acceptodds
Withdrawn from ICLR 2027

Geometric Distributional Control: Learning Progress with Partial Structural Knowledge

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Partial structural knowledge can constrain admissible decisions without specifying how to achieve task progress. We introduce Geometric Distributional Control (GDC), which formulates decision-making under such knowledge as learning and navigating conditional distributions of feasible action segments. The central principle is positive expected progress of a conditional action-segment distribution over supported horizons, rather than positive progress of every segment. This lets waiting or repositioning coexist with goal-directed behavior and extracts directional knowledge from imperfect trajectories without globally optimal demonstrations. Multi-horizon candidate selection lets the learned guidance respond to the temporal scale of the current context. GDC learns a progress-weighted distribution in a compact latent space and combines its score with available analytical structure to guide online geometric optimization. We establish conditions under which distributional approximation and candidate selection retain positive expected segment progress. A conditional objective-gap bound links representation coverage, value approximation, and online search accuracy to decision quality on a data-supported region. Experiments on traffic control and hierarchically coupled nonconvex merit optimization demonstrate effectiveness across changing constraints and problem scales.

Primary area:
learning on time series and dynamical systems
Keywords:
distributional control, partial structural knowledge, adaptive horizons, expected progress, geometric optimization

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 4 reviews, above 27% of the others.
Reviewer_cdmROpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Geometric Distributional Control (GDC), a framework for decision-making under partial structural knowledge. It separates feasibility enforcement, handled by explicitly available constraints, from directional guidance, learned as a progress-oriented distribution over short action segments. The learned guidance is represented in a compact latent space and incorporated into a Riemannian energy-flow procedure with multi-horizon candidate selection. The paper also provides a conditional theoretical analysis linking representation and search accuracy to decision quality, alongside a known-feasibility guarantee.

The method is evaluated on SUMO-based route-progress control and parameterized bilevel/trilevel optimization benchmarks. The reported results show improved arrival rates over several baselines, successful transfer to held-out traffic environments, and normalized merit gaps below 0.65% on the reported multilevel instances. While the framework is conceptually interesting and the empirical results are encouraging, the theoretical guarantees rely on modeling assumptions, and the experiments do not fully isolate the contribution of learned directional guidance from that of the feasibility and candidate-selection mechanisms.

Critical strengths
  1. A genuinely useful problem framing and role separation. The paper takes seriously an information pattern that is common in deployed control but under-served by both MPC and RL: explicit structure tells you what is executable, but not which executable action to take. Framing the learned object as a \emph{progress-tilted conditional distribution over short feasible segments}, and using its score as a local direction rather than as a policy or a value function, is a clean and reusable idea. The ``positive expected progress over a conditional distribution'' criterion is a sensible relaxation that lets waiting/yielding coexist with goal-directed behavior, and it is a meaningful conceptual contribution beyond another policy-fitting recipe.

  2. The empirical evidence is broad and the role separation is exercised, not just asserted. The paper evaluates on two structurally different domains (SUMO closed-loop traffic control and parameterized bilevel/trilevel nonconvex merit optimization), reports closed-loop arrival, collisions and known-constraint residuals, and includes a frozen-model transfer study under changed known scaffolds (blocked maps, changed graph/disturbance parameters). The SUMO multi-horizon results (100% arrival, zero collisions, 9--49 point arrival gain over rule/BC/PPO/SAC on the complex map) and the multilevel gaps below 0.65%0.65\% with explicit box/step feasibility are the kind of evidence that materially supports the claimed role separation. The horizon-use diagnostics (e.g., H=24H=24 or 4848 on 75.2%75.2\% of long-route decisions, shorter horizons near local events) are a nice operational illustration of why multi-horizon candidate pooling matters.

Critical weaknesses
  1. The central theoretical guarantee is conditional on an assumption that the paper does not empirically validate, and the gap between theory and practice is under-quantified. Theorem 4.3 / D.3 identifies the tilted pushforward qβZq_\beta^Z with an idealized maximum-entropy continuation density p⋆∝ZHp^\star \propto Z_H, then concludes that the learned score is an approximate descent direction for the soft progress value VprogsoftV_{\mathrm{prog}}^{\mathrm{soft}}. The authors are commendably explicit that this identification is a modeling assumption, not a fact derived from the training objective.'' But the paper then leans on this result in the abstract and introduction (conditional descent and execution bounds''), while the only empirical diagnostic of positive expected progress is the optional support screen in Appendix B, which the authors themselves describe as a heuristic and not a calibrated confidence bound. There is no reported measurement of E[Δρ]\mathbb{E}[\Delta\rho] under the learned pθZp_\theta^Z on the test contexts, no estimate of the approximation error ϵapprox\epsilon_{\mathrm{approx}}, and no ablation of how sensitive the SUMO/multilevel results are to violating Assumption 4.1 (e.g., in the sparse-TLS setting where stopping is locally correct). As a result, the reader cannot tell whether the directional guarantee is doing real work or whether the empirical success is driven entirely by the known-feasible filter, fallback and candidate ranking. This is consequential because the theory is presented as a main contribution.

  2. Key methodological components are difficult to verify from the main text, and several design choices are not ablated. The reported controller depends on: (i) a compact latent action-segment autoencoder with dz≪Hdud_z \ll H d_u; (ii) a clean-latent denoising predictor whose implied score is used online; (iii) a decoder-pullback Riemannian metric G(z)G(z); (iv) a two-timescale flow with retraction/backtracking and candidate filtering; (v) a selection cost JselJ_{\mathrm{sel}} combining learned energy, predicted progress, effort, known violation and distance to fallback; and (vi) a horizon pool with a runtime candidate-progress gate. The main text reports end results, but the critical ablations are largely absent or relegated to unreadable appendix material: how much does the learned score contribute relative to the known-only/fallback controller under the same filter and horizon pool? How sensitive are the results to dzd_z, to the tilt β\beta, to the horizon pool, and to the candidate-progress gate? The paper reports GDC fixed-$H$'' vs.\ GDC multi-HH'', which isolates horizon pooling, but does not isolate the learned direction from the feasibility scaffold. Given that the paper's central claim is that data supplies the missing direction while structure supplies feasibility, this is the single most important ablation and it is missing from the main results.

