{"id":"174e24ff-8a96-4252-8951-979eae030ce5","arxiv_id":"2502.05677","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A counterfactual surprise metric, Hist-prim with query-centric feedforward prediction and Wasserstein distance, identifies interactive driving scenarios with 0.82+ Spearman correlation to a human-trained reward model.","lead":"This paper proposes a metric that measures how surprising a driving scene is, by changing one agent's past behavior and measuring how much other agents' predicted futures shift. The winning design, which swaps in history motion primitives and uses the Wasserstein distance, ranks interactive nuScenes scenarios with over 0.82 correlation to a reward model learned from human preferences.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (1) computes the distribution shift over all agents, yet the claimed 'surprise potential on others' requires excluding the manipulated target; the 0.827 correlation may reflect self-history sensitivity rather than interactivity.","rationale":"I read the paper as claiming a specific unsupervised metric ranks driving scenarios by their interactivity, with the empirical support being the >0.82 Spearman correlation against a human-preference-trained reward model. The reader's identified weakness—the reward model itself is not validated against held-out human labels—is real and should be addressed. However, the more load-bearing issue is what the metric actually computes. Equation (1) is written for the full joint distribution, and for Hist-prim the intervention changes the target's own history; including the target in the shift measurement makes the metric sensitive to how much a prediction model changes its own output when its input changes, which is not the same as one agent surprising others. The paper never states that the target is excluded, and the wording 'surprise potential on others' implies it should be. A single masked re-computation would settle whether the reported correlation is driven by cross-agent effects. Since this is testable and the current text is ambiguous, the reader's conditional verdict is appropriate; I recommend UNCHANGED, pending the check.","tokens_in":12729,"tokens_out":6469,"duration_ms":60907,"concrete_test":"Recompute the Spearman correlation for the best configuration (Hist-prim + FFP-QC + W2, Table I) after masking the target agent from D: compute W2 only between the predicted distributions of all non-target agents under F(ξ) and F(G(ξ)). If the correlation stays above roughly 0.8, the metric genuinely tracks interaction; if it drops substantially, the headline is an artifact of self-history sensitivity. Report the same masked correlation for the Fig. 5 qualitative examples.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The formal definition in Sec. III-A, S(ξ) := D(F(ξ), F∘G(ξ)), applies D to the full joint predicted distribution over all agents. For the winning counterfactual Hist-prim (Sec. IV-A), G replaces the target agent's history with a motion primitive. That agent's own predicted future therefore appears on both sides of D and will shift even in completely non-interactive scenes, simply because the input history changed. The paper's stated object is 'surprise potential on others' (Abstract, Sec. I) and the qualitative evidence in Fig. 5 is about changes in other, interactive agents, but no step masks or excludes the target agent from D in Eq. (1) or in the reported experiments. As a result, the metric can conflate a direct self-history effect with a cross-agent interaction effect. If the self-effect dominates the Wasserstein distance, the headline 0.827 Spearman correlation with the reward-model ranking may be driven by prediction-model sensitivity to input perturbations rather than by true interactivity. This is an internal-consistency problem in the metric itself, independent of the reward-model validation concern.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a family of \"surprise potential\" metrics for ranking driving scenarios by interactivity, decomposing the metric into three design axes: counterfactual generation, future prediction architecture, and distribution-shift measure. It evaluates many instantiations of these axes on nuScenes, using as ground truth a reward model trained on 5,000 human pairwise preference labels. The proposed configuration Hist-prim + FFP-QC + W2 achieves a Spearman correlation of 0.827 against this reward model's ranking, outperforming rule-based baselines. The paper also demonstrates downstream utility by showing that scenarios ranked as highly surprising correlate with lower planner safety and that upsampling such scenarios improves a learned planner.","tokens_in":12928,"tokens_out":5627,"duration_ms":61185,"significance":"If the central claim holds, the paper provides a practically useful tool for dataset curation in autonomous driving and organizes a previously fragmented design space into a systematic comparison. The Hist-prim counterfactual, the use of query-centric prediction, and the Wasserstein-distance shift measure are sensible contributions, and the downstream planning results are a valuable demonstration. However, the headline evidence is weakened by three load-bearing issues: the reward model is never validated against held-out human labels, the reported 0.827 is a post-selection maximum over a grid evaluated on the same validation set, and the mathematical definition in Eq. (1) does not exclude the manipulated target agent from the distribution shift, creating an internal-consistency problem with the stated goal of measuring \"surprise potential on others.\" These issues are fixable within the manuscript's scope, but they must be addressed before the human-alignment claim can be accepted.","major_comments":[{"comment":"The definition S(ξ) := D(F(ξ), F∘G(ξ)) applies the shift measure D to the full joint predicted distributions over all agents, yet the paper's stated object is \"surprise potential on others\" (Abstract, Section I) and the qualitative evidence in Figure 5 concerns predictions for other agents. With the Hist-prim counterfactual (Section IV-A), the target agent's history is replaced by motion primitives, so that agent's own predicted future appears on both sides of D and can shift even in a non-interactive scene merely because the input history changed. No step in Eq. (1) or in the experimental section masks or excludes the target agent from D. Unless the implementation did exclude the target (which the paper should state explicitly), the reported 0.827 correlation may be driven by self-history sensitivity rather than cross-agent interactivity. I request a formal definition of \"others,\" an explicit target-exclusion version of Eq. (1), and an ablation separating the self-shift from the other-agent shift.","section":"§III-A, Eq. (1)"},{"comment":"The reward model trained on 5,000 pairwise preference labels is treated as the ground-truth ranking, but the paper never validates this model against held-out human labels, for example by reporting pairwise classification accuracy, inter-annotator agreement, or a direct comparison on held-out preference pairs. Figure 6 shows that the Spearman correlation saturates as the reward model is trained on more preference data, but the target being predicted is still the reward model's own ranking; this does not establish that the reward model faithfully represents human intuition. The abstract's claim of \"a correlation of more than 0.82 with human labels\" is therefore not supported by the reported experiments; the actual quantity is a correlation with a learned proxy. The authors should either validate the reward model directly or soften the claim and compute the surprise metric's agreement on held-out human preference pairs.","section":"§III-C, §V-B"},{"comment":"The headline 0.827 correlation is the maximum over a grid of configurations (counterfactual × architecture × shift measure × mode count) evaluated on the same nuScenes validation set, and the best configuration is selected after inspecting those correlations. The paper reports no confidence intervals, no multiple-comparison control, and no separate model-selection split, so the reported value is a post-selection maximum. The small gaps visible in Table I (e.g., 0.827 versus 0.829 for FFP-QC-W2-5 or 0.828 for FFP-QC-W2-15) are unlikely to be meaningful without uncertainty quantification. I recommend reporting bootstrap confidence intervals and using a nested or held-out protocol for the design-space search, or explicitly framing the result as a selected benchmark.","section":"§V-B, Table I"}],"minor_comments":[{"comment":"The name \"Fut-none\" is confusing because it denotes the baseline with no counterfactual future conditioning rather than a future perturbation; a one-sentence clarification would help readers map the label to the definition.","section":"§IV-A"},{"comment":"Equation (2) writes D_L2 = Σ_k ‖μ_k^1 − μ_k^2‖^{1/2}; if the intended quantity is the squared L2 distance between means, the exponent should be 2, and the current expression is not a norm. Please correct the formula or justify the root-distance choice.","section":"§IV-C, Eq. (2)"},{"comment":"The caption refers to \"Hist-prim-Hist-prim,\" which appears to be a typo for the single method \"Hist-prim.