acceptodds
Under review as a conference paper at ICLR 2027

LINC: Label-Free In-Context Network Reconfiguration under Unobserved Failure-Mode Switches

Abstract

Policies trained in a single operating regime often fail when the environment changes without warning—the analogue of a robot that cannot recover after a broken leg, or a network that keeps repairing hubs after the attack has moved to a region. We study this problem as in-context meta-reinforcement learning in a hidden-parameter (contextual) MDP whose latent parameter—the failure mode—is unobserved and may switch inside an episode. We propose LINC, which infers a label-free context from a short window of transition tokens by self-supervised next-step prediction with variance–covariance regularization, and conditions a PPO policy on the detached context together with a size-weighted bridging prior. At test time the agent adapts entirely in-context, without parameter updates or access to task labels or switch signals. On synthetic network-reconfiguration tasks, LINC matches strong contextual baselines when the mode is held fixed, recovers faster and with a shallower transient after unobserved switches, and degrades most gracefully under out-of-distribution mixtures, intensities, and scales.

Primary area:
reinforcement learning
Keywords:
in-context reinforcement learning, hidden-parameter MDP, non-stationary environments, self-supervised representation learning, graph neural networks, network reconfiguration

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wabDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies online adaptation in network reconfiguration when the failure mode is a hidden parameter that can switch within an episode. The environments are synthetic graphs with 20–80 nodes (BA, ER, WS), and there are three failure modes, namely hub attack, regional failure and random failure. At each step the agent adds one edge, and the reward is the increment in network capability Φ\Phi, defined as the average of normalized algebraic connectivity and giant-component fraction.

Critical strengths

The problem is well motivated, since unlabeled failure-mode switches within an episode occur in real infrastructure networks. The comparison is fair in the sense that all learned baselines share the graph encoder, edge-scoring head and transition tokens and differ only in the context pathway. Results use 10 seeds on the same 20 test episodes per mode, and recovery time handles right-censoring with its bias stated openly. The OOD sweeps cover mixture, intensity and scale. The appendix gives full hyperparameters and code is provided. The limitations section is candid about the single switch schedule, the missing specialist bound and the unisolated contribution of the prior.

Critical weaknesses

W1. All ablations and the HH sweep (Section 5.4) use stationary episodes, so they say nothing about switch recovery. The claim that H=40H=40 is worse because stale tokens delay turnover cannot be tested without switches. It is also unclear whether PPO-GRU uses the same H=20H=20 window.

W2. A baseline for memoryless PPO is missing. A true-mode oracle and per-mode specialists are also missing, so the paper never shows that the best repair differs across modes. Finally, the paper does not say whether the baselines use the bridging prior. If they do not, LINC's advantage at N=200N=200 may come from the prior, since Prior-only already beats every other learned baseline there.

W3. The learned baselines are PEARL (2019), ESCP (2022) and PPO-GRU. The authors name SeCBAD (Chen et al., 2022) and Ren et al. (2022) as closest to their setting but do not compare with them, although SeCBAD targets exactly piecewise-stable contexts with abrupt intra-episode changes. MoSS (Wang et al., 2023) and TIGR (IEEE TNNLS), which also learn self-supervised contexts for non-stationary tasks, are missing too.

W4. The components come from CaDM, VICReg, ESCP and Stooke et al. (2021), and the authors state that predictive context learning is not their contribution. Therefore, precise ablation of the contribution of each component should be provided.

Questions for authors

Q1. Do ESCP, PEARL and PPO-GRU use the bridging prior? How does Figure 4(c) change if all baselines use the prior, or if LINC uses β=0\beta=0?

Other questions please refer to the weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer wabD

We sincerely thank the reviewer for the careful and detailed review.

W1. Thank you for raising this point. Because switch patterns cannot be enumerated at training time (cf. Luo et al., 2022), we trained on stationary episodes, so the curves in Figure 5 reflect stationary training only. We are now evaluating the four ablations and the H sweep on the switching episodes of Section 5.2 (General Response, point 3). We also agree that the stale-token explanation for H = 40 is not supported by the current evidence, and we will revise Section 5.4 according to the switching results. Regarding PPO-GRU: it uses the same H = 20 window and transition tokens as LINC, and differs only in that its encoder is trained end-to-end by the PPO loss, without Ldyn\mathcal{L}_{\mathrm{dyn}}, LVC\mathcal{L}_{\mathrm{VC}}, or stop-gradient (Appendix C). We will state this explicitly in Section 5.1.

W2. The memoryless PPO baseline corresponds to the w/o Context ablation in Figure 5, which uses the same encoder, head, and prior but no history; its switching results are included in the evaluation above. Regarding the prior, all baselines use it (General Response, point 1). We agree that showing whether the best repair differs across modes is important, and we will follow up on the oracle and per-mode specialists in this thread.

W3. We thank the reviewer for these references. (i) SeCBAD: we are implementing SeCBAD in our shared framework, keeping the graph encoder, edge-scoring head, prior, and PPO budget fixed and replacing only the context pathway. Because SeCBAD is trained on switching episodes with access to the true context (Chen et al., 2022, Appendix A.6), whereas LINC is trained on stationary episodes without labels, we will report SeCBAD under both training protocols. (ii) MoSS and TIGR: both learn the context with a Gaussian-mixture VAE and are trained with SAC for continuous control (Wang et al., 2023; Bing et al.), whereas all compared methods here use PPO over a variable set of candidate edges. To compare their variational (KL-regularized) context objective under the shared framework, we are adding a KL-regularized variant of LINC (see also our response to Reviewer wNR9, W2). (iii) Ren et al. (2022) also model the switching process explicitly. As the reviewer notes, SeCBAD targets exactly piecewise-stable contexts with abrupt intra-episode changes, and, like LINC, it is trained with PPO; we therefore take SeCBAD as the representative of this class. We will cite TIGR in the revised related work.

W4. We agree that each component should be isolated. The context input, Ldyn\mathcal{L}_{\mathrm{dyn}}, LVC\mathcal{L}_{\mathrm{VC}}, the prior, and the window H are ablated in Figure 5, and these ablations are now being evaluated on switching episodes (W1). The stop-gradient follows the decoupling of representation learning from RL (Stooke et al., 2021): the context encoder is trained only by the prediction loss, so that the policy gradient does not reshape the task-inference representation (Section 4.1). It was not ablated in the submission (Section 6); we are now training a variant without the stop-gradient and will report it if it completes within the discussion period.

Q1. Yes (General Response, point 1). We are adding LINC with β = 0 to the sweep of Figure 4(c).

We hope these responses address your concerns, and we would be happy to answer any further questions.

AuthorsOpenReview ↗

Response to Reviewer wabD (new results)

We thank the reviewer again for the detailed suggestions. The first results are now in the General Response (Update: New Results, 1/2 and 2/2).

W1. On the switching episodes of Section 5.2, against full LINC trained with the same budget and seeds, removing the context input or Ldyn\mathcal{L}_{\mathrm{dyn}} roughly doubles the recovery time, and removing the prior or LVC\mathcal{L}_{\mathrm{VC}} has smaller effects (Table R2). The recovery time increases monotonically with H (0.68, 0.91, 1.00, and 1.47 times that of H = 20 for H = 5, 10, 20, and 40), and the ordering H = 5 < 10 < 40 holds in all 16 further switching settings. This directly tests the stale-token explanation, and Section 5.4 now cites this result.

W2. The memoryless variant (w/o Context) reaches a post-switch capability of 0.769 ± 0.050, versus 0.894 ± 0.020 for full LINC, and takes 2.33 times as long to recover (Table R2).

Q1. Removing the prior lowers AUC@50 by at most 0.039 on the three OOD sweeps, and its contribution at N = 200 is similar to that within the training range (Table R3).

SeCBAD and the KL-regularized variant (W3) and the stop-gradient ablation (W4) are in progress, and we will post them in this thread. We hope these results address your concerns.

Reviewer_wNR9OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes an in-context adaptation approach for network reconfiguration under unobserved failure-mode switches. The method uses a context encoder to infer a latent representation from recent graph transitions, trained through self-supervised prediction with variance-covariance regularization. A PPO policy is then conditioned on this context and a graph-specific structural prior. Experiments on synthetic network environments demonstrate improved adaptation to failure-mode switches and robustness under OOD shifts compared with several contextual RL baselines.

