A novel prompting technique called "Highlighted Chain of Thought" (HoT) significantly improves large language models' ability to explain reasoning while making their answers more verifiable for humans. The two-step approach:
- AI reformulates questions and marks critical facts using XML tags
- Model generates answers referencing these highlighted elements, creating clear logical connections
- Color-coded highlights enable faster human verification of AI reasoning
- Structured approach forces more careful consideration of presented facts, potentially reducing hallucinations
Performance gains:
- Up to 15% improvement across various benchmarks and models
- Compared to traditional CoT methods, HoT showed gains of 1.6 percentage points for arithmetic tasks, 2.58 for question-answering, and 2.53 for logical reasoning
- Most substantial improvements on AQUA (+14.64) and StrategyQA (+15.07) benchmarks
- Tested across five major models including GPT-4o, Gemini-1.5-Pro, and Llama-3.1 variants across 17 different task types
Important limitations:
- Reasoning models showed minimal or negative benefits from HoT techniques
- Smaller models struggled with tagging instructions, often producing incorrect tags
- Moving tags to random phrases significantly impacted accuracy
Human verification paradox:
- Human testers completed verification 25% faster with highlighted answers
- However, highlighting increased trust in AI responses—even incorrect ones
- Humans correctly identified accurate answers 84.5% of the time (vs 78.8% without highlighting)
- Ability to spot errors dropped from 72.2% to 54.8% when highlighting was present
Future directions: Researchers plan to train models to generate HoT answers directly rather than using prompt examples, potentially making the method more effective and broadly applicable.