acceptodds
Under review as a conference paper at ICLR 2027

Discovering Safety Spectra in Autonomous Driving via Constrained Adversarial Self-Play

Abstract

Naturalistic traffic data contain few safety critical interactions, limiting their ability to expose the weaknesses of autonomous driving policies during training. Adversarial scenario generation can target these weaknesses, but it faces two challenges: optimizing only for safety violations may produce implausibly extreme behavior, while scenarios generated against a fixed policy may lose relevance as the protagonist evolves. To address these challenges, we introduce Constrained Self-Play with Adversarial Rationality (Co-SPAR), a closed-loop framework that couples prior-constrained adversary optimization with region-wise protagonist optimization. On the adversary side, a deviation budget around a behavioral prior defines nested adversary classes with different attack strengths. Within each class, the adversary maximizes the constraint decay function (CDF), which measures the protagonist's future violation risk, to identify the worst admissible adversary. On the protagonist side, the resulting robust CDF defines a robust feasible region that certifies safety against every adversary in the class. This region guides the protagonist to reduce violation risk outside it and improve task return inside it while preserving safety. We prove that at each adversarial strength, exact policy iteration monotonically expands the robust feasible region, improves task performance within it, and converges to the optimal CDF. Across strengths, we establish a robust feasible region spectrum and prove that a protagonist optimal at a stronger strength retains its safety certificates against weaker adversary classes. Experiments show Co-SPAR advances the attack–realism Pareto frontier with a controllable adversarial spectrum and learns protagonist policies that outperform adversarial training baselines across nominal and adversarial traffic. Cross-strength evaluations demonstrate the predicted robust feasible region ordering and stronger-to-weaker safety transfer.

Primary area:
reinforcement learning
Keywords:
reinforcement learning, safe reinforcement learning, adversarial training, self-play, autonomous driving, adversarial scenario generation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YbcZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a constrained adversarial self-play framework for autonomous driving. A KL budget around a behavioral prior controls adversarial strength, while CDF-based risk estimates guide learning to prioritize risk reduction or task performance depending on the state. The paper analyzes feasible-region expansion and safety transfer across adversarial strengths, and evaluates attack-realism trade-offs and driving performance in PufferDrive using Waymo Open Motion Dataset scenarios.

Critical strengths
  • Strong empirical evaluation. Experiments on 10,000 training and 1,000 held-out scenarios cover multiple baselines and nominal, adversarial, and cross-strength settings. Co-SPAR improves safe task completion by over 10 percentage point over the strongest baseline, supporting its practical value.

  • Extensive theoretical analysis. Under the stated assumptions, the paper establishes feasible-region expansion, CDF convergence, and safety-certificate transfer across adversarial strengths. These results connect the adversary constraints to safety evaluation and protagonist optimization.

  • The studied problem is important, and the overall framework is reasonable.

Critical weaknesses
  • Gap between theory and implementation. The theory assumes exact policy iteration and worst-case adversary optimization, while the implementation uses learned CDF critics, PPO, empirical average KL control, and decentralized updates. It is unclear how these approximations affect feasible-region estimates and safety preservation, or which guarantees apply to the trained policies.

  • Unclear multi-agent credit assignment. Overlapping neighborhoods can assign the same protagonist violation to several adversaries without distinguishing their contributions. The paper needs stronger justification or targeted experiments showing that these local updates support the joint adversarial objective.

  • Limited distinction from prior work. Adversarial self-play and ego-adversary co-evolution have been studied in prior work, such as ADV-0. Beyond cross-strength safety theory, the differences in optimization objectives and policy updates remain unclear.

Questions for authors

See weakness.

5 comments

We sincerely appreciate your detailed and constructive feedback. Below are our responses to your comments.

W1 (Safety guarantees for learned policies)

Gap between theory and implementation. The theory assumes exact policy iteration and worst-case adversary optimization, while the implementation uses learned CDF critics, PPO, empirical average KL control, and decentralized updates. It is unclear how these approximations affect feasible-region estimates and safety preservation, or which guarantees apply to the trained policies.

