Paper deep dive
Evaluation Is All You Need for Multi-Modal Autonomous Driving
Zeyu He, Shiqi Liu, Ke Chen, Yun Yan, Jinzi Wu, Dianqiao Lei, Sirui Wang, ShuRui Peng, Tao Chen, Zhuo Huang, Yu Wu, Yadong Shao, Zhichao Li, Ke Sun, Yang Guan, Keqiang Li, Shengbo Eben Li
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/28/2026, 3:42:30 AM
Summary
The paper introduces iDriveVLA, a multi-modal autonomous driving framework that addresses the generation-evaluation asymmetry where strong trajectory generators fail to be reliably evaluated. iDriveVLA employs a Safety-aware Scorer and a VLM-guided Modulator to filter unsafe candidates and adaptively weight planning criteria based on scene semantics. Trained via a progressive oracle-aligned strategy, it achieves state-of-the-art performance on the NAVSIM v1 leaderboard.
Entities (8)
Relation Signals (7)
iDriveVLA → achievesscoreon → NAVSIM-v1
confidence 97% · On the public NAVSIM v1 leaderboard, iDriveVLA achieves a new state-of-the-art performance of 94.95 PDMS
iDriveVLA → usescomponent → Safety-aware Scorer
confidence 95% · iDriveVLA introduces a unified trajectory evaluator comprising a Safety-aware Scorer
iDriveVLA → usescomponent → VLM-guided Modulator
confidence 95% · together with a VLM-guided Modulator for scene-adaptive criterion weighting.
Safety-aware Scorer → performsfunction → risk_estimation
confidence 90% · Safety-aware Scorer for quality and risk estimation
VLM-guided Modulator → performsfunction → scene_adaptive_weighting
confidence 90% · VLM-guided Modulator for scene-adaptive criterion weighting.
InternVL3-2B → isusedby → VLM-guided Modulator
confidence 88% · we use a frozen InternVL3-2B... For semantic modulation
DrivoR → isbasefor → iDriveVLA
confidence 85% · For trajectory generation, we adopt the multi-modal planner of DrivoR
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Multi-modal planning is promising for autonomous driving by representing multiple plausible behaviors in ambiguous and long-tail scenarios. Existing methods mainly focus on improving trajectory multi-modality, enhancing trajectory representations, or reshaping the candidate distribution. Nevertheless, we identify a pronounced generation-evaluation asymmetry in multi-modal planning: despite strong oracle performance, existing planners often fail to reliably select the best available candidate, leaving substantial planning potential unrealized. To address this challenge, we propose iDriveVLA, a multi-modal planning framework that improves the candidate trajectory space while enabling more reliable and context-aware trajectory evaluation. Specifically, iDriveVLA introduces a unified trajectory evaluator comprising a Safety-aware Scorer for quality and risk estimation, together with a VLM-guided Modulator for scene-adaptive criterion weighting. We further develop an oracle-aligned progressive training strategy consisting of candidate imitation pretraining, candidate space refinement, and semantic ranking alignment. On the public NAVSIM v1 leaderboard, iDriveVLA achieves a new state-of-the-art performance of 94.95 PDMS, surpassing the human-expert reference.
Tags
Links
- Source: https://arxiv.org/abs/2609.30818v1
- Canonical: https://arxiv.org/abs/2609.30818v1
Trouble viewing inline? Open PDF directly →
Full Text
58,282 characters extracted from source content.