Questions for authors

Q1. Can you report a direct empirical estimate of Eτ∼pθZ(⋅∣o,g,Cknown,H⋆)[Δρ(τ,g)]E_{\tau\sim p_\theta^Z(\cdot|o,g,C_{known},H^\star)}[\Delta\rho(\tau,g)] on the SUMO test contexts, alongside the empirical distribution of Δρ\Delta\rho (including the fraction of negative/zero-progress segments)? This would test Assumption 4.1 rather than assume it. In particular, in the sparse-TLS setting where stopping is locally correct, what fraction of accepted learned segments have Δρ≤0\Delta\rho \le 0?

Q2. What is the measured or bounded approximation error ϵapprox\epsilon_{\mathrm{approx}} between sθs_\theta and the idealized −∇Vprogsoft-\nabla V_{\mathrm{prog}}^{\mathrm{soft}} on the data-supported region? If it cannot be measured directly, can you provide a proxy (e.g., score-matching residual, or a comparison of the learned score against a finite-difference estimate of the tilted density on held-out anchors)?

Q3. The most important missing ablation: under the \emph{same} known-feasible filter, fallback, horizon pool and candidate-ranking budget, how much does the learned score term contribute over (a) a known-only/fallback controller and (b) a controller that uses the predicted progress Δ^ρ\widehat{\Delta}\rho but \emph{not} the learned energy/score? In other words, is the improvement attributable to the learned direction, or to the scaffold/filter/horizon-pooling machinery?

Q4. How sensitive are the SUMO and multilevel results to the latent dimension dzd_z, the tilt β\beta, the horizon pool H\mathcal{H}, and the runtime candidate-progress gate? A small sensitivity table would substantially strengthen the claim that the method is not finely tuned.

Q5. In the multilevel experiments, the online criterion is the penalized joint merit Φ\Phi rather than the exact nested response in Eqs.~(146)/(149). Can you report the nested-response residual x−x0⋆(y)x - x_0^\star(y) (or the trilevel analogue) for the reported GDC iterates, so that the reader can judge how far the solution is from true hierarchical feasibility?

5 comments
AuthorsOpenReview ↗

What fraction of progress is negative or zero, especially near red signals?

Reviewer comment — Critical Weaknesses 1:

The central theoretical guarantee is conditional on an assumption that the paper does not empirically validate.... Reviewer comment — Questions For Authors Q1:

Q1. Can you report a direct empirical estimate of on the SUMO test contexts, alongside the empirical distribution of (including the fraction of negative/zero-progress segments)? This would test Assumption 4.1 rather than assume it. In particular, in the sparse-TLS setting where stopping is locally correct, what fraction of accepted learned segments have ?

Authors’ response.

Yes. On 50,000 held-out H6 segments, 80.846% have positive, 11.974% zero, and 7.180% negative terminal progress, and the mean is positive in every driving category, including red signals (0.764).

  • Why this tests Assumption 4.1. The assumption concerns the expected terminal progress of the tilted distribution, not positivity of every segment. By Eq. 3 of the submission, the tilted mean equals the untilted mean plus a nonnegative variance integral, so the condition holds for every β≥0\beta\ge0 while waits and reversals remain in the data.
  • The stopping case (new in the revision). Of the 6,482 held-out segments that start at a red signal, 92.194% have zero, 0.216% negative, and 7.590% positive progress; their mean stays positive (0.764), and the deployed tilt raises it to 66.413. Where no horizon qualifies, the support screen abstains and λ→1\lambda\to1 recovers the known-feasible controller (Remark D.8 of the submission).
  • Accepted learned segments (new in the revision). Over 240 deployed routes, 60,314 of 60,467 executed decisions (99.74%) select a learned proposal, 13,165 of them under an active red-signal flag. Every selected proposal has positive surrogate terminal progress and zero recorded known violation: mean 136.877 overall (95% CI [134.647, 139.224]) and 129.595 under the red-signal flag (95% CI [126.264, 133.053]).

Evidence: Table 24 (p. 58); Table 26 (p. 59).

AuthorsOpenReview ↗

Q2. What is the approximation error to the idealized score?

Reviewer comment — Questions For Authors Q2:

Q2. What is the measured or bounded approximation error between and the idealized on the data-supported region? If it cannot be measured directly, can you provide a proxy (e.g., score-matching residual, or a comparison of the learned score against a finite-difference estimate of the tilted density on held-out anchors)?

Authors’ response.

We report the proxies you suggest, on 50,000 held-out H6 segments from 160 map groups: the implemented gradient matches central finite differences of the learned energy with mean cosine 0.999994, and the learned score orders held-out pairs with accuracy 0.9765.

  • Score–energy consistency. At 64 anchors, the relative-L2 error against finite differences is 0.000441, and the cosine similarity is 0.999994 on average (minimum 0.999817).
  • Ordering, which is how the score is used. GDC uses the score to compare and refine known-feasible candidates, so ordering error is the relevant error: Spearman/Pearson correlations are 0.9844/0.9734, and pairwise ordering accuracy is 0.9765 over 98,541 informative pairs.
  • Calibration. Predicted mean terminal progress is 27.2697 against an observed 27.2989.

Evidence: Table 25 (p. 59).

AuthorsOpenReview ↗

Q3. What does the learned term add under the same scaffold?

Reviewer comment — Critical Weaknesses 2:

Key methodological components are difficult to verify from the main text, and several design choices are not ablated. ...

Reviewer comment — Questions For Authors Q3:

Q3. The most important missing ablation: under the \emph{same} known-feasible filter, fallback, ...

Authors’ response.

The improvement is attributable to the learned direction, not to the scaffold: with the filter, fallback, horizon pool, and 195-candidate budget identical, the learned score term completes the routes in 18.217 fewer mean steps than the known-only/fallback controller (95% CI [15.886, 20.789]). All arms share GDC's candidate generator, checker, projection, fallback, horizon pool, 240 paired routes, and seeds, and differ only in how the executed candidate is chosen; to the best of our knowledge, no existing method isolates a learned direction inside a fixed feasibility scaffold, so we built them ourselves.

