{"id":"45e63efb-3fa9-4258-8fe3-c4d107c032d1","arxiv_id":"2507.19344","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A bilevel optimization framework with normalizing flows recovers the latent obstacle and optimal trajectories in mean-field games from partial trajectory data, achieving about 0.1 relative L2 error even with fewer than 100 trajectories.","lead":"This paper presents a method to infer both the hidden obstacle and optimal agent trajectories in mean-field games from only partial trajectory observations, using normalizing flows and bilevel optimization. The approach is mesh-free, handles high-dimensional settings, and works with only dozens of sample trajectories, which matters for drone surveying and motion planning.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported ~0.1 obstacle-recovery errors rely on a regularizer R(phi) that injects the ground-truth obstacle mass into Eq. (22); without it, B is identifiable only up to an additive constant, so the central claim is not yet established.","rationale":"The paper's central claim is that solving the bilevel problem (15)-(16) with the regularizer (21)-(22) recovers the hidden obstacle B and optimal agent trajectories from partial trajectory observations, with approximately 0.1 relative L2 obstacle error. For this to be true in a genuine inverse problem, the method must recover B from the trajectories without using information about B* itself. The mass regularizer violates this condition: it matches the integral of the learned obstacle to the integral of the true obstacle, which is unknown by assumption. Because an additive constant in B does not change the lower-level minimizer, trajectory data alone cannot determine that constant; the regularizer is what pins it down, and it does so only because the true mass is inserted into the loss. The statement that the mass can be normalized to one without loss of generality conflates a forward-model gauge choice with an inverse-problem constraint: changing the normalization of B changes the effective lambda_I in the lower-level objective, and lambda_I is fixed in the algorithm. Thus the reported error numbers do not yet demonstrate obstacle recovery from data alone. The reader's conditional verdict is appropriate: the method is plausible and the experiments are internally consistent, but the known-mass prior must either be removed, estimated from data, or explicitly identified as an additional required input before the central claim is accepted. I agree with the reader's weakest-assumption analysis and therefore recommend no change to the verdict.","tokens_in":14319,"tokens_out":11526,"duration_ms":126537,"concrete_test":"Run the Two Bars experiment from Table 1 twice, keeping all settings identical: (i) omit R(phi) from Eq. (22), and (ii) replace the target mass in Eq. (21) by 2 * integral_Omega B* dx while generating data from B*. For (i), compare the recovered obstacle to B* after optimally shifting by an additive constant; if the shifted error remains near 0.05, the regularizer is not the source of accuracy, and if it degrades substantially, the claim depends on knowing the mass. For (ii), check whether the recovered obstacle approximates B*/2 or B*; if it approximates B*/2, the mass regularizer is injecting the target scale, confirming the concern. Also report the unshifted relative L2 errors for both runs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the mass regularizer R(phi) in Eqs. (21)-(22). Adding a constant c to B* leaves the lower-level objective (16) unchanged up to a constant: lambda_I * integral integral (B* + c)(F_theta(x)) p0 dx dt = lambda_I * integral integral B*(F_theta(x)) p0 dx dt + c * lambda_I, which is independent of theta. Hence observed trajectories cannot distinguish B* from B* + c. R(phi) removes this ambiguity by matching the integral of B_phi over Omega to the integral of B* over Omega, i.e., the ground-truth mass of the unknown obstacle. The paper's claim that this is 'without loss of generality' is not valid in the inverse setting: the regularizer requires exactly the quantity the method is supposed to recover. The reported relative L2 errors in Tables 1 and 2 therefore measure reconstruction with an answer-key-dependent prior; without it, the recovered obstacle is at best defined only up to an additive constant, so the ~0.1 error claim is not established. The issue is compounded because in the experiments Omega is a box and some obstacles have support outside Omega (Sec. 4.1), so the regularizer uses the truncated true mass over the domain, an even more target-specific quantity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a bilevel-optimization formulation for the joint inverse problem of recovering an unknown obstacle and optimal agent trajectories in a variational mean-field game, given partial observations of optimal trajectories. The lower level is the normalizing-flow-parameterized forward MFG objective (16); the upper level minimizes the squared distance between predicted and observed