Paper deep dive
YOUDrive: Driving Style as a Composable Direction for End-to-End Autonomous Driving
Guantian Zheng, Lidong Yu, Tianyu Gao, Jiaxing Chen, Jiashu Li, Chen Lv
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:End-to-end driving policies can achieve strong task performance while offering limited control over their driving style. Driving style describes how a vehicle accelerates, brakes, and maintains distance. Policies trained without explicit preference conditioning lack a direct mechanism for adjusting these behaviors at inference. We propose YOUDrive, a vision--language--action planner for continuous and composable driving-style control. Each teacher's behavior is distilled into a task vector, a learned parameter update that induces a style direction in the planner's representation. These directions can be scaled and combined at inference without retraining the backbone. Conditioned on these directions, a flow-matching decoder models a multimodal trajectory distribution and generates kinematically feasible motion alternatives. We introduce the Style Alignment Score (SAS) to describe human-referenced kinematic behavior after safety and task gating. On NAVSIM navtest, a single coefficient monotonically controls assertiveness across six teacher-derived personas, while PDMS remains above 0.90 across the tested individual and blended configurations, reaching 0.954. Further analyses reveal systematic policy-specific kinematics after controlling for scene context, whereas the evaluated discrete style labels provide limited additional explanatory power. Code, checkpoints, and the metric will be released.
Tags
Links
- Source: https://arxiv.org/abs/2610.14802v1
- Canonical: https://arxiv.org/abs/2610.14802v1
Trouble viewing inline? Open PDF directly →
Full Text
57,127 characters extracted from source content.
Expand or collapse full text
YOUDrive: Driving Style as a Composable Direction for End-to-End Autonomous Driving Guantian Zheng1,3 Lidong Yu3 Tianyu Gao4 Jiaxing Chen2 Jiashu Li4 Chen Lv1,† 1AutoMan Lab, Nanyang Technological University 2School of Vehicle and Mobility, Tsinghua University 3Neolix 4City University of Hong Kong gzheng004@e.ntu.edu.sg Abstract End-to-end driving policies can achieve strong task performance while offering limited control over their driving style. Driving style describes how a vehicle accelerates, brakes, and maintains distance. Policies trained without explicit preference conditioning lack a direct mechanism for adjusting these behaviors at inference. We propose YOUDrive, a vision–language–action planner for continuous and composable driving-style control. Each teacher’s behavior is distilled into a task vector, a learned parameter update that induces a style direction in the planner’s representation. These directions can be scaled and combined at inference without retraining the backbone. Conditioned on these directions, a flow-matching decoder models a multimodal trajectory distribution and generates kinematically feasible motion alternatives. We introduce the Style Alignment Score (SAS) to describe human-referenced kinematic behavior after safety and task gating. On NAVSIM navtest, a single coefficient monotonically controls assertiveness across six teacher-derived personas, while PDMS remains above 0.90 across the tested individual and blended configurations, reaching 0.954. Further analyses reveal systematic policy-specific kinematics after controlling for scene context, whereas the evaluated discrete style labels provide limited additional explanatory power. Code, checkpoints, and the metric will be released. †footnotetext: †Corresponding author. 1 Introduction End-to-end driving models learn to predict trajectories from sensor inputs using large-scale driving datasets that pool demonstrations from multiple drivers [3, 9, 12]. These data reflect not only the demands of each scene but also how individual drivers execute their maneuvers. Their patterns of acceleration, braking, and following distance characterize driving style. A policy trained on these demonstrations acquires its own driving behavior, but learning from different drivers does not by itself give users control over that behavior. The challenge is to make driving style adjustable at inference while preserving the policy’s ability to respond to traffic and complete the driving task. One approach is to condition a driving policy on discrete style labels, such as aggressive, normal, and conservative [6]. However, the distribution of scenes and motion states can differ substantially across these categories. When conservative labels are concentrated in low-speed or near-stationary clips, they reflect scene conditions more than driver preferences (Sec. 4.4). This limits their ability to distinguish driver preferences from scene-dependent behavior. For style control, we instead seek a behavioral reference that can be examined under comparable driving conditions. We treat each trained teacher policy as a driver and characterize its style through the trajectories it produces across a common set of scenes. Building on this policy-level view, we propose YOUDrive to transfer teacher behaviors to a shared vision–language–action (VLA) planner and control their expression at inference. For each teacher, we distill its driving behavior into a task vector, a learned parameter update relative to a common base planner [11]. For a given scene, the difference between the updated and base representations defines the teacher’s style direction. A coefficient controls the strength of this direction, allowing interpolation between the two representations. Directions from multiple teachers can also be combined through weighted addition in the same representation space, without retraining the shared planner. A flow-matching decoder translates the resulting style condition into motion by generating control sequences and integrating them through a constrained motion model. This produces multiple kinematically feasible trajectory candidates under the chosen style condition. To characterize the resulting behavior, we introduce the Style Alignment Score (SAS), which measures interpretable, human-referenced kinematic features after safety and task gating. Reporting SAS alongside the unchanged PDMS allows style changes to be examined separately from planning performance. Our NAVSIM evaluation covers both single-teacher control and multi-teacher composition, demonstrating systematic changes in assertiveness while maintaining high planning performance. Our contributions are summarized as follows: • We introduce a task-vector formulation of driving-style control in a shared VLA planner, coupled with a style-conditioned flow-matching decoder. Teacher-derived directions can be scaled and combined at inference without retraining the backbone. • We propose the Style Alignment Score (SAS), which characterizes human-referenced kinematic behavior after safety and task gating, allowing driving style to be reported separately from planning performance. • Our NAVSIM study identifies policy-specific kinematic patterns and demonstrates monotonic assertiveness control across six teacher-derived personas. PDMS remains above 0.90 across the tested individual and blended configurations, reaching 0.954. 