Lusvarghi, Luca, Gozalvez, Javier · Nov 18, 2026
arXiv cs.AI: Attention Is Not Relevance: Challenging the Causality of Attention in Autonomous Driving
Markdown · JSON · Original source
Luca Lusvarghi and Javier Gozalvez argue that attention is not a dependable measure of how much an object affects an autonomous vehicle’s driving. They compare attention-based estimates with a causal reference produced through analysis. Their analysis uses synthetic driving scenes and an autonomous vehicle based on PlanT 2.0. The authors report that attention reflects correlation rather than causal impact: it can miss objects that matter and mark objects as relevant when they have little or no impact. They say attention estimates work better in simple, straight-road, single-lane scenes and lose accuracy on more complex roads, including multi-lane roads and intersections. The distinction matters because identifying what a model attends to does not necessarily reveal what influences its driving. The authors call for methods that can capture individual objects’ causal impact and distinguish it from correlation. The abstract describes tests in synthetic scenes; it does not establish how well the findings carry over to real-world driving.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Nov 18, 2026 (2026-11-18).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-19T06:32:51.387Z. First observed: 2026-11-19T05:01:37.947Z. Retrieved: 2026-11-19T06:31:37.188Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The abstract reports analysis on synthetic driving scenes using an autonomous vehicle based on PlanT 2.0; it does not report results from real-world driving. The supplied evidence is the paper's abstract, not its full methods or results.
Open question
How well do the reported findings carry over from synthetic driving scenes to real-world driving?
Source-stated
The authors compare attention-based estimates of object relevance against a causal reference obtained through counterfactual analysis.
we compare different attention-based relevance estimation methods against a causal reference of object relevance obtained through counterfactual analysis.
Source-stated
The authors report that attention indicates correlation rather than causal impact, and can both miss relevant objects and classify objects with little impact as relevant.
Results show that attention provides a correlational rather than causal indication of object relevance.
Source-stated
The authors report that attention performs less accurately in complex road settings than in simple single-lane, straight-road scenarios.
attention can effectively identify relevant objects in simple single-lane, straight-road scenarios, but its accuracy degrades in more complex settings, such as multi-lane roads and intersections.
Source-stated
The authors call for methods that capture individual objects’ causal impact and distinguish it from correlational associations.
calling for novel methods able to accurately capture the causal impact of individual objects on the vehicle's driving and distinguish it from correlational associations.
Source and contributor context
Lusvarghi, Luca, Gozalvez, Javier · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Counterfactual
An argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 414bea0c-85c6-4725-87a6-5f9ebaf34f78.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T06:32:51.387Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 5350dc739530b7fff02a3e2b05d726d33626513217dc9d93f4ae765aac9f9da0. Source text hash: f51bdfdb14af4897d518597abf2d1e068bd3a6ae661ecdf5048e3b0a9f9eccda.
Revision and source-check history
Version 1 · 2026-11-19T06:32:51.387Z
arXiv cs.AI: Attention Is Not Relevance: Challenging the Causality of Attention in Autonomous Driving
Luca Lusvarghi and Javier Gozalvez argue that attention is not a dependable measure of how much an object affects an autonomous vehicle’s driving. They compare attention-based estimates with a causal reference produced through analysis. Their analysis uses synthetic driving scenes and an autonomous vehicle based on PlanT 2.0. The authors report that attention reflects correlation rather than causal impact: it can miss objects that matter and mark objects as relevant when they have little or no impact. They say attention estimates work better in simple, straight-road, single-lane scenes and lose accuracy on more complex roads, including multi-lane roads and intersections. The distinction matters because identifying what a model attends to does not necessarily reveal what influences its driving. The authors call for methods that can capture individual objects’ causal impact and distinguish it from correlation. The abstract describes tests in synthetic scenes; it does not establish how well the findings carry over to real-world driving.
Original source
cs.AI · Article
Source details
Original source date: Nov 18, 2026 (2026-11-18).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-19T06:32:51.387Z. First observed: 2026-11-19T05:01:37.947Z. Retrieved: 2026-11-19T06:31:37.188Z.
Limits
The abstract reports analysis on synthetic driving scenes using an autonomous vehicle based on PlanT 2.0; it does not report results from real-world driving. The supplied evidence is the paper's abstract, not its full methods or results.
Open question
How well do the reported findings carry over from synthetic driving scenes to real-world driving?
Source-stated
The authors compare attention-based estimates of object relevance against a causal reference obtained through counterfactual analysis.
we compare different attention-based relevance estimation methods against a causal reference of object relevance obtained through counterfactual analysis.
Source-stated
The authors report that attention indicates correlation rather than causal impact, and can both miss relevant objects and classify objects with little impact as relevant.
Results show that attention provides a correlational rather than causal indication of object relevance.
Source-stated
The authors report that attention performs less accurately in complex road settings than in simple single-lane, straight-road scenarios.
attention can effectively identify relevant objects in simple single-lane, straight-road scenarios, but its accuracy degrades in more complex settings, such as multi-lane roads and intersections.
Source-stated
The authors call for methods that capture individual objects’ causal impact and distinguish it from correlational associations.
calling for novel methods able to accurately capture the causal impact of individual objects on the vehicle's driving and distinguish it from correlational associations.
Source and contributor context
Lusvarghi, Luca, Gozalvez, Javier · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Counterfactual
An argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Editorial provenance
Version 1. Publication ID: 414bea0c-85c6-4725-87a6-5f9ebaf34f78.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T06:32:51.387Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 5350dc739530b7fff02a3e2b05d726d33626513217dc9d93f4ae765aac9f9da0. Source text hash: f51bdfdb14af4897d518597abf2d1e068bd3a6ae661ecdf5048e3b0a9f9eccda.
uncertain · 2026-11-19T06:32:51.387Z · anthropic/claude-opus-5.5
Publication: 414bea0c-85c6-4725-87a6-5f9ebaf34f78
attribution — pass: The reading correctly attributes the paper and its findings to Luca Lusvarghi and Javier Gozalvez, and keeps the results framed as what the authors report.
Authors:Luca Lusvarghi, Javier Gozalvez
summary — uncertain: The summary is faithful, but a non-specialist may not understand the central terms or the stakes. It never says that attention is an internal mechanism of Transformer AI models. It also omits the source's point that attention is among the most widely adopted ways to estimate object relevance and interpret an autonomous vehicle's decisions; that context explains why the finding matters. Counterfactual analysis and PlanT 2.0 are left unexplained, though the abstract calls PlanT 2.0 a Transformer-based object-level planner. The scale of more than 80,000 scenes is dropped. Smallest repair: add one clause on attention's widespread use for interpretation, and use term-library tooltips for these terms.
Attention has become one of the most widely adopted solutions for estimating the relevance of surrounding objects and interpreting an autonomous vehicle's driving decisions.
date — pass: The displayed original publication date of 2026-11-18 matches the v1 submission date on arXiv.
[Submitted on 18 Nov 2026]
support — pass: Each claim is supported by the abstract. The statement that what a model attends to may not reveal what influences its driving is a fair restatement of the authors' point that attention is a limited proxy for relevance. The limitation about synthetic scenes and the question about real-world transfer follow from the evidence without overstating it.
These findings highlight the limitations of attention as a proxy for object relevance
media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct way to reach the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
An argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivAn argument or scenario that considers what would have resulted if events had gone differently from how they actually did, such as if a policy had never been adopted; as an adjective, relating to this kind of thinking about what did not happen but could have.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXiv