{"id":"8231f534-7a2b-49d3-900d-d519e76e174f","arxiv_id":"2605.16087","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A transformer-based 3D object detector for autonomous driving is extended with attention-derived saliency maps, perturbation-validated faithfulness checks, uncertainty calibration, and robustness training, then deployed on a prototype vehicle with a real-time XAI monitoring interface.","lead":"The paper builds a perception system for self-driving cars on top of a transformer detector, adding attention-based explanations, uncertainty calibration, and robustness training, then runs it on a prototype vehicle with a live visualization dashboard. A smart generalist might read it to see what concrete engineering steps look like when moving trustworthy-AI ideas from theory into a safety-critical real-world prototype.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Faithfulness of attention weights as explanations rests solely on perturbation consistency tests whose sufficiency for 3D detectors is unproven.","rationale":"The reader's weakest assumption correctly isolates the single point where the trustworthiness argument is least anchored. No other internal inconsistency or missing control appears load-bearing from the abstract and stated claims.","tokens_in":1720,"tokens_out":268,"duration_ms":14457,"concrete_test":"Re-run the perturbation consistency protocol on the same model but replace attention maps with Grad-CAM saliency; if the consistency metric does not drop by >15% relative to attention, the claim that attention specifically yields faithful explanations is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that attention-derived saliency maps constitute faithful explanations of the transformer detector's decisions. The paper states this is validated 'using perturbation-based consistency tests' but supplies no comparison to alternative explanation methods (e.g., integrated gradients, occlusion), no ablation on perturbation strategy, and no analysis of whether the tests can detect known failure modes of attention (spurious focus on background or non-causal tokens). In a safety-critical perception setting, if the tests only measure local consistency rather than causal fidelity, the 'faithful explainability' component of the trustworthy pipeline does not hold.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes a Trustworthy AI perception module for autonomous driving based on a transformer detector. Explanations are derived from attention weights at inference time and validated via perturbation-based consistency tests; an uncertainty estimation and calibration module is integrated along with robustness-enhancing training. Experiments are claimed to demonstrate faithful saliency behavior, improved robustness, and well-calibrated uncertainty. The full pipeline is deployed in a prototype vehicle with a real-time XAI interface visualizing saliency maps, uncertainty state, and documentation artifacts.","tokens_in":1833,"tokens_out":457,"duration_ms":26602,"significance":"If the quantitative results and validation hold, the work is significant for providing one of the few end-to-end implementations of trustworthy AI elements (faithful explainability, calibrated uncertainty, robustness) in a 3D perception system for autonomous driving, including real-vehicle deployment and an operational XAI interface. This bridges theoretical frameworks with practical systems integration and could serve as a reference for safety-critical applications.","major_comments":[{"comment":"Abstract: The abstract reports positive experimental outcomes on faithfulness, robustness, and calibration but supplies no quantitative numbers, baseline comparisons, or details on post-hoc choices (e.g., perturbation types, calibration method). This absence makes it impossible to assess the magnitude or reliability of the claimed improvements.","section":"Abstract"},{"comment":"Explanation validation (referenced in Abstract): The claim that attention-derived saliency maps constitute faithful explanations rests solely on perturbation-based consistency tests. The manuscript provides no comparison to alternative methods (e.g., integrated gradients, occlusion), no ablation on perturbation strategy, and no analysis of whether the tests detect known attention failure modes such as spurious focus on background or non-causal tokens. In a safety-critical 3D detector setting, this leaves the faithfulness component of the trustworthy pipeline insufficiently supported.","section":"Explanation validation"}],"minor_comments":[{"comment":"The supplementary materials link is provided, but the main text would benefit from explicit cross-references to specific quantitative results or figures supporting the 'faithful saliency behavior' claim.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which highlight opportunities to strengthen the presentation of our results. We address each major comment below and commit to revisions that improve clarity and support for the claims without altering the core contributions.","responses":[{"response":"We agree that the abstract would benefit from quantitative details to enable readers to evaluate the scale of improvements. In the revised version we will incorporate representative metrics (e.g., faithfulness consistency scores, robustness gains under perturbation, and expected calibration error) together with concise references to the perturbation strategy and calibration procedure employed.","revision_made":"yes","referee_comment":"[Abstract] Abstract: The abstract reports positive experimental outcomes on faithfulness, robustness, and calibration but supplies no quantitative numbers, baseline comparisons, or details on post-hoc choices (e.g., perturbation types, calibration method). This absence makes it impossible to assess the magnitude or reliability of the claimed improvements."},{"response":"We recognize that the current validation relies exclusively on perturbation consistency and lacks explicit comparisons or failure-mode analysis. We will add (i) a comparison of attention-derived maps against integrated gradients and occlusion, (ii) an ablation on perturbation parameters, and (iii) a targeted examination of attention behavior on background or non-causal regions, including discussion of implications for 3D detection safety.","revision_made":"yes","referee_comment":"[Explanation validation] Explanation validation (referenced in Abstract): The claim that attention-derived saliency maps constitute faithful explanations rests solely on perturbation-based consistency tests. The manuscript provides no comparison to alternative methods (e.g., integrated gradients, occlusion), no ablation on perturbation strategy, and no analysis of whether the tests detect known attention failure modes such as spurious focus on background or non-causal tokens. In a safety-critical 3D detector setting, this leaves the faithfulness component of the trustworthy pipeline insufficiently supported."