Critical strengths
  1. Clear formulation and promising empirical results. The method follows a coherent design, with several components validated through ablation studies. In particular, LINC achieves faster recovery following unobserved failure-mode switches and better performance under the evaluated OOD shifts.
  2. Application of in-context adaptation to network reconfiguration. The paper extends established ideas from contextual RL to network reconfiguration, combined with domain knowledge, addresses the challenge of adapting to unobserved changes in network failure dynamics.
Critical weaknesses
  1. Limited evidence for generalizability beyond the network-reconfiguration setting. The main contributions appear to lie in the graph-specific transition representation and structural prior, whose applicability to other adaptation problems remains unclear. Even within network domains, the experiments consider only synthetic graph families and a single-edge repair action. Furthermore, the contributions of the handcrafted node features and structural prior to OOD generalization are not fully isolated, making it difficult to distinguish generalizable methodological insights from domain-specific design choices.
  2. Insufficient justification and analysis of the proposed components. The rationale for certain design choices could (and should) be developed further. For example, the paper provides limited analysis of why combining dynamics prediction with VICReg regularization is preferable to alternative representation-learning objectives. The choice itself resembles an encoder-decoder architecture, and some common regularizers in this case are sparsity, KL constrained prior and discretization [1].
  3. Restricted experimental evaluation of adaptation. Failure-mode switching is evaluated using only one fixed sequence (hub attack → regional failure → random failure), with predetermined switch times. The OOD evaluations additionally focus on stationary episodes rather than switching environments. Consequently, the experiments do not fully establish robustness to arbitrary switching patterns, and it is hard to evaluate the richness and generalizability of the learned context.
  4. Limited methodological novelty relative to existing contextual RL approaches. The context encoder and context-conditioned policy formulation have been extensively studied in RL, for example in the dynamics adaptation community [2,3,4]. Although the authors acknowledge these connections (line 115-116), the proposed predictive objective (Eq. 3 and 4) largely combine established techniques. Their empirical effectiveness is supported by the ablations, but the methodological contribution beyond adapting existing techniques to network reconfiguration remains unclear.

References

[1] Garrido, Quentin, et al. "Learning latent action world models in the wild." ICML 2026.

[2] Kumar, Ashish, et al. "RMA: Rapid Motor Adaptation for Legged Robots." RSS 2021.

[3] Lee, Kimin, et al. "Context-Aware Dynamics Model for Generalization in Model-Based Reinforcement Learning." ICML 2020.

[4] Xu, Zhiming, et al. "Dynamics Are Learned, Not Told: Semi-Supervised Discovery of Latent Dynamics Geometries for Zero-Shot Policy Adaptation." ICML 2026.

Questions for authors
  1. Could the authors clarify the key methodological contributions relative to existing self-supervised context-learning approaches, for example CaDM [3] ? A comparison result using CaDM objective would be nice, but a mathematical clarification of why the current Eq. 3 & 4 are not merely engineering choices would also be fine.
  2. Which aspects of the proposed method are expected to generalize beyond network reconfiguration? Provide either theoretical or empirical results to support your claim.
  3. Could the authors provide ablations measuring the contributions of individual components to switch recovery, rather than only stationary training performance (i.e., extend the result of Fig.5 left) ?
1 comment
AuthorsOpenReview ↗

Response to Reviewer wNR9

We sincerely thank the reviewer for the detailed review and the constructive suggestions.

W1 (Limited evidence for generalizability). Thank you for this comment. Our scope is network reconfiguration under unobserved failure-mode switches, a setting in which existing learning-based reconfiguration methods assume a fixed failure regime within an episode (Section 2). We do not claim applicability beyond this domain, and we will revise the first sentence of the abstract accordingly. Within this domain, real topologies are being added (response to Reviewer hfLF, W1); the single-edge action follows the edge-addition formulation of Darvariu et al. (2021); and the four node features are chosen by the failure mechanism (Albert et al., 2000; Section 4.2). The node features and the prior are shared by all learned methods (General Response, point 1), and the prior's own contribution to OOD generalization is being evaluated with β = 0.

W2 (Insufficient justification of the components). The regularizers discussed by Garrido et al. [1] limit the information in a latent inferred from (st,st+1)(s_t, s_{t+1}), which could otherwise encode the next state. In LINC, ztz_t is inferred only from transitions before t (Eq. 1), so the prediction target is not an input to the encoder, and the relevant failure is collapse rather than leakage (Section 4.3); the variance term of VICReg (Bardes et al., 2022) is designed to prevent collapse. Removing LVC\mathcal{L}_{\mathrm{VC}} lowers the final episode-average Φ from about 0.81 to 0.73 (Figure 5). We are adding a KL-regularized variant for a direct comparison.