trajectory points. The authors reformulate the bilevel problem as a penalty problem (20) and add a mass regularizer (21)-(22) to address the additive-constant non-identifiability of the obstacle. They report relative L2 obstacle errors around 0.05-0.15 in 2D, 3D, and up to d=10 experiments, and show improved generalization in a scarce-data regime compared with a trajectory-fitting baseline.","tokens_in":14574,"tokens_out":6484,"duration_ms":65156,"significance":"If the claims are eventually supported, the paper would provide a mesh-free, high-dimensional-capable inverse MFG solver with a practical penalty algorithm. The use of normalizing flows for trajectories, the avoidance of hypergradient backpropagation, and the systematic low-data experiments are useful and go beyond simply applying existing solvers. However, the central obstacle-recovery claim is currently contingent on a regularizer that uses the ground-truth obstacle mass, so the quantitative headline result is not yet established.","major_comments":[{"comment":"The mass regularizer R(phi) explicitly uses the ground-truth obstacle integral, R_Omega B*(x) dx, and the paper's own 'Physical Space' paragraph acknowledges that B* is identifiable only up to an additive constant from trajectory data. Since adding a constant c to B* changes the lower-level objective (16) only by a theta-independent constant, observed trajectories cannot distinguish B* from B*+c. The claim that this is 'without loss of generality, because the ground truth obstacle mass can always be normalized to, or regarded as, one' is not valid: rescaling B to have unit mass changes the interaction cost and hence the optimal trajectories, so it is not a free normalization. Consequently the reported relative L2 errors in Tables 1 and 2 and the 'approximately 0.1' claim in the introduction are obtained with an answer-key-dependent prior. In Section 4.1 the Two Bars obstacle has support outside Omega=[-3,3]^2, so the regularizer uses the truncated true mass over Omega, an even more target-specific quantity. The central obstacle-recovery claim is not established unless the authors either estimate the additive constant/mass from data, or reformulate the recoverable quantity as the equivalence class modulo additive constants and report shape errors computed after removing the best constant.","section":"Section 3.5, Eqs. (21)-(22)"},{"comment":"The claimed equivalence between the maximum-likelihood objective and the upper-level objective in (15) is not generally valid. In a Markovian Gaussian transition model, the negative log-likelihood is, up to constants, sum_{k=0}^{K-1} E[||f_{theta_{k+1}}(x_k) - x_{k+1}||^2], where the conditioning state x_k is the observed intermediate trajectory point. The composition F_{theta_k}(x_0) used in the final expression of Eq. (18) is not equal to the conditional prediction f_{theta_{k+1}}(x_k) unless the model already predicts exactly the observed intermediate states, which is not true during training. Thus the interpretation of (22) as 'regularizing maximum likelihood trajectory learning' is at best an approximation. The derivation should either be corrected by defining the upper objective directly as a trajectory matching loss, or the claim of exact equivalence should be softened.","section":"Remark 1, Eq. (18)"},{"comment":"The penalty method relies on the envelope theorem identity grad_phi H(phi) = grad_phi L(theta*; phi) 'for any theta* in argmin'. The lower-level problem (16) is nonconvex because it involves compositions of normalizing flows and an MLP obstacle, and H(phi)=min_theta L(theta;phi) is generally nonsmooth when the argmin is not unique. The identity requires differentiability of the value function and typically uniqueness or at least a regular selection of the optimizer; neither condition is established. If the identity fails, the gradient used in the outer loop is not the true hypergradient. Please state concrete assumptions under which the identity holds, or provide a numerical verification (e.g., comparing the surrogate gradient with a finite-difference hypergradient on a test case).","section":"Section 3.4, Eq. (20) and following paragraph"}],"minor_comments":[{"comment":"The text reads 'In addition, we assuming the underlying MFG system' and should be 'we assume'; also 'Jeffery's divergence' should be 'Jeffreys divergence'.","section":"Section 3.1"},{"comment":"The temporal discretization is stated as 'tk = k Delta t, Delta t = K' which appears to be a typo; it should be Delta t = 1/K to be consistent with the use of Delta t in Eq. (12).","section":"Equations (11)-(12)"},{"comment":"In the description of the cylinder obstacle, the notation lists 'mu_2 = (2.5, 2), mu_2 = (-2.5, 2)' where the second 'mu_2' should be 'mu_3'.","section":"Section 4.2, Three Cylinders"},{"comment":"The captions for Figures 5 and 6 are identical and end with the incomplete phrase 'as data'; each caption should describe the specific obstacle setting and explain the left and right panels.","section":"Figures 5 and 