{"id":"fcc1f892-23b5-4ece-bd09-4659815f12ba","arxiv_id":"2502.03286","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A learned driver model is embedded in closed-loop simulation to produce trajectory predictions conditioned on candidate plans of the automated vehicle.","lead":"This paper builds a traffic simulator where each car is controlled by a behavior model learned from real driving data, so an automated vehicle can ask what happens if it follows a particular plan. It shows one intersection example where choosing to brake versus go first changes the predicted behavior of other drivers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central conditional-prediction claim rests on a single qualitative example; the AIRL policy's reaction to externally imposed AV trajectories is never quantitatively validated, so out-of-distribution generalization remains undemonstrated.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the AIRL policy must generalize to AV-conditioned trajectories that differ from the self-play distribution, and the only evidence is a qualitative example. My reading of Section 2.4 and Figure 4 confirms that the conditional prediction mechanism is an action-level intervention rather than a full planner-consistent trajectory, and no quantitative conditional evaluation is provided. This is not an internal inconsistency; the method is plausible, and the unconditional results in Table 2 are a useful baseline. However, the paper's headline contribution is conditional prediction, and that contribution is currently supported only by one figure with no metrics, baselines, or ablation. The proposed concrete test would directly measure whether conditioning on an externally imposed plan produces realistic multi-agent reactions, thereby settling whether the central claim holds. Since the reader's verdict already flags this as CONDITIONAL, no change to the verdict is needed.","tokens_in":7605,"tokens_out":3238,"duration_ms":34811,"concrete_test":"Select a held-out set of INTERACTION test scenes with interactive situations (e.g., all-way stops and merges). In each scene, designate one vehicle as the AV and feed its recorded ground-truth action sequence from the dataset as the conditioned plan, while simulating all remaining vehicles with the learned AIRL policy for 10 s. Compare the predicted trajectories, collision rate, and off-track rate of the remaining vehicles against the ground-truth recordings, and also compare against the unconditional self-play simulation from Table 2. If the conditional simulation is not at least as accurate as the unconditional one, or if the predicted reactions diverge sharply from ground truth in interaction metrics, then the central claim that the model can evaluate alternative AV strategies is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that predictions are generated by a microscopic simulation in which surrounding agents react to candidate AV trajectories, thereby enabling evaluation of different planning strategies (Abstract, Section 2.4). For this claim to hold, the learned AIRL policy must produce realistic reactions when one agent's actions are replaced by a planner-defined plan rather than by the policy's own actions. During training, all agents use the same shared policy (Section 2.1), so the policy only experiences closed-loop rollouts in which every agent follows the learned policy. The conditional experiment in Figure 4 instead forcibly replaces vehicle ID 6's accelerations with constant braking for the first 5 seconds. This intervention places the other agents' observations outside the training distribution in a way that is never evaluated quantitatively. Table 2 reports only unconditional prediction performance, where all agents follow the learned policy; no conditional metric, baseline, or held-out comparison is provided. The single qualitative example, while suggestive, cannot establish that the model correctly predicts how human drivers would respond to arbitrary candidate AV trajectories. Additionally, Section 2.4 explicitly states that dynamic plan adaptation is left to future work and fixed paths are used, so the abstract's claim that candidate trajectories adapt dynamically is not demonstrated. The absence of quantitative conditional evaluation is the load-bearing gap.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conditional motion prediction framework for automated driving. A behavior policy is trained via Adversarial Inverse Reinforcement Learning on the INTERACTION dataset and then used as a closed-loop simulator in which all vehicles execute the same shared policy. To make conditional predictions, the automated vehicle's actions are replaced by a manually defined plan, and the other agents' reactions are simulated. The authors report unconditional prediction performance (RMSE, collision rate, off-track rate) showing AIRL outperforms behavior cloning, and provide a single qualitative example of conditional prediction at an all-way stop intersection.","tokens_in":7863,"tokens_out":6893,"duration_ms":65925,"significance":"The contribution is potentially significant: if the learned policy produces realistic responses to alternative AV plans, it would enable evaluating cooperative strategies and mitigate the freezing robot problem. The unconditional results are encouraging, particularly the low collision and off-track rates, and the use of AIRL for a closed-loop model is a sensible direction. However, the paper's central conditional claim is not quantitatively evaluated: the single example is illustrative only, with no conditional metrics, baselines, or error bars, and the paper overclaims dynamic plan adaptation. These gaps must be addressed before the contribution can be regarded as demonstrated.","major_comments":[{"comment":"The paper's central claim—that the framework produces meaningful conditional predictions for alternative AV plans—is supported only by a single qualitative