acceptodds
Under review as a conference paper at ICLR 2027

PrePair: Speculative Drafter Training for Generative UI under Semantic Acceptance

Abstract

Generative UI lets large language models answer with interactive interfaces instead of plain text, but an interface spans many more tokens than a text reply, limiting real-time interaction. Speculative decoding accelerates generation by letting a small drafter propose tokens that a large target verifies in parallel, keeping each only if it matches the target's token distribution. In Generative UI, however, we find that the drafter writes many components that are semantically correct but not token-identical to the target's, and some correct ones that the target gets wrong, so per-token acceptance both discards valid drafts and bounds the whole system by the target. In this paper, we lift speculative decoding from per-token acceptance to component-level semantic acceptance, in which the target checks all components against several criteria in one forward pass. Existing drafter training, built for per-token acceptance, imitates the target, which under semantic acceptance both rewrites components the drafter already writes correctly and passes on the target's errors. To overcome this, we propose PrePair, which reinforces accepted components and distills the target only on the divergent tokens of failed ones, located by chain repair, so the two objectives never share a token. With a 1.7B drafter and a 32B target, PrePair raises semantic acceptance over supervised fine-tuning from 81.0% to 87.8%, for a 3.5× end-to-end speedup over the target alone versus 2.3× for EAGLE-3, and on Macaron-A2UI the trained system outscores the 32B target writing directly under an independent evaluator.

Primary area:
generative models
Keywords:
Speculative Decoding, Efficient LLM Inference, Drafter Training, Semantic Verification, Generative UI, Knowledge Distillation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_PHg5OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PrePair, a training framework for speculative drafting in Generative UI under component-level semantic acceptance. The authors identify a limitation of conventional token-level speculative decoding, where semantically valid UI components may be rejected because their token sequences differ from those of the target model.

To address this issue, PrePair combines a tree-structured semantic verifier with a preserve-and-repair training strategy. Accepted components are reinforced through a policy-gradient objective, while rejected components receive selective knowledge distillation at divergent token positions identified through chain repair. The two objectives operate on disjoint supervision positions to avoid conflicting training signals.

Experiments with a Qwen3-1.7B drafter and a Qwen3-32B target demonstrate improvements in semantic acceptance from 81.0% to 87.8% over SFT. The complete system achieves a 3.49× end-to-end speedup over direct target decoding, compared with 2.33× for retrained EAGLE-3, while improving the average Macaron-A2UI quality score from 3.65 to 3.74 over SFT.

Critical strengths
  1. Well-motivated problem formulation and interesting methodology.

The paper identifies a meaningful mismatch between token-level agreement and semantic correctness in structured generation. The observation that many semantically accepted UI components differ from the target's outputs provides clear motivation for reconsidering conventional speculative decoding objectives. The proposed preserve-and-repair framework is intuitive and technically interesting, particularly its separation of reinforcement for accepted components and selective distillation for rejected ones.

  1. Promising efficiency improvements and clear presentation.

The reported 3.49× end-to-end speedup is substantial, and the comparisons with EAGLE-3 and token-level acceptance demonstrate the practical potential of semantic speculative decoding. The paper also evaluates output quality using independent models and provides useful ablations. Overall, the presentation is clear, and the figures effectively illustrate the proposed verification and training procedures.

Critical weaknesses
  1. The contribution of disjoint supervision is not sufficiently isolated from other optimization choices.

A central methodological claim of PrePair is that assigning preservation and correction objectives to disjoint positions avoids conflicting training signals and improves drafter performance. However, the "plain sum" baseline in Section 4.4 modifies several optimization choices simultaneously, including the treatment of negative advantages, the reward margin term, and chain-position weight normalization. Consequently, its substantial performance degradation cannot be attributed specifically to overlapping supervision. Although the existing ablations demonstrate the effectiveness of the overall training strategy, they provide limited evidence for the independent contribution of disjoint supervision, which is important for establishing the novelty of the proposed method relative to existing selective distillation approaches.

Suggested experiments: I encourage the authors to conduct a controlled ablation comparing the original PrePair objective with a variant that additionally applies target distillation to accepted components, while keeping the reward formulation, optimization settings, and training budget consistent. Reporting acceptance rates, output quality, and repair efficiency for both variants would directly clarify whether avoiding distillation on semantically accepted components is responsible for the observed improvements.

  1. The reliability of semantic verification on realistic UI errors remains insufficiently established.

PrePair relies heavily on the semantic verifier to determine which components should be preserved and which require correction. However, Appendix G primarily evaluates three content-related criteria using synthetic defects constructed from 107 accepted content components. While the reported results are encouraging, they provide limited evidence about naturally occurring generation errors, especially those involving dialogue state and interactive functionality. The authors also acknowledge that these context-dependent criteria are less reliable. This raises uncertainty about whether improvements in verifier acceptance consistently reflect better semantic correctness. In particular, although acceptance increases from 81.0% to 87.8%, the L2 task-construction score remains unchanged at 1.95 relative to SFT, suggesting that improvements may be uneven across different aspects of UI quality.