  • (a) Over a known-only/fallback controller (new in the revision). The learned score term contributes 18.217 fewer mean steps (95% CI [15.886, 20.789]; fewer steps on 229 of 240 routes) and lowers mean known residual from 1.57×10−61.57\times10^{-6} to 0. Inside one decision, 194 of the 195 candidates pass the known gate and 27 pass the joint known/progress gate; removing only the learned ranking term changes the executed candidate.
  • (b) Over predicted progress without the learned energy (new in the revision). This is the analytical-progress arm of Table 13, which keeps the predicted-progress term and drops the learned energy and score refinement. In SUMO, where the remaining route distance dtd_t is part of the observation, both arms reach 240/240 with zero collisions and zero residual, and GDC executes a nonfallback candidate on 99.74% of decisions versus 98.39%. Where progress has no analytical form, the learned energy is what improves the result: removing score refinement raises Φ\Phi from 0.101394 to 0.103281 (bilevel) and from 0.175032 to 0.183572 (trilevel), and on real nuPlan logs the learned estimator reduces held-out progress MAE by 58.8–74.1% relative to kinematic integration.

Evidence: Tables 13–14 (p. 54); Appendix H.1 (p. 55); Table 30 (p. 61); Table 38 (p. 64).

Reviewer_sgnyOpenReview ↗
Rating
2(−1)
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper proposes to combine the explicit knowledge of good/bad actions with learned progress indicators for control problems. For learning progress, the authors consider short action segments, label them with a corase progress score and reweight them. Then, a latent diffusion model is trained on this rewighted data to learn progress estimates. At evaluation time, this score is used as one term in latent space gradient flow alonside constaints and rule terms. Using these gradients, the method refines several candidates, and the best one is executed after projecting it into the space of valid/feasible set against known constraints. The proposed method reaches its goal with a 100% success rate without collisions.

Critical strengths

The paper has strong results on the driving benchmark they test on compared to the baselines

Critical weaknesses
  • In my understanding, the proposed method samples multiple actions if they obey the rules and defaults to a fallback option if the rules are violated. It does not look like the baselines have access to such a harness, which might make it unclear if the proposed method works better / safer due to this explicit harness or due to the proposed pipeline

  • An alternative strategy to this method could be to sample candidate actions from a diffusion policy and scoring them for task progress and safety; for instance done by Diffusion-ES (https://arxiv.org/abs/2402.06559). In my understanding the paper emphasizes on not sampling actions from the policy but rather only using the diffusion policy to get a sense of direction of progress. Why is the alternative not a viable path? Are there any emperical comparisons with such an approach?

  • I am not an expert in this domain, but I believe many driving papers (including the paper cited above) use nuPlan to test their algorithms. Testing on popular benchmarks like nuPlan might make the results stronger

Questions for authors
  • Do baselines get safety fallbacks?
  • Comparisons against alternative stragies like Diffusion-ES and on popular benchmarks like nuPlan
8 comments
AuthorsOpenReview ↗

Do the baselines receive the same safety fallback?

Reviewer comment — Critical Weaknesses 1:

In my understanding, the proposed method samples multiple actions if they obey the rules and defaults to a fallback option if the rules are violated. It does not look like the baselines have access to such a harness, which might make it unclear if the proposed method works better / safer due to this explicit harness or due to the proposed pipeline

Reviewer comment — Questions For Authors:

Do baselines get safety fallbacks?

Authors’ response.

In the revision they do. Rerun with exactly GDC's filter, projection, and fallback, PPO and SAC lower their residual but keep the same arrivals, so the harness secures feasibility and does not account for GDC's gains.

  • Baselines with GDC's harness (new in the revision). On the same 240 routes, SAC residual drops from 0.032112 to 0.0000646 and PPO residual from 3.97×10−53.97\times10^{-5} to 9.77×10−79.77\times10^{-7}; arrivals stay at 240/240 and 239/240, with zero collisions.
  • GDC's learned part, with the harness held fixed (new in the revision). With the same 195-candidate generator, filter, projection, fallback, routes, and seeds, GDC completes all 240 routes in 18.217 fewer mean steps than fallback-only execution (95% CI [15.886, 20.789]) and lowers mean known residual from 1.57×10−61.57\times10^{-6} to 0. To the best of our knowledge, no existing method isolates a learned direction inside a fixed feasibility scaffold, so we built this control ourselves.
  • Inside one decision. 194 of the 195 candidates pass the known gate and 27 pass the joint known/progress gate; the learned ranking decides which of them is executed.

Evidence: Tables 13–14 (p. 54); Appendix H.1 (p. 55); Table 17 (p. 56).

AuthorsOpenReview ↗

Is Diffusion-ES a viable alternative?

Reviewer comment — Critical Weaknesses 2:

An alternative strategy to this method could be to sample candidate actions from a diffusion policy and scoring them for task progress and safety; for instance done by Diffusion-ES (https://arxiv.org/abs/2402.06559). In my understanding the paper emphasizes on not sampling actions from the policy but rather only using the diffusion policy to get a sense of direction of progress. Why is the alternative not a viable path? Are there any emperical comparisons with such an approach?

Reviewer comment — Questions For Authors:

Comparisons against alternative stragies like Diffusion-ES and on popular benchmarks like nuPlan

Authors’ response.

Yes, it is a viable alternative, and we added it on the same 240 paired routes, with identical maps, edge routes, and SUMO seeds: both methods reach 240/240 with zero collisions, and GDC attains the lower known-constraint residual, 0 versus 0.000693.

MethodArrivalCollisionMean known residual
Known-only control240/24000.131281
Diffusion-ES240/24000.000693
GDC240/24000

The two methods use the learned model differently. Diffusion-ES samples executed candidates from a diffusion policy and scores them through population search. GDC targets data made of imperfect feasible segments, including waits and reversals, when part of the physics needed to rank them is unknown: it learns a progress-tilted measure over known-feasible segments, uses its score as a local direction inside known-feasible search, and leaves the executed action to the known checker, projection, and fallback.

Evidence: Table 13 (p. 54); Tables 18–19 (p. 56).

