{"id":"665dab91-ed72-4da0-b68f-68e82d82a77b","arxiv_id":"2608.01636","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A forward-inverse dynamic game framework with state-dependent KL regularization and a scene-aware network for inferring agent cost weights from demonstrations.","lead":"This paper combines a KL-regularized dynamic game with a neural inverse-learning module to plan multi-agent trajectories when agent objectives are unknown and interaction strength varies in space. A state-dependent weight adaptively balances optimality against learned behavioral priors, and the authors claim better collision avoidance and cost recovery than existing game-theoretic planners.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's linear Riccati recursion Eq. (13b) has a sign error: the KL term should contribute +λPᵀΣ̃⁻¹(α+μ̃), not (α−μ̃), so the claimed FBNE affine offsets are wrong even under the paper's PSD assumptions.","rationale":"The reader's verdict is REJECT with moderate confidence, and my independent read agrees that the central theoretical claim is not sound as stated. The reader highlighted the PSD-assumption violation as the weakest assumption: the collision-avoidance primitive f₂ in Eq. (26) has an indefinite Hessian, so Proposition 1's premise Qᵢ_t⪰0 is not satisfied in the paper's own CAMP and RM experiments. That is a genuine and load-bearing gap. However, the single most direct mathematical defect is the sign error in Eq. (13b), which makes the value recursion in Theorem 1 algebraically incorrect even under the paper's assumptions. The scalar LQ counterexample above isolates this cleanly: the printed recursion produces the wrong sign for the KL contribution to the affine value coefficient, and since z_t propagates backward into Eq. (12b), the solver's action offsets are not the FBNE offsets. This is not a disagreement with a different community's consensus; it is an internal inconsistency with the paper's own derivation in Appendix C. I also note that the PSD violation is independently fatal for the claimed per-iteration well-posedness guarantee, because an indefinite Qᵢ_t can make Zᵢ_t indefinite and destroy the uniform positive definiteness of Mᵢ,k on which covariance well-posedness rests. The paper's experiments cannot rescue the theory here: no formal verification is provided, and the ablation tables contain internal inconsistencies (e.g., D_par values described in the text do not match the reported advantage of the full model). Both defects support the reader's REJECT verdict, so I recommend no change to the verdict. The good-faith path forward would be to correct Eq. (13b), add an explicit convexification or verify PSD-ness of the actual Qᵢ_t used in experiments, and then re-run the reported comparisons.","tokens_in":20436,"tokens_out":13410,"duration_ms":114992,"concrete_test":"Re-run the one-step scalar LQ verification: set A=B=Q=R=λ=Σ̃=Z_{t+1}=1, μ̃=1, l=r=z_{t+1}=0, compute P=1/3 and α=−1/3 from Eqs. (12), then compare z_t from Eq. (13b) with the linear coefficient of the exact Bellman value V_t(δ)=min_u[½δ²+½u²+½(u−1)²+½(δ+u)²]. The direct value has linear coefficient 1/3; Eq. (13b) gives −1/3. If the mismatch reproduces, the printed recursion is wrong and the affine feedback offsets in Theorem 1 must be corrected before any experimental claims are re-evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing defect is an algebraic sign error in Theorem 1, Eq. (13b). In Appendix C, after substituting μᵢ = −Pᵢδx − αᵢ into the KL term (B) of Eq. (49), the quadratic form (μᵢ−μ̃ᵢ)ᵀ(Σ̃ᵢ)⁻¹(μᵢ−μ̃ᵢ) expands with μᵢ−μ̃ᵢ = −Pᵢδx − (αᵢ+μ̃ᵢ). Collecting the linear coefficient in δx therefore gives +λᵢ Pᵢᵀ(Σ̃ᵢ)⁻¹(αᵢ+μ̃ᵢ), not the printed +λᵢ Pᵢᵀ(Σ̃ᵢ)⁻¹(αᵢ−μ̃ᵢ). A one-step scalar check with A=B=Q=R=λ=Σ̃=Z_{t+1}=1, μ̃=1, and l=r=z_{t+1}=0 yields the true value coefficient z_t = 1/3 by direct Bellman evaluation, while Eq. (13b) gives z_t = −1/3. Because z_t is computed backward and feeds Eq. (12b) for α_t at earlier times, the solver's predicted affine feedback laws are not those of the claimed Gaussian feedback Nash equilibrium, even in the perfectly convex regime where Proposition 1's PSD assumptions hold. Separately, the paper's own cost primitive f₂ = −log‖pᵢ−pⱼ‖² in Eq. (26) has an indefinite Hessian, so the premise Qᵢ_t⪰0 of Proposition 1 is violated in the reported experiments; this second issue reinforces the rejection, but the sign error is the more direct falsifier of the central theorem.