Co-SPAR’s safety certificate applies to a learned protagonist at states where its true robust CDF satisfies the feasibility threshold. In training, a learned adversary with average KL control approximates the worst admissible adversary under the state-wise KL constraint. The critic’s estimate determines which update a state receives. Near the feasibility threshold, an estimation error can switch the update from task improvement to risk reduction or vice versa. Our monotone expansion result applies to the exact region-wise policy improvement step that the decentralized PPO updates approximate.

Analyzing an idealized policy update and approximating it in training has precedent in safe RL. Feasible Policy Iteration [1] assumes accurate CDF estimates in its analysis and learns the CDF in continuous state spaces. Constrained Policy Optimization [2] derives a constrained update and implements it using a local approximation and line search.

[1] Yujie Yang et al. Feasible Policy Iteration for Safe Reinforcement Learning. arXiv:2304.08845, 2023.

[2] Joshua Achiam et al. Constrained Policy Optimization. ICML, 2017.


W2 (Overlapping adversary neighborhoods)

Unclear multi-agent credit assignment. Overlapping neighborhoods can assign the same protagonist violation to several adversaries without distinguishing their contributions. The paper needs stronger justification or targeted experiments showing that these local updates support the joint adversarial objective.

Two adversaries may observe the same violation, but under a factorized joint policy each adversary’s expected update depends on how its own actions affect the expected team return. Consider a vehicle alongside the protagonist just before another vehicle cuts in. If holding its position blocks the protagonist’s escape route while yielding leaves it open, a subsequent collision in both vehicles’ neighborhoods can provide a learning signal for that earlier choice. If the expected team return is the same for every action available to the vehicle alongside the protagonist, its expected gradient contribution at that observation is zero.

In training, each adversary uses a local CDF based on the first protagonist violation in its neighborhood. Using this local CDF in place of the team-wide signal relies on the neighborhood sufficiency assumption in Appendix C.4 (Assumption 2), under which the substitution preserves that adversary’s expected policy update. Under the independent update approximation, the per-agent steps taken from the same joint rollout together approximate a joint policy update.


W3 (Objectives and policy updates in ASP, ADV-0, and Co-SPAR)

Limited distinction from prior work. Adversarial self-play and ego-adversary co-evolution have been studied in prior work, such as ADV-0. Beyond cross-strength safety theory, the differences in optimization objectives and policy updates remain unclear.

Optimization objectives. ADV-0 formulates a zero-sum game over ego return. Its adversary learns a distribution favoring low-return traffic trajectories, and the ego maximizes return under that distribution. ASP trains a teacher to challenge the student while demonstrating a collision-free solution to the same scenario. In Co-SPAR, the adversary maximizes the CDF over a prior-constrained policy class. The resulting robust CDF identifies where the current protagonist is safe against every admissible adversary, while task return remains a separate protagonist objective.

Policy updates. In ADV-0, updating the adversary changes the traffic distribution for the next ego-return update. In Co-SPAR, evaluating the adversary also determines which objective the protagonist pursues at each state. On joint rollouts, the protagonist uses the estimated robust feasible region to improve task return with a CDF penalty inside the region and to reduce risk outside it.

At fixed β\beta, each exact update preserves the current robust feasible region even when the worst admissible adversary is recomputed for the updated protagonist (Theorem 1). On that same region, task value is nondecreasing against the adversary used to compute the update (Theorem 3).

We added this comparison to the Related Work section.

Thank you again for your detailed feedback. Please let us know if any concerns remain after our response or if further clarification would be helpful.

We appreciate your time and would be grateful to hear your thoughts on the explanations provided in our response.

Reviewer_d9DEOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Co-SPAR, a constrained adversarial self-play framework that balances adversarial realism and driving safety. Its main contributions include prior-constrained adversary optimization, region-wise protagonist optimization, and theoretical safety guarantees. Experiments on Waymo driving scenarios demonstrate improved attack–realism trade-offs and driving success over existing baselines, supported by ablation and cross-strength evaluations.

Critical strengths

This paper presents a theoretically grounded and empirically effective framework for safety-aware adversarial self-play in autonomous driving. Its main strengths are:

  1. Strong theoretical foundation.The paper establishes theoretical guarantees on monotonic expansion of robust feasible regions, policy improvement, and safety certificate transfer across adversarial strengths. These results provide a principled foundation for balancing safety and task performance.

  2. Promising empirical performance. Co-SPAR demonstrates a favorable attack–realism trade-off and consistently improves driving success across nominal and adversarial traffic. Notably, it achieves a 75.8% success rate against strong adversaries, outperforming the best baseline by 11.4 percentage points.

Critical weaknesses
  1. The cross-play evaluation would benefit from broader policy-performance comparisons. Beyond CDF distributions and safety violation rates, please report success rate and safe route progress for policies trained at different (\beta) values against opponents of varying strengths. These comparisons would clarify how safety-certificate transfer translates into practical driving performance and reveal potential trade-offs between robustness and task completion.
  2. The contribution of constrained self-play remains unclear. In Table 2(a), full Co-SPAR improves success over fixed-prior training by only 0.8 percentage points under Log Replay and 1.1 points against strong adversaries. Without variability and stronger component isolation, these results do not establish how much of the overall gain comes from constrained self-play.
Questions for authors
  1. Are the improvements over BC Trained consistent across seeds? The authors should compare the results across different random seeds.
  2. Does increasing (\beta) uncover qualitatively different failure modes, or mainly increase the frequency and severity of existing ones? Reporting scenario diversity and failure-type coverage would clarify the value of the generated training data.
  3. How sensitive is the attack–realism frontier to the behavioral prior? Does Co-SPAR retain its advantage with a weaker prior or under a shift in traffic behavior?
  4. During deployment, how should AVs select (\beta) when the adversary strength is unknown?
4 comments

We sincerely appreciate your detailed and constructive feedback. Below are our responses to your comments.

W1 (Cross-strength task performance)

The cross-play evaluation would benefit from broader policy-performance comparisons. Beyond CDF distributions and safety violation rates, please report success rate and safe route progress for policies trained at different (\beta) values against opponents of varying strengths. These comparisons would clarify how safety-certificate transfer translates into practical driving performance and reveal potential trade-offs between robustness and task completion.

Figure 10(a,b) reports SR and SRP for the same 36 training–test strength pairings as the SVR matrix in Figure 9. For every protagonist, both SR and SRP decrease as the test adversary becomes stronger. For the strong-to-weak pairing examined in our safety-transfer analysis, the protagonist trained at β→∞\beta\to\infty achieves 80.8% SR and 93.0% SRP against the β=0.3\beta=0.3 adversary, close to the matched-strength policy’s 81.8% SR and 93.4% SRP.


W2 & Q1 (Contribution of constrained self-play)

The contribution of constrained self-play remains unclear. In Table 2(a), full Co-SPAR improves success over fixed-prior training by only 0.8 percentage points under Log Replay and 1.1 points against strong adversaries. Without variability and stronger component isolation, these results do not establish how much of the overall gain comes from constrained self-play.

Are the improvements over BC Trained consistent across seeds? The authors should compare the results across different random seeds.

SR gains over BC Trained across seeds

BC Trained and Co-SPAR use the same region-wise protagonist update, so their comparison in Table 2(a) measures the benefit of replacing the fixed BC Prior with a learned, prior-constrained adversary. SR gains over BC Trained were consistent across the three seeds, averaging 0.89 percentage points under Log Replay and 1.09 points against strong adversaries.

Evaluation conditionSeedBC Trained SR (%)Co-SPAR SR (%)Gain (pp)
Log Replay183.4783.4784.1484.14+0.67+0.67
Log Replay281.0881.0881.5081.50+0.42+0.42
Log Replay379.5079.5081.0881.08+1.58+1.58
Strong Adv. Avg.175.9575.9577.1577.15+1.20+1.20
Strong Adv. Avg.274.1874.1874.6774.67+0.49+0.49
Strong Adv. Avg.373.9173.9175.4875.48+1.57+1.57

SR gains from updating versus freezing a trained adversary

To isolate the contribution of continued adversary adaptation, we compared updating and freezing a trained adversary from the same protagonist–adversary checkpoint after 52.4M agent steps of joint training. We continued training the protagonist in both runs with the same β\beta, update rule, and remaining interaction budget. Against the same fixed test opponents, continued adaptation raised final protagonist SR by 1.70 percentage points under Log Replay and 1.75 points on Strong Adv. Avg.

Q2 (Failure modes across adversarial strengths)

Does increasing (\beta) uncover qualitatively different failure modes, or mainly increase the frequency and severity of existing ones? Reporting scenario diversity and failure-type coverage would clarify the value of the generated training data.

