{"id":"15b2f6f8-9533-4147-8910-88d7ba7ddc0f","arxiv_id":"2505.08404","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Intention-aware Policy Graphs, applied to 830 nuScenes driving scenes, attribute desires and intentions to an autonomous vehicle and produce interpretable global and local teleological explanations, with a reported 75.7% interpretability and 92.0% reliability.","lead":"The paper explains autonomous vehicle behavior by building probabilistic graphs of inferred desires and intentions from driving data, then answering questions like 'why did the car brake?'. It is demonstrated on the nuScenes dataset and can flag possible dataset or model flaws.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reliability metrics are circular: the state discretizer was explicitly optimized to maximize them (§3.1), so the 75.7%/92.0% headline figures are in-sample and do not independently establish reliable teleological explanations.","rationale":"The reader correctly flags the traffic light red assumption as undermining the legal-compliance analysis, and I do not dispute that. However, the paper's strongest claim is the aggregate reliability of teleological explanations, supported by the 75.7%/92.0% metrics. Section 3.1 openly states the discretizer was optimized for those metrics, making the evaluation circular. This is the single most load-bearing vulnerability: if the metrics are inflated by design choices, the central claim of 'reliable' explanations collapses, regardless of traffic light states. The traffic light assumption only affects one desire and the compliance interpretation; the circularity affects the entire quantitative evaluation. The scene-1084 anomaly and nighttime stop sign gap are useful qualitative demonstrations, but they do not rescue the headline numbers. A concrete, inexpensive check is to compute the metrics with a pre-specified discretizer (no post hoc optimization) and see whether the high values persist. If they do not, the paper's central evidence is overclaimed; if they do, the concern is resolved. The reader's verdict of CONDITIONAL remains appropriate, but the condition should include demonstrating that the metrics are robust to the discretizer choice, not merely resolving the traffic light issue.","tokens_in":11186,"tokens_out":6967,"duration_ms":69183,"concrete_test":"Re-run the full pipeline on the same 830 nuScenes trajectories but with the state discretizer fixed a priori—for example, using the predicates and thresholds from the original IPG paper (Gimenez-Abalos et al., arXiv:2409.19038) or from Hayes & Shah (2017) without any post hoc adjustment—and recompute the aggregate attributed-intention and expected-intention probabilities. If the headline 75.7%/92.0% figures drop by more than, say, five percentage points, the reported reliability is an artifact of the §3.1 optimization rather than a property of the method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—that intentions can be attributed 75.7% of the time with 92.0% certainty of fulfillment—rests on metrics computed over an IPG whose state discretization was selected post hoc to maximize those very metrics. Section 3.1 states: 'we experimented with several state discretisers, this one was optimised for the metrics later presented.' The desires in Section 4.3 are then defined over the chosen predicates, and the same 830 trajectories are used both to build the PG and to report the metrics. This is a circular evaluation: the high attributed-intention probability is partly guaranteed by choosing state features that align with the hand-crafted desires, and the high expected-intention probability is inflated because many desires include common actions (e.g., Lane Keeping includes going straight, braking, accelerating). No held-out scenes, pre-registered discretizer, or sensitivity analysis is provided. Consequently, the 75.7%/92.0% figures do not yet demonstrate that the explanations are reliable in any externally meaningful sense; they demonstrate internal consistency after feature selection. The traffic-light red assumption (§4.3) is a real additional weakness for the legal-compliance sub-claim, but the circular metric is more load-bearing because it undercuts the headline reliability numbers themselves.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a post-hoc, model-agnostic explainability method for autonomous vehicles, building on Intention-aware Policy Graphs (IPGs). It discretizes 830 nuScenes trajectories into interpretable states via eleven hand-crafted predicates and threshold-based action labels, defines safe and unsafe desires, and computes intention values as path probabilities to desire-fulfilling states. It reports a global attributed-intention probability of 75.7% and an expected-intention probability of 92.0%, analyzes how these metrics vary by day/night and rainy conditions, and demonstrates local explanations and question answering on scene-1084, including an identified map inconsistency.","tokens_in":11476,"tokens_out":5404,"duration_ms":55101,"significance":"If the quantitative claims were externally supported, this would be a useful contribution to XAI for autonomous vehicles: the method is model-agnostic, provides human-readable teleological explanations ('what/why/how'), and the scene-level analysis demonstrates a concrete ability to surface dataset and model anomalies. The nighttime stop-sign finding in Table 2 and the scene-1084 map inconsistency are promising qualitative results. However, the current evaluation does not establish the headline