acceptodds
Under review as a conference paper at ICLR 2027

MultiView Safe Decoding:Composing Cross-Horizon Evidence for Safer LLM Generation

Abstract

Lookahead can reveal several useful views of a candidate action, yet a decision based on one view can miss the implications of another. We introduce MultiView Safe Decoding (MVSD), a controller that augments an existing lookahead ranker with evidence composed across horizons. Two shared heads reuse its trace-wise features to predict absolute and pairwise executed losses, and joint calibration turns their predictions into constraints on one executed-loss vector. On a fixed panel, the controller chooses the highest-ranked candidate whose combined bounds satisfy loss and candidate-relative regret tolerances, retaining the original action when none qualifies. We establish marginal validity after action and horizon selection and give a construction in which two views jointly select an action chosen by neither view alone. A baseline-relative improvement bound connects this mechanism to changes in executed loss. Our evaluation separates benchmark performance, decision gains, evidence contributions, and computational cost, with matched-information derived heads and learned fusion as controls. On four safety benchmarks (HarmBench, SALAD-Bench, StrongREJECT, JustEval), MVSD achieves 92.92% SALAD safety and strong multi-metric performance, improving over the reference baseline by +0.48pp. However, this comes at 5 computational cost versus simpler methods. Best-of- remains superior on latency (1.46% HB ASR vs. 1.67%), while InferAligner excels on continuous harmfulness (SR: 0.01750 vs. 0.02559). We identify three failure patterns through qualitative analysis: benign edge case over-refusal (36%), inconsistent multi-view signals on creative requests (44%), and computational waste on all-safe candidate sets (20%). The framework makes cross-horizon composition explicit, demonstrates when it provides value, and honestly characterizes its limitations and trade-offs.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Decoding-Time Safety, Future-Guided Action Selection, Paired-Future Supervision, Uncertainty-Aware Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3eDkOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a safety decoding problem about how evidence from different lookahead horizons can be combined for candidate action selection. The authors propose MVSD, which combines absolute and pairwise loss predictions across horizons into calibrated constraints for selecting the final action.

Critical strengths

1. The paper studies an safety decoding question about how evidence from different lookahead horizons can jointly affect action selection. The formulation is quite insightful. Casting cross-horizon evidence as constraints for action selection is easy to follow and reasonable.

2. MVSD has a simple and practical design. Compared with heuristic fusion, its design philosophy is more principled. The component and decision-level analyses make this method more concrete.

Critical weaknesses

1. I see that the authors use the predictive complementarity across horizons as the main motivation for MultiView. Here the views are based on nested observations, where a longer horizon can contain the shorter horizon information. I am concerned that the observed complementarity may mainly reflect different prediction errors, without showing that the views provide genuinely independent evidence.

2. The identification of the composition effect is not very clean. The main comparisons change both the available evidence and how that evidence is combined. The current results show that different views can provide complementary signals, but do not isolate how much of the gain comes from composition itself. e.g, the constructed examples show that multiple views can jointly support an action that no single view supports, but do not show it can drive the observed gains.

Questions for authors

1. Given that the horizon views are nested, what evidence shows that horizon composition contributes beyond horizon-specific prediction errors?

2. On the actual evaluation set, how often does the selected action depend on evidence from multiple views, and how often do those composition-driven action changes improve over the reference action?

3. How do the current analyses isolate the effect of the composition rule while keeping the fitted predictions and calibration intervals fixed?

4. What uncertainty estimates accompany the common-bank improvements, especially the relatively small benign-loss gain?

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

This paper proposes MVSD to improve LLM generation safety by combining evidence across lookahead horizons. It integrates absolute and pairwise loss predictions with joint calibration and constraint propagation to select qualified response openings. The paper provides theoretical guarantees and evaluates the method through safety benchmarks, a common execution bank, and ablation studies.

Critical strengths
  1. The paper addresses the integration of evidence across different lookahead horizons and proposes a constraint-based framework for safer decoding. The overall approach is clear and interesting.

  2. The paper combines loss prediction, joint calibration, and constraint propagation, supported by theoretical guarantees and analysis. The methodology is well structured.

Critical weaknesses

1. Limited evaluation of safety improvements.
Table 2 primarily reports harmful regret, which remains zero when all candidates are unsafe. The experiments lack an analysis of absolute harmful loss and selection outcomes across different candidate safety groups.

2. Insufficient analysis of cross-horizon evidence composition.
Section 4.3 introduces constraint-deletion analysis but does not report quantitative results. Figure 3 mainly compares separately trained variants without directly measuring the contribution of individual views to decision improvements.

3. Limited experimental settings.
The experiments use a single generator–critic pair and only intervene at the beginning of generation. Evaluations across different model backbones, opening lengths, and multi-step interventions are missing.

Minor Weakness: Labeling error in Figure 2(B).
Figure 2(B) incorrectly presents zB≤0.45z_B \le 0.45 as the conclusion. The correct inference is zB≤0.45z_B \le 0.45 and zA−zB≤−0.36z_A-z_B \le -0.36, which imply zA≤0.09z_A \le 0.09.

Questions for authors

1. Could the authors report the results of both historical configurations and full MVSD on the four main benchmarks, and clarify whether full MVSD uses a single fixed configuration?

2. What explains the differences in SALAD results for InferAligner (89.68% vs. 91.42%) and SafeInfer (88.33% vs. 87.68%) between Table 1 and Table A11?

3. Could the authors provide the actual evaluation sample sizes, paired confidence intervals, and results across different random seeds for Table 2?

4. How many samples require multiple views to meet the qualification criteria, and how does removing individual views affect action selection and executed outcomes?

Reviewer_CrT2OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

This paper constructs multiple views from different lookahead horizons and uses each view to predict both the absolute loss and the pairwise loss. These predictions are converted into constraints and combined across views. Among the openings whose worst-case loss and regret fall below predefined thresholds, the method selects the one ranked highest by the reference critic. The authors theoretically show that the loss and regret of the selected opening are bounded by these thresholds with high probability.

Critical strengths
  1. The paper compares against a wide range of baselines.
  2. The method is supported by theoretical analysis, in particular, Proposition 1.
Critical weaknesses
  1. While the core idea is simple, the paper is written in a way that makes it difficult to parse. The illustrative examples provided in the paper even make it more confusing. For example, the motivating example in the introduction is described only in vague terms (e.g., "one veiw", "another view" a “tight” bound and a “loose” estimate from another), without a clear explanation of how these can occur in practice. As a result, it is difficult to follow and fails to motivate why combining absolute and comparative predictions is useful.

  2. Many terms and notations are introduced without definition or explanation, which makes the details hard to follow. For example, x and y_{<t} are not explained, and Q^J_E and \mathcal{H} are introduced without describing what they represent. It is also unclear what a “view” means though it is used quite a lot throughout the paper, or what the “controller” refers to.

  3. The explanations are overly abstract throughout.

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.