W3 (Restricted experimental evaluation of adaptation). For the switching schedules, please see General Response, point 2. We are also applying the scale and intensity shifts of Section 5.3 to switching episodes.

W4 (Limited methodological novelty). As stated in Section 2, predictive context learning itself is not our claim. Unlike RMA (Kumar et al., 2021), which regresses onto privileged environment parameters, and Xu et al. (2026), whose contrastive positives are segments generated under the same dynamics parameters, LINC uses no mode label or grouping during training. Our contribution is the label-free formulation of intra-episode failure switches in network reconfiguration and the graph-specific design for it (Section 2).

Q1. Thank you for this question. Let x=(gˉt,αt)x = (\bar g_t, \alpha_t) and y=(gˉt+1,rt)y = (\bar g_{t+1}, r_t). For the squared loss in Eq. 3, the optimal readout satisfies

E∥y−E[y∣x,z]∥2=E∥y−E[y∣x]∥2−E∥E[y∣x,z]−E[y∣x]∥2.\mathbb{E}\Vert y - \mathbb{E}[y \mid x, z]\Vert ^2 = \mathbb{E}\Vert y - \mathbb{E}[y \mid x]\Vert ^2 - \mathbb{E}\Vert \mathbb{E}[y \mid x, z] - \mathbb{E}[y \mid x]\Vert ^2 .

The context lowers the loss only through the last term; when the one-step conditional means differ little across modes, a near-constant ztz_t is nearly optimal for Eq. 3 alone, and the variance term in Eq. 4 excludes such solutions. CaDM (Lee et al., 2020) predicts M∈{5,10}M \in \lbrace 5, 10 \rbrace future steps with one context, which assumes that the context is constant over that horizon; under intra-episode switches, the horizon can cross a switch. LINC predicts a single step, which does not require this assumption.

Q2. The transition token and the bridging prior are specific to graphs. The context objective (Eqs. 3–4) uses only transitions, and the argument in Q1 does not depend on the graph structure. We have no empirical results outside network domains and make no claim there.

Q3. Yes; please see General Response, point 3.

We hope these responses address your concerns, and we would be happy to answer any further questions.

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

The paper investigates an agent repairing a damaged graph network by adding a single edge at each timestep. The network is being damaged by an unknown failure type (hub attack, regional outage, random failure); the paper analyzes meta-RL methods for adapting to the updated failure type by learning to identify and update the strategy based on previous timesteps.

The method, called LINC: The agent is represented as a graph neural network encoder learning representations per network; a recurrent network (GRU) sees the last 20 transitions and learns a representation zz predicting only the next graph state and reward, which then conditions an RL-learned policy. The policy is also conditioned by a hand-written prior which favors edges connecting large components of the graph.

LINC is shown to be as strong as standard baselines when dealing with a fixed failure type, and has the smallest drop in performance when dealing with fixed failure switches (recovering within ~9 steps, twice as fast as the next-strongest baseline). Ablations show that context input is the most important component of the method, and the authors also analyze the information encoded in the context.

Critical strengths
  1. The results on the synthetic datasets are convincing, especially over strong baselines like ESCP and PEARL, on failure mode switches and distributional shifts.
  2. The ablations in Fig5 are generally quite careful and demonstrate each of the components are useful during training. Fig 6 / section 5.5 is quite useful analysis to confirm the general intuition that the context captures meaningful damage modes.
Critical weaknesses
  1. The method is tested on synthetic tasks with small network sizes (trained on up to 80 nodes, tested on up to 200 nodes). A single result on a real-world dataset would help alleviate concerns about generality to real-world networks.
  2. The switching episodes vary in a specific order with long and fixed dwell times. These results are not yet shown to generalize to more realistic scenarios; while the authors state this limitation, the takeaway of the paper would be greatly strengthened by results outside of this setting.
Questions for authors
  1. Can you clarify whether the prior is also added to the baselines? If not, would it be possible to see how the baselines work with the prior added (ESCP, PEARL, and PPO-GRU receive the same prior with beta=0.2) and/or LINC without the prior, to isolate the prior's effect on the large-scale result in Figure 4c?
  2. Would it be possible to see a plot which randomizes the damage mode order and switch times (including shorter dwell times)? This will help us understand whether the results are robust to these parameters and potentially over more realistic graphs. (Even better if we could see it on a non-synthetic dataset, of course.)
4 comments
AuthorsOpenReview ↗

Response to Reviewer hfLF

We sincerely thank the reviewer for the positive assessment and the helpful suggestions.

