acceptodds
Under review as a conference paper at ICLR 2027

MARCH: Margin-Aware Reasoning for CBF-Harnessed Multi-Agent Control

Abstract

Control Barrier Functions (CBFs) are widely used to enforce safety by activating a low-level Quadratic Program (QP) to correct high-level planning decisions before execution. This CBF-QP-based architecture is attractive for LLM-based planners and works well in simple single-robot or low-constraint settings, where such corrections are only occasionally needed. However, in swarm systems with multiple coupled constraints, frequent CBF-QP activation can make the low-level CBF-QP hard to solve or even infeasible. This suggests a different perspective: maintaining CBF-QP feasibility is not only a low-level optimization problem; it also depends on whether the high-level LLM planner avoids decisions that repeatedly activate the CBF-QP. Based on this view, we propose Predictive Harness-CBF (Predictive H-CBF), a harness-based framework that uses low-level CBF-QP feedback to revise future high-level decisions generated by an LLM-based planner. Specifically, the harness receives CBF margins and related QP feedback, organizes them together with task context and memory, and generates multiple high-level decision candidates. A predictive evaluator then rolls out these candidates over a short horizon and selects the final decision based on task progress, predicted CBF activation, and QP feasibility. The selected decision guides future high-level LLM output, while the low-level CBF-QP continues to enforce safety during execution. Experiments on 3-D swarm-control tasks show that Predictive H-CBF reduces repeated CBF activation, hard CBF-QP infeasibility, and unnecessary CBF interventions while maintaining competitive coverage performance.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
CBF, Harness agent, Multi-agent system

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_U3NLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces MARCH, a framework that feeds constraint margins and invervention diagnostics from a low-level CBF-QP controller back to an LLM. The LLM generates candidate four-dimensional control parameters, which are then evaluated via short-term prediction to select those that improve the future QP feasibility and the trade-off between safety and performance. Additionally, the framework provides an independent feasibility audit over a finite time horizon.

Critical strengths
  1. The feedback interface is practically useful: it employs CBF-QP diagnostics to guide future planning, as nominal objective adjustments alone cannot resolve infeasibility.
  2. The predictive selection mechanism receives targeted experimental support. Comparing the "Random", "Fixed", and "Progress-only" strategies alongside the full selection strategy is more compelling than simply comparing three complete systems.
Critical weaknesses

1.The second contribution integrates established components, including LLM proposals, memory, predictive evaluation,and CBF-QP filtering. Although the ablations and generator comparisons demonstrate empirical benefits, the distinctive advantage of this particular coupling over simpler CBF-aware predictive planner remains insufficiently characterized. 2. Candidates are evaluated over three control steps, while the selected planning parameters are held for eight. It would be valuable to justify this asymmetric choice of time horizons.

Questions for authors
  1. What distinctive benefits arises from coupling feedback-conditioned LLM proposals with predictive selection?
  2. Why was a three-step prediction horizon chosen for an eight-step holding interval, given the suffix failures reported in Table S39? Could the authors compare H=3 and H=8, with L=8 fixed, on matched test scenes and report feasibility, task performance, and planning latency?
Reviewer_qeKGOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes MARCH, a framework that uses feedback from a low-level CBF-QP to guide high-level LLM planning for multi-agent control. The LLM proposes candidate settings for the nominal controller, while a predictive evaluator uses short-horizon CBF-QP rollouts to select among them. Experiments on 3-D multi-UAV coverage across different difficulty levels and transfer tasks report fewer QP infeasibilities and safety violations compared with reactive filtering and direct-feedback baselines.

Critical strengths
  1. A useful framing of CBF intervention as a planning signal. The paper treats repeated CBF-QP intervention and infeasibility as evidence that high-level decisions may be poorly matched to the low-level safety filter, and feeds this information back into planning rather than only correcting each command reactively. This is an interesting direction, since feasibility depends not only on the current QP but also on the trajectory induced by high-level decisions.
  2. A safety-conscious separation between high-level proposals and low-level control. The LLM does not directly output control inputs; it proposes bounded parameters for the nominal controller, while a deterministic evaluator selects among the candidates using CBF-QP rollouts and the CBF-QP continues to filter executed commands. The selector ablation further suggests that CBF-aware predictive selection, rather than candidate diversity alone, is an important component of the observed improvement over direct feedback.
Critical weaknesses
  1. Serious inconsistencies in the reported experimental results. Some results in Table 4/Figure 7 are incompatible with the paper’s own metric definitions. For example, Easy/Feedback reports 95.1% QP-feasible, 98.6% safe, and 91.2% certified-safe episodes, although certified-safe is their intersection and therefore must be at least 93.7%. Similarly, Easy/Predictive reports 100% safe episodes but different QP-feasible and certified-safe rates (99.2% vs. 99.1%). Appendix T explicitly identifies these contradictions, and Appendix V states that they cannot be reconciled without changing the preserved values. Seven binary rates in Table 4 also fall outside the 0.2-point grid implied by 500 episodes. In addition, Reactive CBF, which does not use an LLM, reports different Easy QP-feasibility results in Table 3 (17.0%) and Table 4 (23.1%) without a sufficiently clear explanation of what changed between these evaluations. Taken together, these issues substantially reduce my confidence in the reliability of the reported experimental evidence.
  2. The specific contribution of the LLM is not convincingly established. While some supplementary experiments show gains from LLM-generated candidates, these gains are not consistent across the matched-prior and control comparisons. In Table S65, for example, CEM and Grid achieve 61.2% and 59.8% Hard QP feasibility, respectively, compared with 59.0% for the LLM, while other matched comparisons show little or no significant advantage. This suggests that a substantial part of the improvement may come from the deterministic CBF-aware predictive evaluator rather than from LLM reasoning itself. Since the LLM is central to the framing of MARCH, stronger evidence isolating its contribution over simpler candidate-generation methods is needed.
Questions for authors