2 Related Work End-to-end and VLA planning. End-to-end driving methods integrate scene understanding and trajectory planning within a learned policy. TransFuser combines camera and LiDAR features for imitation-based driving [3]. UniAD coordinates perception and prediction tasks around planning, while VAD uses vectorized scene representations [9, 12]. VLA approaches use pretrained vision-language models to connect visual observations and language instructions with action generation. EMMA represents driving inputs and outputs in a unified language format [10], whereas AutoVLA encodes trajectories as discrete feasible actions and predicts them autoregressively [27]. For multimodal trajectory generation, DiffusionDrive uses truncated diffusion with multimodal anchors [19], while GoalFlow employs goal-conditioned flow matching [25]. YOUDrive builds on a VLA planner and replaces discrete action decoding with a continuous flow-matching head in a kinematically constrained control space. We focus on steering the resulting trajectory distribution with teacher-derived behavior directions. Personalized autonomous driving. Personalized driving has been studied through learning driving styles from demonstrations and modeling social preferences in vehicle interactions [13, 23]. In end-to-end driving, StyleDrive introduces scenario-aware aggressive, normal, and conservative annotations and corresponding style-conditioned baselines [6]. Its style-modulated PDMS adjusts comfort, progress, and time-to-collision terms to assess style alignment. Although these labels enable categorical control, unequal scene and motion-state distributions across categories complicate the interpretation of individual preferences, as examined in our analysis (Sec. 4.4). StyleVLA conditions trajectory generation on driving-style instructions with physics-informed training objectives [7]. TrajHF fine-tunes generative trajectory models using human feedback to align their outputs with driving preferences [14]. YOUDrive focuses on deriving reusable behavior directions through teacher-policy distillation. Each teacher provides a behavioral reference across scenes, and its induced direction can be scaled and combined with other teacher directions within a shared VLA planner. We also retain PDMS and report human-referenced kinematic behavior through SAS after safety and task gating, so behavioral changes can be examined separately from planning performance. Task vectors and activation steering. Task arithmetic constructs task vectors by subtracting pretrained weights from their fine-tuned counterparts, allowing learned adaptations to be scaled and combined [11]. Model Soups averages fine-tuned model weights [24], while TIES-Merging addresses interference between parameter updates during merging [26]. Activation steering instead intervenes in hidden representations during inference. ITI shifts selected attention-head activations [15], while Persona Vectors identifies activation directions associated with behavioral traits [2]. YOUDrive links parameter updates to representation-level control through teacher-policy distillation. The learned task vector induces a scene-dependent direction in the VLA planning representation. We scale and compose these directions to condition trajectory generation, rather than directly merging weights at inference. 3 Method 3.1 Overview At each planning step, YOUDrive takes multi-view images I, ego state e, and a navigation command g as input and predicts a trajectory τ=((,,,,,))k=1Tτ= ((x_k,y_k, _k) )_k=1^T, where each pose contains a planar position and heading. We use T=10T=10 poses at intervals of Δt=0.5 t=0.5 s. A shared VLA backbone encodes the inputs into content tokens. Each teacher’s behavior is distilled into a task vector, whose induced change in the content representation defines a scene-dependent style direction. The selected direction is scaled, or several directions are combined, to condition a flow-matching decoder (Fig. 1). The decoder generates control sequences that are integrated through a constrained motion model to produce trajectory candidates, from which the final output is selected. Figure 1: YOUDrive pipeline. The VLA backbone provides content tokens, style is a task-vector interpolation applied to those tokens, and a flow-matching decoder samples several feasible candidates before returning one. 3.2 VLA backbone YOUDrive uses Qwen2.5-VL-3B [1] as a shared scene encoder. Multi-view camera images I are encoded as visual tokens, while ego state e and navigation command g are provided as text tokens. After base-policy training, the backbone maps x=(I,e,g)x=(I,e,g) to a content sequence h0(x)=fθ0(x)∈ℝL×dh_0(x)=f_ _0(x) ^L× d. Here θ0 _0 denotes the trained backbone parameters; the trajectory decoder has separate parameters. This sequence provides scene and navigation context for trajectory generation and serves as the common reference for the teacher-specific representation changes defined in Sec. 3.3. 3.3 Style as a direction in VLA content space For each teacher policy p, we train a LoRA adapter [8] on its navtrain trajectories while keeping the base backbone and trajectory decoder fixed. The training objective is the same control-space flow-matching objective used for the base, with teacher rather than human trajectories as targets (Sec. 3.4). Following task arithmetic [11], we treat the adapter’s parameter update ΔWp W_p relative to the shared backbone θ0 _0 as a task vector. Applied to the same scene x, the base and teacher-adapted backbones produce h0(x)h_0(x) and hp(x)=fθ0+ΔWp(x)h_p(x)=f_ _0+ W_p(x), respectively. Their difference defines the teacher-induced style direction Δhp(x)=hp(x)−h0(x). h_p(x)=h_p(x)-h_0(x). (1) The parameter update ΔWp W_p and the representation change Δhp(x) h_p(x) are distinct: the latter depends on the current scene and is the quantity used to control trajectory generation. Scaling and composition. For a single teacher, the decoder receives the interpolated content cp,α(x)=h0(x)+αΔhp(x),α∈[0,1],c_p,α(x)=h_0(x)+α\, h_p(x), α∈[0,1], (2) where α=0α=0 recovers the base policy and α=1α=1 uses the teacher-adapted content. For a blend, we first combine the task vectors with weights βp _p and evaluate the adapted backbone: ΔW W_ β =∑pβpΔWp, = _p _p W_p, (3) cα,(x) c_α, β(x) =h0(x)+α(fθ0+ΔW(x)−h0(x)). =h_0(x)+α (f_ _0+ W_ β(x)-h_0(x) ). The βp _p used for the reported blends sum to one; the overall strength α remains independently adjustable. Because the backbone is nonlinear, the blend-induced representation change need not equal ∑pβpΔhp(x) _p _p h_p(x). We test the resulting behavior under scaling and composition in Sec. 4.5. Why the content space. The VLA content sequence already provides the decoder with scene and navigation context. Interpolating the base and adapted content therefore lets a teacher update change the representation consumed by the same trajectory generator, without training a separate decoder or restricting style to a fixed label. 