Expand or collapse full text
Evaluation Is All You Need for Multi-Modal Autonomous Driving 1School of Vehicle and Mobility & College of AI, Tsinghua University 2Dongfeng Motor Corporation Research & Development Institute †Equal contribution 🖂Corresponding author Abstract Multi-modal planning is promising for autonomous driving by representing multiple plausible behaviors in ambiguous and long-tail scenarios. Existing methods mainly focus on improving trajectory multi-modality, enhancing trajectory representations, or reshaping the candidate distribution. Nevertheless, we identify a pronounced generation–evaluation asymmetry in multi-modal planning: despite strong oracle performance, existing planners often fail to reliably select the best available candidate, leaving substantial planning potential unrealized. To address this challenge, we propose iDriveVLA, a multi-modal planning framework that improves the candidate trajectory space while enabling more reliable and context-aware trajectory evaluation. Specifically, iDriveVLA introduces a unified trajectory evaluator comprising a Safety-aware Scorer for quality and risk estimation, together with a VLM-guided Modulator for scene-adaptive criterion weighting. We further develop an oracle-aligned progressive training strategy consisting of candidate imitation pretraining, candidate space refinement, and semantic ranking alignment. On the public NAVSIM v1 leaderboard, iDriveVLA achieves a new state-of-the-art performance of 94.95 PDMS, surpassing the human-expert reference. 1 Introduction End-to-end autonomous driving (Hu et al., 2023; Guan et al., 2023) has rapidly advanced by jointly optimizing perception, prediction, and planning. More recently, vision-language-action (VLA) models (Jiang et al., 2025; Zhou et al., 2026a) further introduce rich semantic knowledge and reasoning capabilities into driving systems. However, existing planners still struggle with challenging and long-tail scenarios, such as partially occupied lanes, unusual obstacles, temporary road structures, and ambiguous traffic interactions (Yao et al., 2026). A key difficulty in these scenarios is that the appropriate driving behavior is often inherently ambiguous: multiple motion strategies may be feasible, yet they can differ substantially in safety, efficiency, comfort, and interaction with surrounding agents (Liao et al., 2025). As illustrated in Figure 1(a), this has motivated multi-modal planning (Guo et al., 2025; Li et al., 2025b; Guan et al., 2026), where multiple plausible trajectories are generated and subsequently evaluated for final execution. Recent studies have substantially strengthened candidate generation through diffusion-based trajectory modeling (Liao et al., 2025), more expressive trajectory representations (Xing et al., 2026), and training-distribution design for improved candidate coverage (Ang et al., 2026). Despite increasingly strong candidate generation, we find that trajectory evaluation has become a major bottleneck in multi-modal planning. As shown in Figure 1(b), increasing trajectory diversity improves oracle performance by providing more potential solutions, while also introducing more competing behaviors that the evaluator must distinguish. This reveals a generation–evaluation asymmetry: the ability to generate diverse high-quality trajectories advances faster than the ability to reliably evaluate and select among them. This asymmetry is further quantified in Figure 1(c), where the oracle candidate achieves 99.21 PDMS, whereas the learned evaluator reaches only 93.67 PDMS on NAVSIM v1 (Dauner et al., 2024). The substantial gap indicates that considerable planning potential is already contained in the candidate set but remains unrealized. A key limitation is that existing evaluation mechanisms rely mainly on predefined planning signals and lack sufficient safety awareness and scene-adaptive semantic reasoning when comparing behaviorally distinct trajectories. To address this challenge, we propose iDriveVLA, a multi-modal planning framework that improves both candidate quality and trajectory selection. iDriveVLA introduces a trajectory evaluator that combines explicit safety awareness with VLM-guided semantic adaptation, enabling more reliable selection among behaviorally