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a forward–inverse dynamic game framework for multi-agent trajectory planning. The forward module iteratively linearizes nonlinear dynamics and costs around a nominal trajectory and solves a KL-regularized linear–quadratic game with a state-dependent weight λ_i(x_t), claiming a closed-form Gaussian feedback Nash equilibrium (Theorem 1), per-iteration well-posedness (Proposition 1), and Lipschitz continuity of λ_i (Proposition 2). The inverse module uses maximum-entropy IRL with a scene-aware neural network and a physics-informed rule loss to infer agent cost weights from demonstrations. Experiments are reported for cooperative navigation and ramp merging, including ablation studies, sensitivity analysis, and hardware tests.","tokens_in":20906,"tokens_out":11188,"duration_ms":92158,"significance":"If the theoretical results were correct, the adaptive KL-regularization scheme and the scene-aware inverse module would be a useful practical contribution: the state-dependent λ provides an intuitive mechanism to tighten reference-policy adherence near hazards, and the Lipschitz property supports stable updates. The paper is explicit about assumptions and gives proofs for Lemma 1, Theorem 1, and Propositions 1–2, and the ablation design cleanly separates the contributions of the adaptive weight and the learned cost weights. However, the central theorem contains an algebraic sign error, the main well-posedness proposition presupposes a convexity condition violated by the paper's own cost primitive, and the simulation evaluation is self-referential because the 'expert' demonstrations are generated by the same forward simulator used for rollouts. These issues are load-bearing, not cosmetic.","major_comments":[{"comment":"The KL term in Eq. (49) expands with μ_i − μ̃_i = −P_i δx − (α_i + μ̃_i), so the linear coefficient in δx is +λ_i P_i^T(Σ̃_i)^{-1}(α_i + μ̃_i), not +λ_i P_i^T(Σ̃_i)^{-1}(α_i − μ̃_i) as printed in Eq. (13b). A one-step scalar check with A=B=Q=R=λ=Σ̃=Z_{t+1}=1, μ̃=1, and l=r=z_{t+1}=0 gives the true value coefficient z_t=1/3 by direct Bellman evaluation, whereas Eq. (13b) gives z_t=−1/3. Since z_t propagates backward into Eq. (12b), the affine feedback offsets computed by Algorithm 1 are not those of the claimed Gaussian FBNE, even under the PSD assumptions of Proposition 1.","section":"Section IV-A, Eq. (13b) and Appendix C"},{"comment":"Proposition 1 is stated under Q_i_t⪰0, R_ij_t⪰0, and R_ii_t≻0 (Section IV-A, after Eq. (5)). In the experiments, the collision-avoidance primitive f_i2 = −Σ_{j≠i} log(||p_i − p_j||^2) in Eq. (26) is part of c_i_t. In two dimensions the Hessian of −log||p_i − p_j||^2 has one positive and one negative eigenvalue, so Q_i_t is indefinite at generic configurations. Therefore the premise of Proposition 1 is violated by the planner as actually implemented; the proof that Z_i_t⪰0 and Σ_i*_t≻0 does not go through, and the claimed per-iteration well-posedness and uniqueness are not established for the reported experiments. The authors need to either use a convex surrogate, project the quadraticized Hessian, or prove the relevant statements under weaker assumptions.","section":"Section IV-A, Proposition 1, and Section V-A.3, Eq. (26)"},{"comment":"The Conclusion states that 'expert demonstrations are currently obtained from forward simulation rather than naturalistic datasets such as INTERACTION.' Because Algorithm 2 also generates its rollout set with the same forward simulator (Lines 4–8), the quantitative metrics D_par, D_tra, and Coll measure the inverse module's ability to reproduce the simulator's own cost model, not its ability to recover genuine human objectives. The real-robot experiments are qualitative and do not report inverse-game metrics. The paper's claim of validating inference of 'unknown agents' objectives' is therefore unsupported by the presented evidence.","section":"Section VI, Conclusion"}],"minor_comments":[{"comment":"The terminal condition is stated as z_i_T = l_i_T in Eq. (13b), but the base case in Appendix C says z_i_T = 0 provided x̄_T = x_ref; please reconcile the two conventions.","section":"Eq. (13b) and Appendix C"},{"comment":"The text says Ours(fixedλ) 'reduces D_par from 1.18 (KLGame) to 1.49'; since 1.49 is larger than 1.18, the wording should be corrected