Goal Model

Speculative Decoding uses a small, fast draft model to predict multiple tokens in advance, which are then verified in parallel all at once by the large model.

Medusa adds multiple prediction heads on top of a model to predict tokens at different future positions and confirms them through tree-based verification, without requiring an additional draft model, and can improve inference speed by 2–3×.

Traditional soft-label distillation requires the draft model to approximate the target model's full probability distribution at every position, including high-entropy regions, which is inconsistent with the goal of maximizing acceptance length at inference time.

Attack success rate is not an intrinsic property of the target model but a joint product of the target model and the grader.

The core idea of speculative decoding is: use a smaller draft model to quickly generate multiple candidate tokens, then have the target large model verify in parallel in one pass whether these tokens are acceptable.

Goal models are established tools in requirements engineering and business analysis, with common forms including the i* framework and the KAOS method.

The method first builds a goal model for the organization, pre-defining the axes of classification (classification dimensions and criteria) rather than passively discovering them from the data.

The cost of the method is the additional effort required to construct the goal model, which is an upfront cost that must be paid compared to automatic clustering.

The paper has LLMs interpret the textual narrative of each variant in the context of the goal model and assign it to the most appropriate category, using the LLM's semantic reasoning capability to connect low-level process behavior with high-level business objectives.

Co-design treats the draft model and the target model as an integrated system to be designed and trained from the outset, in order to improve the token acceptance rate.