distinct candidates. We further develop an oracle-aligned progressive training strategy that progressively improves candidate generation and selection reliability through three stages: candidate imitation pretraining, candidate space refinement, and semantic ranking alignment. Our main contributions are summarized as follows: • We identify and characterize a generation–evaluation asymmetry in multi-modal planning: strong candidate generators already provide high oracle potential, yet imperfect trajectory evaluation prevents this potential from being fully realized. • We propose iDriveVLA, featuring a trajectory evaluator that integrates a Safety-aware Scorer for low-quality candidate filtering, along with a VLM-guided Modulator for scene-adaptive semantic evaluation. We further develop a progressive training strategy to improve both candidate performance and ranking alignment. • On the public NAVSIM v1 benchmark (Dauner et al., 2024), iDriveVLA achieves 94.95 PDMS and ranks 1st, surpassing the human-expert reference. Figure 1: Core idea. (a) Multiple plausible trajectories may coexist in challenging scenes, making reliable trajectory selection critical. (b) Increasing trajectory diversity improves oracle performance but also introduces more competing behaviors for the evaluator, revealing a generation–evaluation asymmetry in multi-modal planning. (c) This asymmetry leaves substantial oracle potential untapped in the strong multi-modal planner DrivoR (Kirby et al., 2026). In contrast, iDriveVLA improves trajectory evaluation, substantially narrows the oracle gap, and surpasses the human-expert reference on NAVSIM v1 (Dauner et al., 2024). 2 Related Work Vision-Language-Action Planning. Recent studies (Jiang et al., 2025; Zhou et al., 2026a) introduce vision-language models (VLMs) into autonomous driving to leverage their rich semantic priors and scene reasoning capabilities. A common line of work first equips VLMs with driving-specific knowledge through visual question answering or instruction tuning (Chi et al., 2025), and then extends them into VLA models that directly predict driving trajectories via action tokens (Zhou et al., 2025), textual trajectory representations (Hwang et al., 2025), or hidden-state-based action decoders (Li et al., 2026c). Despite their stronger semantic reasoning capabilities, existing VLM-based planners often remain behind strong end-to-end driving models in trajectory-level planning performance (Dauner et al., 2024; Jia et al., 2024). Multi-modal Trajectory Generation. Recent end-to-end autonomous driving methods increasingly adopt multi-modal trajectory generation to capture diverse and uncertain driving behaviors. Vocabulary-based approaches, such as VADv2 (Jiang et al., 2026), discretize the continuous planning space into representative trajectory candidates, providing broad coverage of driving modes. Generative methods, such as GenAD (Zheng et al., 2024) and DiffusionDrive (Liao et al., 2025), further model continuous trajectory distributions to generate diverse candidates with greater flexibility. More recently, proposal-based methods such as iPad (Guo et al., 2025) and DrivoR (Kirby et al., 2026) employ learnable queries to iteratively generate and refine trajectory hypotheses, enlarging the optimization space beyond fixed vocabularies. These advances have substantially improved trajectory diversity and quality; however, the resulting gains are increasingly constrained by the accuracy of downstream trajectory evaluation, limiting their full contribution to overall driving performance. Scorer-Based Autonomous Driving. Multi-modal planning has become a common paradigm in end-to-end autonomous driving, where diverse trajectory hypotheses are generated and subsequently ranked by a trajectory scorer for final execution (Guo et al., 2025). Recent works improve this paradigm by learning more robust and generalizable scoring functions across different candidate distributions (Li et al., 2025b), predicting interpretable planning metrics for candidate ranking (Kirby et al., 2026), or learning structured scoring rules directly from driving demonstrations (Xiong et al., 2025). However, their scoring criteria are primarily derived from geometric or predefined planning signals, with limited capability to adapt candidate preferences according to high-level scene semantics. 