{"id":"28d9ddd8-e97e-4180-91fe-309a8e0488e9","arxiv_id":"2508.00384","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"NIVA, a hierarchical Bayesian multi-agent traffic simulator, disentangles intention and driving-style latents and matches state-of-the-art results on the Waymo dataset.","lead":"NIVA is a new computer model for traffic simulation that generates realistic driving scenes for testing self-driving cars. It learns separate controls for a driver's intention, such as turning or going straight, and driving style, such as aggressive or cautious, making simulations more interpretable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's hard argmax intention assignment trains only one mode per agent, so the claimed intention/style controllability at sampling is not validated.","rationale":"The reader identified the hard assignment k* = argmax q(z_k) as the weakest assumption, and my review agrees that this is the load-bearing concern. The Algorithm 1 hard assignment is not merely an approximation; it is inconsistent with the stated variational objective in Eq. 12, which includes KL regularization over q(z_n) but then discards the distribution except for its mode. This directly threatens the paper's claimed controllability and disentanglement, which is one of the two pillars of the central claim. The empirical headline (competitive Waymo metrics) may survive this concern, since minADE and realism are computed over unlabeled rollouts, but the 'embellishing control' contribution would not. A quantitative test comparing hard vs. soft assignment on intention-conditioned metrics is the right way to settle it. I therefore keep the reader's CONDITIONAL verdict: the paper should not be accepted as-is without either fixing the training objective or providing quantitative evidence that the hard-assignment model preserves multimodality and controllability.","tokens_in":10355,"tokens_out":3283,"duration_ms":36687,"concrete_test":"Retrain or fine-tune NIVA without the argmax hard assignment: use the full variational distribution q(z_n) from Eq. 13 in the ELBO (e.g., sample z_n ~ q(z_n) with Gumbel-Softmax, or sum over k with weights q(z_n^k)). On the Waymo validation set, compare per-intention minADE and intention-conditioned diversity (e.g., average pairwise displacement between trajectories generated with different z tokens) against the Algorithm 1 variant. If the hard-assignment model matches the soft-assignment model on these metrics, the collapse concern is resolved; if not, the controllability claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes 'embellishing control over intentions and driving styles,' which requires that the latent variables z and b meaningfully and independently steer generated behavior after training. The paper's training procedure undermines this. In Algorithm 1, after computing q(z_n^k) by Eq. 13, the algorithm selects k* = argmax_k q(z_n^k) and computes the loss only for that single intention. This is a hard assignment: the generator is trained only on the most probable intention for each agent, despite the variational objective in Eq. 12 including a KL term for the full distribution q(z_n). Consequently, the model may never learn to produce realistic trajectories for the other K-1 intention tokens. At sampling time, Algorithm 2 draws z_n ~ Cat(pi) from the prior, so it can query intentions that were effectively never trained. The only evidence for controllability is the qualitative Fig. 3, with no quantitative validation such as per-intention minADE, intention classification accuracy, or trajectory diversity. If the latent intention space collapses in this way, the claimed disentanglement and controllability, a primary contribution, does not hold.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NIVA, a hierarchical Bayesian model for closed-loop multi-agent traffic simulation. Each agent is associated with a continuous \"behavior style\" latent b_n and a discrete \"intention\" latent z_n, and the observed states are generated through latent motion states s_t with a linear-Gaussian emission model. Training uses a mean-field variational approximation and an EM-style iterative algorithm, while sampling is autoregressive through a decoder-only Transformer with adaptive LayerNorm modulated by the latents. On the Waymo Open Motion Dataset, NIVA reports competitive simulation metrics (Table II), including the best minADE among the listed methods, and Fig. 3 qualitatively shows trajectories that change with different intentions and style values. The paper's central claims are that NIVA unifies open-loop and closed-loop simulation paradigms and provides interpretable, controllable generation of traffic