Grounded Decoding

The core mechanism involves modifying the raw logit scores before the softmax function. By boosting the probabilities of tokens found in the evidence document and penalizing tokens that would introduce unsupported information, the model is steered toward faithful generation.

• Context-Aware Boosting: Tokens with high TF-IDF or semantic similarity to the evidence receive additive bonuses.
• Entropy Maximization Penalty: The model is penalized for selecting high-probability tokens that lack evidential support.
• Inference-Only: Unlike fine-tuning, this requires no weight updates, preserving the model's general capabilities.

A powerful grounding approach that runs two forward passes: one with the evidence context and one without. The final token distribution is derived from the logit difference between the two passes, amplifying signals unique to the evidence.

• Expert vs. Amateur: The evidence-conditioned model acts as the 'expert,' while the unconditioned model is the 'amateur.'
• Hallucination Suppression: Effectively cancels out generic, high-probability tokens that the model would generate from its parametric memory alone.
• Computational Cost: Requires 2x the compute per token, a key trade-off for high-stakes accuracy.

A lightweight method that prepends a learned, continuous vector sequence to the evidence context. This virtual token sequence acts as a steering signal that biases the entire generation toward extractive, grounded behavior without modifying the base model.

• Virtual Tokens: These are not human-readable words but optimized embedding vectors.
• Task-Specific: Different prefixes can be trained for summarization vs. question answering.
• Minimal Latency: Adds negligible inference time compared to full contrastive decoding.

This technique intervenes in the attention mechanism itself by applying a binary or weighted mask that forces every generated token to attend exclusively to the evidence document, ignoring the prompt's instruction tokens or prior generated text.

• Attention Re-weighting: Redistributes attention scores to prioritize source spans.
• Copy Mechanism Enhancement: Increases the likelihood of directly copying entities and figures from the source.
• Implementation: Requires hooks into the model's attention layers, making it framework-dependent but highly precise.

Integrates a Natural Language Inference (NLI) model directly into the decoding loop. For each candidate sequence in the beam, an NLI model scores whether the hypothesis (the candidate) is entailed by the premise (the evidence). Beams with contradiction scores are pruned.

• Factual Pruning: Actively discards hallucinated branches during generation.
• Computational Intensity: Running an NLI model at each decoding step is expensive, often reserved for final answer validation.
• Granular Control: Provides token-level or span-level factual verification.

A method that uses a lightweight, fine-tuned 'critic' model to evaluate the factual consistency of partial sequences during generation. The critic's score is used to re-rank or adjust the logits of the main generator, ensuring only factually consistent tokens survive.

• Critic Model: Often a smaller, distilled model trained on hallucination detection data.
• Plug-and-Play: The critic can be swapped without retraining the main generator.
• Trade-off: Balances the speed of heuristic logit manipulation with the accuracy of explicit verification.