We evaluated the BC Trained protagonist on the same 1,000 scenarios at each adversarial strength, with 32 rollouts per scenario. After excluding initial violations, we classified the first collision for each protagonist in each rollout. Ahead and behind refer to the other vehicle’s position relative to the protagonist. We excluded unclassified collisions (less than 1.3% at each strength) from the counts and shares.

All five collision types already occurred at β=0\beta=0. Stronger adversaries produced more collisions and changed which types predominated. From β=10\beta=10 to β→∞\beta\to\infty, the number of classified collisions rose by 7.4%, while the share involving oncoming or crossing traffic rose from 40.3% to 70.7%.

Adversary strength β\betaSame direction (ahead)Same direction (behind)Same direction (side contact)OncomingCrossingTotal classified collisions
001591712782053521,165
0.30.32661532832114021,315
111,3141845452955252,863
331,9712211,5508421,6836,267
10103,1184382,4141,9712,05910,000
∞\infty1,4324061,3032,9664,63110,738

Q3 (Sensitivity to the behavioral prior)

How sensitive is the attack–realism frontier to the behavioral prior? Does Co-SPAR retain its advantage with a weaker prior or under a shift in traffic behavior?

We reduced the number of WOMD scenarios used to train the BC prior from 80,000 to 4,000 and 800 while keeping the other training settings fixed.

Near a realism score of 0.50, all three Co-SPAR variants reach 55–58% ASR, including the variant whose BC prior used 800 scenarios (1% of the full training set). At similar realism, SAGE reaches about 45% ASR and ASP about 42% (Figure 7 in the revised paper). Reducing the BC training set from 80k to 4k and then 800 scenarios lowered the highest observed realism score from about 0.75 to 0.73 and then 0.69.


Q4 (Selecting β\beta for deployment)

During deployment, how should AVs select (\beta) when the adversary strength is unknown?

When opponent strength is unknown, we treat β\beta as a choice of how much deviation from nominal traffic the protagonist should be prepared to handle. Each budget defines a class of admissible adversaries around the behavioral prior. Theorem 2 shows that a protagonist optimal in robust CDF at a stronger budget retains its safety certificate at every weaker budget for the states certified at the stronger one.

If two budgets both support the driving task and the states of interest remain feasible under the stronger budget, we would favor that budget for its broader adversary coverage. Increasing the budget further can shrink the robust feasible region, so the choice depends on which states the deployment must cover.

A future extension is to train a β\beta-conditioned protagonist and use recent traffic interactions to select its budget at test time.

Thank you again for your thoughtful comments. We would appreciate your feedback on our response and would be happy to clarify any remaining points.

Thank you for your rebuttal. I will maintain my positive score.

Reviewer_fTVnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Co-SPAR, a framework for closed-loop adversarial training of autonomous driving policies. A state-wise KL budget around a behavioral prior defines a family of admissible adversaries. The adversary maximizes a constraint decay function (CDF), while the protagonist uses the resulting robust feasible region to prioritize risk reduction outside the region and task improvement inside it. For a finite-state/action formulation with exact robust evaluation and policy improvement, the paper establishes monotonic CDF improvement, feasible-region expansion, convergence to the optimal CDF, and cross-budget region inclusions. Experiments in PufferDrive on WOMD evaluate attack–realism trade-offs, driving performance, cross-strength interactions, and component ablations.

Critical strengths
  1. Strong empirical performance on safe task completion. Table 1 reports success rates of 82.2% under Log Replay and 75.8% against the strong-adversary average, compared with 69.7% and 64.4% for the strongest corresponding baseline. Safe route progress also improves. These are meaningful gains on held-out scenarios, supported by multiple training seeds and repeated rollouts. The results support a useful safety–progress trade-off, although Co-SPAR does not achieve the lowest violation rate in Table 1.

  2. A coherent formulation linking adversary restrictions and safety-aware policy improvement. Defining a family of prior-constrained adversaries makes the scope of the idealized safety guarantee explicit. Combining this with robust CDFs yields a consistent account of feasible regions and their dependence on adversarial strength. The attack–realism evaluations and cross-strength analyses complement the main driving results rather than reporting task return alone.

