Reasons-Based Trajectory Evaluation Framework

• Reasons-Based Trajectory Evaluation Framework is an approach that judges trajectories by how well they reflect human reasons beyond traditional outcome metrics.
• It integrates stakeholder justifications using formal evaluation functions and weighted aggregation to capture legality, efficiency, and comfort.
• It supports dynamic ethical decision-making by triggering replanning when key reason scores fall below predefined stakeholder thresholds.

Searching arXiv for papers on reasons-based trajectory evaluation and closely related trajectory-level evaluation frameworks. Searching arXiv for "reasons-based trajectory evaluation framework automated vehicles Meaningful Human Control". Searching arXiv for exact title "Assessing the Alignment of Automated Vehicle Decisions with Human Reasons". The Reasons-Based Trajectory Evaluation Framework is an approach to trajectory assessment in which a candidate trajectory is judged not only by outcome variables such as safety, comfort, or task completion, but by the reasons that justify choosing it. In the automated-vehicle literature, the framework is formulated as a transparent evaluation layer that scores trajectories by how well they track the reasons of relevant human agents, thereby operationalizing the tracking condition of Meaningful Human Control (MHC) (Suryana et al., 31 Jul 2025). In related work, the same process-oriented orientation appears in supervision layers for replanning, reason-responsiveness tests for vision-language driving models, and trajectory-level auditing frameworks in agentic AI, all of which reject purely endpoint-based evaluation in favor of reason-sensitive assessment of the path itself (Suryana et al., 31 Jul 2025).

1. Motivation and conceptual scope

The framework arises from a mismatch between how many autonomous systems are evaluated and how human practical judgment is ordinarily exercised. In automated driving, prevailing approaches are commonly organized around rigid rules, hand-built heuristics, optimization objectives, collision avoidance, trajectory smoothness, and kinematic feasibility. The reasons-based framework argues that this is inadequate for everyday ethical conflict, where legality, efficiency, comfort, safety, and social expectations can point in different directions. The canonical examples are not rare trolley-style dilemmas but ordinary cases such as overtaking a cyclist on a no-passing road, rolling stops at empty junctions, or choosing between strict compliance and socially sensible behavior (Suryana et al., 31 Jul 2025).

This critique has a close analogue in model evaluation. CARE-Drive observes that standard driving evaluation is mostly outcome-based, checking collision rates, trajectory error, comfort metrics, or whether a maneuver looks correct, but not whether a model’s decision changes when the relevant human reasons change. The concern is that a system may produce a plausible explanation that is merely post hoc rationalization, thereby creating false confidence in a safety-critical setting (Suryana et al., 17 Feb 2026). A reasons-based framework therefore asks a different question: not merely whether a trajectory is acceptable in hindsight, but whether the trajectory is responsive to the human-relevant considerations that justify it.

Within this conception, trajectory evaluation is not identical to planning. The framework does not replace the planner or controller; it assesses candidate or executed trajectories in a way that is ethically interpretable. This distinction is central: the contribution is a formal evaluation layer over existing planning algorithms, rather than a new motion-planning algorithm.

2. Normative foundation and formal structure

The normative core of the framework is the tracking condition of Meaningful Human Control. Tracking is interpreted as responsiveness to the reasons of relevant human agents, including policymakers, drivers or passengers, cyclists, pedestrians, and, more generally, anyone affected by the decision. To operationalize this requirement, the framework models each agent’s reasons as formal evaluation functions and aggregates them across time, across reasons within an agent, and across agents (Suryana et al., 31 Jul 2025).

Let the agent set be $\mathcal{H} = \{h_1,\dots,h_n\}$, and let each agent $h_i$ have a reason set $\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$. Each reason is represented by a per-time-step function

$$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$

where the value $1$ denotes full satisfaction of that reason at that time step and lower values denote poorer alignment.