W1 (synthetic networks). We trained on BA, ER, and WS graphs; the BA and WS generators reproduce the heavy-tailed degree distributions and small-world clustering observed in real infrastructure networks (Barabási & Albert, 1999; Watts & Strogatz, 1998; Section 5.1). We agree that a result on real networks would strengthen the paper, and we are evaluating the trained policies without retraining on the IEEE 118-bus power grid and a communication backbone from the Internet Topology Zoo, under the same failure processes.

W2 / Q2 (mode order, switch times, and dwell times). Thank you for this suggestion. We are evaluating all six mode orders, random switch times, and dwell times of 10, 20, and 40 steps (General Response, point 2), and will provide the requested plot.

Q1 (prior in the baselines). Yes, ESCP, PEARL, and PPO-GRU use the same prior with β = 0.2 (General Response, point 1), so Figure 4(c) already compares all methods with the prior. For completeness, we will also report LINC with β = 0 on the same scale sweep.

We hope these responses address your concerns, and we would be happy to answer any further questions.

AuthorsOpenReview ↗

Response to Reviewer hfLF (new results)

We thank the reviewer again for the suggestions. The requested results are now in the General Response (Update: New Results, 1/2 and 2/2).

W1. We evaluated all methods without retraining on the IEEE 118-bus power system and the Deltacom backbone from the Internet Topology Zoo, under the same failure processes (Table R5). Among the learned methods, LINC attains the highest value in all three stationary failure modes and under switching on both graphs, with margins of 0.032 to 0.043 over the strongest baseline. Prior-only is the highest on Deltacom under hub attacks, and we report this in the appendix.

W2 / Q2. Across all six mode orders, random switch times, and dwell times of 40, 20, and 10 steps, LINC attains the highest post-switch capability, and its margin over the strongest baseline grows as the dwell time decreases (Table R1). The requested plot is added as Figure 7 (Appendix D). The switching schedule is also evaluated on the two real topologies (Table R5).

Q1. Removing the prior lowers AUC@50 by 0.025 at N = 200, which is similar to its effect within the training range (0.016 and 0.021 at N ≤ 80). The prior therefore does not account for the large-scale result in Figure 4(c) (Table R3).

We hope these results address your questions.

I appreciate the larger-scale evaluation on real networks.

I think prior-only is still quite strong, e.g. in Table 7. It would be helpful to have Prior-only in Table 3.

I also don't understand how Figure 3 and Table 3 can be consistent, e.g. LINC at .903, ESCP and PEARL tying at .861 don't look like the values in Figure 3.

I will maintain my rating of weak acceptance.

AuthorsOpenReview ↗

Response to Reviewer hfLF (Prior-only in Table 3; Figure 3 vs.

We thank the reviewer for the follow-up and for maintaining the positive assessment.

Prior-only in Table 3. Below is Prior-only on the same switching episodes as Table 3 (post-switch capability; Prior-only involves no training, so there is no std over seeds). We will add this column to Table 3.

SettingLINCPrior-only
Schedule of Section 5.20.9030.673
All six mode orders0.8960.677
Random switch times0.8970.687
Dwell time 40 / 20 / 100.890 / 0.840 / 0.7590.672 / 0.623 / 0.548
N = 80 / 120 / 160 / 2000.885 / 0.862 / 0.837 / 0.8120.677 / 0.669 / 0.666 / 0.660
Damage cap 0.35 / 0.450.804 / 0.6940.563 / 0.463

LINC is above Prior-only in all twelve settings, by 0.152 to 0.241. Prior-only exceeds PPO-GRU at N = 160 and 200 and at dwell time 10, and PEARL at N = 200, and is below all learned methods elsewhere. Under switching, the gap between LINC and Prior-only is larger than under the stationary failures on the real topologies (Table 7).