Reference: Yang, B., Su, H., Gkanatsios, N., Ke, T.-W., Jain, A., Schneider, J., and Fragkiadaki, K. (2024). Diffusion-ES: Gradient-free planning with diffusion for autonomous and instruction-guided driving. CVPR, pp. 15342–15353.

AuthorsOpenReview ↗

Evaluation on nuPlan

Reviewer comment — Critical Weaknesses 3:

I am not an expert in this domain, but I believe many driving papers (including the paper cited above) use nuPlan to test their algorithms. Testing on popular benchmarks like nuPlan might make the results stronger

Reviewer comment — Questions For Authors:

Comparisons against alternative stragies like Diffusion-ES and on popular benchmarks like nuPlan

Authors’ response.

We added a real-world nuPlan Boston experiment with a matched baseline. On 12 held-out test recordings, GDC's learned progress estimator reduces error relative to the baseline by 58.8%, 67.5%, and 74.1% at 3, 6, and 12 s.

HorizonTest segmentsBaseline MAELearned MAEReduction
3 s10,6540.55900.230158.8%
6 s9,2830.94300.306767.5%
12 s7,0921.60540.416674.1%
  • The baseline. It is the estimate a known-structure controller would otherwise use: kinematic integration of the same recorded actions, scored on exactly the same test segments.
  • Data and protocol. 234 parent recordings, about 79.2 hours of driving, split by recording before segment extraction (210 training, 12 validation, 12 test). The 95% recording-bootstrap intervals for the reductions are [56.4%, 62.1%], [64.1%, 70.9%], and [69.1%, 78.1%].
  • Multiple horizons. On 61,200 same-context candidate pools, learned joint scoring across horizons captures 75.8% of the oracle-achievable improvement.

Evidence: Table 29 (p. 60); Tables 30–32 (p. 61).

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

This paper proposes Geometric Distributional Control (GDC), a framework for decision-making when explicit structural knowledge can rule out infeasible actions but does not specify which feasible action makes progress toward the goal. From offline known-feasible action segments labeled with weak progress certificates, GDC learns a latent representation of action segments and a progress-tilted latent density, where segments are reweighted by the exponential of their clipped progress. The goal is positive expected progress over the segment distribution rather than progress of every segment, so that waiting or repositioning can coexist with goal-directed behavior. At deployment, the score of this density is combined with known-constraint and task potentials in a Riemannian gradient flow on the latent space, and candidates from multiple horizons are decoded, filtered, ranked, and projected onto the known-feasible action set, with a fallback when no candidate qualifies. The authors provide conditional descent, feasibility, and approximate optimality statements, and evaluate GDC on SUMO route control and penalized bilevel and trilevel nonconvex optimization, reporting full arrival without collisions and small gaps to offline references.

Critical strengths

S1. The problem setting is well motivated. Many practical controllers can verify feasibility but lack a complete model or value function to choose among feasible actions, and the paper clearly articulates this intermediate information pattern. The idea of positive expected progress, rather than progress of every segment, is intuitive and allows necessary waiting or repositioning behavior.

S2. The design separates feasibility from direction in a principled way. Known constraints are enforced through filtering, projection, and a fallback, so execution feasibility does not depend on the accuracy of the learned model. The authors are also transparent about which properties hold by construction and which rely on modeling assumptions.

Critical weaknesses

W1. The core contribution is difficult to isolate. The method combines several components, including exponential tilting, a latent autoencoder, a diffusion-based score model, a decoder-pullback Riemannian metric, multi-horizon candidate selection, filtering, projection, and a fallback, and the authors acknowledge that exponential tilting and adding potentials to a learned score are not new. Apart from the fixed-horizon versus multi-horizon comparison, the paper does not provide ablations of these components, so it remains unknown which of them drives the performance.

W2. The experimental evidence could be strengthened. In SUMO, both GDC variants reach full arrival, and SAC also reaches full arrival on the sparse map, so the benchmark may be close to saturated. It should also be clarified whether the RL and imitation baselines are given the same known-feasible filter, projection, and fallback as GDC. In the multilevel experiments, results are reported for a single held-out instance per dimension without variance, the online baselines are local solvers, and comparing wall time with SCIP, a global solver with certification, is not a like-for-like comparison.

Questions for authors

Q1. Could the authors provide ablations removing or replacing the main components (learned score, Riemannian metric, latent representation, progress tilt, and the predicted progress term) to clarify which parts drive the performance?

Q2. Could the authors report SUMO results where the baselines use the same known-feasible filter, projection, and fallback as GDC, and multilevel results over multiple held-out instances per dimension, compared against a multistart local solver under the same time budget?

4 comments
AuthorsOpenReview ↗

Q1. Can the main components be removed or replaced?

Reviewer comment — Critical Weaknesses W1:

W1. The core contribution is difficult to isolate. The method combines several components, including exponential tilting, a latent autoencoder, a diffusion-based score model, a decoder-pullback Riemannian metric, multi-horizon candidate selection, filtering, projection, and a fallback, and the authors acknowledge that exponential tilting and adding potentials....

Reviewer comment — Questions For Authors Q1:

Q1. Could the authors provide ablations removing or replacing the main components (learned score, Riemannian metric, latent representation, progress tilt, and the predicted progress term) to clarify which parts drive the performance?

Authors’ response.

Yes. We ablated each of the five components with everything else held fixed; the fixed- versus multi-horizon comparison was in the submission (Table 1), and the five ablations below are new in the revision. Removing the learned score alone costs 18.217 mean steps per route (95% CI [15.886, 20.789]).

ComponentReplaced byWith the replacementGDC
Learned scoreFallback only, same scaffold270.171 mean steps; residual 1.57×10−61.57\times10^{-6}251.954 mean steps; residual 0
Riemannian metricEuclidean latent update, same step normProgress gain 2.3781 in 0.166 s2.4334 in 0.072 s, same 73 calls
Latent representationDirect search over 24 action coordinates393 objective calls in 0.296 s73 calls in 0.072 s
Progress tiltUnweighted β=0\beta=0Held-out mean progress 27.29944.378, higher in every driving category
Predicted-progress termIts weight, and each other ranking weight, halved or doubledAll eleven settings reach the goal with zero collisions in 188–190 steps190 steps

To the best of our knowledge, no existing method isolates a learned direction inside a fixed feasibility scaffold, so we built these controls ourselves; they share GDC's candidate generator, known checker, projection, fallback, routes, and seeds. In the multilevel benchmark, removing score refinement likewise raises Φ\Phi from 0.101394 to 0.103281 (bilevel) and from 0.175032 to 0.183572 (trilevel).