6"},{"comment":"The semi-group notation for the flow maps is confusingly written: after defining Phi^{t_c}_{t_b} composed with Phi^{t_b}_{t_a}, the formula for F_k as a composition of Phi^{t_k}_{t_{k-1}} should be checked and stated more clearly, including the indexing of the flow subnetworks f_{theta_i}.","section":"Section 3.3, flow decomposition"},{"comment":"The baseline is described as 'naive (minimizing l2 trajectory error)', but the upper objective in (15) is also an l2 trajectory error; the distinction between the naive baseline and the proposed method is the penalty term in (20), and this should be stated explicitly.","section":"Section 4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's formulation builds heavily on the authors' own prior work [33] (trajectory-based forward MFG with normalizing flows) and [62] (bilevel inverse MFG), so the novelty statement should clarify what is new beyond combining these components. The central numerical claim is weakened by the answer-key-dependent mass regularizer, but the issue is addressable within the scope of a revision by reframing the identifiability result and adding experiments that do not use the true obstacle mass. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2507.19344.\n\nThe paper proposes a genuinely new formulation: inverse MFG as a bilevel problem in trajectory space, with normalizing flows for agent paths and an MLP for the obstacle. That is a real step beyond the Eulerian, mesh-based inverse MFG work (their own Ref. [62]) and the Gaussian-process approaches that need collocation points. The low-data experiments are the most convincing part: with tens of trajectories the method still recovers plausible obstacles and generalizes better than MLE trajectory fitting. If the method works as advertised, it opens a useful direction for high-dimensional configuration-space inference.\n\nBut the central recovery claim has a load-bearing soft spot. The physical-space identifiability section correctly notes that B and B+c give the same trajectories. The fix, Eq. (21)-(22), matches the integral of the parametrized obstacle to the integral of the true obstacle over the domain. That regularizer requires the very quantity the inverse problem is supposed to deliver: the total mass of the hidden obstacle. Calling this 'without loss of generality' because the mass 'can always be normalized to one' is not valid in the inverse setting; normalizing the true mass to one means you know it. As a result, the reported ~0.1 relative L2 errors are measured relative to an answer-key-dependent prior. Without that prior, the obstacle is only identified up to an additive constant. This does not kill the paper, but it changes the claim: the method recovers obstacle shape up to an unknown constant, and the error numbers should reflect that. The authors need to either fix the gauge in a way that does not require ground-truth mass, or report errors for the unidentifiable component separately.\n\nOther issues are more minor. Remark 1's 'equivalence' with MLE is fine up to an irrelevant variance scale, but the paper doesn't mention the scaling. No code, no error bars, no comparison with existing inverse MFG methods. The high-dimensional demonstrations are essentially 2D obstacles extended constantly, so 'high-dimensional' is oversold.\n\nOn balance: the problem is important, the formulation is new, and the empirical story is internally consistent on synthetic data. The identifiability flaw is serious enough that the headline claim should not be taken at face value, but it is fixable in a revision. I'd send it to peer review, with referees asked to focus on the regularizer and the error metric.","headline":"A genuinely new trajectory-based inverse MFG formulation whose headline obstacle-recovery claim is undermined by a regularizer that requires the ground-truth obstacle mass.","tokens_in":15146,"tokens_out":3753,"would_cite":false,"duration_ms":36542,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A16","49M37","68T07","49N90"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that partial observations of optimal agent trajectories are enough to jointly recover a hidden spatial obstacle and the full optimal paths, with about 0.1 relative L2 error in obstacle reconstruction across low- and…","keywords":["mean-field games","inverse problems","bilevel optimization","normalizing flows","trajectory inference","obstacle recovery","mesh-free methods","high-dimensional"],"falsifier":"Generate trajectories from a forward mean-field game with a ground-truth obstacle $B^\\ast$, then run the method with the mass regularizer set to a different value, for example as if $\\int_\\Omega B^\\ast=2$ instead of 1. If the recovered obstacle is not the original up to that scale, or if the relative $L^2$ error jumps far above 0.1, then the scale-normalization assumption is doing essential work. A second check: use trajectories that avoid an entire region and see whether the method correctly reports no obstacle there or instead invents