acceptodds
Withdrawn from ICLR 2027

BRIDGE-MO: Risk-Aware Witness-Guided Optimization for Multi-Objective Joint Target Attainment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Pareto-based multi-objective optimization is widely used in scientific design to balance competing objectives. However, Pareto optimality is often poorly matched to joint target attainment, where success requires all properties to cross fixed thresholds simultaneously. Pareto methods may continue improving already-satisfied objectives or expanding the trade-off frontier while a bottleneck property remains below target. To address this issue, we introduce Bayesian Risk-Aware Intervention-Driven Guided Exploration for Multi-Objective Optimization (BRIDGE), an optimizer that represents alternative routes to joint attainment through a compact frontier of witness points. An LLM proposes structured intervention cards conditioned on these witnesses, while posterior bottleneck screening and constrained reweighting guide candidate selection with independent exploration and real-evaluation feedback. We derive a conditional, one-step Gaussian-process certificate for risk-gated candidates under an idealized posterior model. Across six finite-pool scientific optimization tasks spanning two to eleven objectives, BRIDGE achieves strong joint-margin trajectories and competitive target-attainment and Pareto performance, supported by module ablations and model-sensitivity analyses. Our results suggest that BRIDGE provides a practical and theoretically grounded framework for scientific optimization focused on simultaneous target satisfaction instead of trade-off expansion.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Bayesian Optimization, AI for Science

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_6aurOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

BRIDGE-MO addresses the problem of fixed target multi-objective joint achievement, where all targets must simultaneously meet a predetermined threshold. Method: Maintain a compact multi witness frontier while retaining evaluated candidates with different near feasible deficit profiles. LLM proposes structured intervention cards for individual witnesses; The posterior tail risk (LCVaR) is then used to screen the degree to which each path can be advanced under uncertainty. The candidate selection combines the risk feasibility continuation score, KL constraint card reweighting, and independent exploration channel. A one-step Gaussian process certificate was derived under an idealized posterior model. The first joint hits with relative goal orientation, Pareto and LLM assisted baselines were reported on six limited pool scientific tasks (with target numbers ranging from two to eleven), supplemented by module and language model ablation.

Critical strengths

Witness the forefront clearly retaining complementary bottleneck profiles, rather than collapsing progress to a single incumbent; This design choice is highly consistent with the minimum margin structure achieved jointly, and is consistently reflected in the early hit behavior reported. Separate language proposals from posterior risk screening to keep the LLM output at a modifiable decision level and avoid contaminating the surrogate model with pseudo observations.

Critical weaknesses

W1. Theoretical guarantee is still a one-step conditional certificate under idealized GP posterior and accurate risk calculation. Accumulated regret, adaptive hyperparameter re fitting, and numerical approximation errors only obtain optional corrections; The main claim is that as a result, it has not been established that the actual algorithm inherits the display improvement relative to the bottleneck GP-UCB.

W2. The main experiment and ablation protocol deliberately used different QM9 target sets. The first hit number on the eleven target main table cannot be compared with the five target ablation package, making it impossible to transfer module contribution claims to the headline benchmark.

W3. Several strong baselines (TAMO, MOEA/D-LLM, qParEGO on some tasks) have a success rate of zero or close to zero within a fixed budget. This raises the question that the experimental pool or initialization protocol may be abnormally advantageous for the proposed risk gating search, rather than having broad representativeness.

W4. Ablation indicates that removing LLM cards or hot start only produces limited delay, while removing batch diversity or cutting-edge risk filtering has a significant impact. The language component appears to be auxiliary rather than core, weakening the claim that witness condition intervention cards are primarily driven by performance.

W5. Each method task uses only five seeds, with several initial hits having a standard deviation exceeding 20-30 units. Only five trajectories make the reported mean advantage lack statistical robustness, especially in high variance tasks such as PFD and Tartarus emitter.

W6. All tasks are limited candidate pools; No continuous or high-dimensional design space experiments were provided. The practical value of multi witness and risk screening mechanisms for truly black box scientific design has therefore not been tested.

W7. The lower tail risk functional relies solely on the working joint distribution recovered through residual shrinkage approximation. Non sensitivity analysis quantifies how the misspecification of the dependency structure affects the level of acceptance or the resulting certificate.

There are several numerical patterns in the table: the first hit mean often ends in. 4 or. 2, the success score is uniformly reported as x/5, and several baseline entries share the exact same 101.0 ± 0.0 pattern. Although it cannot be concluded that there is fraud, these patterns reduce the credibility of statistical reports and require the disclosure of the original trajectories for each sub track.

Questions for authors

Can the QM9 results of the 11th target be recalculated under the same target protocol as the module ablation, in order to measure the contributions of each component on the headline task?

What is the empirical frequency of LLM card alignment with positive covariance on six tasks, and what is the actual frequency of expected certificate improvement in KL constraint step?

Why have several recent LLM assisted and amortized baselines collapsed to zero success on multiple pools, while the proposed method has been successful on every seed?