or the numbers checked against Table I.","section":"Section V-B, 4-agent CAMP"},{"comment":"There are several typographical errors, including 'admits admits' in Theorem 1, 'positon exchange' in Section V-A.2, and 'T rajectories' in Fig. 8; a careful proofread is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"For the editor: the sign error in Eq. (13b) is readily reproducible and is decisive in my assessment; I would not regard it as a mere typographical issue because the proof explicitly derives the wrong coefficient. The PSD-premise violation and the self-referential evaluation reinforce the rejection. If the authors can correct the recursion, re-prove the theory under assumptions that cover their cost model, and redo the evaluation with naturalistic demonstrations, a substantially revised manuscript could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this paper has a genuinely useful idea—making the KL regularization weight state-dependent—but the main theorem as written is wrong, and I would not trust the experimental claims as support. It deserves a serious referee, but not acceptance on this version.\n\nWhat's new: Eq. (14) replaces KLGame's fixed lambda_i with a state-dependent lambda_i(x_t) that tightens reference-policy adherence near obstacles and other agents. That's a plausible, practically motivated extension. The scene-aware network (Eq. 19) with cross-attention over trajectories and maps is a reasonable way to recover cost weights. Proposition 2 (Lipschitz continuity of lambda_i) is fine. There are real hardware experiments, and the paper is honest in the conclusion that expert trajectories come from forward simulation rather than naturalistic data.\n\nThe soft spots are load-bearing. The sign error in Theorem 1 identified by the stress-test is real. From Appendix C, mu_i - mu_tilde_i = -P_i delta_x - (alpha_i + mu_tilde_i), so the linear KL term in the z_t recursion should be +lambda P_i^T Sigma_tilde^{-1}(alpha_i + mu_tilde_i), not +lambda P_i^T Sigma_tilde^{-1}(alpha_i - mu_tilde_i). The scalar check gives z_t = 1/3 by direct Bellman evaluation versus -1/3 from Eq. (13b). Since z_t propagates backward and feeds alpha_t, the claimed FBNE affine strategies are not the equilibrium of the local game, even in the perfectly convex regime. That alone falsifies the central claim. In addition, Proposition 1 assumes Q_i_t >= 0, but the paper's own collision cost f_2 = -log||p_i - p_j||^2 has an indefinite Hessian, so that premise is violated in the reported experiments. The experiments also contradict themselves: the text claims collision-free results, but Table I shows Ours with Coll = 2 in the 4-agent CAMP scenario. And the ablation text saying Ours(fixed lambda) \"reduces D_par from 1.18 to 1.49\" is backwards.\n\nSo the idea has merit, and the engineering is real, but the theory as written is not sound and the empirical story is not coherent. This is not a desk reject—the state-dependent weighting is worth refereeing and the errors are fixable. If it crossed my desk, I would send it to review and expect a major revision or reject-and-resubmit.","headline":"The state-dependent KL weighting is a genuinely good idea, but the central FBNE theorem has a load-bearing sign error and the paper's own collision cost violates the theorem's PSD premise; this version should be rejected, though the idea deserves a serious revised look.","tokens_in":21433,"tokens_out":3350,"would_cite":false,"duration_ms":28706,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a state-dependent KL-regularized dynamic game, solved by freezing the regularization coefficient along a nominal trajectory, has a unique Gaussian feedback Nash equilibrium at each iteration and that the adaptive…","keywords":["multi-agent trajectory planning","feedback Nash equilibrium","KL-regularized dynamic game","state-dependent regularization","maximum-entropy inverse reinforcement learning","scene-aware cost inference","collision avoidance"],"falsifier":"Evaluate the Hessian of the collision-avoidance cost $f_2 = -\\log(\\lVert p_i - p_j \\rVert^2)$ at a typical near-collision configuration; if it has a negative eigenvalue, the Proposition 1 premise $Q_{i,t} \\succeq 0$ fails. Then run the two-agent value recursion (13a) with that cost: if $Z_{i,t}$ loses positive semidefiniteness at any step, the claimed unique Gaussian FBNE per iteration is not assured for the