one.","tokens_in":1856,"feed_emoji":"🧭","tokens_out":1855,"duration_ms":74908,"temperature":0.7,"pith_summary":"The paper seeks to solve an inverse mean-field game: from partial observations of optimal agent trajectories, recover both the unseen spatial obstacle and the full optimal trajectories. It formulates the problem as a bilevel optimization whose upper level fits a normalizing-flow-parametrized trajectory model to the data, while the lower level solves the forward mean-field game for a candidate neural-network-parametrized obstacle. The paper argues that this recovers the obstacle with roughly 0.1 relative L2 error across two-dimensional, three-dimensional, and up to ten-dimensional settings, including disjoint and multi-modal obstacles. It also interprets the bilevel constraint as a mean-field prior on maximum-likelihood trajectory learning, which improves generalization when only dozens of trajectories are available.","feed_headline":"Bilevel solver recovers obstacles from trajectory data alone","feed_subtitle":"A normalizing-flow bilevel optimization jointly infers the obstacle and optimal paths, even with dozens of trajectories.","key_machinery":"The load-bearing object is the bilevel program (15)--(16): an upper-level loss $\\frac{1}{KN}\\sum_{n=1}^N\\sum_{k=1}^K\\|F^\\ast_k(x_0^n;B)-x_k^n\\|_2^2$ matched against trajectory samples, subject to a lower-level minimizer $\\theta^\\ast$ of the discretized forward mean-field-game loss $\\hat L(\\theta;\\phi)$, which contains kinetic energy, obstacle interaction $\\sum_{m,k} B_\\phi(F_{\\hat\\theta_k}(x_m))$, and terminal divergence $D(P_1,F_{\\hat\\theta_K\\#}P_0)$. Trajectories are parametrized as compositions of normalizing-flow maps built from the semigroup property $F_k=f_{\\theta_k}\\circ\\cdots\\circ f_{\\theta_1}$, and obstacles are parametrized by an MLP $B_\\phi$. The penalty method replaces the constraint with $\\lambda_P[\\hat L(\\theta;\\phi)-H(\\phi)]_+$, and the envelope theorem gives the hypergradient as $\\nabla_\\phi H(\\phi)=\\nabla_\\phi \\hat L(\\theta^\\ast;\\phi)$, so the method never backpropagates through the lower-level optimization dynamics.","core_discovery":"On the paper's own terms, the central discovery is that the hidden obstacle and the optimal agent trajectories can be inferred jointly from trajectory samples alone, without density measurements or spatial grids. The key is to couple the data-fitting objective to a lower-level variational mean-field-game problem: for a candidate obstacle, the lower level computes the optimal trajectory flow by minimizing kinetic, interaction, and terminal costs, and the upper level then penalizes the mismatch between this flow and the observed trajectories. A mass regularizer pins down the obstacle's scale, and the penalty reformulation with the envelope theorem makes the bilevel program tractable with automatic differentiation. Because the trajectories are represented by normalizing flows rather than a mesh, the method extends to high dimensions where grid-based inverse mean-field-game solvers become intractable.","pith_inferences":["The mass regularizer assumes the integral of the true obstacle over the domain is known and normalized to one; absent that knowledge, any constant shift $B^\\ast+c$ is indistinguishable from the trajectory data, so the reported relative errors measure recovery of the shape at a preset scale rather than recovery of absolute magnitude.","Because only the product $\\lambda_I B_\\phi$ enters the lower-level objective through the interaction cost, the method can be expected to identify obstacle features only in regions that sampled trajectories actually probe, with unsampled regions filled by the MLP's inductive bias rather than by data.","The same bilevel template could infer other spatially varying mean-field-game parameters, such as drift fields or interaction kernels, by replacing the obstacle network with the appropriate parameterization while keeping the flow-based trajectory representation.","A natural testable extension is partial-time observation, where each trajectory is seen only on subintervals: the semigroup decomposition should let the upper level match only the observed temporal windows, though obstacle identifiability may degrade accordingly."],"forward_implications":["With trajectory samples only, the obstacle can be recovered at roughly 0.1 relative $L^2$ error across two-bar, flower, cylinder, castle, mountain, and Gaussian obstacle benchmarks.","Obstacle recovery and trajectory prediction succeed for disjoint and multi-modal obstacles and in dimensions 5 and 10, where mesh-based inverse mean-field-game methods are intractable.","When training data are scarce (as few as 50 to 3,000 trajectories), the mean-field-regularized bilevel objective generalizes better than plain maximum-likelihood or $L^2$ trajectory fitting, which overfit.","The bilevel training can be viewed as maximum-likelihood trajectory learning regularized by mean-field-game consistency, so the same method doubles as a regularized