Critical weaknesses
  1. The independent benefit of continuously adapting the adversary is not isolated. The paper motivates continuous adversary updates as necessary to track the protagonist’s evolving safety weaknesses. However, BC Trained in Table 2(a) removes both adversarial optimization and adversary adaptation, making it difficult to distinguish the benefit of introducing an adversarial opponent from that of continuously updating it. A controlled comparison against a frozen, previously trained adversary is missing.

  2. Sensitivity to key hyperparameters of the region-wise update is not evaluated. Although the paper examines adversarial strength, it reports only fixed settings for the feasibility threshold, barrier-shape parameter, and penalty cap. The feasibility threshold determines the switch between task improvement and risk reduction, while the penalty cap controls the relative emphasis on safety. It remains unclear whether the reported benefits persist across reasonable parameter choices and how these parameters should be selected.

Questions for authors
  1. What is the independent contribution of continuous adversary adaptation? Could the authors compare jointly updating both policies with freezing the adversary and updating only the protagonist, starting from the same protagonist–adversary checkpoint? Please keep adversarial strength, the protagonist update rule, and the subsequent environment-interaction budget fixed. Reporting final success and safety violation rates, together with the frozen adversary’s attack effectiveness over training, would directly test whether a fixed adversary becomes less useful as the protagonist improves.

  2. How robust is the method to its key hyperparameters? Could the authors explain how the feasibility threshold, barrier-shape parameter, and penalty cap were selected, and provide a small sensitivity study prioritizing the feasibility threshold and penalty cap? Reporting success and safety violation rates while holding other settings fixed would clarify whether the observed safety–performance trade-off requires careful tuning.

4 comments

We sincerely appreciate your detailed and constructive feedback. Below are our responses to your comments.

W1 & Q1 (Continuous adversary adaptation)

The independent benefit of continuously adapting the adversary is not isolated. The paper motivates continuous adversary updates as necessary to track the protagonist’s evolving safety weaknesses. However, BC Trained in Table 2(a) removes both adversarial optimization and adversary adaptation, making it difficult to distinguish the benefit of introducing an adversarial opponent from that of continuously updating it. A controlled comparison against a frozen, previously trained adversary is missing.

What is the independent contribution of continuous adversary adaptation? Could the authors compare jointly updating both policies with freezing the adversary and updating only the protagonist, starting from the same protagonist–adversary checkpoint? Please keep adversarial strength, the protagonist update rule, and the subsequent environment-interaction budget fixed. Reporting final success and safety violation rates, together with the frozen adversary’s attack effectiveness over training, would directly test whether a fixed adversary becomes less useful as the protagonist improves.

After 52.4M agent steps of joint training, we resumed from the same checkpoint and continued training the protagonist with either a frozen adversary or one that continued to learn. Both runs used the same β\beta, protagonist update rule, and remaining interaction budget (800M agent steps in total).

We evaluated the resulting protagonists against the same fixed test adversaries. Across all three test conditions, final SR was 1.70–1.76 percentage points higher with continued joint updates than with a frozen adversary. SRP was also higher and SVR lower in each condition. The tables report means ±\pm standard deviations across three seeds.

Log Replay

ProtagonistSR ↑\uparrowSRP ↑\uparrowSVR ↓\downarrow
Random initialization1.44±0.131.44 \pm 0.138.49±0.318.49 \pm 0.3193.31±0.0893.31 \pm 0.08
Shared checkpoint63.54±7.7163.54 \pm 7.7185.78±2.1985.78 \pm 2.1917.35±2.5617.35 \pm 2.56
Joint policy updates81.63±0.8681.63 \pm 0.8693.13±0.7293.13 \pm 0.727.37±0.337.37 \pm 0.33
Frozen adversary79.93±0.3979.93 \pm 0.3992.68±0.0892.68 \pm 0.087.74±0.877.74 \pm 0.87

BC Prior

ProtagonistSR ↑\uparrowSRP ↑\uparrowSVR ↓\downarrow
Random initialization1.46±0.151.46 \pm 0.158.53±0.448.53 \pm 0.4493.14±0.1093.14 \pm 0.10
Shared checkpoint63.65±7.7563.65 \pm 7.7585.87±2.1685.87 \pm 2.1616.84±2.2416.84 \pm 2.24
Joint policy updates81.33±1.0281.33 \pm 1.0293.05±0.8093.05 \pm 0.807.01±0.447.01 \pm 0.44
Frozen adversary79.57±0.5779.57 \pm 0.5792.51±0.1792.51 \pm 0.177.37±0.637.37 \pm 0.63

Strong Adv. Avg.