\"","section":"Figure 7 caption"},{"comment":"The phrases \"correlation with human labels\" and \"correlation with human annotation\" should be replaced by \"correlation with the reward-model ranking\" unless direct human-label validation is added.","section":"Abstract and Conclusion"},{"comment":"The axes labeled \"nominal scenario ξ\" and \"counterfactual scenario G(ξ)\" do not map obviously to the method names in the figure; a small legend or table clarifying which side of each comparison is used for each entry would improve readability.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of the journal and the design-space study is valuable, but the headline claim of human alignment is currently supported only through an unvalidated learned proxy and a post-selection maximum. I am not recommending rejection; the study is substantially useful and the issues can be addressed with additional experiments and a more careful formal definition. The target-exclusion question in Eq. (1) should be treated as a correctness issue rather than a presentation issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers a solid, systematic comparison of counterfactual surprise metrics for identifying interactive driving scenarios. The Hist-prim counterfactual—replacing the target agent's history with data-derived motion primitives—is new and sensible, and the exhaustive sweep over counterfactual type, prediction architecture, and distribution shift metric is exactly the kind of empirical grounding this subfield needs. The ablation table and downstream planner experiments give the work real practical value. The writing is clear, and the qualitative examples in Fig. 5 do show that interactive agents' futures shift while non-interactive ones don't.\n\nThat said, the headline number deserves caution. First, the reward model trained on 5,000 pairwise preferences is never validated against held-out human labels. The paper shows correlation saturation with more data, but that doesn't tell us whether the model actually tracks human judgment. Second, the 0.827 Spearman is the best over a grid of configurations, selected on the same validation set, with no multiple-comparison correction. The pattern is fairly consistent across modes, so it's not a single fluke, but the reported number is still a post-selection maximum. Third, and most important, Eq. (1) computes the distribution shift over all agents, but the paper's stated object is \"surprise potential on others.\" For Hist-prim, the target agent's own history is replaced, so its own predicted future appears on both sides of D and will shift even in completely non-interactive scenes. Nothing in the formulation or the experiments masks the target agent out of the distribution shift. The metric therefore conflates a self-history sensitivity effect with true cross-agent interactivity. The correlations could still be meaningful—interactive scenes likely produce both effects—but the current numbers don't isolate what the paper claims to measure.\n\nThe paper is honest about some limitations (e.g., prediction model instability, small dataset) but doesn't flag this target-exclusion issue. It's fixable: recompute the metric with the target agent excluded from D, or at least report the decomposition. As is, the central claim is plausible but not fully supported.\n\nWho is this for? Researchers working on scenario curation and AV benchmarking will find the design-space analysis and the Hist-prim idea worth engaging with. It deserves a serious referee: the empirical scope and the potential practical impact justify the review time, but I'd send it back for a substantial revision rather than accept it in this form.","headline":"A genuinely useful comparative study of counterfactual surprise metrics, but the headline 0.82 correlation is weakened by an unvalidated reward model and a metric that doesn't actually isolate 'surprise on others'.","tokens_in":13497,"tokens_out":1835,"would_cite":true,"duration_ms":21102,"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 establishes that a driving scenario's interactivity can be quantified by its surprise potential—the shift, under a counterfactual replacement of one agent's past, in the predicted future distributions of others—and shows that…","keywords":["surprise potential","interactivity","counterfactual reasoning","autonomous driving","scenario curation","distribution shift","Wasserstein distance","human preference"],"falsifier":"Test the metric against rankings from human pairwise labels that were not used in training the reward model: if the Spearman correlation of Hist-prim+FFP-QC+W2 with those held-out labels is substantially below 0.82, then the claimed alignment with human intuition is actually alignment with the reward model's approximation of it.","tokens_in":12505,"feed_emoji":"🚗","tokens_out":6838,"duration_ms":58853,"temperature":0.7,"pith_summary":"Most driving logs are uneventful, so benchmarking an autonomous vehicle requires finding the rare scenarios where road users genuinely interact. This paper