Defective Decoding Process
In general, LLMs employ the Transformer architecture [32] and generate content in an autoregressive manner, where the prediction of the next token is conditioned on the previously generated token sequence. Such a scheme could accumulate errors [105]. Besides, during the decoding process, top-p sampling [28] and top-k sampling [27] are widely adopted to enhance the diversity of the generated content. Nevertheless, these sampling strategies can introduce “randomness” [113], [136], thereby increasing the potential of hallucinations
ENTITY
2 - AI
INTENT
2 - Unintentional
TIMING
1 - Pre-deployment
Risk ID
mit41
Domain lineage
3. Misinformation
3.1 > False or misleading information
Mitigation strategy
1. Implement advanced **Contrastive Decoding** techniques, such as Decoding by Contrasting Layers (DoLa) or Context-Aware Decoding (CAD), to dynamically modulate the token logit distribution during autoregression. This process contrasts output probabilities to amplify contextually faithful or factual tokens while suppressing the selection of overconfident yet potentially erroneous alternatives, directly mitigating the accumulation of decoding errors. 2. Employ **Uncertainty-Aware Beam Search** or **Factual Nucleus Sampling** as refined search strategies. Uncertainty-Aware Beam Search prunes high-uncertainty sequences early in the generation process, thereby steering the model away from less reliable paths that lead to hallucinations, while Factual Nucleus Sampling biases token selection toward options grounded in factual evidence. 3. Systematically calibrate and constrain **decoding hyperparameters**, specifically reducing Temperature and adjusting Top-p/Top-k thresholds (or utilizing dynamic alternatives like Min-p sampling). This action directly reduces the stochasticity introduced by sampling mechanisms, compelling the model to select from a smaller set of the most probable tokens to ensure higher coherence and factual adherence.