The corresponding trajectory-level reason score is the time average

$$F_{ib}(T_a, \mathcal{E}) = \frac{1}{p+1} \sum_{l=0}^{p} f_{ib}(s_{al}, \mathcal{E}_l, t_l).$$

Reasons are then aggregated within each agent as

$$S_i(T_a) = \sum_{b=1}^{m_i} \alpha_{ib} F_{ib}(T_a, \mathcal{E}),$$

where $\alpha_{ib} \in [0,1]$ weights reason $b$ for agent $i$. Across agents, the unbalanced global score is

$h_i$0

with $h_i$1 denoting the importance assigned to agent $h_i$2.

The final score introduces a balance term: $h_i$3 The stated role of $h_i$4 is to penalize excluding agents. The ideal weight distribution is typically uniform, with

$h_i$5

The framework reports that equal weights $h_i$6 yield $h_i$7, whereas a skewed assignment with $h_i$8 and $h_i$9 yields $\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$0 (Suryana et al., 31 Jul 2025). The significance of this design is that ethical acceptability is not reduced to maximizing one stakeholder’s satisfaction; the balance function explicitly discourages structurally ignoring any relevant agent.

3. Canonical case study: overtaking a cyclist on a no-passing road

The main case study is a real-world-inspired overtaking scenario in which an automated vehicle follows a cyclist traveling at about $\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$1 on a two-lane rural road with double solid yellow lines, road width $\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$2, lane width $\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$3 per lane, speed limit $\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$4, and no oncoming traffic (Suryana et al., 31 Jul 2025). The scenario is chosen because it exposes a genuine normative conflict rather than a simple safety violation.

Three agents are modeled. The policymaker prioritizes regulatory compliance and lane discipline. The driver prioritizes time efficiency and reduced delay. The cyclist prioritizes safety and comfort, especially lateral clearance. In the simulation, each agent has one reason, so $\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$5.

Four candidate trajectories are compared. Trajectory 1 is a Small-Gap Overtake, favoring driver efficiency. Trajectory 2 is a Medium-Gap Overtake, intended as a more balanced alternative. Trajectory 3 is a Large-Gap Overtake, favoring cyclist safety and comfort at the cost of extended lane encroachment. Trajectory 4 is Conservative Following, which avoids overtaking and is fully compliant with the law but least efficient.

The reason functions are scenario-specific. The policymaker score is a legality or lane-compliance function based on lateral displacement relative to the lane centerline, with full score for compliance and exponential decay under lane-rule violation. The driver score depends on how long the automated vehicle remains close behind the cyclist; satisfaction remains high until waiting becomes too long, after which the score decays exponentially. The cyclist score is the product of a spatial safety term and a temporal comfort term, capturing both lateral clearance and the discomfort of prolonged close following.

Under equal weights, the paper reports that Trajectory 1 achieves the highest overall score and Trajectory 4 the lowest (Suryana et al., 31 Jul 2025). This does not mean the framework endorses a single universal answer. Rather, the weight-sensitivity analysis shows that the preferred trajectory changes substantially with the priority distribution. The ternary plot reported in the paper identifies regions where Trajectory 1 is selected when driver efficiency is prioritized, Trajectory 3 when cyclist concerns dominate, and Trajectory 4 when policymaker or legal compliance dominates. The authors also report abrupt changes in the chosen trajectory under small weight changes, indicating threshold effects in ethical AV decision-making. A common misconception is therefore that the framework encodes one fixed moral ranking; in fact, its explicit purpose is to reveal how the ranking depends on stated normative priorities.

4. Supervision, thresholding, and replanning

A closely related development converts reasons-based evaluation into a real-time supervision framework integrated with planning and control. In that formulation, the architecture has three components: global motion planning using A* search with motion primitives, Model Predictive Control (MPC) for kinodynamic trajectory tracking, and a Human Reasons-based Supervision Framework that continuously evaluates the current trajectory and triggers replanning when stakeholder reason scores drop below threshold (Suryana et al., 31 Jul 2025).