example. There are no conditional metrics, no baselines (e.g., a reactive model that ignores the AV plan), no error bars, and no systematic evaluation across scenarios. Because the AIRL policy is trained in closed-loop settings where all agents follow the same policy, imposing an external plan as in Figure 4(b) places other agents out of distribution; the paper does not test whether the resulting reactions are realistic or consistent with human behavior. Without such evidence, the conditional-prediction claim is not demonstrated. Concretely, the authors should add quantitative conditional evaluations: in several scenarios, compare conditional rollouts against a reactive baseline, measure the distribution of predicted actions under conditioning against expert action statistics, and report how often the conditioned plan leads to collisions or off-track events rather than removing those vehicles.","section":"Section 3.1 (Conditional Prediction paragraph, Figure 4) and Section 2.4"},{"comment":"The abstract states that \"our approach allows the candidate trajectories to adapt dynamically during the prediction rollout,\" but Section 2.4 explicitly says \"we leave this for future research and use manually defined fixed paths for the A V during prediction.\" The Introduction similarly claims the framework allows the A V to adapt its planned trajectory dynamically. This dynamic adaptation is neither implemented nor tested in the manuscript. The authors should either implement a reactive planner or revise the abstract and introduction to match the fixed-path implementation actually used.","section":"Abstract and Section 2.4"},{"comment":"The sentence \"Vehicles that reach the end of their assigned route, leave the track, or collide are removed from the scene and the simulation is continued with the remaining vehicles\" introduces a bias into the reported statistics. Removing vehicles that leave the track or collide excludes the most severe failures from the RMSE computation, potentially understating errors. This is especially problematic for conditional prediction, where a conditioned plan that induces collisions would be silently censored rather than reported as a failure. Please report results with and without removal, or at least analyze how this design choice affects the RMSE and the collision/off-track rates, and specify how RMSE is computed for vehicles that are removed before the prediction horizon ends.","section":"Section 3 (first paragraph of Experiments)"},{"comment":"The claim that the reward offset c \"promot[es] survival without changing the optimal behavior with respect to the discriminator model\" requires qualification in the presence of terminal states. If episodes terminate on off-track or collision, adding a positive constant c to every step changes the value of surviving and can alter the optimal policy by making the agent more averse to risky states. The manuscript does not provide a theoretical justification or empirical evidence that the learned policy's behavior is unaffected by c. Please clarify this point or justify the choice of c more rigorously.","section":"Section 2.2, Eq. (4) and the Modifications paragraph"}],"minor_comments":[{"comment":"The table reports RMSE after 10 s, but the text does not define how RMSE is computed (e.g., mean Euclidean position error over time? per vehicle? how are vehicles matched between prediction and ground truth?). Please specify the metric precisely.","section":"Section 3.1 (Table 2)"},{"comment":"The paper mentions \"Several example scenarios are available at [website]\" but the main text includes only one conditional example. Including additional conditional examples or quantitative summaries in the paper would strengthen the presentation of the central contribution.","section":"Section 3.1 (Conditional Prediction paragraph)"},{"comment":"The ground-truth panel (c) corresponds to the unconditional prediction (a), not to the conditional prediction (b). The authors should clarify that the ground truth does not validate the counterfactual, and that the conditional prediction is an illustration rather than a quantitatively evaluated result.","section":"Figure 4"},{"comment":"The policy decoder outputs the mean and standard deviation of the next acceleration and steering angle, but the distribution family is not stated. Please specify (e.g., independent Gaussians) so that Eq. (3) and the action likelihood are unambiguous.","section":"Section 2.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's topic fits well with venues on interactive prediction and planning for automated driving, and the unconditional results are solid. However, the central conditional-prediction claim is currently supported only by a single qualitative example, and the abstract overstates dynamic plan adaptation. The proposed experiments are feasible and would substantially improve the manuscript, so major revision rather than rejection seems appropriate. I would also emphasize to the authors that the evaluation protocol for conditional prediction needs to be designed carefully, especially regarding the removal of collided or off-track vehicles."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this is a clean, honest paper that extends the authors' prior AIRL driver model to conditional prediction by simulation. The integration is new: learning a behavior policy with AIRL and then rolling it out in closed loop while one agent's actions are replaced by a planned trajectory. The unconditional evaluation is solid as far as it goes: seven seeds, RMSE plus collision/off-road rates, and it clearly beats behavior cloning on robustness. The writing is clear, and the paper is upfront that dynamic plan adaptation is future work rather than overselling it.