Two decoders are better than one: A new way to make LLMs trustworthy
Knowledge triplets and dual decoders team up to stop LLM hallucinations
https://arxiv.org/abs/2411.07870
🎯 Original Problem:
LLMs often generate ungrounded or hallucinated content when adapted to specific domains, lacking factual accuracy and verification from source materials. This becomes critical in business applications where responses must be based on verified context and domain-specific knowledge.
-----
🔧 Solution in this Paper:
→ A post-processing algorithm converts knowledge triplets from RAG context and LLM output into graphs, verifying and correcting hallucinations by comparing subject-relation-object patterns
→ TrustfulLLM, a dual-decoder architecture, shares weights between two decoders - one processes the grounding context while the other handles user prompts
→ Cross-attention mechanism fuses guided context with natural language generation, ensuring outputs remain grounded in verified information
→ The system integrates with Microsoft's commercial applications, using intent detection to filter queries and retrieving relevant internal documents
-----
💡 Key Insights:
→ Knowledge triplets from RAG context serve as effective verification tools for LLM outputs
→ Dual-decoder architecture reduces irrelevant entities in generated text by 11.1% compared to baseline models
→ Cross-attention between context and prompt helps maintain factual consistency
-----
📊 Results:
→ TrustfulLLM + HC achieved ROUGE-L: 0.55, METEOR: 0.51
→ Perfect Groundedness score of 5.00 with GPT-Similarity: 4.68
→ BERTScore: 0.93, outperforming all baseline models including GPT-4
Share this post