}],"tokens_in":1386,"tokens_out":409,"duration_ms":21465,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's core contribution is showing that a transformer-based 3D detector can be wrapped with attention-derived saliency, perturbation checks, calibration, and robustness training, then run live in a prototype car with a dashboard interface. That deployment step is the part worth noting; most XAI work stops at benchmarks, so getting it into a vehicle is concrete engineering work that matters for safety monitoring discussions.","headline":"This is a straightforward engineering integration paper that gets a real-vehicle prototype running with XAI and uncertainty tools, but adds no new methods and supplies almost no quantitative evidence.","tokens_in":2386,"tokens_out":158,"would_cite":false,"duration_ms":15412,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[],"headline":"XAI/robustness pipeline for 3D vehicle perception unrelated to RS forcing chain or J-cost","alignment":"orthogonal","rationale":"Paper's core machinery (cross-attention extraction from CMT transformer, perturbation-based faithfulness tests, KL/von-Mises uncertainty head, masked-modal training, post-hoc TS/PS calibration) is standard engineering for trustworthy perception; no overlap with RS theorems on distinction-to-spacetime forcing, J(x) = ½(x + x⁻¹) − 1, φ-ladder, 8-tick periodicity, or parameter-free constants. Domain (cs.RO) lies outside RS scope.","tokens_in":50715,"confidence":"high","tokens_out":151,"duration_ms":6645,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A transformer detector for autonomous driving yields faithful explanations from its attention weights at inference time, plus calibrated uncertainty and robustness improvements, all deployed in a prototype vehicle with a real-time interface","keywords":["trustworthy AI","explainable AI","autonomous driving","perception","transformer detector","uncertainty calibration","robustness","saliency maps"],"falsifier":"A controlled test in which the attention-derived saliency maps are compared against a ground-truth importance measure obtained by systematically ablating input regions and measuring change in the detector's output scores; systematic mismatch would falsify the faithfulness claim.","tokens_in":2599,"feed_emoji":"🚗","tokens_out":742,"duration_ms":22570,"temperature":0.7,"pith_summary":"The paper sets out to show that a complete trustworthy-AI pipeline can be built for 3D perception in driving by taking a transformer detector, extracting explanations directly from its attention maps, adding an uncertainty-calibration module, and applying robustness training. It then validates the explanations with perturbation consistency tests, measures gains in robustness and calibration, and moves the entire system onto a real vehicle where an interface displays the saliency maps, uncertainty state, and documentation artifacts live. A sympathetic reader would care because current deep networks for scene understanding remain black boxes that conflict with safety standards and make debugging or oversight difficult; if the pipeline works, it supplies one concrete route from abstract trustworthy-AI guidelines to an operational perception stack.","feed_headline":"Attention weights from a driving detector give real-time explanations","feed_subtitle":"Prototype vehicle runs a full pipeline of faithful saliency, calibrated uncertainty, and robustness training with a live XAI display.","key_machinery":"Attention weights extracted from the transformer detector at inference time, used as the source of saliency explanations and validated by perturbation consistency tests, together with a separate uncertainty-calibration module and robustness training.","core_discovery":"Building on a transformer-based detector, explanations are derived from the attention mechanism at inference time and validated for faithfulness using perturbation-based consistency tests. An uncertainty estimation and calibration module is integrated, robustness-enhancing training methods are applied, and the resulting system is shown to produce faithful saliency behavior, improved robustness, and well-calibrated uncertainty estimates. The full set of trustworthy-AI elements is finally deployed in a prototype vehicle together with an XAI interface that visualizes documentation artifacts, model uncertainty state, and saliency maps in real time.","pith_inferences":["The same attention-extraction pattern could be tested on other transformer architectures used for 3D detection to see whether faithfulness holds across detector families.","If the real-time interface is kept, it might serve as a template for logging artifacts required by future automotive safety standards.","Extending the uncertainty calibration to multi-modal sensor fusion would be a direct next step that the current single-detector pipeline leaves open."],"forward_implications":["Explanations become available at inference time with no extra forward passes required.","The perception module can be monitored in real time for uncertainty spikes that may indicate out-of-distribution inputs.","Robustness training reduces performance drop under common perturbations such as noise or occlusion.","The deployed XAI interface supplies a single screen that combines saliency, uncertainty, and model documentation for human oversight."],"fun_headline_variants":["Transformer attention validated for faithful explanations in driving AI","Prototype vehicle deploys attention based XAI with uncertainty calibration","Robustness training yields calibrated uncertainty in perception transformer","Real time saliency maps from attention mechanism in vehicle prototype","Faithfulness tested via perturbations for AI perception module deployment"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Attention weights from the transformer at inference time give faithful accounts of the model's actual decisions, with faithfulness checked only through the described perturbation tests.","fun_headline_variants_meta":{"raw":{"variants":["Transformer attention validated for faithful explanations in driving AI","Prototype vehicle deploys attention based XAI with uncertainty calibration","Robustness training yields calibrated uncertainty in perception transformer","Real time saliency maps from attention mechanism in vehicle prototype","Faithfulness tested via perturbations for AI perception module deployment"]},"model":"grok-4.3","cost_usd":0.005229,"raw_usage":{"total_tokens":2456,"prompt_tokens":675,"num_sources_used":0,"completion_tokens":74,"cost_in_usd_ticks":52290500,"prompt_tokens_details":{"text_tokens":675,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1707,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":675,"tokens_out":74,"duration_ms":12458,"temperature":1.0,"reasoning_tokens":1707,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-25T06:26:27.483586+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A controlled test in which the attention-derived saliency maps are compared against a ground-truth importance measure obtained by systematically ablating input regions and measuring change in the detector's output scores; systematic mismatch would falsify the faithfulness claim.","supporting_citations":[],"review_version":2}