Evidence: Table 1 (p. 7); Tables 13–14 (p. 54); Table 16 (p. 55); Table 24 (p. 58); Table 27 (p. 59); Table 28 (p. 60); Table 38 (p. 64).

AuthorsOpenReview ↗

Q2. Can safety access, optimization variation, and compute be compared fairly?

Reviewer comment — Critical Weaknesses W2:

W2. The experimental evidence could be strengthened. In SUMO, both GDC variants reach full arrival, and SAC also reaches full arrival on the sparse map, so the benchmark may be close to saturated. It should also be clarified whether the RL and imitation baselines are given the same known-feasible filter, projection, and fallback as GDC. In the multilevel experiments, results are reported for a single held-out instance per dimension without variance, the online baselines are local solvers, and comparing wall time with SCIP, a global solver with certification, is not a like-for-like comparison.

Reviewer comment — Questions For Authors Q2:

Q2. Could the authors report SUMO results where the baselines use the same known-feasible filter, projection, and fallback as GDC, and multilevel results over multiple held-out instances per dimension, compared against a multistart local solver under the same time budget?

Authors’ response.

Yes, all three are now reported. With GDC's harness, PPO and SAC lower their residual but keep their arrivals; over 160 held-out multilevel instances, GDC attains a lower Φ\Phi than both online controllers in every instance; and under the time budget, every GDC run satisfies the step limits along its whole path, while no multistart L-BFGS-B run does.

  • SUMO baselines with the same harness (new in the revision). Frozen PPO and SAC, wrapped in exactly GDC's filter, projection, and fallback on the 240 paired routes: SAC residual drops from 0.032112 to 0.0000646 (paired difference −0.032048, 95% CI [−0.037245, −0.027144]) and PPO residual from 3.97×10−53.97\times10^{-5} to 9.77×10−79.77\times10^{-7}, while arrivals stay at 240/240 and 239/240. The harness secures feasibility; it does not supply the direction.
  • A less saturated SUMO suite (new in the revision). On 60 newly generated maps with changed closures, changed geometry, and longer routes, against DTQL, FQL, GFP, EDA, and ReBRAC-v2 (1,440 executions), GDC has the lowest mean known residual in every category, and all 15 paired intervals lie below zero. On long routes, FQL, GFP, EDA, and ReBRAC-v2 reach 0/80 goals, and GDC reaches 80/80.
  • Multiple held-out instances (new in the revision). Twenty per family/dimension group (160 in total), with frozen GDC models: mean normalized gap 0.342% (bilevel) and 0.398% (trilevel), versus 44.437% / 54.311% for Known and 41.942% / 63.926% for Nested, with zero box and step violations.
  • Multistart solver under the same time budget (new in the revision). The matched 30-second comparison with multistart L-BFGS-B is in Table 36. Checked against the step limits at every accepted iterate, GDC is feasible at 100% of iterates (347/347 bilevel, 330/330 trilevel) along the whole path in 10/10 runs per family; L-BFGS-B is feasible at 69.8778% and 72.5962% of iterates and in 0/10 runs per family. Credited only with iterates reached along a step-feasible path from the initial point, GDC attains the lower Φ\Phi in all 20 pairs. SCIP is kept as an offline reference in Table 3 and Figure 4.

Evidence: Table 3 (p. 9); Table 17 (p. 56); Tables 20–23 (p. 57); Table 33 (p. 62); Table 36 (p. 63); Appendix M.6 (p. 64); Table 39 (p. 65).

Thank you for the detailed response and new experiments. My main concerns remain.

Q1. The new controls do not show that the learned components drive performance: in Table 14, the analytical-progress arm (learned ranking and refinement removed) is slightly faster than the full controller at less than half the per-decision cost, and the β = 0 arm is also faster.

Q2. The harness clarification, transfer suite, and 160 instances are appreciated. However, the transfer baselines are still not harness-matched, and in the requested time-matched comparison multistart L-BFGS-B is lower in all 20 pairs.

I keep my rating at 2.

AuthorsOpenReview ↗

Follow-up to Reviewer oxCz: learned ranking under a fixed scaffold, harness, and step-limit feasibility

Thank you for engaging with the new experiments. We answer both points directly.

Q1. The learned term decides which feasible segment is executed, and it is what carries over where progress has no analytical form. The analytical-progress arm reads SUMO's exact remaining route distance dtd_t, which a system with partly unknown physics does not provide; its completion steps are in Table 14. With the scaffold held fixed:

  • The executed segment. Inside one retained decision, with candidates, checker, and fallback fixed, a different segment is executed without the learned ranking term (Appendix H.1, p. 55).
  • Fallback use. The full controller executes a nonfallback candidate on 99.74% of decisions, compared with 98.39% for the analytical arm (Table 13, p. 54).
  • Where progress has no analytical form. In the multilevel benchmark, score refinement lowers Φ\Phi from 0.103281 to 0.101394 (bilevel) and from 0.183572 to 0.175032 (trilevel) (Table 38, p. 64); on real nuPlan logs, the learned estimator reduces held-out progress MAE by 58.8–74.1% relative to kinematic integration (Table 30, p. 61).
  • β=0\beta = 0. This arm keeps the learned score model and the joint gate and differs from GDC in the tilt alone; the tilt's role is distributional, and on 50,000 held-out segments it raises mean terminal progress from 27.299 to 44.378 and raises it in every driving category (Table 24, p. 58).

Q2. The harness does not supply direction, and credited only with step-feasible iterates, GDC is lower than L-BFGS-B in all 20 pairs. Wrapped in GDC's filter, projection, and fallback, PPO and SAC lower their residual while their arrival and collision counts do not change (Table 17, p. 56). The long-route failures in the transfer suite are failures of direction: FQL, GFP, EDA, and ReBRAC-v2 reach 0/80 goals with stopped fractions 0.9913, 0.9815, 0.4362, and 0.8510, against 80/80 and 0.0060 for GDC (Table 23, p. 57). Diffusion-ES, the harness-matched alternative we added, runs with explicit projection and filtering on the same 240 routes and records mean known residual 0.000693, while GDC records zero (Table 18, p. 56).