behaviors.","tokens_in":10711,"tokens_out":7424,"duration_ms":79587,"significance":"If the latent variables genuinely and independently steer generated behavior, NIVA would be a valuable contribution: it is parameter-efficient (1.0M parameters), competitive on a standard benchmark, and it connects probabilistic open-loop prediction with closed-loop simulation under one generative model. The paper is also transparent in presenting its probabilistic assumptions and closed-form Gaussian computations. However, the contributions as stated are only partially supported by the current evidence: the control claim rests on a training procedure that may not train all intention modes, and the paper provides no quantitative controllability evaluation. The benchmark results are plausible but their significance is mostly confirmatory with respect to existing simulators.","major_comments":[{"comment":"The training objective is inconsistent with the stated ELBO and with the sampling procedure. After computing q(z_n^k) via Eq. (13), Algorithm 1 selects k* = argmax_k q(z_n^k) and computes the loss only for that single intention. This means the generator is trained primarily on the most probable intention per agent rather than on the full categorical distribution q(z_n) used in Eq. (12). At sampling time, Algorithm 2 draws z_n ~ Cat(pi), so it can query intention tokens that may have been trained only rarely or never for a given history. The paper provides no quantitative evidence that all K intentions produce realistic behaviors: Fig. 3 is qualitative, and there are no per-intention minADE, intention classification accuracy, controllability metrics, or trajectory diversity metrics. The abstract's claim of \"embellishing control over intentions and driving styles\" is therefore not currently supported. Please either train with the full ELBO expectation over q(z_n), use a soft/random assignment during training, or provide explicit validation that all codebook entries are meaningfully learned and controllable.","section":"Section III-E, Algorithm 1, Eq. (12)-(13), Algorithm 2"},{"comment":"The posterior precision in Eq. (10) is mis-specified: P = (σ²_s_t^{-1} + A^T(ε I)A)^{-1} should be (σ²_s_t^{-1} + A^T ε^{-2} A)^{-1} to be consistent with the marginal covariance A(σ²_s_t I)A^T + ε²I in Eq. (9) and with the posterior mean formula A^T(ε^{-2} I)o_t. As printed, the covariance term has inconsistent dimensions and would lead to incorrect posterior updates. The same issue appears in Algorithm 2, where P^{-1} is written as Σ^{-1}_{s_t} + A^T A, omitting the ε^{-2} factor. Please correct the derivation and the algorithm, or state explicitly that ε is set to 1 in all experiments.","section":"Equations (9)-(10) and Algorithm 2"},{"comment":"The generative process described in Section III-A does not match the autoregressive sampling in Algorithm 2. Equations (3)-(4) define an open-loop prior p(s_t | O≤Th, z_n) and a per-step emission p(o_t | S_t, M, b_n) with no dependence on previously sampled observations o_{<t}. Algorithm 2 instead conditions NIVA on O_{<Th+t} at each time step, and the temporal self-attention in Eq. (7) uses keys o_j for j=1,...,t-1, which for t > Th include simulated observations. This is a genuine closed-loop autoregressive model, but the formal generative model and the variational objective in Eq. (12) are written for a different, non-autoregressive process. Please make the formal generative process and the training objective consistent with the implemented sampler, or clearly explain how Eq. (3)-(4) are intended to represent the autoregressive model actually used.","section":"Section III-A vs. Algorithm 2"}],"minor_comments":[{"comment":"The text says \"In V, we compare performance...\" but should refer to Table V; also the dash for Nblocks=3 should be explained in the caption or text as a training failure due to GPU memory.","section":"Section IV-E(b), Table V"},{"comment":"The equations for the adaptive-norm block contain a likely typo: \"Q^n_{t,3} ← δ1(b_n)⊙\" is incomplete, and the subscript in the second cross-attention key/value \"m_j + R_n j\" is missing a comma. Please clarify the residual structure.","section":"Section III-C, Eq. (8)"},{"comment":"The noise vector ν_n is sampled once per agent before the t loop and then reused for every time step. This is an unusual reparameterization that couples all timesteps through the same noise; the authors should state whether this is intentional and discuss its effect on trajectory diversity.","section":"Algorithm 2"},{"comment":"The claimed unification of sequence-to-sequence prediction and next-token-prediction closed-loop models is not formally developed beyond the hierarchical generative process; a short derivation or explicit mapping to NTP would make this contribution clearer.","section":"Introduction, Contributions"},{"comment":"NIVA's minADE is the best among the listed methods, but its realism, interactive, and map-based scores are all slightly below KiGRAS. The word \"competitive\" is appropriate, but the conclusion's statement that the model \"encodes driving styles effectively\" is not backed by quantitative results and should be softened or supported.