Reviewer_N8CWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BRIDGE-MO addresses fixed-target multi-objective optimization over finite candidate pools, where success requires every objective to meet a prescribed threshold. The method maintains a small, diverse set of evaluated witnesses and uses LLM-generated intervention cards to suggest repairs to their remaining deficits. Per-objective Gaussian processes, coupled posterior samples, lower-tail bottleneck risk, and KL-constrained reweighting guide selection, supplemented by independent exploration and batch diversity. The theory provides conditional one-step regret certificates and expected-certificate improvement under positively aligned card scores. Experiments cover six scientific pools with two to eleven objectives, using a reported budget of 100 evaluations and five seeds. BRIDGE-MO achieves the lowest mean first-attainment count on all six tasks. Component and language-model ablations use a separate five-objective QM9 protocol.

Critical strengths
  • Combining multiple evaluated witnesses, posterior risk screening, and revisable intervention cards provides an interpretable mechanism for using scientific priors while retaining experimental feedback.

  • Across six heterogeneous candidate pools, BRIDGE-MO attains all targets in every reported seed and achieves the lowest mean first-hit count, providing promising empirical evidence.

Critical weaknesses
  • The incremental contribution beyond existing constrained and target-oriented BO is insufficiently established. Requiring zj(x)≥0z_j(x)\geq 0 for every objective is exactly black-box feasibility under constraints cj(x)=−zj(x)≤0c_j(x)=-z_j(x)\leq 0, while maximizing the bottleneck is a reference-point maximin scalarization. Existing constrained BO already addresses feasibility search without a feasible incumbent; BMOO additionally compares infeasible designs through their constraint-violation vectors, which is directly relevant to preserving complementary deficit profiles. Hanaoka is an appropriate included baseline, but the paper lacks comparisons against direct bottleneck-based acquisition under the same surrogate and a violation-aware constrained method. It also does not specify whether methods such as qLogNEHVI receive the targets through constraints or suitable reference points. Consequently, the current evidence does not determine how much improvement comes from the proposed witness/card machinery versus aligning acquisition more directly with the evaluation objective.

  • The experiments do not yet establish statistically robust gains attributable to the central mechanisms. Table 1 reports five-seed means and standard deviations without paired uncertainty estimates or tests. Some advantages over the strongest competitors are small relative to the reported variability: PFD is 41.6±24.541.6\pm24.5 versus MORBO’s 48.6±12.648.6\pm12.6, and Sandwich is 21.4±11.821.4\pm11.8 versus 24.2±13.424.2\pm13.4. These summaries do not establish statistical significance. Moreover, component ablations are restricted to a separate five-objective QM9 protocol, and there is no explicit, otherwise matched K=1K=1 control for the central multi-witness claim. Removing LLM cards changes the mean first hit from 12.2 to 14.2 in that protocol, without uncertainty for this difference. Thus, the evidence does not yet isolate the contributions of multiple witnesses and language guidance to the larger gains on the primary tasks.

Questions for authors
  1. Under the same objective surrogates, representations, initialization, and evaluation budget, how does BRIDGE-MO compare with direct Monte Carlo acquisition on min⁡jzj(x)\min_j z_j(x), the bottleneck-UCB reference used in the theory, or a constrained BO method that accounts for violation profiles before finding a feasible point? Please also specify the target/constraint encoding, hypervolume reference points, finite-pool adaptations, and relevant surrogate or backbone settings of the existing baselines. These details would help establish the incremental algorithmic contribution.

  2. Do the gains persist on the primary task protocols when using a single witness, or when disabling all language guidance, while keeping the risk, exploration, and batch-diversity components fixed? The existing five-objective QM9 ablations are useful, but do not establish these contributions on the eleven-objective main task or the other benchmarks. Please also provide the complete objective, threshold, and scale specification for the five-objective ablation protocol.

  3. Please provide paired per-seed first-hit results, uncertainty estimates for the differences, and success-by-budget statistics against the strongest relevant baselines. Table 1 explicitly codes unsuccessful runs as 101; please distinguish this budget-penalized score from uncensored time to attainment and assess the robustness of the comparisons to the treatment of unsuccessful runs. How were the target thresholds selected, and what fraction of each frozen pool is jointly feasible?

  4. How do the continuation levels from Eq. (8) determine the eligible support and ranking in Eqs. (61)–(62)? Please clarify the distinction between “expected joint margin” and “expected bottleneck margin,” and reconcile the LCB-based witness boundaries in Eqs. (3)–(5) with the observed-margin boundaries in Eq. (60). Which conditions of Theorem 1 are enforced or empirically verified in the implementation, given its finite posterior samples, largest KL-feasible reweighting step, and deterministic diversity-based selection?

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

This paper proposes an LLM-driven approach to identifying feasible solutions to highly constrained, expensive problems. It combines ideas from black-box optimisation, LLMs, and Bayesian experimental design. The algorithm is demonstrated across a range of scientific problems and some theoretical bounds on regret are provided.

Critical strengths

The motivating problem is sensible and there is a wide range of different problems considered in the experiments.

Critical weaknesses
  1. The problem is not new, and the relevant literature is missing. The paper frames “fixed-target multi-objective joint-attainment optimisation” as a new problem class. But requiring the worst normalised margin to be non-negative is the same as looking for points in the intersection of the objectives’ level sets. Finding such regions is well studied under names like level-set estimation, excursion-set estimation and feasible-region identification. Examples include the SUR of Chevalier et al. (2014) for excursion sets. It also maps directly onto constrained BO with probability of feasibility (Gardner et al., 2014), but considering every objective is treated as a constraint. None of this work is discussed, and the related-work section only covers MOBO, target-oriented methods and LLM optimisers. For the paper to be a fair contribution, it must position itself against this literature and benchmark against some of these methods.

  2. Key algorithmic components already exist without reference. Several building blocks are presented as new but have clear precedents. Gradually expanding a set of trusted regions from known-good points is the core idea of safe BO (e.g., SafeOpt, Sui et al., 2015, and follow-ups). Using CVaR or other risk measures in BO acquisition and decision rules, including multi-objective settings, is also established (e.g Daulton et al., 2022). The paper should say clearly what is new relative to these methods rather than presenting the combination as new from the ground up.

  3. The chosen baselines are weak. Most baselines are built for Pareto-front recovery. The paper itself argues early on that such methods are mismatched to joint target attainment, so their poor first-hit times are unsurprising and say little about BRIDGE-MO. To be convincing, the evaluation needs simpler methods that target joint feasibility directly under GP surrogates, for example: (i) maximising the joint probability that all objectives meet their thresholds, as in constrained BO; (ii) level-set or excursion-set acquisition functions (e.g., LSE or SUR-based); and most importantly (iii) a bottleneck GP-UCB rule: run UCB on each normalised margin and take the minimum across objectives. The authors define this exact rule in Appendix A.2 as their theoretical reference point, yet it is never run as a baseline. This is the most natural comparison and it is missing. If BRIDGE-MO does not clearly beat these simple, cheap baselines, it is hard to justify its complexity.

  4. The paper is very hard to follow. I read the paper several times and still found the key ideas unclear, which is a concern on its own. The paper leans on jargon (“witness”, “intervention card”, “certificate”, “continuation level”, “net certified repair”, “route”) that is used many times before it is defined. The introduction jumps into these concepts early, and the contribution list is close to impenetrable without having read Sections 3 and 4. Figure 1 tries to show everything at once with a short caption that doesn’t explain the panels, so it overwhelms rather than explains. I’d strongly encourage a rewrite. Introduce the problem and a simple baseline first, then add each component with a one-line motivation, and redesign Figure 1 as a simple diagram with a caption that walks through it.

  5. Algorithm complexity and missing detail. The full loop is given very briefly in Section 3.5, and I found it hard to work out what happens each round. Important parts are left vague in the main text, most notably how the LLM actually produces cards and how a card turns into a score over candidates. The method also has many moving parts and hand-set constants (Table 3 and Appendix C: witness growth schedule, distance weights, path priors and floors, KL budget, safety mixing weight, shortlist proportions, refresh rules, and score weights that change after the first hit). This raises questions about robustness and how much tuning each task needed.

  6. Unclear value of the LLM. I don’t understand why the LLM adds value. For general black-box objectives, it isn’t clear what an LLM could propose that a decision rule over the GP posterior couldn’t do automatically. This needs empirical justification.

  7. Ablations are not enough to justify the design. All ablations use a separate five-objective QM9 protocol, not the main benchmarks. Each complicated component (multi-witness frontier, risk-based continuation gate, KL-constrained card reweighting, safety mixing, diversity selection) should be justified against a simpler alternative across several tasks, starting from the bottleneck GP-UCB baseline and adding components one at a time. A GP-only variant (no LLM at all, keeping the frontier and risk screening) on all six tasks would be especially informative.

  8. Theory. I’m not an expert in regret analysis and can’t fully verify the theorem. I did try to read the appendix proofs, but they are very dense and read more like a log of steps than carefully presented maths, it seems to add length and notation more than insight.

Questions for authors
  1. How does your problem setting differ from level-set or excursion-set estimation over the intersection of the objectives’ feasible regions, or from constrained BO with every objective treated as a constraint? Why are these literatures not discussed?

Why is the bottleneck GP-UCB rule from Appendix A.2 not included as a baseline? How does BRIDGE-MO compare to it, and to maximising joint probability of feasibility, on all six tasks? 3. What is new about the risk-gated expansion compared with SafeOpt-style safe-set expansion and existing CVaR/MVaR-based BO? 4. Concretely, how does the LLM produce cards? What exactly is in the prompt for each task, and how are its suggestions turned into candidate scores? Could you show example cards and whether they were scientifically sensible? 5. What can the LLM supply that a posterior-based decision rule cannot? Can you show a task where removing the LLM causes a statistically significant drop?

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

Related papers

Discussion (0)

Sign in to comment.