3.4 Flow-matching trajectory decoder Our decoder is a flow-matching action head for a VLA. Generative planners such as DiffusionDrive and GoalFlow model multiple plausible trajectories [19, 25]. YOUDrive retains this capability while using teacher-derived directions as continuous, composable conditions for trajectory generation. The DiT/AdaLN-based head [22] cross-attends to the style-conditioned VLA content sequence c of Eqs. 2 and 3 and generates multiple candidates. In the content-interpolation configuration, the separate AdaLN style input is kept neutral; the teacher direction acts through the content sequence. We decode in a control space: writing uk=(ak,ωk)u_k=(a_k, _k) for longitudinal acceleration and yaw rate, each candidate corresponds to a control sequence z=(u1,…,uT)∈ℝT×2z=(u_1,…,u_T) ^T× 2. The training target z1z_1 is obtained by inverting the target trajectory through the unicycle model and standardizing the two control channels. Base training uses human trajectories; persona distillation uses the corresponding teacher trajectories. The initial speed for target inversion is estimated from the first target displacement. Rectified-flow objective. Following rectified flow and flow matching [21, 20], for z0∼(0,I)z_0 (0,I) and target z1z_1 we construct zt=(1−t)z0+tz1,t∈[0,1],z_t=(1-t)\,z_0+t\,z_1, t∈[0,1], (4) with target velocity z1−z0z_1-z_0. The decoder vϕv_φ, with parameters ϕφ separate from the backbone θ0 _0, regresses that velocity conditioned on the content sequence c through cross-attention: ℒFM=t,z0,z1‖vϕ(zt,t,c)−(z1−z0)‖2.L_FM=E_t,\,z_0,\,z_1 \|\,v_φ(z_t,t;c)-(z_1-z_0)\, \|^2. (5) The training objective also includes a lightly weighted control-jerk penalty (λj=0.01 _j=0.01) on the estimated clean controls, weighted by t for flow times t>0.5t>0.5. The same objective is used for base training and persona distillation. Sampling. At inference we draw N=16N=16 independent initial samples and integrate z˙t=vϕ(zt,t,c) z_t=v_φ(z_t,t;c) using K=100K=100 Euler steps: zt+1K=zt+1Kvϕ(zt,t,c).z_t+ 1K=z_t+ 1K\,v_φ(z_t,t;c). (6) The sampled set can represent distinct maneuvers under the same scene and style condition instead of averaging them into a single point estimate. After reversing control normalization, we cap acceleration and yaw rate as a¯k=clip(ak,−amax,amax) a_k=clip(a_k,-a_ ,a_ ) and ω¯k=clip(ωk,−ωmax,ωmax) ω_k=clip( _k,- _ , _ ). We use amax=5a_ =5 m/s2, ωmax=1 _ =1 rad/s, and vmax=25v_ =25 m/s. From initial ego speed v0v_0 (provided by e at inference), heading ψ0=0 _0=0, and position p0=(0,0)p_0=(0,0), the unicycle model integrates the controls as vk v_k =clip(v0+Δt∑i=1ka¯i, 0,vmax),ψk=ψk−1+ω¯kΔt, =clip\! (v_0+ t _i=1^k a_i,\,0,v_ ), _k= _k-1+ ω_k t, (7) pk p_k =pk−1+vk(cosψk,sinψk)Δt, =p_k-1+v_k( _k, _k)\, t, yielding candidates within the specified kinematic bounds. Finally, a prediction-based PDM selector [4, 5] ranks the N candidates using the current scene, map context, and constant-velocity forecasts of surrounding agents, without using their logged future trajectories. The selected candidate is the planner output on which style and planning performance are evaluated. 4 Experiments 4.1 Setup Our planner uses Qwen2.5-VL-3B [1] as its scene encoder: multi-view images, ego status, and the navigation command are tokenized into content tokens summarizing both the scene and the requested maneuver. On top we run the flow-matching decoder of Sec. 3.4, conditioned on the content tokens through cross-attention; it flows in a control space of longitudinal acceleration and yaw rate under a unicycle model, integrated with K steps and sampled N=16N=16 times per frame. We warm-start from the released AutoVLA_PDMS_89 checkpoint [27] and fine-tune the language backbone jointly with the decoder on NAVSIM navtrain, vision encoder frozen and the human trajectory as target; the result is our “base” (α=0α=0). Each persona adapter is distilled from a teacher planner—GTRS [18], Hydra-MDP [17], WoTe [16], TransFuser [3], GoalFlow [25], DiffusionDriveV2 [19]—and applied through Eq. 2. Each is a LoRA [8] module (r=64r=64, αLoRA=128 _LoRA=128, dropout 0.050.05) on the attention and MLP projections; the base VLM (including the vision tower) and the flow head are frozen, so only the LoRA weights train per persona. We evaluate on NAVSIM [5] navtest (about 12k scenarios, 8 shards) and report scenario-weighted PDMS (v1), alongside our Style Alignment Score (SAS), which places each policy on an assertiveness axis H and a proximity axis P relative to the human (Sec. 4.2). 4.2 SAS: Style Alignment Score Gate, then score. Safety and driving style answer different questions and should not share one axis, so we evaluate a policy in two stages. First a gate: a trajectory must pass NAVSIM’s safety and task terms (no-collision, drivable-area compliance, time-to-collision, ego progress); style is scored only for trajectories that clear the gate, and style never modulates the safety verdict. Second, we ask how the passing motion aligns in style with a target driver. SAS and PDMS are thus reported on separate axes: we leave PDMS unchanged and report SAS beside it. This separation is one of reporting—neither score modulates the other—not a claim of statistical independence; indeed a persona that improves ego-progress can also raise PDMS. Atoms and human-referenced normalization. We summarize a trajectory by physically interpretable kinematic atoms: dynamics a+,a−,jℓ,jt\a^+,a^-,j_ ,j_t\ (peak accel/decel, longitudinal and lateral jerk) and proximity thw,ttc,dmin,dvru\thw,ttc,d_ ,d_vru\ (min time-headway, time-to-collision, object and VRU gap). For a policy d and atom f we take the per-scene value and its median mf(d)m_f(d) over the split. To compare heterogeneous units against the human driver, we apply a signed z-score zf(d)=mf(d)−mf(human)σf,σf=stdd′mf(d′),z_f(d)\;=\; m_f(d)-m_f(human) _f, _f=std _d m_f(d ), (8) where σf _f is taken over the evaluated policies (the human, our base, and the six teachers); thus z=0z=0 is human, z>0z>0 is “more than human”, and one unit is one cross-policy standard deviation. SAS is therefore a within-study, relative diagnostic: it measures how far apart policies sit on a fixed evaluation set, not a portable benchmark. Its validity rests on independent evidence (per-model separability and monotone response to α), not on a universal scale. Two style axes. We aggregate the atoms into a primary Assertiveness axis and a secondary Proximity axis, H(d)=1||∑f∈zf(d),P(d)=1||∑f∈(−zf(d)),H(d)= 1|D|\! _f z_f(d), P(d)= 1|P|\! _f (-z_f(d) ), (9) with D the dynamics atoms and P the proximity atoms (equal weight). H>0H>0 denotes harder, jerkier driving than the human; P>0P>0 denotes tighter gaps. SAS reports style as (H,P)(H,P) with H as the primary, validated axis. What makes SAS valid. A style metric must (i) separate drivers and (i) respond monotonically to a style control. We show that H is strongly per-model separable (Sec. 4.4) and is driven monotonically by the single strength α (Sec. 4.5), and we report a per-atom variance ratio η2η^2 next to the axes so the metric states its own discriminative power. On navtest the proximity atoms vary little across policies; we therefore keep P secondary and return to it as a limitation. 