trajectory model.","The mesh-free normalizing-flow parametrization makes the approach applicable to inverse problems in configuration spaces, a setting the paper names as a plausible future application."],"supporting_citations":[{"why":"Supplies the trajectory-based mean-field-game formulation and the normalizing-flow parametrization of flow maps on which the bilevel problem is built.","marker":"[33]"},{"why":"Formulates inverse mean-field games as bilevel optimization in an Eulerian, mesh-based setting that this paper extends to a mesh-free trajectory setting.","marker":"[62]"},{"why":"Provides the high-dimensional neural mean-field-game solver framing and the Gaussian obstacle test setting used for the high-dimensional experiments.","marker":"[57]"},{"why":"Supplies the NSF-CL normalizing-flow architecture used to parametrize agent trajectories in the experiments.","marker":"[25]"},{"why":"Establishes the variational mean-field-game formulation that underlies the lower-level objective.","marker":"[43]"},{"why":"Provides the envelope theorem used to compute hypergradients in the penalty method without backpropagating through the lower-level solver.","marker":"[9]"}],"fun_headline_variants":["Bilevel optimization jointly infers paths and hidden obstacles","Sparse trajectory data alone recovers hidden MFG obstacles","Normalizing-flow bilevel solver reveals obstacles from sparse paths","Joint trajectory-obstacle inference in inverse mean-field games","Bilevel inverse MFG: obstacles from trajectories without density data"],"cache_read_input_tokens":17152,"weakest_assumption_plain":"The method assumes the total amount of obstacle in the domain is known and set to one; if that total is unknown, any obstacle shifted by a constant explains the trajectory data equally well, so the reported recovery accuracy holds only at that chosen scale.","fun_headline_variants_meta":{"raw":{"variants":["Bilevel optimization jointly infers paths and hidden obstacles","Sparse trajectory data alone recovers hidden MFG obstacles","Normalizing-flow bilevel solver reveals obstacles from sparse paths","Joint trajectory-obstacle inference in inverse mean-field games","Bilevel inverse MFG: obstacles from trajectories without density data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1579,"prompt_tokens":864,"completion_tokens":715,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":480,"completion_tokens_details":{"reasoning_tokens":634}},"tokens_in":480,"tokens_out":715,"duration_ms":7049,"temperature":1.0,"reasoning_tokens":634,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:55:16.173930+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate trajectories from a forward mean-field game with a ground-truth obstacle $B^\\ast$, then run the method with the mass regularizer set to a different value, for example as if $\\int_\\Omega B^\\ast=2$ instead of 1. If the recovered obstacle is not the original up to that scale, or if the relative $L^2$ error jumps far above 0.1, then the scale-normalization assumption is doing essential work. A second check: use trajectories that avoid an entire region and see whether the method correctly reports no obstacle there or instead invents one.","supporting_citations":[{"cited_title":"Bridging mean-field games and normalizing flows with trajectory regularization.Journal of Computational Physics, 487:112155, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the trajectory-based mean-field-game formulation and the normalizing-flow parametrization of flow maps on which the bilevel problem is built."},{"cited_title":"A bilevel optimization method for inverse mean-field games*","cited_arxiv_id":null,"evidence_quote":"Formulates inverse mean-field games as bilevel optimization in an Eulerian, mesh-based setting that this paper extends to a mesh-free trajectory setting."},{"cited_title":"A machine learning framework for solving high-dimensional mean field game and mean field control problems.Proceedings of the National Academy of Sciences, 117(17):9183–9193, 2020","cited_arxiv_id":null,"evidence_quote":"Provides the high-dimensional neural mean-field-game solver framing and the Gaussian obstacle test setting used for the high-dimensional experiments."},{"cited_title":"Neural spline flows.Advances in Neural Information Processing Systems, 32:7511–7522, 2019","cited_arxiv_id":null,"evidence_quote":"Supplies the NSF-CL normalizing-flow architecture used to parametrize agent trajectories in the experiments."},{"cited_title":"Mean field games.Japanese journal of mathematics, 2(1):229– 260, 2007","cited_arxiv_id":null,"evidence_quote":"Establishes the variational mean-field-game formulation that underlies the lower-level objective."},{"cited_title":"Theory of maxima and the method of lagrange.SIAM Journal on Applied Mathematics, 20(3):343–357, 1971","cited_arxiv_id":null,"evidence_quote":"Provides the envelope theorem used to compute hypergradients in the penalty method without backpropagating through the lower-level solver."}],"review_version":2}