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable candidate for a robotics venue, but the core interpretability/control contribution is currently under-validated. The authors should be pushed to add quantitative controllability experiments (e.g., intention-classification accuracy, per-intention rollout quality, style-conditioned diversity) and to reconcile the formal generative model with Algorithm 2. The Eq. (10) typo is easily fixable but should be corrected before any implementation is built on it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's the short version: NIVA is a reasonable, incremental simulator paper with competitive Waymo Sim Agents numbers, but its central pitch—interpretable control over intentions and driving styles—is not actually validated, and the training algorithm gives me a concrete reason to doubt it.\n\nWhat's new: the architecture is a combination of known components (relative spacetime encoding, decoder-only Transformer, DiT-style adaptive norm, a linear Gaussian emission) organized as a hierarchical Bayesian model with continuous style and discrete intention latents. The closed-form marginalization over the latent motion state is a neat touch, and the benchmark table shows the model lands at or near the top on minADE and kinematic realism. The ablations on attended neighbors and data scale are informative.\n\nThe soft spots. First, Algorithm 1 selects a single intention per agent via argmax of q(z) and trains the generator only on that one. At sampling, Algorithm 2 draws z from the prior, so the model can be asked to generate trajectories for an intention token that was never trained for that agent/history. The only evidence for controllability is one qualitative figure. No per-intention metrics, no classification accuracy, no diversity measure. The stress-test note is on target: the disentanglement claim is unsupported. Second, the 'unification' of seq2seq and NTP models is a framing story, not a demonstrated result—fine as motivation, but it shouldn't be sold as a contribution. Third, Eq. (10) has a precision-matrix typo (ε I should be ε^{-2} I), and Algorithm 2 repeats the same inconsistency. Minor but confusing. Fourth, no error bars or code, and some hyperparameters (π, ε) are left underspecified.\n\nNone of this kills the paper. The core model is plausible and the benchmark numbers are useful. But a serious revision needs to either fix the hard-assignment training (e.g., train all K intentions with a soft assignment or add an auxiliary loss) or provide quantitative evidence that all sampled intentions are realistic.\n\nFor you: read this if you care about closed-loop traffic simulation. It deserves peer review—the model and evaluation are substantive, and the issues are addressable. I'd probably accept after major revision.","headline":"Reasonable incremental simulator, but the controllability claim is not backed by the training procedure or the evaluation.","tokens_in":11124,"tokens_out":2361,"would_cite":false,"duration_ms":23112,"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":"NIVA is a hierarchical Bayesian multi-agent traffic simulator that learns separate latents for driving style and intention and matches the top reported realism scores on the Waymo Open Motion Dataset.","keywords":["multi-agent traffic simulation","closed-loop simulation","hierarchical Bayesian model","variational inference","trajectory prediction","driving style","intention modeling","autonomous driving"],"falsifier":"Generate many rollouts from the same history with the same style code but two different intention codes, and measure the distance between the empirical distributions of their endpoints; if the two destination distributions are statistically indistinguishable, or a classifier cannot predict which intention code was used above chance, then the intention latent is inert and the disentanglement claim fails.","tokens_in":10222,"feed_emoji":"🚗","tokens_out":11534,"duration_ms":100039,"temperature":0.7,"pith_summary":"Closed-loop traffic simulators that condition each agent's next action on the evolving scene are becoming the standard for testing autonomous vehicles, but they tend to be black boxes. NIVA is a probabilistic simulator