reliability claims: the state discretizer was explicitly optimized for the reported metrics, the 'expected intention probability' is definitionally tied to the intention value being thresholded, and the legal-compliance sub-claim rests on an unverified traffic-light assumption. The qualitative findings and the formal plan-based explanation mechanism should be preserved in a revision, but the quantitative claims need reframing and additional validation.","major_comments":[{"comment":"The central quantitative claim is evaluated in-sample in a way that is acknowledged in Section 3.1: the state discretizer was 'optimised for the metrics later presented', and Section 4.3 defines desires over those same predicates. Since the same 830 trajectories are used both to construct the policy graph and to compute the 75.7% and 92.0% headline metrics, these numbers reflect post-hoc feature selection rather than an independent assessment of explanation reliability. The paper should add a held-out evaluation, a pre-registered or fixed discretizer chosen without access to the evaluation scenes, and a sensitivity analysis over discretization thresholds and the commitment threshold C.","section":"Section 3.1 and Section 5.1"},{"comment":"The 'expected intention probability' is a within-graph statistic that is directly derived from the intention definition: Id(s) is the probability of reaching a desire-fulfilling state from s, and the metric averages Id(s) over states with Id(s)>C=0.5. High values such as 91.6% and 92.0% are therefore re-expressions of the same quantity used to define the intention, not independent evidence that the inferred intentions are reliable. The paper should relabel this quantity as internal consistency of the graph, or validate it against ground-truth labels such as human-annotated compliance judgments or an independently known set of traffic violations.","section":"Section 3.2 and Section 5.1"},{"comment":"The 'Approach Traffic Light' desire is defined under the assumption that a relevant traffic light is red whenever the vehicle is near it (Section 4.3), because nuScenes does not provide signal state. This makes the legal-compliance interpretation unverifiable: the reported 16.4% unfulfilled cases could be green-light passes, dataset annotation errors, or genuine violations, and the analysis as presented cannot distinguish these alternatives. The paper should either use a dataset with traffic-light phase information or explicitly restrict the claims to 'approach behavior' without labeling it as compliance or violation.","section":"Section 4.3 and Section 5.1"},{"comment":"The threshold-based action labelling heuristic and the breadth of the desire definitions substantially inflate the 'any desire' attribution metric. For example, Lane Keeping includes GoStraight, Gas, and Brake, so many ordinary states trivially support it; the 75.7% union metric over safe and unsafe desires is therefore not a strong interpretability result by itself. Reporting the distribution of attributed desires, an ablation with narrower Ad sets, or a comparison against a baseline with randomly paired desires would clarify how much information the attribution actually carries.","section":"Section 4.2 and Section 5.1"}],"minor_comments":[{"comment":"In the sentence 'we assess the interpretability and reliability of the vehicle's behaviour using two metrics over a desired [6]', 'a desired' appears to be a typo for 'a desire d'.","section":"Section 5.1"},{"comment":"The state transition in Table 3 uses 'FrontObject(Yes)' and 'FrontObject(No)', but the predicate introduced in Table 1 is 'ObjectsNearby'; please use consistent feature names.","section":"Table 3"},{"comment":"In the sentence 'instead, it opts to break', 'break' should be 'brake'.","section":"Section 5.3"},{"comment":"The captions say 'first metrics from the right' when referring to the global aggregate values; please label these values directly in the figures or in the caption to avoid ambiguity.","section":"Figures 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline major revision. The main obstacles are the acknowledged post-hoc optimization of the state discretizer (Section 3.1) and the definitional nature of the reliability metric (Section 5.1); both are fixable by reframing the claims and adding held-out or externally validated evaluation. The qualitative vulnerability findings are the strongest part of the paper and should be retained. I do not see a novelty-disclosure issue, but the quantitative evaluation needs substantial rework before the central claims can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the application: the IPG machinery is from the authors' prior work, but turning it loose on nuScenes with AV-specific desires (stop signs, crosswalks, pedestrians) yields two interesting findings that stand on their own. The nighttime stop sign gap and the scene-1084 map inconsistency are concrete, falsifiable observations about the dataset and its annotations. The question-answering walk-through in Section 5.3 is the best part of the paper; it shows the method producing explanations that a human can actually inspect and reason about, which is more than most XAI papers offer.\n\nI agree with the stress-test note that the headline numbers are circular, though I'd phrase it slightly differently. The expected intention probability is high because the intention value is defined as the probability of reaching desire-fulfilling states; reporting that states with high intention value usually fulfill the desire is close to restating the definition. The post-hoc selection of the state discretizer makes this worse, since the features were chosen, by the authors' own admission, to make these metrics look good. So the 75.7% and 92.0% figures should not be read as evidence that the explanations are reliable in any externally meaningful sense. They are internal consistency checks. The traffic-light red assumption is also a genuine soft spot, but the authors flag it, and the legal-compliance reading is clearly presented as conditional. I don't think this is a fatal flaw for the paper's real contribution, which is the demonstration that intention-based explanations can surface dataset and model vulnerabilities.\n\nThere is no user study, no held-out scenes, no sensitivity analysis for C or the discretization thresholds. Those are real omissions, but the paper is honest about them: the discussion names the need for a user study and the domain knowledge required to define desires. The citation pattern looks fine; the reliance on [6] is legitimate since that is the underlying method.\n\nWho is this for? Researchers working on XAI for autonomous driving, especially anyone interested in dataset auditing or human-readable explanations of AV behavior. It is a proof of concept, not a rigorous validation. A serious referee should push on the circularity and ask for a pre-registered discretizer or at least a sensitivity analysis, but the qualitative findings and the scene analysis deserve to see the light of day. I would send it to peer review.","headline":"Applies the authors' own IPG method to AV driving with hand-crafted desires; the qualitative scene analysis is genuinely useful, but the headline metrics are largely self-referential, so the 'reliable explanations' claim needs tempering.","tokens_in":11966,"tokens_out":993,"would_cite":false,"duration_ms":12032,"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":"A graph of inferred desires can explain 75.7% of autonomous vehicle behavior, with 92.0% certainty that attributed intentions are fulfilled, and it exposes dataset blind spots like a nighttime stop-sign gap and a map error.","keywords":["Explainable AI","Autonomous driving","Intention-aware Policy Graphs","Teleological explanation","Desires and intentions","nuScenes dataset","Dataset vulnerability","Human-centric XAI"],"falsifier":"In the scenes where the ego vehicle approaches a traffic light but does not slow or stop, read the actual light color from the camera images in those frames; if a substantial share are green, the red-light assumption is false and the compliance conclusions drawn from the 16.4% unfulfilled 'Approach Traffic Light' cases do not hold.","tokens_in":11013,"feed_emoji":"🚗","tokens_out":9876,"duration_ms":87101,"temperature":0.7,"pith_summary":"This paper tries to establish that the behavior of an autonomous vehicle can be explained in human terms—by the desires and intentions we attribute to it—without opening its internal model. The authors turn recorded driving trajectories into a probabilistic graph of discrete states and maneuvers, overlay hand-crafted desires (stopping for pedestrians, turning, ignoring stop signs), and compute an intention value for every state. On the nuScenes urban driving dataset, they report that 75.7% of states admit an attributed intention at their commitment threshold, and that attributed intentions are fulfilled with 92.0% certainty. They argue this makes teleological explanation quantitative, usable for checking legal compliance, and able to surface vulnerabilities in datasets and models. The contribution is a practical, post-hoc explainability tool rather than a new theory of intention.","feed_headline":"One graph explains 75.7% of autonomous vehicle behavior","feed_subtitle":"Attributed intentions are fulfilled with 92.0% certainty, and the same graph exposes dataset blind spots.","key_machinery":"The central object is an Intention-aware Policy Graph: a directed graph whose nodes are semantic driving states and whose edges are labeled maneuvers with transition probabilities, augmented by a desire set and per-state intention values. The intention value $I_d(s)$ sums the probabilities of all paths from state $s$ that reach a state satisfying desire $d$ and execute a fulfilling action; this single quantity carries the explanation. Two global metrics summarize it: the attributed intention probability, the share of states where $I_d(s)$ exceeds the commitment threshold $C=0.5$, and the expected intention probability, the probability that an attributed intention is actually fulfilled. The graph also supports answering the telic questions 'what do you intend in state $s$?', 'why would you do action $a$ in $s$?', and 'how do you plan to fulfill intention $I$ from $s$?'