acceptodds
Under review as a conference paper at ICLR 2027

Local-to-Collective Predictability from Heterogeneous Multi-Agent Interactions

Abstract

Collective behavior in multi-agent systems emerges from repeated local interactions, yet existing approaches lack an explicit representation of how heterogeneous local interaction changes propagate to collective-level degradation. To address this gap, this work introduces (), an agent-level representation of how strongly an agent's spatiotemporal interaction dynamics support prediction of its future motion. Building on this representation, learns \IP from ego-centric interaction histories and selectively aggregates agent-level predictability through learned relevance into a Collective Predictability Score (). Both representations are learned without using degradation labels or collective-state metrics as training targets. Experiments show that distinguishes heterogeneous local interaction conditions and achieves stronger degraded-agent localization than simple local-statistics baselines, while relevance-weighted captures collective degradation structure consistently across multiple interaction dynamics. These results establish local interaction predictability as a learned intermediate representation that connects distributed agent-level dynamics to collective degradation.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Multi-Agent Systems, Collective Behavior, Interaction Modeling, Representation 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_wgpaOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Local-to-Collective Predictability (L2CP), a framework for relating heterogeneous local interactions to collective degradation in multi-agent systems. The method learns an agent-level Interaction Predictability (IP) score from ego-centric spatiotemporal interaction histories and aggregates these scores through learned relevance weights into a Collective Predictability Score (CPS).

Both representations are trained through future-trajectory prediction and related auxiliary objectives without directly using degradation labels. Experiments on simulated Boids, Vicsek, and Cucker–Smale systems evaluate degraded-agent localization and the correlation between CPS and future collective degradation.

Critical strengths
  1. The manuscript is generally well organized, and the motivation and methodology are easy to follow. Understanding how heterogeneous local interaction relate to collective degradation is an interesting problem with potential applications in swarm monitoring and distributed multi-agent systems.

  2. The experiments examine agent-level degradation localization, collective degradation alignment, relevance-weighted aggregation, and generalization across different simulated interaction dynamics. This provides several complementary perspectives on the proposed representations rather than relying on a single evaluation metric.

Critical weaknesses

1. Limited methodological novelty and insufficient differentiation from existing representation-learning approaches. Although the paper introduces the concepts of Interaction Predictability (IP) and Collective Predictability Score (CPS), the underlying methodology primarily combines established components: interaction-history encoding, associative memory, MLP-based scalar prediction, attention-weighted aggregation, and trajectory-prediction supervision.

Specifically, IP is obtained by applying a sigmoid prediction head to learned agent representations, while CPS is a softmax-weighted average of individual IP scores. The trajectory decoder and associated learning objectives also largely follow conventional prediction-based representation learning.

While this combination is reasonable, I do not see a sufficiently distinct methodological innovation that explains why the proposed framework should be considered a substantial advance beyond existing interaction-aware representation learning or attention-based aggregation.

More importantly, the manuscript does not clearly establish what new information IP and CPS capture that cannot be obtained from existing learned trajectory representations, prediction errors, or uncertainty estimates. This limits the originality and broader methodological significance of the contribution.

2. The learning objective and experimental evaluation do not convincingly establish the claimed interpretation and advantages of IP/CPS.

The manuscript interprets IP as the degree to which local interaction histories support future-motion prediction. However, the proposed learning objectives do not sufficiently establish that the resulting scalar has this specific interpretation.

In particular, the predictive log-variance contains both independently learned variance estimates and additive IP/CPS corrections. Since the variance head can potentially compensate for these corrections, the NLL objective alone does not uniquely identify IP or CPS as measures of interaction predictability. The directional ranking objective further encourages IP to follow prediction-error ordering, making its distinction from a learned prediction-difficulty score unclear.

This issue also affects the interpretation of the learned relevance weights. While attention weights can improve aggregation, their learned values do not necessarily identify agents that are genuinely more influential in the evolution of collective degradation.

The experimental evidence is not sufficient to resolve these concerns. In Table 2, agent-level degradation localization achieves an AUROC of only 0.596, a relatively modest improvement over the simple statistical baselines. Moreover, the collective-level comparisons primarily use trajectory-prediction uncertainty rather than stronger architecture-matched or representation-matched alternatives.

I would particularly encourage comparisons with attention-pooled prediction errors, learned uncertainty-based scores, and similarly sized representation-learning models trained under the same prediction objective. These would help determine whether the reported improvements come from a genuinely useful predictability representation or simply from additional learned transformations and aggregation.

Finally, the experiments are restricted to simulated flocking dynamics with controlled degradation mechanisms. Although the evaluation across three simulators is useful, it provides limited evidence of generalization to more complex or realistic multi-agent interaction settings.

Questions for authors
  1. What distinguishes IP from a learned predictor of trajectory error or uncertainty, and CPS from standard attention-weighted aggregation of agent-level uncertainty estimates?

  2. Given that the learnable variance head can compensate for the IP/CPS terms in the Gaussian NLL, what guarantees that these scores acquire the intended predictability semantics? Do the authors have evidence that the representation remains stable or interpretable across different initializations?

  3. Could the authors clarify whether architecture-matched baselines using prediction errors, learned uncertainty, or standard attention pooling were evaluated? What evidence establishes that IP/CPS provides benefits beyond these simpler alternatives?

  4. Are there results evaluating L2CP under unseen degradation mechanisms or different agent population sizes, rather than only varying conditions within the same families of simulated interaction dynamics?