Figure 3 and Table 3. The two differ in the models, the episodes, and the metric. (i) Table 3 is evaluated with models retrained for the rebuttal with the same configuration and training budget, on a newly generated set of switching episodes. As stated in Section 5.1, the stationary, switching, and distribution-shift evaluations use separately generated episode sets, so absolute values are not comparable across tables and figures; the training-free Prior-only is also higher on these episodes (0.673, versus about 0.62 in Figure 3(a)). (ii) Figure 3(b, c) reports the maximum drop and the recovery time, whereas Table 3 reports the post-switch capability, the mean of Φt\Phi_t over the 40 steps after each switch, which also includes the recovered part of the window. On the episodes of Table 3, the orderings are the same as in Figure 3(b, c): LINC has the smallest drop among the learned methods, and the recovery time orders LINC < ESCP < PEARL < PPO-GRU. The gap between ESCP and PEARL is smaller on these episodes (recovery times 1.62 and 1.77 times that of LINC), which is why the two tie in post-switch capability. We will state the models, the episode set, and the metric explicitly in the caption of Table 3.

We thank the reviewer again for the careful reading.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their time and constructive feedback. We first address the points raised by more than one reviewer; individual responses are posted under each review, and new results will be added to this thread as they become available.

1. Prior and history window in the baselines (hfLF, wabD, wNR9). All learned baselines (ESCP, PEARL, PPO-GRU) use the same bridging prior with β = 0.2 and the same H = 20 window as LINC; this is implemented in the base agent class shared by all learned agents (supplementary code). Together with the shared graph encoder, edge-scoring head, and transition tokens noted by Reviewer wabD, the methods differ only in the context pathway. We apologize for not stating this explicitly and will do so in the revised Section 5.1. To isolate the contribution of the prior itself, we are evaluating LINC trained with β = 0 on the three OOD sweeps of Section 5.3, against full LINC trained with the same budget and seeds.

2. Switching schedules (hfLF, wNR9). In addition to the schedule of Section 5.2 (80-step dwell), we are evaluating all methods with (i) all six mode orders, (ii) randomly sampled switch times, and (iii) dwell times of 10, 20, and 40 steps, i.e., shorter than, equal to, and twice the history window H.

3. Ablations on switching episodes (wabD, wNR9). Since switch patterns cannot be enumerated at training time (cf. Luo et al., 2022), we trained on stationary episodes, and Figure 5 accordingly reports training curves. We are evaluating the four ablations and the H sweep on the switching episodes of Section 5.2, against full LINC trained with the same budget and seeds.

Further experiments in progress: evaluation on real infrastructure topologies without retraining (hfLF, wNR9); SeCBAD as an additional baseline (wabD); a KL-regularized context variant (wabD, wNR9); an ablation of the stop-gradient (wabD).

All changes in the revised manuscript will be highlighted in blue.

AuthorsOpenReview ↗

Title: Update: New Results (1/2)

We thank the reviewers for their patience. Below and in the follow-up post (2/2) we report the first set of new results; the remaining experiments (SeCBAD, the KL-regularized variant, and the stop-gradient ablation) will be posted in this thread as they complete. All methods are evaluated on the same test tasks (graphs and switch schedules), with 30 episodes per setting and seed unless stated otherwise; we report mean ± std over seeds.

For the switching experiments we report post-switch capability, the mean of Φt\Phi_t over the window after each switch (40 steps, or the dwell time for the dwell-time schedules), averaged over all switches in an episode. It reflects both the depth of the drop and the speed of recovery, and it remains defined when the dwell time is shorter than the recovery horizon.

Table R1. Switching schedules (hfLF, wNR9). Post-switch capability, 10 seeds.

SettingLINCESCPPEARLPPO-GRU
Schedule of Section 5.20.903 ± 0.0160.861 ± 0.0170.861 ± 0.0220.791 ± 0.037
All six mode orders0.896 ± 0.0140.856 ± 0.0240.839 ± 0.0310.753 ± 0.031
Random switch times0.897 ± 0.0110.861 ± 0.0210.837 ± 0.0220.755 ± 0.034
Dwell time 400.890 ± 0.0160.851 ± 0.0240.811 ± 0.0270.743 ± 0.042
Dwell time 200.840 ± 0.0160.781 ± 0.0210.706 ± 0.0280.642 ± 0.047
Dwell time 100.759 ± 0.0210.667 ± 0.0270.635 ± 0.0330.539 ± 0.045
Switching, N = 800.885 ± 0.0120.845 ± 0.0240.822 ± 0.0280.741 ± 0.047
Switching, N = 1200.862 ± 0.0170.799 ± 0.0240.771 ± 0.0290.677 ± 0.045
Switching, N = 1600.837 ± 0.0120.748 ± 0.0240.708 ± 0.0280.610 ± 0.045
Switching, N = 2000.812 ± 0.0180.712 ± 0.0210.646 ± 0.0260.543 ± 0.047
Switching, damage cap 0.350.804 ± 0.0230.725 ± 0.0250.717 ± 0.0420.617 ± 0.052
Switching, damage cap 0.450.694 ± 0.0230.614 ± 0.0260.597 ± 0.0450.498 ± 0.051

LINC attains the highest post-switch capability in every setting; for each of the six mode orders, its margin over the strongest baseline lies between 0.031 and 0.047. The margin over ESCP increases as the dwell time decreases (0.039, 0.059, and 0.092 for dwell times of 40, 20, and 10 steps) and as the network grows (0.040 at N = 80 and 0.100 at N = 200).

Table R2. Ablations on switching episodes (wabD, wNR9). Schedule of Section 5.2, 8 seeds; all models are trained with the same budget and seeds as full LINC (H = 20). Recovery time is given relative to full LINC.

ModelPost-switch capabilityRelative recovery time
Full LINC0.894 ± 0.0201.00
w/o Context (memoryless PPO)0.769 ± 0.0502.33
w/o Ldyn\mathcal{L}_{\mathrm{dyn}}0.801 ± 0.0251.95
w/o Prior (β = 0)0.851 ± 0.0241.11
w/o LVC\mathcal{L}_{\mathrm{VC}}0.882 ± 0.0141.16

Removing the context input or Ldyn\mathcal{L}_{\mathrm{dyn}} lowers post-switch capability by 0.125 and 0.093 and roughly doubles the recovery time. Removing the prior or LVC\mathcal{L}_{\mathrm{VC}} has smaller effects (0.043 and 0.012; the latter is within one standard deviation).

H sweep (wabD, wNR9). On the same episodes, the recovery time relative to H = 20 is 0.68, 0.91, 1.00, and 1.47 for H = 5, 10, 20, and 40, increasing monotonically with H. The ordering H = 5 < 10 < 40 also holds in all 16 further switching settings (each of the six mode orders and the other ten rows of Table R1). This supports that a longer window delays turnover after a switch, the mechanism stated in Section 5.4, which now cites this result.

Tables R1 and R2 are added to Appendices D and E of the revised manuscript (Tables 3 and 4); all changes in the revised manuscript are highlighted in blue. The results on the prior, the context statistics, and real topologies follow in the next post (2/2).

AuthorsOpenReview ↗

Title: Update: New Results (2/2)

Continuing from the previous post (1/2), with the same evaluation protocol.

Table R3. The bridging prior (hfLF, wabD, wNR9). Decrease in AUC@50 when the prior is removed (full LINC minus LINC trained with β = 0), 8 seeds, on the OOD sweeps of Section 5.3.

SweepDecrease in AUC@50
Network scale, N = 40 / 80 / 120 / 160 / 2000.016 / 0.021 / 0.022 / 0.011 / 0.025
Attack mixture, α = 0 to 10.026 – 0.039
Attack intensity, damage cap 0.15 to 0.550.027 – 0.035

Removing the prior lowers AUC@50 by at most 0.039, and its contribution at N = 200 (0.025) is similar to that at N ≤ 80 (0.016 and 0.021), so the prior does not explain the large-scale behavior of LINC. Since all learned methods share the prior (General Response, point 1), their differences lie in the context pathway. We have revised the two sentences in Sections 5.3 and 5.4 accordingly.

Table R4. Context statistics (wNR9). Statistics of ztz_t over stationary episodes of each failure mode (16 dimensions, 8 seeds).

ModelMean per-dimension stdDimensions with std < 0.01Between-mode share of variance
Full LINC0.46800.465
w/o LVC\mathcal{L}_{\mathrm{VC}}0.1761.10.431
w/o Ldyn\mathcal{L}_{\mathrm{dyn}}0.50700.094

Without LVC\mathcal{L}_{\mathrm{VC}}, the per-dimension std falls below 40% of that of full LINC and some dimensions become nearly constant (partial collapse); without Ldyn\mathcal{L}_{\mathrm{dyn}}, the context varies but carries little information about the failure mode (cf. our response to Q1 of Reviewer wNR9).