The navigation problem is written with stakeholder reason constraints: $\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$6 subject to dynamics, safety, control constraints, and

$\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$7

Here the stakeholder set is

$\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$8

and the reason functions satisfy

$\mathcal{R}_i = \{r_{i1},\dots,r_{im_i}\}$9

The trigger condition is explicitly threshold-based: $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$0 When the minimum stakeholder score drops below its threshold, replanning is invoked. In the reported cyclist-overtaking implementation, the thresholds are $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$1, $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$2, $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$3, $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$4, and $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$5 for all stakeholders (Suryana et al., 31 Jul 2025).

The experimental contrast is sharp. Without supervision, the automated vehicle follows behind the cyclist, exhibits stop-and-go behavior, reaches the goal in about $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$6 seconds, aligns well with the policymaker reason, but the driver reason decays sharply to zero and cyclist comfort declines over time. With the replanner, the vehicle follows initially, the driver reason falls below the threshold of $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$7 at around $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$8 seconds, replanning is triggered, the vehicle briefly enters the opposite lane to overtake, and the goal is reached in about $$f_{ib}(s_{al}, \mathcal{E}_l, t_l): (s_{al}, \mathcal{E}_l, t_l) \rightarrow [0,1],$$9 seconds (Suryana et al., 31 Jul 2025). The paper interprets this as evidence that rigid rule-following can be ethically inappropriate in context, and that reason-based supervision can prompt adaptation when the current trajectory ceases to track relevant human reasons.

5. Adjacent driving frameworks and alternative notions of “reasons”

Several neighboring frameworks evaluate trajectories through explicit justificatory structure, though they differ in what counts as a reason and how it is operationalized. CARE-Drive evaluates reason-responsiveness in vision-LLMs behaviorally rather than through direct trajectory scoring. It compares baseline and reason-augmented decisions under controlled contextual perturbations and uses a two-stage process consisting of prompt calibration and contextual evaluation. In the cyclist overtaking scenario, role-only prompts produced a $1$0 overtaking rate in calibration, while reason-augmented prompting increased overtaking, and the calibrated configuration was reported as gpt-4.1 with Tree-of-Thought (Suryana et al., 17 Feb 2026). CARE-Drive therefore measures whether explicit human reasons causally influence observable decision behavior, not merely whether the final maneuver is acceptable.

WoTE provides a different form of reasons-based evaluation grounded in predicted future consequences. It uses a BEV world model to imagine future bird’s-eye-view states for candidate trajectories and a reward model decomposed into imitation reward and simulation rewards for no collisions, drivable area compliance, time-to-collision, comfort, and ego progress. The final trajectory is the one with the highest predicted reward, and the paper explicitly characterizes the framework as reasons-based because the decision path is candidate trajectory $1$1 predicted future scene $1$2 interpretable rewards $1$3 selection (Li et al., 2 Apr 2025).

Other driving papers ground trajectory evaluation in formal safety or multi-criteria structure. One framework uses logged vehicle trajectories to determine whether the subject vehicle is in a collision unavoidable situation, thereby creating an objective reference for comparing TTC, PCM, and MPrISM without contamination from background-vehicle prediction assumptions (Yan et al., 2024). Another proposes a unified safety-comfort-efficiency objective in mixed traffic, with adaptive elliptical safety zones, Shoelace-formula overlap area, jerk-based comfort constraints, and travel time as the efficiency term (Naidja et al., 1 Sep 2025). These frameworks are reasons-based in a broader engineering sense: they evaluate trajectories through explicit grounds such as collision inevitability, comfort, and efficiency, even when they do not formalize stakeholder reasons in the MHC sense.