that explicitly factors each agent's behavior into a continuous driving-style variable and a discrete intention, then autoregressively samples future motion through a two-stage generative process: an open-loop prior conditioned only on history and intention, followed by closed-loop resampling conditioned on map features and other agents. The paper argues that this hierarchical Bayesian formulation unifies earlier sequence-to-sequence trajectory predictors with newer next-token-prediction simulators, and that the learned latents remain semantically meaningful. On the Waymo Open Motion Dataset, NIVA matches the realism metrics of leading baselines while reporting the lowest minimum average displacement error and the best kinematic realism, with a one-million-parameter model.","feed_headline":"Simulator splits style from intention, matches top traffic models","feed_subtitle":"The hierarchical Bayesian NIVA also controls aggressiveness and destination, enabling targeted AV testing.","key_machinery":"The machinery is a three-level hierarchical generative model. Agent count is drawn from a Poisson prior; each agent's style $b_n$ is drawn from a standard Gaussian; each agent's intention $z_n$ is drawn from a categorical prior. Conditioned on history $O_{\\le T_h}$ and intention, temporal self-attention blocks parameterize an open-loop Gaussian prior over a high-dimensional latent state $s_t$. Then adaptive-norm Transformer blocks—Transformer blocks whose layer-normalization scale and shift are MLP outputs of the style variable, conditioning attention on behavior—resample $s_t$ by cross-attending to map features and other agents, and a linear Gaussian emission model $p(o_t|s_t)$ yields closed-form marginal and posterior distributions for observations and latents. Training alternates between a variational E-step that updates the posterior of $b$ and $z$ (with hard intention assignment) and an M-step that maximizes the ELBO with respect to generative parameters.","core_discovery":"The central discovery is that a hierarchical Bayesian generative process with two latent levels—a continuous style $b_n$ drawn from a standard Gaussian and a discrete intention $z_n$ drawn from a categorical prior—can drive a decoder-only Transformer whose adaptive layer-normalization parameters are modulated by those latents, producing closed-loop multi-agent trajectories that are both realistic and coachable. The generative process first samples a latent trajectory $\\hat{s}_{T_h+t}$ from an open-loop prior conditioned only on history and intention (Eq.~3), then resamples the observation $o_{T_h+t}$ from a closed-loop distribution conditioned on map features, other agents' latent states, and the style variable (Eq.~4). A linear Gaussian emission model makes the marginal predictive distribution over $o_t$ and the posterior over the latent state $s_t$ available in closed form, which is what lets the model sample efficiently (Algorithm~2) and condition on ground-truth observations when they are available. Training maximizes an evidence lower bound with a mean-field variational family, and the E-step assigns each agent the single intention with highest posterior probability, i.e. $k^*=\\arg\\max_k q(z_k)$. The paper reads this as a Bayesian unification of open-loop sequence-to-sequence prediction and closed-loop next-token-prediction simulation, and its experimental section supports the claim that the disentangled latents control destination and aggressiveness in the generated scenarios.","pith_inferences":["A testable extension of the paper's framework would replace the hard intention assignment $k^*=\\arg\\max_k q(z_k)$ with a soft mixture-weighted objective or a Gumbel-softmax relaxation; if the hard assignment is collapsing multimodal intentions, the soft version should improve diversity metrics such as the realism score without hurting minADE.","The Poisson process for agent count is stated but then fixed to the observed $N$ in experiments, so the framework does not yet simulate variable traffic density; a natural extension is to learn the arrival rate from data and couple it to the style prior, making scenario generation open-ended.","The unification claim could be made quantitative by deriving the open-loop objective of a sequence-to-sequence model as a special case of the NIVA ELBO when the closed-loop resampling step is ignored; the paper gestures at this but does not show it formally."],"forward_implications":["Autonomous-vehicle developers could use NIVA to generate large batches of scenarios with chosen intentions and driving styles, enabling targeted testing of rare or adversarial situations such as aggressive cut-ins.","The closed-form marginal and posterior