4.3 Style and safety on navtest (v1) Table 1 reports PDMS with its sub-scores on navtest together with the SAS assertiveness axis H, for the base policy and, for each of the six teachers, the native teacher planner and our distilled persona. Every persona is applied at full strength (α=1α=1); the continuous behavior as α varies is the subject of Sec. 4.5, and persona blends, which reach operating points beyond any single teacher, are shown in Tab. 5(b). At full strength the personas span a wide assertiveness range, from smooth teachers close to the human (H≈0.05H≈0.05–0.150.15 for GTRS, Hydra-MDP) to an aggressive persona (DiffusionDriveV2, H≈3.3H≈3.3), while our personas keep PDMS in a narrow band (0.900.90–0.940.94) around the base. The sub-scores localize the change. The hard-safety components that gate the score (no-collision NC and drivable-area compliance DAC) stay essentially flat, whereas ego-progress (EP), time-to-collision (TTC), and comfort (Comf) move with H. Each distilled persona lands near its native teacher’s H, so the adapter transfers the teacher’s assertiveness rather than a generic average: where a native planner sits at one fixed point on the assertiveness axis, the matching persona sweeps a controllable range around it (Sec. 4.5), and the complementary proximity axis is deferred to Sec. 4.7. PDMS sub-scores SAS config α PDMS↑ NC DAC EP TTC Comf H human anchor – – – – – – – 0.00 base (ours) 0 0.912 98.8 96.6 86.7 96.3 100.0 0.48 each teacher’s native style (α=α=T) and our persona steered to it (α=1α=1) GTRS T 0.899 99.3 97.6 83.2 96.1 100.0 −0.01-0.01 ↪ ours 1.0 0.903 97.5 96.6 87.6 93.6 100.0 0.05 Hydra-MDP T 0.865 98.3 96.0 78.7 94.6 100.0 +0.05+0.05 ↪ ours 1.0 0.907 98.0 96.6 87.7 94.4 100.0 0.13 WoTe T 0.883 98.5 96.8 81.9 94.9 99.9 0.16 ↪ ours 1.0 0.905 98.4 95.6 86.9 95.5 100.0 0.17 TransFuser T 0.840 97.7 92.8 79.2 92.8 100.0 0.06 ↪ ours 1.0 0.900 98.6 95.7 85.8 95.5 100.0 0.10 GoalFlow T 0.903 98.4 98.3 85.0 94.6 100.0 1.93 ↪ ours 1.0 0.935 99.4 97.9 89.0 98.0 100.0 1.99 DiffusionDriveV2 T 0.881 98.2 96.2 82.2 94.7 100.0 2.72 ↪ ours 1.0 0.941 99.5 98.5 89.5 98.1 100.0 3.26 Table 1: PDMS with its sub-scores (no-collision NC, drivable-area compliance DAC, ego-progress EP, time-to-collision TTC, comfort Comf) and the SAS assertiveness axis H (Eq. 9) on NAVSIM navtest. For each teacher we list its native behavior (α=α=T) and, below it (↪ ours), our persona steered to that teacher at full strength (α=1α=1); base is the same model with no persona (α=0α=0), and the human driver sits at H=0H=0. Native-teacher PDMS are the reported navtest v1 numbers from each planner’s paper [3, 17, 25, 19, 16]; GTRS reports v2 EPDMS only, so its v1 row is our own evaluation. 4.4 Style is a per-model signal, not a label A style axis is only meaningful if style is a real, systematic signal rather than scene noise. Here a driver is a behaviorally self-consistent policy—a model, in the sense of Sec. 1—so its style is the kinematic signature it produces across scenes. On a fixed set of tokens we collect trajectories for 16 such drivers (the human reference, external planners, and our base) and run a one-way variance decomposition of per-token kinematics with driver identity as the factor (Tab. 2). Driver identity explains a mean η2η^2 of 0.27, up to 0.46 for lateral jerk, so how a driver accelerates, brakes, and jerks is strongly driver-specific. The one exception is average speed (η2=0.005η^2=0.005), set by road and traffic—hence a weak style coordinate on navtest. This supports treating style as a continuous axis organized by model-as-driver rather than coarse labels. Figure 2 makes the split visible in atom space—dynamics atoms fan out across policies while proximity atoms collapse near the human—so we take assertiveness H as primary and keep proximity P secondary (Sec. 4.7). kinematic feature η2η^2(driver)↑ ICC(1) F avg. speed 0.005 0.005 20 peak accel 0.275 0.288 1631 peak decel 0.287 0.300 1735 long. jerk 0.445 0.461 3451 lat. jerk 0.460 0.476 3663 lat. accel (max) 0.167 0.176 862 mean 0.273 0.286 – Table 2: Variance decomposition of per-token kinematics with driver identity as the factor (16 drivers, common tokens). Control features are driver-dominated, while average speed is scene-locked. In contrast, the coarse ANC label of [6] carries little style signal once environment is accounted for (Sec. 4.4). Figure 2: Style geometry in atom space (human-referenced signed z; the dotted ring marks the human reference, radius grows with assertiveness on the dynamics half H and with gap-tightening on the proximity half P). Left: native teachers (α=α=T); right: our personas. Both fan out along the dynamics atoms (the primary axis H); on the proximity atoms the cross-policy spread is small, so P is near-degenerate on navtest and secondary, with a teacher/persona sign mismatch we treat as an open problem (Sec. 4.7). Discrete labels do not encode style beyond speed. If model-as-driver is the right unit for style, the natural alternative—hand-assigned categorical labels—should be tested with the same kinematic atoms. We take the aggressive/normal/conservative (ANC) labels released by StyleDrive [6] for the navtest scenes and evaluate them with our kinematic atoms, on the human log-replay trajectory of each scene. This is the fair test: StyleDrive assigns ANC from each clip’s own ego kinematics, so measuring the labels against those same kinematics asks exactly what the label is supposed to encode. Grouping the human trajectories by ANC, the raw between-group effect size on the assertiveness atoms is near zero (ε2=0.006 ^2=0.006 for peak acceleration, 0.0120.012 for peak deceleration, ≤0.05≤ 0.05 for jerk, 0.0630.063 aggregated to H); the only atom ANC separates is average speed (ε2=0.088 ^2=0.088), and the proximity atoms it appears to move follow directly from speed. As a positive control, grouping the same trajectories by a speed band recovers large effects on exactly these atoms (ε2=0.38 ^2=0.38–0.400.40 for peak accel/decel), so ANC’s near-zero effect is a true null, not a dead test. The residual signal also vanishes once environment is accounted for: regressing each atom on the scene environment (scenario, turn, lead, road), ANC adds almost nothing (ΔR2(ANC∣env)=0.037 R^2(ANC )=0.037 on H, 0.0000.000 on P; Tab. 3), whereas the same regression leaves driver identity a strong organizer (ΔR2=0.232 R^2=0.232 on H, over six times the label’s). What ANC does track is speed: its composition follows a six-bin speed partition (Cramér’s V=0.34V=0.34), and conservative is largely a stationary-scene detector—median speed 0.450.45 m/s, 77.5%77.5\% of its scenes near-stationary (<2<2 m/s), and 87%87\% of fully stopped scenes labelled conservative. A plain speed partition with no style design separates our atoms as well as or better than ANC on nearly every atom. The label thus re-describes driving context (how fast, what scene), not an independent per-model style—the failure mode that motivates model-as-driver. ANC label driver identity raw effect size, peak accel 0.006 0.275 raw effect size, peak decel 0.012 0.287 raw effect size, assertiveness H 0.063 0.270 ΔR2(⋅∣env) R^2(· ) on H 0.037 0.232 ΔR2(⋅∣env) R^2(· ) on P 0.000 0.065 Table 3: Two complementary diagnostics of what organizes driving style, on the navtest atoms. The ANC column groups the human log-replay trajectories by StyleDrive’s label; the driver column groups the 16 policies’ trajectories by identity. Top: raw between-group effect size (ε2 ^2, rank-based, for ANC; η2η^2, one-way ANOVA, for driver). Bottom: incremental adjusted R2R^2 after regressing out the scene environment, penalizing the differing degrees of freedom (3 ANC classes vs. 16 drivers). After control, model identity remains a strong organizer, unlike the label; P is weak for both (Sec. 4.7). 4.5 Controllable, composable style Figure 3: Style controllability in BEV. Each row shows one scene decoded at increasing style strength α (columns, 0→10→ 1): an assertive persona (DiffusionDriveV2, top) and a cautious persona (GTRS, bottom). Trajectory color encodes α (00 is the base output, 11 is full persona style); the black curve is the fixed base anchor. A single coefficient sweeps the decoded trajectory continuously, and every α stays safe (gate-passing). Ego is outlined in red, other agents in blue. Sweeping α moves the assertiveness axis H and its kinematics from the base toward each teacher, so one coefficient spans a range rather than switching between presets (Tabs. 4 and 4; decoded trajectories in Fig. 3). For the DiffusionDriveV2 persona, raising α from 0 to 1 drives H monotonically to 3.263.26 while PDMS stays high and peaks near α=0.75α=0.75—strong style expression at no cost to safety; the dose–response is smooth and monotone across the whole persona set (Fig. 4). Multi-modality is common enough to matter on navtest: among scenes with sufficient driver coverage, about 7.4% are laterally bimodal and 13.9% longitudinally bimodal (brake or go). There a single-mode regressor returns a path between the modes, whereas the sampling decoder draws distinct maneuvers and commits to one, so style survives where averaging would erase it. Figure 4: Dose–response: as the single strength α rises, the dynamics atoms (peak deceleration and longitudinal jerk shown) move monotonically from the base toward each teacher’s operating point, so α is a continuous style coefficient rather than a discrete selector. The smoother personas (lower panel, zoomed) separate cleanly at small magnitudes. α PDMS↑ H 0.00 0.912 0.48 0.25 0.926 0.53 0.50 0.941 1.52 0.75 0.947 2.81 1.00 0.941 3.26 (a) DiffusionDriveV2 (aggressive): H climbs to 3.263.26. α PDMS↑ H 0.00 0.912 0.48 0.25 0.919 0.27 0.50 0.923 0.18 0.75 0.916 0.10 1.00 0.903 0.05 (b) GTRS (smooth): H falls as driving turns more conservative. Table 4: Style following as α increases (full sweep in Tab. 7). One coefficient α steers the base toward each teacher in opposite directions on the assertiveness axis H—the aggressive DiffusionDriveV2 persona (left) rises while the smooth GTRS persona (right) falls—with PDMS high throughout. Task-vector properties of the persona direction. A construction that is linear in content space need not stay linear or additive at the measured output, so we test these two task-vector properties directly on the style axis. Linearity we have already seen as the dose–response of Sec. 4.5, where H moves monotonically with α (Figs. 4 and 4); we quantify it as the fit R2R^2 of H against α over the full sweep. Composition asks whether a blend ∑iwiΔi _iw_i _i matches the weighted sum of the single-persona shifts (Tab. 5(a)), letting personas reach new operating points beyond any single teacher. Composition is not just a property but a capability: a convex blend yields a new operating point at no extra training cost and reaches places no single teacher does. Table 5(b) lists measured blends at α=1α=1: the strongest, DiffusionDriveV2/GTRS, attains PDMS=0.954PDMS=0.954, above every single persona (0.9410.941, Tab. 1), while blends without DiffusionDriveV2 occupy a milder band (H≈1.05H≈1.05), so blends sweep a continuous style range at uniformly high safety. Same-direction blends track the weighted-sum prediction to within ∼10% \!10\%, whereas aggressive++smooth blends are super-additive: the aggressive direction dominates, so measured H exceeds the linear prediction rather than averaging toward the milder teacher. property measurement result linearity in α R2R^2 of H vs. α 0.930.93 additive composition error vs. weighted sum (same-dir.) 6%6\% (a) Task-vector properties: linearity is the R2R^2 of H against α over the aggressive teachers (where H has room to move); composition is the relative error between a measured blend and the weighted sum of single-persona shifts, over the same-direction blends of (b). persona blend (weights) PDMS↑ H P ddv2 ++ gtrs (60/40) 0.954 +2.69+2.69 +2.45+2.45 ddv2 ++ gf ++ hy (40/40/20) 0.950 +2.42+2.42 +2.37+2.37 ddv2 ++ gf (60/40) 0.949 +2.96+2.96 +2.27+2.27 ddv2 ++ gf ++ tf ++ gtrs (35/35/15/15) 0.940 +2.13+2.13 +2.46+2.46 gf ++ gtrs (50/50) 0.930 +1.05+1.05 +2.35+2.35 tf ++ gf (50/50) 0.921 +1.06+1.06 +2.33+2.33 (b) Representative blends at α=1α=1, a convex mixture ∑iwiΔi _iw_i _i of single-persona directions (codes: ddv2 = DiffusionDriveV2, gf = GoalFlow, tf = TransFuser, gtrs = GTRS, hy = Hydra-MDP). No retraining; the best, ddv2 ++ gtrs, hits 0.9540.954 vs 0.9410.941 for the strongest single persona (Tab. 1). H/P are the SAS axes (human at 00); full set (all α) in Appendix A. Table 5: Persona directions behave as reusable, composable task vectors. 