Suggested experiments: I encourage the authors to evaluate the verifier on a representative sample of naturally occurring accepted and rejected components, using independent semantic correctness annotations. Reporting false-accept and false-reject rates, particularly for state continuity and interactivity, would provide stronger evidence of verifier reliability. An additional breakdown of acceptance improvements against task-level quality would help clarify the relationship between verifier judgments and actual UI correctness.

Questions for authors

Q1. Can the authors provide a controlled ablation that isolates the contribution of disjoint supervision from other optimization choices? In particular, how does PrePair compare with a variant that also distills accepted components under otherwise identical training settings?

Q2. How reliable is the semantic verifier on naturally occurring UI generation errors? Could the authors provide false-accept and false-reject statistics for state continuity and interactivity, and clarify why the higher semantic acceptance rate does not lead to an improvement in the L2 task-construction score?

Reviewer_PVXiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper “PrePair: Speculative Drafter Training for Generative UI under Semantic Acceptance” presents an interesting idea that uses speculative decoding for Generative UI. The work identifies shortcomings in the existing methods: per-token acceptance both discards valid proposals, which bounds generation. The work instead advocates for semantic acceptance criteria.

Critical strengths
  1. The idea presented in the paper is reasonable: Accepted components receive an objective that preserves them, while divergent tokens receive gradients towards the target.

  2. The authors have provided some ablations for their proposed method (Section 4.4). This builds some confidence with regards to the proposed technique.

Critical weaknesses
  1. The writing of the paper is not good and I found the paper hard to read the paper. In general, a scientific paper needs to be self-contained and properly structured. I feel the paper has been written assuming the readers/reviewers already know all the content. I will try to give a few examples that illustrate this point, although many exist throughout the paper:

    1. [Line 103, 319]: I don’t know what SFT means. Could the authors expand on that ? It seems like an important detail.
    2. [Line 108] Tree-structured semantic verifier: Why is it necessary ? Is it the main idea ? If it is the main idea, the introduction should have explained it in detail and given reasons for its benefits. The only line I see in the Introduction section is [Line 078], which does not inform me much. If it is not the main idea, the introduction should not have mentioned it.
    3. [Line 188] This entire section was very hard to follow through. It uses terms such as trunk, branch and readout position before defining or motivating them. A lot of key concepts are packed together without giving proper motivation. For example: “zero-cost hard gates (schema validity, resolvable references, no verbatim repetition)”. How were these gates selected ? Why only these and not something else ? I feel a lot of critical insights are swept under the rug, assuming that the reader will automatically understand the presented content.

    I think the paper can improve from a thorough re-write. The authors should decide what to prioritize and focus on how they present the technical content.

  2. Lack of human preference evaluation and qualitative results. The claims of the paper rely on LLM judges, so it is hard to judge whether the proposed method is actually doing well. The author has also not sown rendered interfaces or example components that motivate the method (like showing repaired components). Maybe visually showcasing cases where the method is better than the baselines can help.

  3. The results lack statistical significance. Error bars in the result section (over different seeds) would improve the confidence in the results.

Questions for authors

Refer to the Weakness section.

Reviewer_sye6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

For generative UI task where the output has component structure, the paper proposes Prepair, which uses the target model to provide component-level judgement instead of token-level judgement for speculative decoding. For component-level judgement, the To better adapt the draft model for the task, Prepair trains the draft. For sequence that needs repairing, Prepair asks the target to only generate the fix tokens and use the remaining tokens from the draft. Prepair also reinforce the correct responses based on relative advantages. The experiments show that prepair outperforms token-level method with lower cost.

Critical strengths

(1) The paper utilize the component structure of generative UI and proposes a new method to address the issue of token-level method.

Critical weaknesses

(1) The evaluation part is not comprehensive enough. The experiment only uses one model, which is not enough to show whether the method is generalizable. Besides, the method only compares with internal baselines and eagle and there are existing speculative decoding baselines that also based on semantic units instead of tokens. For example, [1] also proposes to judge the acceptance based on semantic spans.

[1] Parallel Prefix Verification for Speculative Generation

(2) The generation of repair seems to require much more computation cost than using the target model to rewrite the target and multiple rounds needs to be sequential. Can the author compare the computation cost of both methods?

Questions for authors

(1) The results that prepair outperforms the target model on some dimensions are interesting. Is this because the target model are better at judging instead of predicting a single token?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.