\n\nThe soft spot is exactly where the title points. The conditional-prediction claim is the whole contribution, and it is supported by one qualitative figure. There are no conditional metrics, no baselines, no error bars, and no code. The scenario in Figure 4 is suggestive—forcing the right-turning vehicle to brake does change the predicted trajectories of two other vehicles in the expected direction—but it does not establish that the model correctly predicts human responses to arbitrary candidate AV plans. The policy was trained in closed-loop rollouts where every agent follows the same learned policy, so overwriting one agent's actions with an external plan puts the other agents' observations out of the training distribution. That generalization may be fine, but it is unevaluated. The practice of removing collided or off-track vehicles also flatters the reported statistics, though this is a standard trick and only a minor concern here.\n\nI disagree with any suggestion that the paper is circular. The behavior model is trained on INTERACTION and evaluated on held-out INTERACTION data, which is normal supervised evaluation. The hand-tuned reward constant c=5 does not directly control the conditional outcome. The authors cite the relevant prior work, including their own; that is not a flaw.\n\nBottom line: the idea is sensible, the unconditional experiments are above average for this subfield, and the central gap is a missing evaluation rather than a demonstrated flaw. I would send it to peer review, but I would push hard for quantitative conditional evaluation—multiple scenarios, at least one baseline that ignores the conditioning, and error bars over seeds. As it stands, it is a workshop-quality demonstration of concept, not a fully supported method.\n\nRecommendation: engage with it, but treat the conditional claim as unproven until the experiments appear.","headline":"A plausible integration of AIRL into conditional prediction, but the central claim rests on a single qualitative example.","tokens_in":8337,"tokens_out":2056,"would_cite":false,"duration_ms":18329,"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":"When a self-driving car changes its plan, the simulation changes how other drivers are predicted to react.","keywords":["conditional prediction","automated driving","multi-agent behavior modeling","adversarial inverse reinforcement learning","traffic simulation","closed-loop prediction","cooperative planning","INTERACTION dataset"],"falsifier":"Record or collect real-world cases in which a driver takes an alternative action, such as braking at an all-way stop instead of proceeding first, run the conditional simulation with that action as the conditioned plan, and compare the simulated trajectories of the other agents against the ground-truth trajectories from those cases. A quantitative divergence on such a conditional benchmark would show that the model generalizes to conditioned plans only in appearance, not in distribution.","tokens_in":7419,"feed_emoji":"🚗","tokens_out":7746,"duration_ms":69756,"temperature":0.7,"pith_summary":"Modular driving stacks usually predict what other vehicles will do and then plan around that forecast, which leaves no room for maneuvers that depend on other drivers reacting to the car's own behavior. This paper argues that prediction and planning should be coupled: the automated vehicle queries the prediction model with a candidate trajectory, and the model simulates the scene under that assumption, letting other agents respond step by step. The authors show that a closed-loop simulation in which every surrounding vehicle is controlled by a behavior policy learned from human demonstrations produces scene-consistent forecasts and can propagate the effect of a changed plan. For example, if the ego vehicle brakes instead of going first at an intersection, an oncoming driver is predicted to proceed earlier. If this holds, a planner could evaluate cooperative strategies like merging or yielding by comparing the predictions each candidate plan induces.","feed_headline":"Simulation shows how drivers react to a car's alternative plans","feed_subtitle":"The model predicts how surrounding traffic would yield or proceed under each candidate plan.","key_machinery":"The mechanism is a closed-loop traffic simulation in which a graph-based behavior model, trained with Adversarial Inverse Reinforcement Learning, controls every non-ego vehicle. Agent-centric observations encode vehicles and road elements within 30 meters as polylines, passed through message passing and cross-attention to produce the next acceleration and steering angle; conditioning is implemented by overriding the ego vehicle's executed action with its candidate plan at each simulation step. The AIRL reward is modified with a constant survival bonus and action noise during discriminator training to keep the simulated drivers on track, and Proximal Policy Optimization with Generalized Advantage Estimation updates the shared policy.","core_discovery":"The central claim is that a single behavior policy trained via Adversarial Inverse Reinforcement Learning and executed simultaneously for all vehicles in a stepwise traffic simulation is sufficient to generate conditional predictions that capture bidirectional interactions between an automated vehicle and its surroundings. To condition a forecast on a candidate plan, the automated vehicle's state is updated along the plan during the rollout while all other agents follow the learned policy; because each predicted state depends on the previous states of every vehicle, the other drivers' simulated trajectories shift in response to the plan, and the plan itself can in principle adapt to those reactions. The paper demonstrates this with an all-way-stop intersection where replacing the ego vehicle's motion with constant braking changes the predicted progress of the oncoming vehicle and the vehicle behind it, while leaving uninvolved vehicles unaffected.","pith_inferences":["The paper validates conditioning qualitatively on one intersection; a natural next step is a quantitative benchmark that compares conditional predictions against ground-truth cases where drivers actually took the conditioned action.","Because the interaction graph and shared-policy design do not depend on which agent is conditioned, the same machinery could simulate hypothetical actions of multiple agents or of non-vehicle road users.","If the conditional forecasts generalize beyond the training distribution, a planner could use them to decide when to assert right-of-way, proceeding when the simulation predicts that other drivers will yield."],"forward_implications":["A planner can score candidate plans by simulating each one and using the resulting conditional predictions to choose maneuvers that other drivers are predicted to accommodate.","Because all agents share the same learned policy and react to each other's updated states, the predictions remain scene-consistent over a 10-second horizon, avoiding the frozen-robot failure of purely reactive prediction.","The stepwise rollout lets the ego plan be revised during prediction, so planning strategies can be built as trees or receding-horizon searches rather than single fixed trajectories.","Conditioning only changes the agents that are actually affected by the ego vehicle's alternative plan, so the model does not need to be retrained for each new query.","Training in closed-loop simulation with AIRL yields markedly lower collision and off-track rates than behavior cloning after 10 seconds, which is what makes the conditional rollouts usable for planning."],"supporting_citations":[{"why":"It supplies the adversarial inverse reinforcement learning objective that reconstructs a reward from demonstrations and trains the behavior policy.","marker":"[11]"},{"why":"It establishes the graph-based AIRL driving model whose architecture the paper adapts for closed-loop conditional prediction.","marker":"[10]"},{"why":"It provides the INTERACTION dataset of real intersection, roundabout, and merging scenarios used for training and evaluation.","marker":"[17]"},{"why":"It introduces the conditional behavior prediction idea of producing forecasts conditioned on an agent's plan, which this work realizes through simulation.","marker":"[5]"},{"why":"It articulates the freezing robot problem that motivates coupling prediction with planning.","marker":"[4]"},{"why":"It supplies Proximal Policy Optimization, the reinforcement learning algorithm used to maximize the surrogate AIRL reward.","marker":"[13]"},{"why":"It provides Generalized Advantage Estimation, which computes the policy-gradient advantages used in the RL updates.","marker":"[12]"},{"why":"It supplies the cross-attention mechanism that lets each agent's embedding attend to surrounding agents and road polylines.","marker":"[14]"},{"why":"It provides the shared-policy multi-agent reinforcement learning setup that lets a single policy control every simulated vehicle simultaneously.","marker":"[7]"}],"fun_headline_variants":["Simulation predicts driver responses to AV's planned moves","Conditional prediction: traffic sim shows reactions to AV plans","AIRL-powered simulation forecasts driver reactions to plans","Interactive traffic simulation anticipates driver responses","How other drivers react: simulation conditioned on AV plans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned driver policy must respond realistically to ego trajectories that differ from what the training drivers actually did, even when those conditioned plans push the situation outside the distribution of the recorded demonstrations.","fun_headline_variants_meta":{"raw":{"variants":["Simulation predicts driver responses to AV's planned moves","Conditional prediction: traffic sim shows reactions to AV plans","AIRL-powered simulation forecasts driver reactions to plans","Interactive traffic simulation anticipates driver responses","How other drivers react: simulation conditioned on AV plans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000185,"raw_usage":{"total_tokens":1241,"prompt_tokens":781,"completion_tokens":460,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":397,"completion_tokens_details":{"reasoning_tokens":388}},"tokens_in":397,"tokens_out":460,"duration_ms":5027,"temperature":1.0,"reasoning_tokens":388,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T05:14:59.266470+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record or collect real-world cases in which a driver takes an alternative action, such as braking at an all-way stop instead of proceeding first, run the conditional simulation with that action as the conditioned plan, and compare the simulated trajectories of the other agents against the ground-truth trajectories from those cases. A quantitative divergence on such a conditional benchmark would show that the model generalizes to conditioned plans only in appearance, not in distribution.","supporting_citations":[{"cited_title":"Graph-based adversarial imitation learning for predicting human driving behavior,","cited_arxiv_id":null,"evidence_quote":"It establishes the graph-based AIRL driving model whose architecture the paper adapts for closed-loop conditional prediction."},{"cited_title":"Identi- fying driver interactions via conditional behavior prediction,","cited_arxiv_id":null,"evidence_quote":"It introduces the conditional behavior prediction idea of producing forecasts conditioned on an agent's plan, which this work realizes through simulation."},{"cited_title":"Unfreezing the robot: Navigation in dense, interacting crowds,","cited_arxiv_id":null,"evidence_quote":"It articulates the freezing robot problem that motivates coupling prediction with planning."},{"cited_title":"Modeling interaction-aware driving behavior using graph-based representations and multi-agent reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"It provides the shared-policy multi-agent reinforcement learning setup that lets a single policy control every simulated vehicle simultaneously."}],"review_version":1}