3 Method 3.1 Diagnosing the Trajectory Evaluation Bottleneck Given multi-view visual observations ℐI and the current ego state e, a multi-candidate trajectory generator produces K future trajectory hypotheses: =ii=1K=πgen(ℐ,),i∈ℝT×3,T=\ τ_i\_i=1^K= _gen(I, e), τ_i ^T× 3, (1) where T denotes the planning horizon and each trajectory consists of future positions and headings. Figure 2: PDMS distributions of the selected and oracle candidates from the base planner. The oracle distribution is strongly concentrated near high PDMS, whereas the selected candidates exhibit a substantially heavier low-score tail. Generating multiple candidates increases behavioral diversity and improves the chance that a high-quality trajectory is included in T. However, the final planning performance is determined not only by what the generator can produce, but also by whether the trajectory evaluator can identify a good candidate from the resulting set. To diagnose this issue, we analyze the strong multi-modal driving model DrivoR (Kirby et al., 2026) by comparing the downstream PDMS of its selected trajectory with that of the oracle candidate, defined as the highest-PDMS trajectory within the same candidate set. As shown in Figure 2, oracle candidates are strongly concentrated near high PDMS, whereas selected candidates exhibit a broader distribution with a heavier low-score tail. This reveals a clear generation–evaluation asymmetry: high-quality trajectories are often already available, but imperfect evaluation prevents their potential from being fully realized. As candidate diversity increases, this issue may become more pronounced because the evaluator must distinguish among more competing behaviors. A simple analysis of this candidate-competition effect is provided in Appendix A. These observations motivate us to shift attention from candidate generation alone toward more reliable trajectory evaluation. Appendix D.2 further show that safe, high-quality trajectories may already exist among geometrically similar candidates, yet the evaluator can still select an unsafe alternative. This exposes its insufficient sensitivity to fine-grained safety risks and interaction semantics, directly motivating our safety-aware semantic trajectory evaluator. Figure 3: Overview of the iDriveVLA framework. The candidate generator produces multiple trajectory hypotheses from multi-view observations and the ego state. Our evaluator consists of a safety-aware Scorer and a VLM-guided Modulator. The Scorer predicts criterion-wise planning scores and safety risks, while the Modulator generates scene-adaptive criterion weights. Unsafe candidates are filtered before score aggregation, and the highest-scoring trajectory is selected. The framework is progressively optimized through three training stages. Flame and snowflake icons denote trainable and frozen modules, respectively. 3.2 Safety-Aware Semantic Trajectory Evaluation Motivated by the generation–evaluation asymmetry identified above, we improve candidate evaluation with a Safety-aware Semantic Evaluator. As shown in Figure 3, it consists of a Safety-aware Scorer for candidate quality and safety estimation and a VLM-guided Modulator for scene-adaptive evaluation. Their outputs are combined through safety filtering and score aggregation to select the final trajectory. Safety-aware Scorer. For each candidate trajectory i τ_i, the Scorer first constructs a scene-conditioned trajectory representation: i=πscore(i,,ℐ). c_i= _score ( τ_i, e,I ). (2) Following (Dauner et al., 2024), we consider six structured planning criteria, ℳ=NC,DAC,TTC,EP,DDC,Comf.,M=\NC,DAC,TTC,EP,DDC,Comf.