-T5 (11B) · CICERO_v2 zero-shot | F1: 60.22 -> 69.35 | Helped |
| scale from Flan-T5-3B to Flan-T5-11B with ReX-GoT | ReX-GoT · CICERO_v2 | F1: +4.97% | Helped |
Surprising findings
- Forward reasoning beats backward exclusion on CICERO (mostly single-choice) while backward exclusion beats forward reasoning on the multi-choice-only CICERO_v2.
Limitations
- Relies on the uncertain selection of the optimal GoT path via voting, with limited detail on path selection.
- Performance gains on multi-choice questions with two correct answers remain weakest.
Experimental setup
Datasets: CICERO CICERO_v2 CICERO-Multi
Benchmarks: CICERO CICERO_v2 CICERO-Multi
Metrics: F1 Exact Match
Baselines: CCID MCCI TEAM Flan-T5+Prompt Flan-T5+CoT Flan-T5+GoT
Builds on: chain-of-thought prompting TEAM Flan-T5
Closest papers by meaning
- Reasoning Models Are Test Exploiters: Rethinking Multiple Choice (ICML 2026 · similarity 0.87)
- Complex Sequential Question Answering: Towards Learning to Converse Over Linked Question Answer Pairs with a Knowledge Graph (AAAI 2018 · similarity 0.86)
- Argument Graph Uncertainty: Quantifying Uncertainty from the Logical Structure of Reasoning Chains (NeurIPS 2026 · similarity 0.85)
- Learning to Clarify: Multi-turn Conversations with Action-Based Contrastive Self-Training (ICLR 2025 · similarity 0.85)
- Listening to the Wise Few: Query–Key Alignment Unlocks Latent Correct Answers in Large Language Models (NeurIPS 2026 · similarity 0.85)
- CORECODE: A Common Sense Annotated Dialogue Dataset with Benchmark Tasks for Chinese Large Language Models (AAAI 2024 · 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.