Can Task‑Aware Decoding Tame LLM Hallucinations? Insights from IJCAI 2024

Background

Large language models (LLMs) generate fluent text but often produce inaccurate or fabricated statements, known as hallucinations. Retrieval‑augmented Generation (RAG) mitigates hallucination by injecting external factual knowledge via a retrieval step, yet the LLM can still hallucinate even when retrieved documents are correct.

Sources of Hallucination

Data‑induced Hallucination

Errors, omissions, outdated information, or domain‑sparse training data limit the factual coverage of LLMs. Common remedies include extensive data cleaning and knowledge editing, but these approaches either lack scalability or require additional modules.

Training‑induced Hallucination

LLMs are trained as single‑directional autoregressive models, which introduces attention deficiencies and exposure bias. Supervised fine‑tuning (SFT) or reinforcement learning from human feedback (RLHF) can amplify hallucination when annotation data conflict with the model’s internal knowledge. Architectural or objective changes have limited generality.

Inference‑induced Hallucination

Decoding strategies with high temperature increase the probability of low‑frequency tokens, raising hallucination risk. Attention bottlenecks and insufficient context modeling also contribute.

Layer‑Contrast Decoding (DoLa)

DoLa exploits the observation that lower transformer layers encode syntactic patterns while higher layers contain factual knowledge. By contrasting logits from higher and lower layers, DoLa amplifies factual signals and suppresses purely linguistic ones.

Task‑aware Decoding (TaD)

TaD is a plug‑and‑play decoding method that leverages the difference between a pre‑fine‑tuned LLM and its supervised fine‑tuned counterpart. The difference forms a “knowledge vector” that re‑weights token probabilities toward task‑relevant factual knowledge.

Key steps:

Run the pre‑fine‑tuned LLM on an input and record token logits.

Run the fine‑tuned LLM on the same input and record logits.

Compute the element‑wise difference between the two logits to obtain the knowledge vector.

Adjust the original LLM’s token probabilities by adding the knowledge vector, thereby favoring tokens whose probability increased after fine‑tuning.

Knowledge Vector

The knowledge vector quantifies how token‑level conditional probabilities shift from the generic pre‑training distribution p_θ to the task‑specific fine‑tuned distribution p_ϕ. It captures the transition from broad world knowledge to downstream‑task expertise and is used to steer generation toward more accurate, domain‑relevant answers, especially when fine‑tuning data are scarce.

Experimental Evaluation

TaD was evaluated on multiple LLM families (including LoRA‑adapted and Adapter‑based models) across tasks such as multiple‑choice QA, complex reasoning, and chain‑of‑thought benchmarks. Across all settings TaD consistently outperformed baseline RAG and other contrastive decoding methods. Gains were larger when the proportion of fine‑tuning data was reduced, demonstrating robustness in low‑resource scenarios.

Real‑World Deployment

TaD was integrated with a RAG pipeline in an enterprise knowledge‑question‑answer system covering thousands of business scenarios. The combined TaD+RAG service reduced hallucination‑induced errors and lowered operational overhead.

Future Outlook

While hallucination cannot be fully eliminated under the current language‑model paradigm, combining system‑level approaches (RAG, agents, memory modules) with model‑level interventions like TaD offers a promising direction. Deeper integration of external knowledge and continued research on low‑hallucination decoding are expected to improve the reliability of LLM‑driven applications.

References

Hallucination is Inevitable: An Innate Limitation of Large Language Models

A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

Unveiling the Causes of LLM Hallucination and Overcoming LLM Hallucination

Editing Large Language Models: Problems, Methods, and Opportunities

ACL 2023 Tutorial: Retrieval‑based Language Models and Applications

Theoretical Limitations of Self‑Attention in Neural Sequence Models

Sequence level training with recurrent neural networks

Discovering language model behaviors with model‑written evaluations

Dola: Decoding by Contrasting Layers Improves Factuality in Large Language Models

Bert Rediscovers the Classical NLP Pipeline

Retrieval‑Augmented Generation for Large Language Models: A Survey

TaD: A Plug‑and‑Play Task‑Aware Decoding Method to Better Adapt LLM on Downstream Tasks

Inference‑time Intervention: Eliciting Truthful Answers from a Language Model

Beyond RAG: Building Advanced Context‑Augmented LLM Applications