\, corresponding to collision avoidance, drivable-area compliance, time to collision, ego progress, driving-direction compliance, and comfort. A metric head predicts the criterion-wise scores: i=σ(πmetric(i)), v_i=σ ( _metric( c_i) ), (3) where i=vimm∈ℳ v_i=\v_i^m\_m denotes the normalized planning scores of candidate i τ_i. In parallel, a lightweight safety head predicts the collision and off-road risks: (picol,pioff)=σ(πsafe(i)), (p_i^col,p_i^off )=σ ( _safe( c_i) ), (4) which are used to suppress candidates with high predicted risk before final ranking. VLM-guided Modulator. Fixed criterion weights cannot capture the context-dependent preference among different driving behaviors. We therefore introduce a semantic Modulator that adapts the evaluation weights according to the current scene. Given the front-view image and a predefined driving prompt, a frozen VLM extracts a high-level semantic representation: =VLM(Ifront,prompt). z=VLM (I_front,prompt ). (5) Conditioned on z and the candidate features ii=1K\ c_i\_i=1^K, the DiT Modulator predicts a scene-adaptive criterion-weight vector: ∼πdit(⋅∣,ii=1K),=αmm∈ℳ. α _dit (· z,\ c_i\_i=1^K ), α=\ _m\_m . (6) The Modulator changes only how different planning criteria are weighted and does not modify the candidate trajectories themselves. Safety Filtering and Score Aggregation. For each candidate, we first construct a safety mask from the predicted collision and off-road risks. Unsafe candidates are removed from consideration, while the remaining candidates are ranked by aggregating the criterion-wise scores predicted in equation 3: Mi=[max(picol,pioff)≤δ],Si=Mi∑m∈ℳαmvim,M_i=I [ (p_i^col,p_i^off )≤δ ], S_i=M_i _m _mv_i^m, (7) where Mi∈0,1M_i∈\0,1\ denotes the safety mask. The final trajectory is then selected as ^=argmaxiSi. τ= τ_ _iS_i. 3.3 Oracle-Aligned Progressive Training To jointly improve candidate quality and selection reliability, we train iDriveVLA in three progressive stages: candidate imitation pretraining, candidate space refinement, and semantic ranking alignment. These stages progressively establish diverse candidates, improve their downstream planning quality, and align trajectory ranking with scene context. Stage 1: Candidate Imitation Pretraining. We first jointly train the trajectory generator and Scorer, while disabling the VLM-guided Modulator. Given the expert trajectory ⋆ τ , the generator is optimized with a best-of-K imitation objective: ℒtraj=mini=1,…,K‖i−⋆‖1.L_traj= _i=1,…,K \| τ_i- τ \|_1. (8) This objective encourages the candidate set to contain trajectories close to expert behaviors while preserving diverse motion hypotheses. Meanwhile, the Scorer is supervised using candidate-level planning metrics and safety labels. Let gimg_i^m denote the ground-truth value of criterion m∈ℳm , and define bicol=[giNC=0]b_i^col=I[g_i^NC=0] and bioff=[giDAC=0]b_i^off=I[g_i^DAC=0]. The evaluator loss is: ℒeval=1K∑i=1K[∑m∈ℳBCE(vim,gim)+BCE(picol,bicol)+BCE(pioff,bioff)].L_eval= 1K _i=1^K [ _m BCE(v_i^m,g_i^m)+BCE(p_i^col,b_i^col)+BCE(p_i^off,b_i^off) ]. (9) Combining the trajectory imitation objective with the evaluator supervision in equation 9, the overall Stage-1 objective is: ℒS1=ℒtraj+λevalℒeval,L_S1=L_traj+ _evalL_eval, (10) where λeval _eval balances trajectory imitation and structured evaluator supervision. Stage 2: Candidate Space Refinement. Although Stage 1 establishes a diverse candidate space, the generated distribution may still contain trajectories that are suboptimal under the downstream planning objective. We therefore further refine the candidate generator according to planning rewards. Inspired by reward-guided trajectory generation in diffusion-based planning (Zheng et al., 2025a), we introduce a reward-weighted supervised refinement (RFT) objective to emphasize training samples associated with higher downstream planning quality. Specifically, let R⋆∈[0,1]R ∈[0,1] denote the PDMS reward of the expert trajectory ⋆ τ . Building on the best-of-K imitation objective in equation 8, we weight each training sample according to the downstream quality of its expert trajectory: wRFT=exp(βR⋆),ℒRFT=wRFTℒtraj,w_RFT= (β R ), _RFT=w_RFTL_traj, (11) where β controls the strength of reward-based refinement. This exponential weighting can also be interpreted as reward-based exponential tilting of the supervised training distribution under a KL-regularized objective; we provide the corresponding derivation in Appendix B. The complete Stage-2 objective is: ℒS2=ℒRFT+λevalℒeval.L_S2=L_RFT+ _evalL_eval. (12) This stage places greater optimization emphasis on expert trajectories with higher downstream planning quality while retaining structured evaluator supervision. Stage 3: Semantic Ranking Alignment. Even with improved candidate quality, reliable trajectory selection requires understanding scene-dependent