.","core_discovery":"The paper's central claim is that teleological explanation—explanation in terms of attributed goals rather than causes—can be made quantitative for autonomous vehicles. Observed trajectories are discretized into states described by natural-language predicates (speed, steering, lane position, nearby crosswalks, pedestrians, stop areas) and actions labelled as maneuvers; these populate a policy graph whose edges carry transition probabilities. Each hand-crafted desire $d$, safe or unsafe, specifies a set of relevant states and a set of fulfilling actions, and the intention value $I_d(s)$ is computed as the probability, summed over all paths from $s$, of reaching a state where $d$ is fulfilled. Across 830 scenes, the authors report that at threshold $C=0.5$ an intention can be attributed to 75.7% of states and that attributed intentions are fulfilled with 92.0% certainty. The same graph enables concrete failure diagnoses: at night the vehicle never manifests stop-sign intentions, and in scene 1084 the intended lane change appears later than the observed maneuver because the map omits a real lane divider.","pith_inferences":["The headline 75.7% and 92.0% figures measure coverage of the hand-crafted desire vocabulary, not of all possible driving intentions; adding desires such as obstacle avoidance or emergency braking would change both numbers, a testable consequence of the framework.","The unfulfilled 'Approach Traffic Light' cases (16.4%) conflate lawful green-light passes with violations because the color is assumed red; a dataset with signal annotations would let this distinction be resolved directly.","The same intention metrics could be monitored online: streaming vehicle states could be flagged when an attributed intention is unlikely to be fulfilled, before the maneuver completes, turning explanation into a real-time safety signal.","The scene-1084 diagnosis suggests the method can also audit the map itself; comparing computed intentions against observed maneuvers across all scenes might reveal further map errors beyond the single reported case."],"forward_implications":["Any autonomous driving system can be audited after the fact using only its observed trajectories and map data, without access to its planner, reward function, or training labels.","Compliance can be assessed per desire: for example, stop-sign approach intentions are fulfilled 83.6% of the time when attributed, and the remaining cases can be inspected as rolling stops, violations, or dataset errors.","The same policy graph doubles as a dataset-quality probe, as shown by the nighttime disappearance of stop-sign intentions and the map-divider omission in scene 1084.","Local queries at any state produce plans and purpose attributions that support counterfactual reasoning, such as explaining why braking at a given state serves a weaker lane-change intention than the optimal accelerating turn."],"supporting_citations":[{"why":"Supplies the Intention-aware Policy Graph formalism and the attributed/expected intention metrics used throughout the experiments.","marker":"[6]"},{"why":"Provides the driving trajectories, map geometry, and object annotations from which the policy graph is built.","marker":"[2]"},{"why":"Grounds the discretisation of continuous states and actions into semantic natural-language predicates.","marker":"[11]"},{"why":"Supplies the folk-conceptual model of desires and intentions that the teleological explanations adopt.","marker":"[16]"},{"why":"Motivates intent-based explanation as the route to trustworthiness, legal auditing, and accountability.","marker":"[7]"},{"why":"Supports the premise that purpose-oriented explanations are the most satisfying form for human users.","marker":"[10]"}],"fun_headline_variants":["Graph explains 75.7% of AV decisions with 92% certainty","AI explains 75.7% of autonomous driving, 92% certain","Teleological graph: 75.7% of AV states get goals, 92% met","Graph explains 75.7% of AV behavior, exposes blind spots","Policy graph reveals intentions behind 75.7% of AV moves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The compliance analysis assumes that every traffic light the vehicle approaches is red, because the dataset records no signal color; if some of those lights are green, the unfulfilled 'Approach Traffic Light' cases become lawful passes rather than violations, and the reported compliance distinction collapses.","fun_headline_variants_meta":{"raw":{"variants":["Graph explains 75.7% of AV decisions with 92% certainty","AI explains 75.7% of autonomous driving, 92% certain","Teleological graph: 75.7% of AV states get goals, 92% met","Graph explains 75.7% of AV behavior, exposes blind spots","Policy graph reveals intentions behind 75.7% of AV moves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001052,"raw_usage":{"total_tokens":4402,"prompt_tokens":910,"completion_tokens":3492,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":3390}},"tokens_in":526,"tokens_out":3492,"duration_ms":22881,"temperature":1.0,"reasoning_tokens":3390,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:54:54.733473+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In the scenes where the ego vehicle approaches a traffic light but does not slow or stop, read the actual light color from the camera images in those frames; if a substantial share are green, the red-light assumption is false and the compliance conclusions drawn from the 16.4% unfulfilled 'Approach Traffic Light' cases do not hold.","supporting_citations":[{"cited_title":"How the Mind Explains Behavior: Folk Explanation, Mean- ing and Social Interaction","cited_arxiv_id":null,"evidence_quote":"Supplies the folk-conceptual model of desires and intentions that the teleological explanations adopt."},{"cited_title":"Why Interpreting Intent Is Key for Trust- worthiness in the Age of Opaque Agents","cited_arxiv_id":null,"evidence_quote":"Motivates intent-based explanation as the route to trustworthiness, legal auditing, and accountability."},{"cited_title":"People Attribute Purpose to Autonomous Vehicles When Explaining Their Behavior: Insights from Cognitive Science for Explainable AI","cited_arxiv_id":"2403.08828","evidence_quote":"Supports the premise that purpose-oriented explanations are the most satisfying form for human users."}],"review_version":1}