ProtagonistSR ↑\uparrowSRP ↑\uparrowSVR ↓\downarrow
Random initialization1.41±0.121.41 \pm 0.128.42±0.368.42 \pm 0.3693.49±0.1193.49 \pm 0.11
Shared checkpoint59.18±7.5559.18 \pm 7.5581.32±2.2181.32 \pm 2.2123.63±1.4323.63 \pm 1.43
Joint policy updates74.67±1.1574.67 \pm 1.1587.67±0.9587.67 \pm 0.9515.50±0.4515.50 \pm 0.45
Frozen adversary72.92±1.0772.92 \pm 1.0787.18±0.6787.18 \pm 0.6715.64±0.2715.64 \pm 0.27

Training-time attack effectiveness

The frozen adversary's training-time ASR falls from 17.35% early in training to 7.86% late in training as the protagonist continues to learn. Training-time ASR also fell under joint updates.

Training stageJoint policy updatesFrozen adversary
Early (updates 111–160)19.46±3.5519.46 \pm 3.5517.35±1.7217.35 \pm 1.72
Middle (updates 751–800)8.17±0.788.17 \pm 0.788.01±0.658.01 \pm 0.65
Late (updates 1477–1526)8.44±0.618.44 \pm 0.617.86±0.367.86 \pm 0.36

W2 & Q2 (Hyperparameter sensitivity)

Sensitivity to key hyperparameters of the region-wise update is not evaluated. Although the paper examines adversarial strength, it reports only fixed settings for the feasibility threshold, barrier-shape parameter, and penalty cap. The feasibility threshold determines the switch between task improvement and risk reduction, while the penalty cap controls the relative emphasis on safety. It remains unclear whether the reported benefits persist across reasonable parameter choices and how these parameters should be selected.

How robust is the method to its key hyperparameters? Could the authors explain how the feasibility threshold, barrier-shape parameter, and penalty cap were selected, and provide a small sensitivity study prioritizing the feasibility threshold and penalty cap? Reporting success and safety violation rates while holding other settings fixed would clarify whether the observed safety–performance trade-off requires careful tuning.

We use ϵ=0.1\epsilon=0.1 as a tolerance on the estimated CDF when defining the feasible region. Within this region, α=1.2\alpha=1.2 keeps the CDF penalty small in the interior and makes it rise more steeply near the boundary, where its weight reaches wmax⁡=1.0w_{\max}=1.0. We vary each parameter separately while holding the other settings fixed.

Varying ϵ\epsilon or α\alpha changes mean SR by at most 1.9 percentage points and SVR by at most 1.0 point relative to the defaults across all three conditions. The main trade-off comes from the penalty cap. Raising wmax⁡w_{\max} from 1.01.0 to 1.51.5 lowers SVR by 1.2–1.4 points at an SR cost of 2.9–3.5 points. Lowering wmax⁡w_{\max} to 0.50.5 raises SVR in every condition without improving SR. These results support wmax⁡=1.0w_{\max}=1.0 for safe task completion and a higher cap when reducing violations takes priority.

Log Replay

Protagonist settingSR ↑\uparrowSRP ↑\uparrowSVR ↓\downarrow
Default82.2±1.782.2 \pm 1.793.4±1.193.4 \pm 1.17.5±0.87.5 \pm 0.8
ϵ=0.05\epsilon=0.0580.8±1.080.8 \pm 1.092.5±0.592.5 \pm 0.58.3±1.08.3 \pm 1.0
ϵ=0.15\epsilon=0.1582.0±0.282.0 \pm 0.293.3±0.193.3 \pm 0.18.1±0.38.1 \pm 0.3
α=1.1\alpha=1.180.9±1.080.9 \pm 1.092.8±0.792.8 \pm 0.77.9±0.77.9 \pm 0.7
α=1.3\alpha=1.381.4±0.981.4 \pm 0.993.2±0.393.2 \pm 0.37.8±1.77.8 \pm 1.7
wmax⁡=0.5w_{\max}=0.581.7±1.481.7 \pm 1.493.6±0.493.6 \pm 0.49.9±1.19.9 \pm 1.1
wmax⁡=1.5w_{\max}=1.579.3±1.279.3 \pm 1.291.8±0.691.8 \pm 0.66.1±0.46.1 \pm 0.4