For the time-matched comparison (Table 36, p. 63), an online controller must also keep every executed step within the step limits, so we reran the same 20 pairs with every trajectory saved and checked each accepted iterate (Appendix M.6, p. 64; Table 39, p. 65):

FamilyMethodFeasible iteratesPath-feasible runs
BilevelGDC100% (347/347)10/10
BilevelL-BFGS-B69.8778%0/10
TrilevelGDC100% (330/330)10/10
TrilevelL-BFGS-B72.5962%0/10

Credited only with iterates reached along a step-feasible path from the initial point, GDC attains the lower Φ\Phi in all 20 pairs; in every L-BFGS-B run, the first step-limit excess occurs at the first or second accepted iterate of the first start. Among online controllers held to the same constraints, GDC attains a lower final Φ\Phi than Known and Nested in all 160 held-out instances (Table 33, p. 62).

Reviewer_8nniOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces GDC, a controller for settings where known constraints can reject actions but cannot rank them. Known constraints define the feasible set of short action segments; a scalar progress certificate is attached to each offline segment; the empirical segment measure is exponentially tilted toward higher certificate values; an encoder/decoder compresses segments into a low-dimensional latent space; a denoising score model is fit on the tilted latent measure; and online, that score is one term in a projected Riemannian flow, with feasibility handled by explicit filtering, projection, and a fallback. The stated contributions are the "positive expected segment progress" criterion (allowing waits and reversals), multi-horizon candidate selection, and conditional local-descent / coverage-controlled optimality / feasible-execution results. Experiments are SUMO route following on custom perturbed grids and a parametric quartic-well "bilevel/trilevel" family.

Critical strengths

The one idea worth keeping is the division of labour: a hand-written checker, projection, and fallback own feasibility, and learning is confined to a local direction among survivors. That is the right instinct, and the frozen-model transfer run is the kind of test this decomposition should be judged by. I have nothing else of substance to add here.

Critical weaknesses

1. There is no new mechanism, only a new vocabulary. Exponential tilting toward a segment-level return is reward-weighted regression (Peters & Schaal, 2007; Peters et al., 2010; Peng et al., 2019), and the paper's own Eq. 8 is standard denoising score matching with importance weights, which it says itself. The online rule (Algorithm 1) is: propose K segments, discard infeasible ones, rank the survivors by an energy, take the first action, replan — that is sampling-based MPC with a safety filter. Naming the objects "fiber", "scaffold", "certificate", and "merit" does not change what they are, and the paper never isolates what the combination buys over its own parts.

2. The theory assumes the thesis it is meant to support. Assumption 4.1 assumes positive expected progress — the paper's central claim — and it is never checked anywhere in the paper. Eq. 3 is the variance identity of the tilted family, dμ/dβ=Var≥0d\mu/d\beta = \mathrm{Var} \ge 0: it shows the tilt raises the mean relative to the untilted data, not that the tilted mean is positive, which is the only case that matters. Theorem 4.3 is Cauchy–Schwarz against an idealized score, and the authors concede the identification is "a modeling assumption, not a fact derived from the training objective." Theorem 4.5 assumes the online potential tracks the oracle. Proposition 4.7 holds because the method projects or falls back. εapprox\varepsilon_{approx}, αval\alpha_{val}, and hnh_n are never measured. These are statements about the assumptions, not about GDC.

3. The experiments cannot separate the method from a competent baseline. In SUMO the route and the remaining route distance dtd_t are inputs, and the certificate is exactly the drop in dtd_t over the segment — the information needed to act is already in the observation, and the paper's own fallback (target speed, brake for leaders, no discretionary lane changes) should finish most routes. So "Rule safety" arriving at 91% with mean violation 0.148, while GDC arrives at 100% everywhere with violation 10−710^{-7} and zero collisions, reads as weak baselines and a deliberately crippled rule. Multi-horizon — one of the two advertised contributions — moves arrival from 1.000 to 1.000 and progress from 1.000 to 1.000; its entire measured effect is changing violation from 1.6×10−51.6\times10^{-5} to 0. There is no ablation of β\beta, of dzd_z, or of the five hand-set weights in Eq. 143, no controller latency, and no reported failure.

The second domain is worse. Nested response constraints are "penalized rather than enforced" (F.3), so the "bilevel/trilevel" task is a single-level box-constrained quartic program with d≤12d \le 12 and analytic gradients. No standard solver given the same information — L-BFGS-B, IPOPT, an augmented-Lagrangian or trust-region method — appears; only the authors' own "Known" and "Nested" scaffolds, which GDC beats fivefold. The SCIP wall-clock ratios are not a comparison, since SCIP solves the exact bilevel program and GDC optimizes a different penalized objective; on the one instance where SCIP certifies quickly (bilevel, d=6d=6) it is twice as fast as GDC. A 0.09–0.64% gap to a multistart reference that is not a global optimum supports no claim at all.

4. The paper is fluent and unverifiable. The same three sentences about "role separation" are restated in the abstract, the introduction, Section 2, Section 5, the conclusion, and the appendix, while not one decision is traced concretely end to end. Table 1 is Tables 4 and 5 reporting identical numbers; the 90-versus-100-episode SAC discrepancy is buried in a footnote; 40 pages of appendix formulas contain no worked example. No code, no maps, no environment. The AI-use statement is consistent with all of this, and I take the disclosure at face value — the consequence is a paper that asserts its contribution in every paragraph and demonstrates it in none.