planner as implemented.","tokens_in":20231,"feed_emoji":"🚗","tokens_out":6165,"duration_ms":52950,"temperature":0.7,"pith_summary":"This paper tries to establish that multi-agent trajectory planning under unknown goals can be treated as a forward–inverse dynamic game in which a state-dependent KL-regularization weight balances optimal control against a learned reference policy. The forward game is solved by linearizing dynamics and costs around a nominal trajectory and freezing the regularization coefficient per iteration, yielding a local linear–quadratic game. The paper claims each such local game has a unique Gaussian feedback Nash equilibrium with well-defined covariances and value recursions (Proposition 1), and that the adaptive weight is Lipschitz in the joint state (Proposition 2). The inverse module infers each agent's cost weights from demonstrations using maximum-entropy IRL with a scene-aware network and a physics-informed rule loss. If correct, this gives a principled way to make game-theoretic planners safer in dense, safety-critical scenes without committing to full rationality or fixed imitation strength.","feed_headline":"Adaptive game weights steer robots safely through dense traffic","feed_subtitle":"KL regularization that tightens near obstacles and rivals cuts collisions to zero in merging tests","key_machinery":"The mechanism that carries the argument is the state-dependent regularization coefficient $\\lambda_i(x_t) = \\lambda_i^{\\min} + (\\lambda_i^{\\max} - \\lambda_i^{\\min}) \\exp\\!\\left(-(d_i^{\\mathrm{obs}}(x_t))^2/(2\\sigma^2)\\right)$, where $d_i^{\\mathrm{obs}}(x_t)$ is the distance from agent $i$ to the nearest obstacle or other agent. Evaluated on the nominal trajectory and frozen during each inner Riccati recursion, this coefficient makes the local game quadratic and well-posed while letting proximity tighten or relax adherence to the reference policy. The analytic Gaussian policy from Lemma 1 and the coupled Riccati-style recursions for the gains and value parameters are the objects that carry the equilibrium computation.","core_discovery":"The central claim is that adding a state-dependent weight $\\lambda_i(x_t)$ to the KL-regularized dynamic game preserves the tractable structure needed for feedback Nash equilibrium computation while adapting imitation strength to local risk. The paper derives a Gaussian feedback equilibrium for the frozen local game: each agent's optimal policy is Gaussian with mean $-P_{i,t}\\delta x_t - \\alpha_{i,t}$ and covariance $\\lambda_{i,t}(M_{i,k,t})^{-1}$, and it proves per-iteration well-posedness and Lipschitz continuity of the weight. Alongside this, the inverse module recovers cost weights from demonstrated trajectories via a scene-aware network trained with maximum-entropy IRL and a rule penalty, and the experiments report lower collision counts and trajectory errors than fixed-$\\lambda$ and KKT-based baselines, including zero collisions in a five-agent ramp-merging scenario.","pith_inferences":["The PSD assumption gap suggests a convex surrogate for the log-barrier, or an explicit projection of $Q_{i,t}$ onto the positive semidefinite cone, would restore Proposition 1's guarantee for the same planner.","The state-dependent $\\lambda_i$ can be read as a safety-margin proxy; a testable corollary is that collision counts should fall monotonically as $\\lambda_i^{\\max}$ rises in dense scenes, which the 20-agent ablation partially supports.","The inverse module's scene-aware weighting should be tested on naturalistic human-driving data, which the paper itself lists as future work, to separate the benefit of map context from the benefit of extra training data."],"forward_implications":["Near obstacles and dense interactions, agents automatically follow the learned reference policy more closely, while in open areas they can optimize more freely.","Each outer iteration of the solver is a frozen linear–quadratic game with a unique Gaussian feedback Nash equilibrium, so the iterative scheme stays well-posed before convergence.","Cost weights inferred by the scene-aware inverse network are structurally consistent with the forward solver through the rule-based penalty, improving parameter and trajectory recovery.","In the reported experiments, the full adaptive method achieves zero collisions in the five-agent ramp merging scenario and lower collision counts than fixed-lambda