argues that interactivity can be measured by 'surprise potential': how much another agent's predicted future would shift if one agent's past were replaced by a different but plausible behavior. The authors decompose this idea into three design choices—how the counterfactual is generated, which prediction model produces the future distributions, and which distance measures the shift—and test the combinations on nuScenes. Their best configuration, which swaps the ego vehicle's history for a motion primitive and measures the Wasserstein shift between the predicted futures, reaches a Spearman correlation of about 0.83 with rankings from a reward model trained on human pairwise preferences. That clearly outperforms rule-based filters such as time-to-collision or distance counts, which stay below 0.3.","feed_headline":"Surprise metric matches human interactivity rankings at 0.82","feed_subtitle":"Counterfactual motion primitives plus Wasserstein shift outrank rule-based filters for finding rare interactive scenes.","key_machinery":"The surprise potential is defined as $S(\\xi) = D(F(\\xi), F(G(\\xi)))$, where $F$ is a trajectory prediction model outputting a distribution over all agents' futures, $G$ is a counterfactual generator that edits the scenario, and $D$ is a distribution-shift measure. The winning instantiation uses $G$ = Hist-prim (replacing the target agent's history with collision-free, drivable motion primitives), $F$ = a query-centric feedforward transformer with a Gaussian-mixture prediction head, and $D$ = $W_2$, the 2nd-order Wasserstein distance computed between the modes of the two GMMs with an optimal-transport coupling. Its role is to compute, for every scenario, a scalar score that orders scenes by how strongly the counterfactual intervention propagates through the prediction model.","core_discovery":"The central discovery is that replacing the target agent's recorded history with a set of motion primitives (short, behavior-level trajectory segments such as left turns or lane changes) and then measuring how much the predicted future trajectories of other agents move is a strong indicator of a scenario's interactivity. On the nuScenes validation set, the best instance of this family—Hist-prim for counterfactual generation, a feedforward query-centric predictor (FFP-QC), and the second-order Wasserstein distance ($W_2$) between Gaussian-mixture predictions—achieves a Spearman correlation of 0.827 with the ranking produced by a reward model learned from 5,000 human pairwise comparisons. The same configuration also yields an AUC-ROC of 0.86 when separating the top 10% of interactive scenarios from the rest. The paper further establishes that the measure is not just descriptive: high-surprise scenarios are exactly where planners have lower time-to-collision, and upsampling high-surprise scenes during planner training improves downstream safety metrics.","pith_inferences":["The paper's 0.827 figure measures agreement with a reward model trained on human preferences, not with held-out human labels; validating the reward model against held-out annotator pairs would show how much of the number is genuine human alignment versus the model's own bias.","The superiority of history perturbation over future conditioning suggests a transferable principle: interventions on inputs are more robust than conditioning on outputs because the latter leaks information about the nominal scenario; this may guide interaction scoring in other multi-agent settings such as pedestrian groups or warehouse robots.","The design-space result that the Wasserstein distance beats KL and L2 for GMM shifts could be exploited in cheaper online deployment, for example with sparse transport plans or fixed mode matching, at some cost to the reported correlation.","Because even a single-mode predictor works, the metric could be applied at fleet scale with lightweight trajectory forecasters, making large-scale interactive-log mining feasible without heavy generative models."],"forward_implications":["Autonomous-vehicle teams can automatically rank raw driving logs by interactivity and curate small test sets focused on the rare critical scenarios, instead of relying on the imbalanced natural distribution.","Rule-based filters such as time-to-collision, minimum distance, and maximum acceleration are poor substitutes for interaction identification, all falling below 0.3 Spearman correlation on nuScenes.","Planner validation on curated high-surprise buckets exposes safety differences that are hidden in the full dataset, since mean time-to-collision decreases monotonically as surprise potential increases.","Training a planner on data upsampled by surprise score, with sampling weights in the intermediate range, improves final planner metrics, while