4.6 Ablation Study We ablate the two design choices that most directly shape controllable style: how trajectories are decoded and how much sampling inference uses. Unless noted, each variant is trained and evaluated on the same navtest split at matched capacity; we report PDMS with its safety sub-scores (no-collision NC, drivable-area compliance DAC) next to the reachable style span ΔH H—the range of the assertiveness axis a single model covers across its personas and α∈[0,1]α∈[0,1] (per-persona values in Appendix B, full sweep in Appendix A). Decoder family. We replace the flow-matching decoder with the backbone’s discrete action codebook, keeping conditioning and training data fixed. The codebook preserves safety but quantizes trajectories onto a fixed set of anchors, which caps the reachable style range and leaves only coarse, quantized style control (Tab. 6(a)); the flow-matching decoder recovers a smooth, wider style axis at equal safety. Samples and steps. Sweeping the two inference knobs, both PDMS and the style span ΔH H saturate by N=16N=16, K=20K=20 (Tab. 6(b)); controllability thus lives in the persona direction, not the sampler, so we default to N=16N=16, K=100K=100. decoder PDMS↑ NC↑ DAC↑ EP↑ TTC↑ Comf↑ ΔH H discrete codebook 0.891 0.984 0.956 0.818 0.980 0.999 0.8 flow matching (ours) 0.912 0.988 0.966 0.867 0.963 1.000 3.2 (a) Decoder family (conditioning fixed). Both keep hard safety (NC/DAC) and PDMS close, but the discrete codebook caps the reachable style span (ΔH=0.8 H=0.8 vs 3.23.2) with only coarse, quantized control; flow matching yields a smooth, wider style axis at equal safety. setting PDMS↑ ΔH H candidates N (steps K=100K=100) N=4N=4 0.883 2.5 N=8N=8 0.899 3.1 N=16N=16 (ours) 0.912 3.2 N=64N=64 0.913 3.2 steps K (candidates N=16N=16) K=5K=5 0.864 2.3 K=20K=20 0.910 3.2 K=100K=100 (ours) 0.912 3.2 (b) Inference budget. PDMS and ΔH H saturate by N=16N=16, K=20K=20 and narrow only when the sampler is starved (N=4N=4, K=5K=5), so the default N=16N=16, K=100K=100 leaves little on the table. Table 6: Ablations: decoder family (a) and inference budget (b). Sub-scores as in Tab. 1. 4.7 Limitations SAS is validated along assertiveness (H); proximity (P) is reported only as a secondary, preliminary axis. On navtest the proximity atoms vary far less across policies than the dynamics atoms (cross-policy std 0.140.14–0.290.29 vs 0.570.57–1.521.52), so social style is near-degenerate on this split, dominated by open-road, low-interaction scenes. Within this narrow range every native teacher loosens gaps (P<0P<0) while every persona tightens them (P>0P>0, Fig. 2), so the persona direction does not reproduce the teacher’s sign on P; with the signal this close to degenerate, whether P is independently controllable remains open. Resolving it requires interaction-dense, reactive closed-loop evaluation (e.g. nuPlan reactive, CARLA), where headway and gap have room to vary; we leave this to future work. 5 Conclusion We presented YOUDrive, which treats driving style as a measurable, composable control axis on a shared backbone—each driver a reusable direction—rather than an average in the data or a label pasted onto it. The framework rests on three parts: a flow-matching decoder that commits to a feasible trajectory rather than averaging, and a low-rank task vector per driver that one coefficient scales and composes without retraining(expression); the Style Alignment Score (SAS), which reports style on its own terms after gating hard safety (measurement); and, on NAVSIM, evidence that this axis is a systematic per-model signal, unlike a coarse label, with a single coefficient tracing a wide style range while PDMS stays above 0.90 (validation). Its main limit: SAS is validated along assertiveness alone; the proximity axis is near-degenerate on navtest, so whether social style is independently controllable remains open, calling for interaction-dense, reactive closed-loop evaluation. Validating style against human preference is the next step. References [1] S. Bai et al. (2025) Qwen2.5-VL technical report. arXiv preprint arXiv:2502.13923. Cited by: §3.2, §4.1. [2] R. Chen, A. Arditi, H. Sleight, O. Evans, and J. Lindsey (2025) Persona vectors: monitoring and controlling character traits in language models. arXiv preprint arXiv:2507.21509. Cited by: §2. [3] K. Chitta, A. Prakash, B. Jaeger, Z. Yu, K. Renz, and A. Geiger (2023) TransFuser: imitation with transformer-based sensor fusion for autonomous driving. IEEE TPAMI. Cited by: §1, §2, §4.1, Table 1, Table 1. [4] D. Dauner, M. Hallgarten, A. Geiger, and K. Chitta (2023) Parting with misconceptions about learning-based vehicle motion planning. In Conf. Robot Learn. (CoRL), Cited by: §3.4. [5] D. Dauner et al. (2024) NAVSIM: data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, Cited by: §3.4, §4.1. [6] R. Gao et al. (2025) StyleDrive: towards driving-style aware benchmarking of end-to-end autonomous driving. arXiv preprint arXiv:2506.23982. Cited by: §1, §2, §4.4, Table 2. [7] Y. Gao, D. Hua, M. Piccinini, F. R. Schäfer, K. Moller, L. Li, and J. Betz (2026) StyleVLA: driving style-aware vision language action model for autonomous driving. arXiv preprint arXiv:2603.09482. Cited by: §2. [8] E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, and W. Chen (2022) LoRA: low-rank adaptation of large language models. In ICLR, Cited by: §3.3, §4.1. [9] Y. Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang, et al. (2023) Planning-oriented autonomous driving. In CVPR, Cited by: §1, §2. [10] J. Hwang, R. Xu, H. Lin, W. Hung, J. Ji, K. Choi, D. Huang, T. He, P. Covington, B. Sapp, et al. (2024) EMMA: end-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262. Cited by: §2. [11] G. Ilharco, M. T. Ribeiro, M. Wortsman, L. Schmidt, H. Hajishirzi, and A. Farhadi (2023) Editing models with task arithmetic. In ICLR, Cited by: §1, §2, §3.3. [12] B. Jiang, S. Chen, Q. Xu, B. Liao, J. Chen, H. Zhou, Q. Zhang, W. Liu, C. Huang, and X. Wang (2023) VAD: vectorized scene representation for efficient autonomous driving. In ICCV, Cited by: §1, §2. [13] M. Kuderer, S. Gulati, and W. Burgard (2015) Learning driving styles for autonomous vehicles from demonstration. In IEEE Int. Conf. Robot. Autom. (ICRA), Cited by: §2. [14] D. Li, C. Li, Y. Wang, J. Ren, X. Wen, P. Li, L. Xu, K. Zhan, P. Jia, X. Lang, N. Xu, and H. Zhao (2025) Learning personalized driving styles via reinforcement learning from human feedback. arXiv preprint arXiv:2503.10434. Cited by: §2. [15] K. Li, O. Patel, F. Viégas, H. Pfister, and M. Wattenberg (2023) Inference-time intervention: eliciting truthful answers from a language model. In NeurIPS, Cited by: §2. [16] Y. Li et al. (2025) End-to-end driving with online trajectory evaluation via BEV world model. In ICCV, Cited by: §4.1, Table 1, Table 1. [17] Z. Li et al. (2024) Hydra-MDP: end-to-end multimodal planning with multi-target hydra-distillation. In CVPRW, Cited by: §4.1, Table 1, Table 1. [18] Z. Li et al. (2025) Generalized trajectory scoring for end-to-end multimodal planning. In ICCV, Cited by: §4.1. [19] B. Liao et al. (2025) DiffusionDrive: truncated diffusion model for end-to-end autonomous driving. In CVPR, Cited by: §2, §3.4, §4.1, Table 1, Table 1. [20] Y. Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le (2023) Flow matching for generative modeling. In ICLR, Cited by: §3.4. [21] X. Liu, C. Gong, and Q. Liu (2023) Flow straight and fast: learning to generate and transfer data with rectified flow. In ICLR, Cited by: §3.4. [22] W. Peebles and S. Xie (2023) Scalable diffusion models with transformers. In ICCV, Cited by: §3.4. [23] W. Schwarting, A. Pierson, J. Alonso-Mora, S. Karaman, and D. Rus (2019) Social behavior for autonomous vehicles. Proc. Natl. Acad. Sci. (PNAS) 116 (50), p. 24972–24978. Cited by: §2. [24] M. Wortsman, G. Ilharco, S. Y. Gadre, R. Roelofs, R. Gontijo-Lopes, A. S. Morcos, H. Namkoong, A. Farhadi, Y. Carmon, S. Kornblith, and L. Schmidt (2022) Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time. In Int. Conf. Mach. Learn. (ICML), Cited by: §2. [25] Z. Xing et al. (2025) GoalFlow: goal-driven flow matching for multimodal trajectory generation in end-to-end autonomous driving. In CVPR, Cited by: §2, §3.4, §4.1, Table 1, Table 1. [26] P. Yadav, D. Tam, L. Choshen, C. Raffel, and M. Bansal (2023) TIES-Merging: resolving interference when merging models. In NeurIPS, Cited by: §2. [27] Z. Zhou et al. (2025) AutoVLA: a vision-language-action model for end-to-end autonomous driving. arXiv preprint arXiv:2506.13757. Cited by: §2, §4.1. Appendix A Full Style Control and Composition Results Tabs. 7 and A report the complete results for individual persona sweeps and persona blends, respectively. All results are scenario-weighted means over eight navtest shards. Alongside PDMS and the SAS axes H and P, we include the individual kinematic and proximity measurements, including dimensions with weak separation on this split. The scope column distinguishes human and base references (anchor), individual persona sweeps (sweep), and blends (combo). In sweep, α varies the strength of a single persona and α=Tα=T denotes the native teacher. In combo, the configuration name gives the mixing weights of the teacher task vectors, while α controls their overall strength. The full individual sweeps in Tab. 7 extend the selected examples in Tab. 4. For all six personas, H changes monotonically with α, though in different directions. DiffusionDriveV2 raises H from 0.480.48 to 3.263.26, while GTRS lowers it from 0.480.48 to 0.050.05. The corresponding peak acceleration and longitudinal jerk rise from 0.530.53 and 0.890.89 to 2.662.66 and 4.124.12 for DiffusionDriveV2, but fall to 0.240.24 and 0.270.27 for GTRS. Average speed varies less than these dynamics measurements, consistent with its weak driver-specific signal in Sec. 4.4. PDMS stays above 0.900.90 across the tested persona strengths, but does not change monotonically with α. The proximity axis P also does not follow the native teachers: it is positive for the distilled personas and negative for the native teachers. We keep this as a limitation, as discussed in Sec. 4.7. The blend results in Appendix A give a fuller view of the composition reported in Tab. 5(b). Changing the mixing weights can change both H and PDMS. For DiffusionDriveV2/TransFuser at α=1α=1, increasing the DiffusionDriveV2 weight from 30%30\% to 70%70\% changes H from 0.740.74 to 2.922.92 and PDMS from 0.9010.901 to 0.9480.948. A fixed blend can also be scaled: DiffusionDriveV2/GTRS (60/40) moves from H=0.48H=0.48 at α=0α=0 to 0.920.92 at α=0.5α=0.5 and 2.692.69 at α=1α=1, with PDMS reaching 0.9540.954. These output changes are not necessarily additive, consistent with the nonlinear blended representation in Sec. 3.3. All tested blends retain PDMS above 0.900.90. scope teacher α PDMS↑ H P v¯ v a+a^+ a−a^- jℓj_ jtj_t thw ttc lead dmind_ dvrud_vru anchor human – 0.9480 +0.00 +0.00 4.6181 0.2446 0.3059 0.2898 0.1026 3.7850 4.3915 12.9060 4.5340 12.0630 anchor base(ours) – 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 sweep gtrs 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 sweep gtrs 0.25 0.9187 +0.27 +1.75 4.8300 0.3713 0.3331 0.4590 0.5258 3.2140 3.5045 11.6580 4.4240 11.9630 sweep gtrs 0.5 0.9230 +0.18 +2.03 4.9096 0.3341 0.2454 0.3701 0.4722 3.1210 3.3810 11.5660 4.3940 11.9520 sweep gtrs 0.75 0.9158 +0.10 +2.38 4.9346 0.2753 0.1638 0.3118 0.4139 3.0620 3.2390 11.1695 4.3610 11.8570 sweep gtrs 1 0.9030 +0.05 +3.01 4.9977 0.2449 0.1326 0.2728 0.3721 2.8625 3.0440 10.8545 4.2760 11.7870 sweep gtrs T 0.899 -0.01 -2.02 4.4350 0.2160 0.1840 0.2520 0.2160 4.0270 4.6615 14.1960 4.9570 12.7700 sweep hydra_mdp 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 sweep hydra_mdp 0.25 0.9191 +0.28 +1.69 4.8250 0.3668 0.3568 0.4559 0.5304 3.2820 3.5350 11.7510 4.3980 11.9590 sweep hydra_mdp 0.5 0.9232 +0.21 +1.78 4.8747 0.3171 0.2840 0.3916 0.5004 3.2250 3.4960 11.6775 4.4050 11.9580 sweep hydra_mdp 0.75 0.9198 +0.16 +2.12 4.8833 0.2932 0.2337 0.3416 0.4722 3.1280 3.3625 11.3740 4.3740 11.9080 sweep hydra_mdp 1 0.9067 +0.13 +2.40 4.9509 0.2756 0.2133 0.3013 0.4412 3.0530 3.2420 11.2290 4.3500 11.8650 sweep hydra_mdp T 0.8650 +0.05 -1.99 4.5170 0.2250 0.1700 0.2510 0.3690 3.9170 4.6515 14.1240 4.9690 12.8480 sweep wote 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 sweep wote 0.25 0.9164 +0.32 +1.64 4.8201 0.4228 0.3759 0.6068 0.5135 3.2790 3.5840 11.6820 4.4490 11.8890 sweep wote 0.5 0.9192 +0.29 +1.70 4.8399 0.3552 0.3358 0.5485 0.5483 3.2160 3.4620 11.6325 4.4530 11.9950 sweep wote 0.75 0.9152 +0.23 +1.88 4.8764 0.3007 0.3231 0.5390 0.4679 3.2080 3.4450 11.4520 4.3930 11.9430 sweep wote 1 0.9051 +0.17 +2.09 4.8702 0.2523 0.2993 0.5254 0.4028 3.1610 3.3650 11.4460 4.3900 11.8680 sweep wote T 0.8830 +0.16 -2.41 4.3269 0.2693 0.3496 0.4682 0.3516 4.1780 4.9650 14.3820 4.8985 12.8190 sweep ddv2 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 sweep ddv2 0.25 0.9262 +0.53 +1.55 4.7682 0.5200 0.6112 0.8234 0.6351 3.2960 3.6320 11.8680 4.4540 11.9040 sweep ddv2 0.5 0.9406 +1.52 +1.48 4.8019 1.4282 1.3115 2.2109 1.0144 3.2270 3.6350 12.1850 4.4710 12.0180 sweep ddv2 0.75 0.9467 +2.81 +2.30 4.8995 2.5549 1.9862 3.7266 1.8909 3.0130 3.3460 11.8520 4.4440 11.7650 sweep ddv2 1 0.9407 +3.26 +2.33 4.8513 2.6606 2.1951 4.1235 2.5209 3.0275 3.2980 12.0275 4.4460 11.7610 sweep ddv2 T 0.8810 +2.72 -0.42 4.8577 2.0381 1.9625 4.1628 1.9892 3.4170 3.9205 13.5680 4.9230 12.6070 sweep goalflow 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 sweep goalflow 0.25 0.9235 +0.50 +1.42 4.7899 0.4476 0.6154 0.7507 0.6462 3.3200 3.6700 11.8880 4.4930 11.9030 sweep goalflow 0.5 0.9191 +0.86 +1.67 4.7600 0.7489 0.8253 1.3588 0.8111 3.3030 3.6760 11.8095 4.4290 11.7950 sweep goalflow 0.75 0.9256 +1.50 +1.73 4.8250 1.3531 1.2791 2.2369 1.0610 3.2000 3.5730 11.7940 4.4070 11.9710 sweep goalflow 1 0.9354 +1.99 +2.01 4.8689 1.7530 1.5476 3.1498 1.2741 3.1210 3.4700 11.6940 4.4230 11.8470 sweep goalflow T 0.9030 +1.93 -1.26 4.6061 1.6850 1.7739 3.7930 0.7634 3.7250 4.3440 13.8280 4.8980 12.7450 sweep transfuser 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 sweep transfuser 0.25 0.9153 +0.29 +1.58 4.8027 0.3722 0.3703 0.4809 0.5427 3.3120 3.5900 11.8520 4.4420 11.9100 sweep transfuser 0.5 0.9158 +0.20 +1.64 4.8546 0.3340 