Questions for authors
  1. Where is μβ(c)>0\mu_\beta(c) > 0 measured, and on what fraction of decisions? If the direction is only credited for the sign of progress, this is the number the paper needs.
  2. Replace the learned score with (a) greedy decrease of dtd_t under the same filter and (b) L-BFGS-B on the same parametric program with the same gradients. Do any of the reported gains survive?
  3. The rule controller reports 0.148 mean violation while running the same checker that defines the feasible set, on a task where the route is an input. How is that baseline configured, and how were PPO and SAC tuned?
  4. Multi-horizon changes nothing measurable except the violation term. What do I lose by deleting the horizon pool and fixing H=12H=12?
  5. Will the maps, benchmark generators, and training code be released?
6 comments
AuthorsOpenReview ↗

*Reviewer comment — Critical Weaknesses 1:*

Critical Weaknesses 1:

  1. There is no new mechanism, only a new vocabulary. Exponential tilting ....

Authors’ response.

This characterization is incorrect. GDC learns a direction only on the support that known structure certifies as feasible, uses it only inside known-feasible search, and leaves every executed action to the known checker, projection, and fallback; with that structure held fixed, the learned direction alone shortens routes by 18.217 mean steps (95% CI [15.886, 20.789]). Each equivalence in the review leaves out exactly this learned part:

Equivalence in the reviewWhat it leaves outMeasured
Tilting = reward-weighted regressionRWR fits the executed policy; GDC tilts a measure over known-feasible segments and uses only its scoreHeld-out mean progress 27.299 → 44.378
Eq. 8 = weighted denoising score matchingThe weight is a signed terminal-progress certificate on feasible segments, waits and reversals includedPairwise ordering accuracy 0.9765 on 98,541 held-out pairs
Algorithm 1 = sampling MPC with a safety filterSampling MPC needs a complete candidate cost; GDC learns the missing ranking from dataWithout the learned ranking, a different candidate is executed

Many influential methods derive much of their contribution from how established components are combined for a specific problem, and their own papers describe them this way. AlphaGo trained policy and value networks by a combination of supervised and reinforcement learning and introduced a search algorithm that combines Monte Carlo simulation with these networks (Silver et al., 2016). DQN combined Q-learning with deep convolutional networks and stabilized the combination with experience replay and a periodically updated target network (Mnih et al., 2015). DDPM built on a connection between diffusion probabilistic models and denoising score matching with Langevin dynamics (Ho et al., 2020). Diffuser reinterpreted classifier-guided sampling and image inpainting as planning strategies (Janner et al., 2022), and Diffusion Policy represented a visuomotor policy as a conditional denoising diffusion process executed with receding-horizon control (Chi et al., 2023). Each of these works also contributed the design choices that made its combination work; in each, the central advance was matching known ingredients to a problem structure and showing that the coupling works.

GDC follows this pattern. Exponential tilting, denoising score matching, projection, fallback control, and sampling-based search are established tools; GDC's contribution is how it couples them for one information pattern, in which known structure can certify or repair feasibility but cannot rank feasible directions because part of the physics is unknown.

References: Chi et al. (2023), RSS; Ho, Jain, and Abbeel (2020), NeurIPS 33; Janner et al. (2022), ICML; Mnih et al. (2015), Nature, 518:529–533; Silver et al. (2016), Nature, 529:484–489.

What the combination buys. The revision removes one part at a time and holds the rest fixed:

  • SUMO, same scaffold (new in the revision). Against fallback-only execution with the same candidate generator, filter, projection, fallback, 240 paired routes, and seeds, GDC is faster on 229 of 240 routes and lowers mean known residual from 1.57×10−61.57\times10^{-6} to 0.
  • Multilevel program. In the submission, GDC already lowered Φ\Phi by 70.6–86.6% relative to Known, its own scaffold without the learned direction. In the revision, score refinement alone lowers Φ\Phi from 0.103281 to 0.101394 (bilevel) and from 0.183572 to 0.175032 (trilevel).
  • Against five recent learned controllers (new in the revision). On 60 newly generated maps, all 15 paired GDC-minus-baseline residual intervals lie below zero; on long routes, FQL, GFP, EDA, and ReBRAC-v2 reach 0/80 goals, and GDC reaches 80/80.
  • On real driving logs (new in the revision). On recording-disjoint nuPlan logs, the learned estimator reduces held-out progress MAE by 58.8–74.1% relative to kinematic integration.

Evidence: Tables 2–3 (p. 9); Tables 13–14 (p. 54); Tables 22–23 (p. 57); Table 24 (p. 58); Table 25 (p. 59); Table 30 (p. 61); Table 38 (p. 64).

AuthorsOpenReview ↗

*Reviewer comment — Critical Weaknesses 2:*

Critical Weaknesses 2:

2. The theory assumes the thesis it is meant to support. ...

Reviewer comment — Questions For Authors 1:

Where is measured, and on what fraction of decisions? If the direction is only credited for the sign of progress, this is the number the paper needs.

Authors’ response.

The premise is incorrect: Assumption 4.1 is measured directly in the revision, and it holds. Mean terminal progress is positive before any tilting in every driving category, and 60,314 of 60,467 executed decisions (99.74%) select a learned proposal with positive surrogate terminal progress.

By Eq. 3 of the submission, tilting never lowers the mean, so a positive untilted mean is exactly what Assumption 4.1 requires. On 50,000 held-out H6 segments from 160 map groups (new in the revision):

Driving categoryUntilted mean progressDeployed tilt (β=1\beta=1)
Red signal0.76466.413
Leader-constrained42.72955.005
Free driving31.44344.300
Other low speed6.08778.309
Overall27.29944.378
  • On what fraction of decisions. Over 240 deployed routes, 99.74% of decisions execute a learned proposal, 13,165 of them under an active red-signal flag; each has positive surrogate terminal progress and zero recorded known violation (mean 136.877, 95% CI [134.647, 139.224]).
  • Theorem 4.3. Its measurable counterparts on held-out data are Spearman correlation 0.9844 with observed progress, pairwise ordering accuracy 0.9765 over 98,541 pairs, and mean cosine 0.999994 between the implemented gradient and finite differences of the learned energy.
  • Proposition 4.7. It holds because the method projects or falls back, and that is its purpose: execution stays known-feasible however accurate the learned score is.

Evidence: Table 24 (p. 58); Tables 25–26 (p. 59).