driving preferences. Therefore, we optimize the VLM-guided Modulator as a semantic ranking policy, while freezing the trajectory generator, Scorer, and VLM backbone. Conditioned on the semantic representation z and candidate features ii=1K\ c_i\_i=1^K, it samples multiple semantic ranking profiles: (g)∼πdit(⋅∣,ii=1K),g=1,…,G. α^(g) _dit (· z,\ c_i\_i=1^K ), g=1,…,G. (13) Each sampled profile (g) α^(g) produces candidate scores Si(g)S_i^(g) according to equation 7. We convert these scores into a soft selection distribution over the top-k candidates and compute the corresponding planning reward: Rg=∑i∈ℐk(g)ωi(g)ℛPDMS(i),ωi(g)=exp(Si(g))∑j∈ℐk(g)exp(Sj(g)).R_g= _i _k^(g) _i^(g)R_PDMS( τ_i), _i^(g)= (S_i^(g)) _j _k^(g) (S_j^(g)). (14) where ℐk(g)I_k^(g) denotes the indices of the top-k candidates ranked by Si(g)S_i^(g), and ℛPDMS(i)R_PDMS( τ_i) denotes the PDMS score obtained by evaluating candidate trajectory i τ_i (Guan et al., 2021). The rewards defined in equation 14 are standardized within each sampled group and used to optimize the Modulator with STAPO (Liu et al., 2026): STAPO(θ) _STAPO(θ) =g[min(ρg(θ)Ag,clip(ρg(θ),1−ϵlow,1+ϵhigh)Ag)], =E_g [ ( _g(θ)A_g,clip ( _g(θ),1- _low,1+ _high )A_g ) ], (15) Ag A_g =Rg−μRσR,ρg(θ)=πdit((g)∣,ii=1K)πditold((g)∣,ii=1K), = R_g- _R _R, _g(θ)= _dit ( α^(g) z,\ c_i\_i=1^K ) _dit_old ( α^(g) z,\ c_i\_i=1^K ), where θ denotes the parameters of the Modulator, μR _R and σR _R denote the within-group mean and standard deviation of the sampled rewards, respectively, and ρg(θ) _g(θ) denotes the importance ratio between the current and behavior policies, computed following (Li et al., 2026c). STAPO further masks abnormal gradients induced by outlier policy updates, thereby stabilizing the optimization process. By optimizing semantic ranking policies with downstream planning rewards, the Modulator learns to adapt candidate preferences according to scene semantics (Li, 2023). 4 Experiments 4.1 Settings Experimental Setup. For trajectory generation, we adopt the multi-modal planner of DrivoR (Kirby et al., 2026) and follow its standard configuration. For semantic modulation, we use a frozen InternVL3-2B initialized with the pretrained weights from ReCogDrive (Li et al., 2026c). All models are trained on 8 NVIDIA A800 GPUs following a three-stage training procedure, with RADAR (Ren et al., 2026) used as the optimizer. In Stage 1, we train the model for 25 epochs with a learning rate of 2×10−42× 10^-4. In Stage 2, the learning rate is reduced to 1×10−51× 10^-5, and the model is trained for an additional 30 epochs. In the third stage, we train the VLM-conditioned diffusion module for 25 epochs with a learning rate of 1×10−51× 10^-5. Further training and optimization details are provided in Appendix C.2. Dataset. The model is trained primarily on NAVSIM (Dauner et al., 2024), a planning-oriented autonomous driving dataset from OpenScene. NAVSIM contains approximately 103K training samples and 12K test samples. To improve data diversity, we augment its real-world training set with 134K simulated samples generated by SimScale (Tian et al., 2026). These samples are obtained by perturbing ego trajectories and rendering them in reactive environments. Metrics. We evaluate trajectory quality through closed-loop simulation in NAVSIM, reporting the Predictive Driver Model Score (PDMS) for NAVSIM v1 and the Extended Predictive Driver Model Score (EPDMS) for NAVSIM v2. Detailed definitions and formulations are provided in Appendix C.1. Table 1: Closed-loop evaluation on NAVSIM v1. Higher values indicate better performance. Method NC↑ DAC↑ TTC↑ Comf.↑ EP↑ PDMS↑ PDM-Closed (Dauner et al., 2023) 94.6 99.8 89.9 86.9 99.9 89.1 Human driver (Dauner et al., 2024) 100.0 100.0 100.0 99.9 87.5 94.8 E2E Methods UniAD (Hu et al., 2023) 97.8 91.9 92.9 100.0 78.8 83.4 LTF (Chitta et al., 2023) 97.4 92.8 92.4 100.0 79.0 83.8 PARA-Drive (Weng et al., 2024) 97.9 92.4 93.0 99.8 79.3 84.0 DriveX-S (Shi et al., 2025) 97.5 94.0 93.0 100.0 79.7 84.5 World4Drive (Zheng et al., 2025b) 97.4 94.3 92.8 100.0 79.9 85.1 DRAMA (Yuan et al., 2024) 98.0 93.1 94.8 100.0 80.1 85.5 DiffusionDrive (Liao et al., 2025) 98.2 96.2 94.7 100.0 82.2 88.1 iPad (Guo et al., 2025) 98.6 98.3 94.9 100.0 88.0 91.7 DrivoR (Kirby et al., 2026) 99.0 98.9 96.7 