0.2893 0.4664 0.4407 3.2710 3.5230 11.7300 4.4420 11.9550 sweep transfuser 0.75 0.9112 +0.14 +1.97 4.8675 0.3081 0.2427 0.4518 0.3705 3.2220 3.4175 11.6560 4.3890 11.8780 sweep transfuser 1 0.9004 +0.10 +2.14 4.8629 0.2834 0.2142 0.4582 0.3248 3.1855 3.3540 11.3445 4.3690 11.8400 sweep transfuser T 0.8400 +0.06 -1.58 4.5679 0.2320 0.2474 0.4006 0.2626 3.8750 4.5360 13.9250 4.8690 12.7710 Table 7: Full v1 navtest results for per-teacher α sweeps, aggregated across eight shards with weights proportional to the number of scenarios. The human serves as the SAS reference (H=P=0H=P=0), while the base is the shared policy at α=0α=0; α=Tα=T denotes the native teacher. H and P are the SAS axes defined in Sec. 4.2. Sub-metrics: v¯ v mean speed; a+/a−a^+/a^- peak acceleration/deceleration; jℓ/jtj_ /j_t longitudinal/lateral jerk; thw/ttc minimum headway/time-to-collision; lead/dmin/dvrud_ /d_vru minimum lead/object/VRU distances. scope teacher α PDMS↑ H P v¯ v a+a^+ a−a^- jℓj_ jtj_t thw ttc lead dmind_ dvrud_vru combo ddv2+goalflow 50/50 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+goalflow 50/50 0.5 0.9520 +1.06 +1.73 4.8950 1.0565 0.9788 1.7424 0.7410 3.3390 3.7605 12.0920 4.4030 11.6645 combo ddv2+goalflow 50/50 1 0.9500 +2.64 +2.45 4.8309 2.4876 1.8331 3.6637 1.7195 3.0040 3.4865 12.5640 4.3570 11.6665 combo ddv2+transfuser 70/30 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+transfuser 70/30 0.5 0.9437 +1.06 +1.76 4.7260 1.0065 0.9937 1.6054 0.8264 3.3110 3.9040 12.8625 4.2290 11.8385 combo ddv2+transfuser 70/30 1 0.9480 +2.92 +2.76 4.8494 2.5248 2.0674 3.7133 2.1054 3.1130 3.3615 12.1140 4.1460 11.7070 combo ddv2+transfuser 50/50 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+transfuser 50/50 0.5 0.9456 +0.71 +1.61 4.8358 0.7557 0.6825 1.1434 0.6745 3.2300 3.8975 12.5060 4.4110 11.7720 combo ddv2+transfuser 50/50 1 0.9475 +2.11 +1.86 4.9691 2.0326 1.5691 2.8817 1.3920 3.0725 3.6615 11.9565 4.4330 11.8725 combo ddv2+transfuser 30/70 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+transfuser 30/70 0.5 0.9089 +0.46 +2.04 4.7204 0.5644 0.4788 0.7054 0.5833 3.1725 3.4600 11.2675 4.3330 11.9235 combo ddv2+transfuser 30/70 1 0.9008 +0.74 +2.74 4.9728 0.7749 0.5997 1.1111 0.7917 3.0730 3.4105 11.2915 4.2370 11.5775 combo transfuser+goalflow 50/50 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo transfuser+goalflow 50/50 0.5 0.9248 +0.55 +1.92 4.6543 0.5435 0.6262 0.8534 0.6315 3.3145 3.8360 11.8075 4.2920 11.6380 combo transfuser+goalflow 50/50 1 0.9206 +1.06 +2.33 4.7809 1.0321 0.9105 1.5680 0.8838 3.2060 3.6705 11.0475 4.2370 11.5960 combo ddv2+gtrs 60/40 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+gtrs 60/40 0.5 0.9521 +0.92 +1.79 4.8853 0.8642 0.8730 1.4049 0.7874 3.1620 3.5695 11.8470 4.4390 11.9000 combo ddv2+gtrs 60/40 1 0.9543 +2.69 +2.45 5.0393 2.4151 1.9628 3.5283 1.8284 2.9325 3.2810 11.6675 4.3950 11.8510 combo ddv2+hydra 60/40 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+hydra 60/40 0.5 0.9499 +0.92 +1.93 4.8883 0.8822 0.8653 1.4252 0.7692 3.1735 3.5015 11.7565 4.4490 11.8000 combo ddv2+hydra 60/40 1 0.9519 +2.68 +2.60 5.0141 2.3923 1.9392 3.4845 1.8634 2.8890 3.2330 11.8015 4.3970 11.7950 combo ddv2+goalflow 60/40 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+goalflow 60/40 0.5 0.9452 +1.24 +1.65 4.8364 1.1836 1.1556 1.8826 0.8407 3.1810 3.5640 11.9270 4.4530 11.9990 combo ddv2+goalflow 60/40 1 0.9494 +2.96 +2.27 4.9343 2.6060 2.0955 3.9844 2.0162 3.0575 3.3890 11.8915 4.4170 11.7530 combo ddv2+goalflow+hydra 40/40/20 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+goalflow+hydra 40/40/20 0.5 0.9387 +0.91 +1.98 4.9062 0.8805 0.8821 1.4372 0.7460 3.1440 3.5290 11.6810 4.4320 11.7880 combo ddv2+goalflow+hydra 40/40/20 1 0.9496 +2.42 +2.37 5.0126 2.3134 1.7531 3.3477 1.5544 2.9870 3.3260 11.6890 4.4140 11.7960 combo ddv2+goalflow+wote 45/45/10 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+goalflow+wote 45/45/10 0.5 0.9319 +1.03 +1.96 4.8539 0.9849 0.9920 1.6205 0.7741 3.1700 3.5400 11.6570 4.4300 11.7750 combo ddv2+goalflow+wote 45/45/10 1 0.9490 +2.61 +2.41 4.9875 2.3998 1.8981 3.5852 1.6949 2.9790 3.2710 11.9215 4.4540 11.7420 combo ddv2+goalflow+gtrs 40/40/20 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+goalflow+gtrs 40/40/20 0.5 0.9465 +0.91 +1.98 4.9022 0.8959 0.8772 1.4354 0.7268 3.1600 3.5270 11.7180 4.4330 11.7760 combo ddv2+goalflow+gtrs 40/40/20 1 0.9459 +2.46 +2.42 5.0072 2.3323 1.7972 3.4049 1.5534 2.9580 3.2970 11.5530 4.4290 11.7800 combo ddv2+goalflow+transfuser+gtrs 35/35/15/15 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo ddv2+goalflow+transfuser+gtrs 35/35/15/15 0.5 0.9377 +0.81 +1.97 4.8538 0.7870 0.8131 1.2800 0.7112 3.1810 3.4970 11.6360 4.4230 11.7960 combo ddv2+goalflow+transfuser+gtrs 35/35/15/15 1 0.9402 +2.13 +2.46 5.0196 2.0699 1.5952 3.0159 1.3398 3.0240 3.3060 11.4245 4.4000 11.7130 combo goalflow+gtrs 50/50 0 0.9120 +0.48 +0.08 4.6051 0.5282 0.6151 0.8865 0.4750 3.7640 4.3710 12.8275 4.5030 12.0780 combo goalflow+gtrs 50/50 0.5 0.9246 +0.52 +1.94 4.8540 0.5334 0.5611 0.8342 0.6375 3.2320 3.5480 11.5610 4.4120 11.7490 combo goalflow+gtrs 50/50 1 0.9299 +1.05 +2.35 4.9193 0.9743 0.9576 1.6231 0.8446 3.1040 3.4010 11.2400 4.3550 11.7250 Table 8: Full v1 navtest results for persona blends. Each configuration specifies the teacher task vectors and their mixing weights, with α controlling the overall strength. We report PDMS, SAS axes H and P, and the individual kinematic and proximity measurements, following the same definitions as Tab. 7. All configurations recover the shared base at α=0α=0. persona discrete codebook flow (ours) base (α=0α=0) 0.5 0.5 ddv2 0.8 3.3 goalflow 0.7 2.0 wote 0.1 0.2 transfuser 0.1 0.1 hydra_mdp 0.0 0.1 gtrs 0.0 0.1 ΔH H 0.8 3.2 Table 9: Per-persona assertiveness H for the decoder ablation in Tab. 6(a). Both decoders use the same style conditioning. The base is evaluated at α=0α=0 and each persona at α=1α=1. The bottom row reports the measured assertiveness span ΔH H. Appendix B Per-Persona Ablation Breakdown Tab. 9 gives the per-persona assertiveness scores behind the decoder comparison in Tab. 6(a). The style span ΔH H is the difference between the largest and smallest measured H values across the evaluated persona settings. We show the base (α=0α=0) and each persona at full strength (α=1α=1), together with the resulting span. Flow matching covers a wider assertiveness range than the discrete codebook (ΔH=3.2 H=3.2 versus 0.80.8). Most of this difference comes from the more assertive personas: DiffusionDriveV2 reaches H=3.3H=3.3 with flow matching but 0.80.8 with the codebook, while GoalFlow reaches 2.02.0 and 0.70.7, respectively. The smoother personas remain close to the human reference under both decoders. Thus, the wider range in Tab. 6(a) mainly reflects the decoder’s ability to express the stronger teacher-induced dynamics, rather than a similar increase for every persona.