AuthorsOpenReview ↗

*Reviewer comment — Critical Weaknesses 3:*

Critical Weaknesses 3:

3. The experiments cannot separate the method from a competent baseline. In ....

Reviewer comment — Questions For Authors 2:

Replace the learned score with (a) greedy decrease of under the same filter and (b) L-BFGS-B on the same parametric program with the same gradients. Do any of the reported gains survive?

Authors’ response.

The claim is incorrect, and the gains survive both replacements the review asks for. The competent rule the review describes is our fallback-only arm, run inside GDC's own scaffold, and GDC still completes the routes in 18.217 fewer mean steps (95% CI [15.886, 20.789]).

  • The competent rule (new in the revision). Target speed, braking for leaders, no discretionary lane change: this arm finishes all 240 routes, as the review expects, and GDC is faster on 229 of them and lowers mean known residual from 1.57×10−61.57\times10^{-6} to 0. To the best of our knowledge, no existing method isolates a learned direction inside a fixed feasibility scaffold, so we built these arms ourselves; they share GDC's candidate generator, checker, projection, fallback, routes, and seeds.

  • (a) Greedy decrease of dtd_t (new in the revision). This is the analytical-progress arm in Table 13. It reads SUMO's exact remaining route distance, and GDC, which learns progress from data, matches it at 240/240 arrivals with zero residual. Where progress has no analytical form, the learned direction is decisive: on recording-disjoint real nuPlan logs, it reduces held-out progress MAE relative to kinematic integration by 58.8%, 67.5%, and 74.1% at 3, 6, and 12 s.

  • Stronger learned baselines (new in the revision). Against DTQL, FQL, GFP, EDA, and ReBRAC-v2, run frozen on 60 newly generated maps (1,440 executions), GDC has the lowest mean known residual in every category, and all 15 paired intervals lie below zero. On long routes, FQL, GFP, EDA, and ReBRAC-v2 reach 0/80 goals, DTQL arrives with mean known residual 0.230078, and GDC reaches 80/80 with 1.89×10−51.89\times10^{-5}.

  • (b) L-BFGS-B on the same program and gradients (new in the revision). Ipopt, differential evolution with L-BFGS-B polishing, and 32-start L-BFGS-B now run on the same penalized merit Φ\Phi, bounds, and initial states. An online controller must also keep every executed step within the step limits. Held to that requirement, every GDC iterate is feasible and no L-BFGS-B run is:

FamilyMethodFeasible iteratesPath-feasible runs
BilevelGDC100% (347/347)10/10
BilevelL-BFGS-B69.8778%0/10
TrilevelGDC100% (330/330)10/10
TrilevelL-BFGS-B72.5962%0/10

Credited only with iterates reached along a step-feasible path from the initial point, GDC attains the lower Φ\Phi in all 20 pairs.

The reference gap and SCIP. The gap to the multistart reference now supports a paired claim over 160 held-out instances, reported with mean, median, and P90. With one frozen model per configuration, GDC attains a lower final Φ\Phi than Known and Nested in every instance, with mean normalized gap 0.342% / 0.398% (bilevel / trilevel) versus 44.437% / 54.311% for Known and 41.942% / 63.926% for Nested. SCIP is kept as an offline reference incumbent in Table 3 and Figure 4.

Evidence: Table 3 (p. 9); Tables 13–14 (p. 54); Tables 22–23 (p. 57); Table 30 (p. 61); Table 33 (p. 62); Tables 34 and 36 (p. 63); Appendix M.6 (p. 64); Table 39 (p. 65).

Comments on the paper

4 comments
AuthorsOpenReview ↗

General response to all reviewers

We thank the four reviewers. Each reviewer's questions are answered point by point in that reviewer's own thread, and each answer can be read on its own. In the answers, In the submission marks results that were already in the submitted version, and New in the revision marks experiments and text added in response to the reviews. The new experiments are reported in Appendices G–N; Appendix G.1 states the common statistical conventions, and Table 12 maps each new experiment to the claim it tests. All new or materially revised text is in blue. The main text remains within 10 pages, and the abstract is unchanged.

For easier reading, all of our responses are also collected in a single PDF, GDC_Response_to_Reviewers.pdf, in the supplementary material, with the same numbering as here. The supplementary material also includes an anonymized evaluation package with the maps, routes, and seeds; the per-route and per-instance outcomes; the GDC inference code; the training and evaluation code of the five added learned baselines (DTQL, FQL, GFP, EDA, and ReBRAC-v2); and a script that recomputes the reported numbers.

Dear Reviewers,

Less than two days remain before the discussion period ends on November 18 (AoE). If you have not yet replied to the author rebuttal, please do so as soon as possible, indicating which of your concerns are resolved and which remain.

Best regards, Area Chair

Dear Reviewer 8nni,

I am writing to you specifically because your review gives a clear rejection, so it is especially important for the decision to know whether your concerns still stand after the rebuttal.

Please read the authors' response and reply as soon as possible, indicating which of your technical concerns are resolved and which remain, so that the authors have a chance to address them within the discussion period.

Also, please keep your reply focused on the technical points and professional in tone. Some remarks in your review, such as "a deliberately crippled rule" and the comment on the authors' AI-use statement, go beyond technical assessment.

Best regards, Area Chair

AuthorsOpenReview ↗

Why we are withdrawing: Reviewer 8nni's score of 1 on every criterion, unsupported accusations, and no response

We are withdrawing this submission, and we want to state the reason clearly.

Reviewer 8nni gave the paper a score of 1 on every criterion, together with claims that question our integrity without any evidence, including that our baseline was "deliberately crippled". We do not believe this score reflects a good-faith evaluation of the work. We answered every point in the review with new experiments and provided the evaluation package the reviewer requested, but the reviewer did not take part in the discussion at any point. The relevant parts of the review are included in the supplementary material for the record.

We sincerely thank the Area Chair for their help and coordination throughout the discussion, even though Reviewer 8nni still did not respond. We also thank the other reviewers for their time and constructive technical questions, which led to new experiments and a substantially stronger paper.

We respectfully ask the Senior Area Chair and the Program Chairs to examine this reviewer's conduct and to consider whether this reviewer should continue to review for ICLR, so that other authors are not subjected to the same treatment.

Thank you.

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

Related papers

Discussion (0)

Sign in to comment.