100.0 90.0 93.7 RAP (Feng et al., 2026) 99.1 98.9 96.7 100.0 90.3 93.8 CLOVER (Ang et al., 2026) 99.1 99.0 96.9 100.0 91.7 94.5 TOAD (Xu et al., 2026) 99.0 99.3 96.8 100.0 91.8 94.7 VLA Methods UniVLA (Wang et al., 2025) 96.9 91.1 91.7 96.7 76.8 81.7 FSDrive (Zeng et al., 2025) 98.2 93.8 93.3 99.9 80.1 85.1 AutoVLA (Zhou et al., 2025) 98.4 95.6 98.0 99.9 81.9 89.1 DriveVLA-W0 (Li et al., 2026b) 98.7 99.1 95.3 99.3 83.3 90.2 AdaThinkDrive (Luo et al., 2026c) 98.4 97.8 95.2 100.0 84.4 90.3 SpanVLA (Zhou et al., 2026b) 99.1 97.1 95.2 100.0 86.3 90.3 ReCogDrive (Li et al., 2026c) 97.9 97.3 94.9 100.0 87.3 90.8 ELF-VLA (Luo et al., 2026a) 98.9 98.1 96.0 100.0 85.3 91.0 SGDrive (Li et al., 2026a) 98.6 97.8 96.2 100.0 85.8 91.1 FLARE (Xie et al., 2026) 98.5 98.4 96.0 100.0 86.0 91.4 CLEAR (Xing et al., 2026) 99.1 98.8 97.2 99.6 89.7 93.7 ChainFlow-VLA (Wang et al., 2026) 99.2 99.0 97.2 99.9 91.9 94.8 iDriveVLA 98.9 99.3 96.4 100.0 92.9 94.9 Table 2: Closed-loop evaluation on NAVSIM v2. All results are computed using the original NAVSIM v2 evaluation code without the human-behavior filtering fix. Method NC↑ DAC↑ DDC↑ TLC↑ EP↑ TTC↑ LK↑ HC↑ EC↑ EPDMS↑ E2E Methods Ego Status (Cao et al., 2025) 93.1 77.9 92.7 99.6 86.0 91.5 89.4 98.3 85.4 64.0 TransFuser (Chitta et al., 2023) 96.9 89.9 97.8 99.7 87.1 95.4 92.7 98.3 87.2 76.7 DiffusionDrive (Liao et al., 2025) 98.2 95.9 99.4 99.8 87.5 97.3 96.8 98.3 87.7 84.5 Hydra-MDP++ (Li et al., 2025a) 98.5 98.5 99.5 99.7 87.4 97.9 95.8 98.2 75.7 85.6 CLOVER (Ang et al., 2026) 99.4 99.3 99.5 99.8 86.9 98.9 95.2 98.3 75.5 87.2 VLA Methods DriveVLA-W0 (Li et al., 2026b) 98.5 99.1 98.0 99.7 86.4 98.1 93.2 97.9 58.9 86.1 ReCogDrive (Li et al., 2026c) 98.3 95.2 99.5 99.8 87.1 97.5 96.6 98.3 86.5 83.6 LaST-VLA (Luo et al., 2026b) 98.7 97.9 99.2 99.7 90.3 98.2 96.6 98.3 86.3 87.1 iDriveVLA 98.9 99.3 99.0 96.8 91.6 98.4 85.3 98.3 77.9 87.4 4.2 Main Results As shown in Table 1, iDriveVLA achieves the highest PDMS on NAVSIM v1, reaching 94.9 and ranking 1st among all compared methods. It outperforms both TOAD, the strongest existing end-to-end driving method, and ChainFlow-VLA, the best-performing VLA-based approach. These results demonstrate that improving trajectory evaluation can effectively unlock the potential of multi-modal trajectory planning, rather than relying solely on stronger trajectory generation. Notably, iDriveVLA also surpasses the human-driver reference score of 94.8 PDMS. This further suggests that incorporating safety-aware evaluation and VLM-based semantic modulation can improve the planner’s ability to identify high-quality trajectories and push overall driving performance beyond the human-reference level on this benchmark. To further evaluate iDriveVLA under a more comprehensive driving protocol, we report results on NAVSIM v2, which extends PDMS with additional criteria for driving-direction and traffic-light compliance, lane keeping, and temporal comfort. As shown in Table 2, iDriveVLA achieves state-of-the-art performance with an EPDMS of 87.4. Compared with LaST-VLA and CLOVER, iDriveVLA maintains competitive safety and compliance performance while achieving higher ego progress, indicating robust trajectory selection under the more demanding evaluation protocol. 4.3 Ablation Studies We further analyze the contributions of the three key components of iDriveVLA in Table 3. Starting from the DrivoR baseline with a PDMS of 93.7, introducing the Safety-aware Scorer improves performance to 93.9 by suppressing risky candidates without sacrificing driving efficiency. Candidate space refinement further raises PDMS to 94.2, demonstrating that refinement fine-tuning improves the quality of the candidate set before trajectory selection. Finally, adding the VLM-guided Modulator achieves the best overall performance by incorporating high-level scene semantics into trajectory evaluation, enabling more context-adaptive candidate evaluation. Table 3: Ablation study of the proposed components on NAVSIM v1. Safety Gate denotes the Safety-aware Scorer, Refinement denotes candidate space refinement fine-tuning, and Modulator denotes the VLM-guided Modulator. All variants are trained exclusively on NAVSIM data. Components Planning Metrics PDMS Safety Gate Refinement Modulator NC↑ DAC↑ TTC↑ Comf.