Table R5. Real infrastructure topologies (hfLF, wNR9). Trained policies, without retraining, on the IEEE 118-bus power system (118 nodes, 179 edges) and the Deltacom backbone (113 nodes, 161 edges), the Internet Topology Zoo network with 100 to 200 nodes whose mean degree is closest to that of the training graphs. Only the healthy graph is replaced; failures remain simulated. Stationary: AUC@50, 10 episodes per mode and seed. Switching: post-switch capability, schedule of Section 5.2. Learned methods: 10 seeds; Prior-only involves no training.

GraphSettingLINCESCPPEARLPPO-GRUPrior-only
IEEE 118Hub attack0.848 ± 0.0210.809 ± 0.0130.770 ± 0.0310.744 ± 0.0500.837
IEEE 118Regional failure0.577 ± 0.0370.538 ± 0.0190.498 ± 0.0230.470 ± 0.0500.550
IEEE 118Random failure0.830 ± 0.0180.787 ± 0.0210.754 ± 0.0360.719 ± 0.0480.795
IEEE 118Switching0.874 ± 0.0360.842 ± 0.0250.792 ± 0.0300.673 ± 0.0410.696
DeltacomHub attack0.713 ± 0.0420.674 ± 0.0130.639 ± 0.0310.609 ± 0.0510.759
DeltacomRegional failure0.592 ± 0.0290.553 ± 0.0240.516 ± 0.0300.486 ± 0.0550.548
DeltacomRandom failure0.754 ± 0.0450.715 ± 0.0200.672 ± 0.0400.644 ± 0.0480.710
DeltacomSwitching0.650 ± 0.0320.617 ± 0.0290.565 ± 0.0310.446 ± 0.0370.609

Among the learned methods, LINC is the highest in all eight settings, with margins of 0.032 to 0.043 over the strongest baseline. Prior-only is strong on both graphs: LINC is above it in seven settings (in three of them within one standard deviation) and below it on Deltacom under hub attacks, where the margin of LINC over the learned baselines (0.039) is similar to the other settings.

Tables R3–R5 are added to Appendices F–H of the revised manuscript (Tables 5, 6, and 7); all changes in the revised manuscript are highlighted in blue.

AuthorsOpenReview ↗

Summary of the Discussion for the Area Chair

Dear Area Chair,

We thank you and the reviewers for the discussion. We summarize the main concerns and how they were addressed; all new results are in the General Response (Update: New Results, 1/2 and 2/2) and Appendices D–H of the revised manuscript, with changes highlighted in blue.

Concerns shared by reviewers.

  • Prior in the baselines (hfLF, wabD, wNR9). All learned baselines use the same prior and window as LINC. Removing the prior from LINC lowers AUC@50 by at most 0.039, with a similar contribution at every network scale, so the prior does not explain the large-scale results (Table R3).
  • Single switching schedule (hfLF, wNR9). LINC attains the highest post-switch capability under all six mode orders, random switch times, dwell times of 40, 20, and 10 steps, networks of up to N = 200, and damage caps of up to 0.45; its margin over ESCP grows as the dwell time decreases and as the network grows (Table R1).
  • Ablations on switching episodes (wabD, wNR9). Removing the context input or Ldyn\mathcal{L}_{\mathrm{dyn}} roughly doubles the recovery time; the recovery time increases monotonically with H (Table R2). Context statistics show partial collapse without LVC\mathcal{L}_{\mathrm{VC}} and little mode information without Ldyn\mathcal{L}_{\mathrm{dyn}} (Table R4).
  • Real topologies (hfLF, wNR9). Without retraining, on the IEEE 118-bus system and the Deltacom backbone, LINC is the highest among the learned methods in all settings; Prior-only is the highest on Deltacom under hub attacks, which we report (Table R5).

Reviewer-specific points.

  • hfLF (weak accept, maintained) acknowledged the real-network evaluation; we added Prior-only to the switching table and clarified the difference between Figure 3 and Table 3 (models, episode sets, and metric).
  • wabD asked for the memoryless baseline (added, Table R2), switching ablations and the H sweep (added), and clarification of the prior and PPO-GRU (Section 5.1). SeCBAD, the KL-regularized variant, and the stop-gradient ablation did not complete within the discussion period; we will include them in the next revision.
  • wNR9 stated two concerns: experiments on larger and more realistic topologies, which we have provided, and methodological novelty. On the latter we agree that the context objective is not new and have revised the contribution statement accordingly: the contribution is the problem of online identification of an unobserved intra-episode failure-mode switch in network reconfiguration, and a graph-specific integration of existing components for it, with evidence for each component.

We thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.