Clarifications

  1. Can the authors explain the inconsistencies in Table 4/Figure 7 identified in Appendices T and V, and clarify how these values were obtained? Can the affected results be regenerated directly from the underlying episode records? Also, why do the Reactive CBF results, which do not involve an LLM, differ between Table 3 and Table 4?
  2. What distinguishes the experimental settings where LLM-generated candidates show a clear advantage (e.g., Tables S64 and S66) from the matched-prior and control comparisons where this advantage largely disappears (e.g., Tables S41, S56, and S65)? In particular, is the improvement attributable to the LLM itself or to differences in the information available to the candidate generators?

A matched comparison between the LLM and a strong non-LLM generator such as CEM, using identical inputs, candidate budgets, and evaluation protocols on the main benchmark, would help isolate the marginal contribution of the LLM.

Suggestion: The paper could be better organized, particularly in the appendix, where related experiments and analyses could be grouped more systematically to make the results easier to follow and interpret.

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

The paper proposes an LLM-based planning approach for swarm control systems with bounded constraints. It proposes Predictive Harness-CBF that refines the high-level plan with short-horizon candidate evaluation. The paper also mathematically derived a feasibility audit for bounded-input dynamics. Experiments under 3-D swarm tasks demonstrate the effectiveness of the proposed method.

Critical strengths

The paper focuses on the important problem of safety within swarm systems. It connects a high-level planner LLM with a low-level controller through a CBF-QP-based architecture. The failure modes for different naive system designs are well established, serving as the motivations for the proposed method. The conducted 3-D simulation swarm tasks further demonstrate the effectiveness of the proposed method.

Critical weaknesses

I am not an expert in this area and am unable to assess the correctness of the mathematical derivations within the paper. To me, the general framework is reasonable, but I do have concerns especially on the predictive selection module.

  1. For each LLM candidate, the evaluator simulates a few steps ahead with a dynamics model and the same CBF-QP used at execution. The paper does not say whether this dynamics model is the simulator itself. If it is not, the mismatch between prediction and execution is never measured: there is no comparison of predicted and realized margins, interventions, or infeasibility, and no analysis of how often the score guard admits candidates that later become infeasible. If it is, the evaluator has near-oracle foresight, since the rollout reproduces almost exactly what will happen. The reported gains then show what selection achieves with perfect prediction, not with a realistic predictor, which flatters Predictive H-CBF against Reactive and Feedback H-CBF, which have no lookahead. Since all experiments are in simulation, this second case seems likely, and the sim-to-real gap (wind, delays, state-estimation noise, unmodeled quadrotor dynamics) remains untested. I suggest the authors clarify the rollout model and evaluate under model mismatch, for example by adding process noise, drag or actuation delay to the environment but not the evaluator.
  2. The candidates are scored with a fixed formula, with all parameters hand-set as a human heuristic. It is not sure whether these formulas would be robust to different settings and benchmarks. The paper reports no sensitivity analysis, so we cannot tell whether performance holds across a range of values or depends on a narrow choice. I suggest the authors report how the parameters were selected, include ablations over different numbers, and test on settings that differ more substantially, such as other swarm sizes, dynamics or constraint sets.
  3. I am not sure about this, but is simulation usually enough for this research area? Only real-world scenarios have safety issues, and many factors cannot be reproduced in a simulation environment. Also, the paper contains no related works, which is confusing. It also does not compare with methods rather than its own variants.
Questions for authors

Please see the above weaknesses. Beside them, I also have a practical concern toward this hierarchical paradigm. For drones, what I expected is a high-frequency response, especially for safety-critical settings. Using an LLM is very slow and not responsive. How frequently will the LLM be queried? If LLMs are frequently queried, the agents can only move slowly. What realistic tasks could the proposed method be targeted for?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.