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

This paper proposes L2CP, a framework that connects local interaction histories with collective degradation in multi-agent systems. It learns an agent-level Interaction Predictability score and combines these scores through learned relevance weights into a Collective Predictability Score. Training uses future-motion prediction objectives. Experiments on three simulated collective dynamics examine degraded-agent localization, collective degradation alignment, and the contribution of different model components.

Critical strengths

The paper asks a useful question: can local interaction histories reveal changes that matter to the collective? Keeping both agent-level scores and a collective readout makes the framework more interpretable than reporting a single system-level anomaly score.

The experimental analysis goes beyond a single headline comparison. The authors report multiple dynamics, ablations, nine model seeds, and additional analyses separating within-sequence and between-sequence associations. These results help clarify where the method works and where its evidence is weaker.

Critical weaknesses
  1. The practical benefit over simple statistics is not established. In Table 12, mean neighbor count achieves correlations of 0.834, 0.779, and 0.581, exceeding L2CP’s 0.688, 0.600, and 0.512 on all three collective metrics. This comparison deserves attention in the main paper. Describing CPS as a representation rather than a degradation predictor does not fully resolve the issue, since degradation alignment is the main evidence used to validate that representation.
  2. The results provide limited evidence of anticipating changes within a collective. Table 11 reports within-sequence correlations of only 0.252, 0.228, and 0.148, substantially below the pooled results. Meanwhile, degradation parameters remain fixed within each sequence and are largely driven by a shared latent quality variable. CPS may therefore mainly distinguish persistent simulation conditions. Comparisons against persistence of the current collective state, together with time-varying disturbances, would make the predictive interpretation more convincing.
  3. The distinction between interaction predictability and estimated prediction difficulty remains unclear. IP is explicitly ranked using trajectory errors and coupled to predictive variance. However, the experiments do not compare it with a matched model that directly estimates future prediction error or uncertainty from the same interaction history. The localization result is also modest: AUROC is 0.596, compared with 0.571 for velocity disagreement. Since most local baselines are instantaneous, this does not yet isolate the benefit of the proposed representation from the benefit of using temporal information.
  4. Degradation information enters hyperparameter selection. Appendix D.9 states that the coefficients are selected using validation collective-degradation alignment. Although degradation metrics are excluded from gradient targets, they still influence the final model. This distinction should be explicit in the main paper. Table 15 also shows substantial sensitivity to nearby settings, making results selected using prediction-only criteria particularly relevant.
Questions for authors

What happens when hyperparameters are selected solely by validation trajectory prediction performance? Were the trajectory baselines given comparable tuning opportunities?

How much predictive value does CPS add after accounting for the current collective state and simple local statistics? A persistence baseline would be especially informative here.

With the same encoder and interaction memory, how does IP compare with directly predicting future trajectory error or uncertainty? Please also report uncertainty estimates for the localization AUROC differences.

Reviewer_nVb9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes L2CP. It learns a per-agent Interaction Predictability (IP) score from ego-centric interaction histories and aggregates the scores into a Collective Predictability Score (CPS) via softmax relevance weights. Training uses only future-trajectory objectives: MSE, a Gaussian NLL whose variance is coupled to IP and CPS, and a ranking loss that orders IP by agent trajectory error. No degradation labels are used as training targets. The method is evaluated on synthetic Boids swarms (20 agents) with sensing delay, motion perturbation and field-of-view degradation, plus Vicsek and Cucker–Smale dynamics. 1−IP localizes degraded agents with AUROC 0.596, against 0.512–0.571 for local statistics. Relevance-selected agent subsets recover CPS–degradation alignment faster than random subsets. 1−CPS correlates with future fragmentation and connectivity loss (Spearman 0.688/0.600) more strongly than sample variance from four trajectory-prediction models, and is comparable to AgentFormer on order loss.

Critical strengths

Relevant problem with a clean training setup. Linking heterogeneous local interaction changes to collective-level breakdown is an important question for swarm and multi-robot systems. The method learns its representations from future-motion prediction alone, and the controlled simulator degrades interactions (delay, heading noise, field of view) without changing the underlying interaction policy.

Critical weaknesses

Baseline selection does not support the main claims. The collective-level comparison (Table 3) uses only uncertainty from trajectory predictors, which were not designed to anticipate collective degradation. The method is not compared against the approaches the paper itself positions against: collective-state monitoring and fault/anomaly detection. Simple aggregate local statistics, a persistence baseline, and a degradation-supervised upper bound are also absent. Several main results have no comparison method at all: the severity sweep (Fig. 5), the cross-dynamics evaluation (Table 4), and the relevance analysis (Fig. 4), which compares only against random selection, a near-trivial baseline given CPS = Σαₙ·IPₙ.

Questions for authors
  1. Comparisons for the remaining main results. Could you add comparison methods to the severity sweep (Fig. 5) and the cross-dynamics evaluation (Table 4), for instance the same local-statistic baselines? For Fig. 4, how does relevance-based selection compare with non-trivial selection rules, such as ranking agents by IP or by local statistics, rather than random selection only?
  2. Baselines for collective degradation. Why were only trajectory-prediction uncertainties chosen as baselines in Table 3, given that Sec. 2 positions L2CP against collective-state monitoring and fault/anomaly detection? Could you add the following to the main comparison? (a) Representative methods from the two directions the paper positions against in Sec. 2: collective-state monitoring and multi-agent fault/anomaly detection. (b) Simple aggregate local statistics computed from the same ego-centric observations. Appendix Table 12 suggests mean neighbor count already exceeds 1−CPS on all three metrics. (c) A model supervised directly on degradation, as an upper bound. Please also be clear about whether the baselines receive the same observation history as L2CP, and explain why Trajectron++ correlates negatively on all three metrics. I would like to adjust my scorer if authors answered my questions properly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.