6. Generalization beyond driving: process-aware and trajectory-level evaluation

The reasons-based orientation extends well beyond automated vehicles. In deep research agents, full-trajectory auditing replaces end-to-end answer scoring with a process-aware framework that decomposes the trajectory into atomic sub-queries, actions, and claims, and scores hallucinations via the PIES Taxonomy. Its composite score is

$1$4

where the four terms correspond to explicit summarization, implicit summarization, explicit planning, and implicit planning hallucination (Zhan et al., 30 Jan 2026). The substantive move is analogous: diagnosis is grounded in the reasons for failure along the trajectory, not only in the terminal report.

Tool-augmented LLM agents are evaluated similarly in TRACE, which uses an evidence bank that accumulates prior context and supports multi-dimensional scoring of correctness, efficiency, hallucination, and adaptivity (Kim et al., 3 Oct 2025). A separate TRACE framework for sabotage monitoring in long-horizon LLM-agent trajectories implements a TIJ loop—Triage, Inspect, Judge—with persistent evidence memory and reports aggregate $1$5 and recall $1$6 on SHADE-Arena (Mittapalli et al., 5 Jun 2026). In vision-language reasoning, TRACE decomposes tasks into Auxiliary Reasoning Sets and evaluates sampled reasoning paths through consistency-based metrics such as PMC, PDC, PZC, GMC, and the First Failure Step (Imani et al., 5 Dec 2025). In iterative scientific design, LEAPBench argues that the relevant unit is the full adaptive process and introduces best-so-far AUC as a trajectory metric, reporting that switching from final-outcome to trajectory scoring changes the best-model decision on $1$7 of biology tasks at matched horizons (Zhang et al., 14 May 2026).

Trajectory-level evaluation also appears in data selection and reasoning diagnostics. Rank-Surprisal Ratio evaluates whether reasoning trajectories are both informative and aligned for student-model distillation, with an average Spearman correlation of $1$8 with post-training performance (Yang et al., 20 Jan 2026). TRACED evaluates reasoning quality through latent-space displacement and curvature, associating correct reasoning with high progress and low curvature, and hallucinations with low progress and high curvature (Jiang et al., 11 Mar 2026). Taken together, these works suggest a broader methodological convergence: endpoint scoring is increasingly supplemented by frameworks that analyze why a trajectory is good or bad, where it first fails, and which intermediate considerations drove the outcome.

7. Limitations, unresolved choices, and methodological significance

The framework’s interpretability depends on commitments that remain open. The automated-vehicle formulation does not solve weight assignment: it uses weights over agents and reasons, but does not determine how those weights should be chosen ethically. Equal weighting is used for simplicity, yet the framework itself shows that small weight changes can alter the preferred trajectory abruptly (Suryana et al., 31 Jul 2025). Nor does it guarantee that the mathematical reason functions perfectly capture human reasons; the formalization assumes legality, comfort, or efficiency can be translated faithfully into evaluable functions.

The current case studies are also simplified. The overtaking scenario is a single-AV, single-cyclist setting and does not yet cover dense traffic, negotiation among many agents, or long-horizon planning. The supervision variant relies on fixed thresholds, which are interpretable but coarse, and the reason models are scenario-specific. That framework detects misalignment and triggers reconsideration, but does not itself solve the problem of selecting the best alternative trajectory among competing ethical options (Suryana et al., 31 Jul 2025). CARE-Drive adds a further caution: behavioral sensitivity to reasons does not prove internal causal reasoning, since the framework is explicitly model-agnostic and infers reason-responsiveness from systematic output changes under prompt and context variation (Suryana et al., 17 Feb 2026).

These limitations do not negate the methodological significance of reasons-based trajectory evaluation. Rather, they define its distinctive contribution. The framework shifts evaluation from black-box outcome approval to explicit justification in terms of stakeholder reasons, formal criteria, or process-level evidence. In automated driving, that means assessing whether a maneuver tracks legality, efficiency, comfort, and safety in a manner compatible with Meaningful Human Control. In adjacent fields, it means replacing final-answer or final-outcome scoring with structured auditing of the reasoning path itself. The unifying claim is that a trajectory is not fully understood until the reasons that support it, and the reasons that undermine it, are made explicit.