extreme upsampling overfits to a few scenarios.","A single prediction mode is sufficient for high correlation, suggesting that the distribution-shift signal is robust to the amount of multi-modality captured by the predictor."],"supporting_citations":[{"why":"Provides the nuScenes dataset, the real-world driving logs, validation set, and the scenario distribution used in every experiment.","marker":"[2]"},{"why":"Establishes the surprise concept in driving and the idea that surprise manifests as a deviation from nominal behavior, which the paper builds on.","marker":"[6]"},{"why":"The KL-divergence counterfactual baseline that Fut-pred builds on and that the paper's best configuration outperforms.","marker":"[7]"},{"why":"Introduces the history-removal intervention, the ancestor of Hist-rmv and of the causal-intervention interpretation used here.","marker":"[10]"},{"why":"Another history-manipulation counterfactual baseline that helps define the design space for the counterfactual generator.","marker":"[11]"},{"why":"Supplies the query-centric scene representation whose relative-position embeddings the best prediction model uses.","marker":"[22]"},{"why":"Source of motion primitives as trajectory tokens, inspiring the Hist-prim and Fut-prim counterfactuals.","marker":"[33]"},{"why":"Provides the theoretical basis for the second-order Wasserstein distance and the optimal-transport coupling between Gaussian mixture modes.","marker":"[38]"}],"fun_headline_variants":["Surprise potential hits 0.82 correlation with human interactivity rankings","Surprise metric outperforms rule-based filters for finding interactive scenes","Surprise-based metric finds rare interactive scenarios for AV testing","Surprise potential ranks interactive driving scenes like humans do","Counterfactual surprise measure beats rule-based filters for interactive scenes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the reward model trained on 5,000 human pairwise preferences is an accurate oracle for interactivity across the full nuScenes validation set, and it never checks that reward model against held-out human labels.","fun_headline_variants_meta":{"raw":{"variants":["Surprise potential hits 0.82 correlation with human interactivity rankings","Surprise metric outperforms rule-based filters for finding interactive scenes","Surprise-based metric finds rare interactive scenarios for AV testing","Surprise potential ranks interactive driving scenes like humans do","Counterfactual surprise measure beats rule-based filters for interactive scenes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3263,"prompt_tokens":941,"completion_tokens":2322,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":2237}},"tokens_in":557,"tokens_out":2322,"duration_ms":14699,"temperature":1.0,"reasoning_tokens":2237,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:22:29.601509+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Test the metric against rankings from human pairwise labels that were not used in training the reward model: if the Spearman correlation of Hist-prim+FFP-QC+W2 with those held-out labels is substantially below 0.82, then the claimed alignment with human intuition is actually alignment with the reward model's approximation of it.","supporting_citations":[{"cited_title":"nuScenes: A multimodal dataset for autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Provides the nuScenes dataset, the real-world driving logs, validation set, and the scenario distribution used in every experiment."},{"cited_title":"Identifying driver interac- tions via conditional behavior prediction,","cited_arxiv_id":null,"evidence_quote":"The KL-divergence counterfactual baseline that Fut-pred builds on and that the paper's best configuration outperforms."},{"cited_title":"Interpretable trajectory prediction for au- tonomous vehicles via counterfactual responsibility,","cited_arxiv_id":null,"evidence_quote":"Introduces the history-removal intervention, the ancestor of Hist-rmv and of the causal-intervention interpretation used here."},{"cited_title":"Leveraging neural network gradients within trajectory optimization for proactive human-robot interactions,","cited_arxiv_id":null,"evidence_quote":"Another history-manipulation counterfactual baseline that helps define the design space for the counterfactual generator."},{"cited_title":"Query- centric trajectory prediction,","cited_arxiv_id":null,"evidence_quote":"Supplies the query-centric scene representation whose relative-position embeddings the best prediction model uses."},{"cited_title":"Villani et al","cited_arxiv_id":null,"evidence_quote":"Provides the theoretical basis for the second-order Wasserstein distance and the optimal-transport coupling between Gaussian mixture modes."}],"review_version":1}