and KKT baselines in a 20-agent scenario."],"supporting_citations":[{"why":"Supplies the KL-regularized dynamic game baseline with a fixed coefficient that this paper extends to a state-dependent weight.","marker":"[4]"},{"why":"Provides the iterative linear-quadratic approximation scheme for nonlinear multiplayer games used to freeze the local game each outer iteration.","marker":"[26]"},{"why":"Supplies the maximum-entropy forward and inverse dynamic game formulation that motivates the inverse module.","marker":"[14]"},{"why":"Provides the KKT-based inverse game baseline and the formulation of objective inference from partial observations.","marker":"[7]"},{"why":"Provides the maximum-entropy IRL principle used to score expert trajectories against rollouts.","marker":"[31]"},{"why":"Provides the deep maximum-entropy IRL framework adapted for the scene-aware cost-weight network.","marker":"[29]"},{"why":"Provides the vectorized map encoding used in the scene encoder for map context.","marker":"[30]"}],"fun_headline_variants":["Adaptive game weights eliminate collisions in robot merging tests","State-dependent game weights cut robot collisions to zero","Adaptive KL game weights balance imitation and safety for robots","Forward-inverse game framework with adaptive weights for merging robots","Adaptive rationality in multi-agent games yields safer merging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the local quadratic cost Hessians are positive semidefinite; the paper's own collision-avoidance log-barrier has an indefinite Hessian, so this premise is not met in the experiments, and if it fails the per-iteration uniqueness guarantee can collapse.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive game weights eliminate collisions in robot merging tests","State-dependent game weights cut robot collisions to zero","Adaptive KL game weights balance imitation and safety for robots","Forward-inverse game framework with adaptive weights for merging robots","Adaptive rationality in multi-agent games yields safer merging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000627,"raw_usage":{"total_tokens":2873,"prompt_tokens":890,"completion_tokens":1983,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":1906}},"tokens_in":506,"tokens_out":1983,"duration_ms":11962,"temperature":1.0,"reasoning_tokens":1906,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:06:59.674569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the Hessian of the collision-avoidance cost $f_2 = -\\log(\\lVert p_i - p_j \\rVert^2)$ at a typical near-collision configuration; if it has a negative eigenvalue, the Proposition 1 premise $Q_{i,t} \\succeq 0$ fails. Then run the two-agent value recursion (13a) with that cost: if $Z_{i,t}$ loses positive semidefiniteness at any step, the claimed unique Gaussian FBNE per iteration is not assured for the planner as implemented.","supporting_citations":[{"cited_title":"Blending data-driven priors in dynamic games,","cited_arxiv_id":null,"evidence_quote":"Supplies the KL-regularized dynamic game baseline with a fixed coefficient that this paper extends to a state-dependent weight."},{"cited_title":"Efficient iterative linear-quadratic approximations for nonlinear multi- player general-sum differential games,","cited_arxiv_id":null,"evidence_quote":"Provides the iterative linear-quadratic approximation scheme for nonlinear multiplayer games used to freeze the local game each outer iteration."},{"cited_title":"Maximum-entropy multi-agent dynamic games: Forward and inverse solutions,","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-entropy forward and inverse dynamic game formulation that motivates the inverse module."},{"cited_title":"Online and offline learning of player objectives from partial observations in dynamic games,","cited_arxiv_id":null,"evidence_quote":"Provides the KKT-based inverse game baseline and the formulation of objective inference from partial observations."},{"cited_title":"Maximum entropy inverse reinforcement learning","cited_arxiv_id":null,"evidence_quote":"Provides the maximum-entropy IRL principle used to score expert trajectories against rollouts."},{"cited_title":"Vectornet: Encoding hd maps and agent dynamics from vectorized rep- resentation,","cited_arxiv_id":null,"evidence_quote":"Provides the vectorized map encoding used in the scene encoder for map context."}],"review_version":2}