BC Prior

Protagonist settingSR ↑\uparrowSRP ↑\uparrowSVR ↓\downarrow
Default82.3±1.482.3 \pm 1.493.6±1.093.6 \pm 1.07.0±0.77.0 \pm 0.7
ϵ=0.05\epsilon=0.0580.5±1.180.5 \pm 1.192.4±0.692.4 \pm 0.67.8±0.77.8 \pm 0.7
ϵ=0.15\epsilon=0.1581.7±0.081.7 \pm 0.093.2±0.193.2 \pm 0.17.8±0.47.8 \pm 0.4
α=1.1\alpha=1.180.7±0.980.7 \pm 0.992.8±0.792.8 \pm 0.77.4±0.77.4 \pm 0.7
α=1.3\alpha=1.381.1±0.781.1 \pm 0.793.1±0.393.1 \pm 0.37.5±1.77.5 \pm 1.7
wmax⁡=0.5w_{\max}=0.581.6±1.381.6 \pm 1.393.6±0.393.6 \pm 0.39.5±0.79.5 \pm 0.7
wmax⁡=1.5w_{\max}=1.578.8±1.378.8 \pm 1.391.7±0.691.7 \pm 0.65.7±0.55.7 \pm 0.5

Strong Adv. Avg.

Protagonist settingSR ↑\uparrowSRP ↑\uparrowSVR ↓\downarrow
Default75.8±1.375.8 \pm 1.388.4±0.988.4 \pm 0.915.1±0.415.1 \pm 0.4
ϵ=0.05\epsilon=0.0573.9±1.273.9 \pm 1.287.2±0.787.2 \pm 0.716.1±0.716.1 \pm 0.7
ϵ=0.15\epsilon=0.1575.1±0.275.1 \pm 0.287.9±0.487.9 \pm 0.416.0±0.516.0 \pm 0.5
α=1.1\alpha=1.174.1±0.674.1 \pm 0.687.4±0.587.4 \pm 0.515.8±0.915.8 \pm 0.9
α=1.3\alpha=1.374.7±0.774.7 \pm 0.787.8±0.187.8 \pm 0.115.7±1.115.7 \pm 1.1
wmax⁡=0.5w_{\max}=0.575.4±1.075.4 \pm 1.088.5±0.588.5 \pm 0.517.4±0.217.4 \pm 0.2
wmax⁡=1.5w_{\max}=1.572.3±1.472.3 \pm 1.486.2±0.786.2 \pm 0.713.9±0.213.9 \pm 0.2

Thank you again for your feedback. We have added the hyperparameter sensitivity results to the revised paper (Appendix D.4, Table 9). Please let us know if you have any remaining concerns or questions.

We appreciate your time and would be grateful to hear whether our response and additional experiments have addressed your concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for the Area Chair and Reviewers

Dear Area Chair and Reviewers,

Thank you for your time and effort in reviewing our paper.

We summarize the additional experiments and explanations provided in response to the main concerns below.

  • Reviewer fTVn (adversary adaptation and hyperparameter sensitivity): We compared continued adaptation with a frozen adversary from the same checkpoint, using the same protagonist update rule. Continued adaptation improves final SR by 1.70 to 1.76 pp across three evaluation conditions. We also provided a sensitivity study of the region-wise update parameters, now included in Appendix D.4 of the revised paper.
  • Reviewer d9DE (empirical evaluation): We provided additional comparisons across seeds and adversarial strengths, collision-type analysis, and experiments with priors trained on less data.
  • Reviewer YbcZ (theoretical scope and relation to prior work): We explained the scope of the guarantees, the assumption supporting local updates, and the differences from ASP and ADV-0.

After considering our rebuttal, both d9DE and YbcZ reaffirmed their positive assessments. We respectfully ask the Area Chair to consider the additional evidence and full discussion alongside the revised paper.

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.