for the emission model mean the same trained parameters support both open-loop prediction and closed-loop simulation, so a single model can serve trajectory forecasting and scenario generation.","The one-million-parameter size of the model (versus tens of millions for some baselines) suggests the disentangled hierarchical prior is an efficient inductive bias, and the model degrades gracefully when trained on as little as 1 percent of the data.","Because the ELBO objective reduces to minimizing the data log-likelihood plus KL penalties on style and intention posteriors, the framework supplies a principled objective for future simulators that want to condition on interpretable factors.","If the latent factors are truly disentangled, the model should allow interpolation between aggressive and conservative driving styles, giving a continuous dial for scenario difficulty."],"supporting_citations":[{"why":"Supplies the Waymo Open Motion Dataset used to train and evaluate NIVA, including the 1.1-second history and 8-second future trajectories.","marker":"[2]"},{"why":"Defines the Waymo Sim Agents Challenge benchmark and the realism, kinematic, interactive, and map-based metrics that Table II reports.","marker":"[5]"},{"why":"Supplies the closed-loop formulation for unrolling multi-agent simulations that NIVA extends with latent style and intention variables.","marker":"[7]"},{"why":"TrafficBotsV1.5 is a primary baseline in Table II representing conditional-VAE closed-loop simulation.","marker":"[12]"},{"why":"BehaviorGPT is the next-token-prediction baseline that NIVA claims to unify with open-loop sequence models.","marker":"[13]"},{"why":"The Diffusion Transformer block with adaptive layer normalization is the direct inspiration for NIVA's style-conditioned attention blocks.","marker":"[23]"}],"fun_headline_variants":["NIVA splits style and intention in closed-loop simulation","Bayesian NIVA simulator controls style and intentions","NIVA: hierarchical Bayesian traffic simulation with controls","Probabilistic multi-agent sim separates style and intent"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"NIVA's controllability relies on the assumption that fixing each agent to a single intention via the argmax in training preserves the multimodality of real driver intentions; if that hard assignment collapses the latent space, the style and intention controls will not behave as advertised.","fun_headline_variants_meta":{"raw":{"variants":["NIVA splits style and intention in closed-loop simulation","Bayesian NIVA simulator controls style and intentions","NIVA: hierarchical Bayesian traffic simulation with controls","Probabilistic multi-agent sim separates style and intent"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1363,"prompt_tokens":960,"completion_tokens":403,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":342}},"tokens_in":576,"tokens_out":403,"duration_ms":4520,"temperature":1.0,"reasoning_tokens":342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:10:33.435318+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate many rollouts from the same history with the same style code but two different intention codes, and measure the distance between the empirical distributions of their endpoints; if the two destination distributions are statistically indistinguishable, or a classifier cannot predict which intention code was used above chance, then the intention latent is inert and the disentanglement claim fails.","supporting_citations":[{"cited_title":"Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the Waymo Open Motion Dataset used to train and evaluate NIVA, including the 1.1-second history and 8-second future trajectories."},{"cited_title":"The waymo open sim agents challenge,","cited_arxiv_id":null,"evidence_quote":"Defines the Waymo Sim Agents Challenge benchmark and the realism, kinematic, interactive, and map-based metrics that Table II reports."},{"cited_title":"Trafficsim: Learning to simulate realistic multi-agent behaviors,","cited_arxiv_id":null,"evidence_quote":"Supplies the closed-loop formulation for unrolling multi-agent simulations that NIVA extends with latent style and intention variables."},{"cited_title":"Behaviorgpt: Smart agent simulation for autonomous driving with next-patch prediction,","cited_arxiv_id":null,"evidence_quote":"BehaviorGPT is the next-token-prediction baseline that NIVA claims to unify with open-loop sequence models."},{"cited_title":"Scalable diffusion models with transform- ers,","cited_arxiv_id":null,"evidence_quote":"The Diffusion Transformer block with adaptive layer normalization is the direct inspiration for NIVA's style-conditioned attention blocks."}],"review_version":1}