↑ EP↑ PDMS↑ Base Planner 99.0 98.9 96.7 100.0 90.0 93.7 – ✓ 99.1 98.9 96.7 100.0 91.1 93.9 +0.2 ✓ ✓ 99.0 99.0 96.6 100.0 91.3 94.2 +0.5 ✓ ✓ ✓ 99.1 99.1 96.6 100.0 91.6 94.4 +0.7 4.4 Qualitative Results Figure 4: Comparison between DrivoR and iDriveVLA. Red and green curves denote the DrivoR and iDriveVLA trajectories, respectively. The paired camera and bird’s-eye-view panels show how the predicted paths differ under varied road geometry and surrounding traffic. Figure 5: Qualitative comparison of trajectory selection without and with the VLM-guided Modulator. Red and green curves denote the selected trajectories without and with the Modulator, respectively. The Modulator changes the ranking of existing candidates according to the surrounding road structure and traffic context. We further visualize the trajectories selected by iDriveVLA and the base planner DrivoR in representative driving scenarios, as shown in Figure 4. The red and green curves denote the trajectories selected by DrivoR and iDriveVLA, respectively. In Figures 4(a) and 4(b), DrivoR selects trajectories that lead to collisions or departures from the drivable area, whereas iDriveVLA avoids these failures while maintaining forward progress. As Figures 4(c) and 4(d) show, DrivoR exhibits overly conservative behavior during turning and lane-changing maneuvers, resulting in limited progress. In contrast, iDriveVLA selects smoother and more appropriate trajectories that remain within the drivable area while advancing efficiently. These examples qualitatively demonstrate that context-aware trajectory evaluation enables iDriveVLA to select safer and more effective behaviors across diverse driving scenarios. We also investigate the effect of the VLM-guided Modulator on trajectory selection in Figure 5. With the candidate set held fixed, the Modulator alters only the selection preference. In Figure 5(a), the trajectory selected with modulation, shown in green, follows the lane structure more closely while maintaining greater clearance from nearby obstacles. The remaining three cases further show that scene-level semantic guidance helps avoid overly conservative choices while preserving the existing safety calibration. Overall, the Modulator improves driving progress without introducing additional collisions or drivable-area violations in these examples. Additional comparisons with oracle trajectories are provided in Appendix D.2, showing that iDriveVLA can identify optimal or near-optimal candidates. 5 Conclusion Our work revisits multi-modal planning from the perspective of trajectory evaluation rather than candidate generation alone. Our results show that a considerable portion of planning performance is already contained in the candidate set, but remains inaccessible when the planner cannot reliably distinguish high-quality trajectories from suboptimal ones. Building on this observation, we develop iDriveVLA to jointly improve candidate quality and trajectory selection through safety-aware scoring, semantic modulation, and oracle-aligned training. Extensive experiments on NAVSIM v1 and v2 demonstrate consistent gains across overall planning quality, safety, compliance, and driving progress, with iDriveVLA achieving state-of-the-art performance on both benchmarks. Despite these improvements, a substantial gap between selected and oracle performance still remains, indicating that the potential of existing candidate sets is far from fully exploited. We hope iDriveVLA can serve as a step toward establishing trajectory evaluation as a more explicit research direction in multi-modal autonomous driving and inspire future work on more reliable, adaptive, and semantically informed trajectory evaluation. AI use statement Generative AI tools were employed in a limited supporting role during the preparation of this manuscript. Their use was restricted to improving linguistic clarity and presentation, assisting with routine code debugging and small code revisions, and helping prepare or refine illustrative schematic figures. They were not used to generate scientific content or to make substantive research decisions, including the formulation of research questions or hypotheses, conceptual or theoretical development, mathematical reasoning, methodology and experimental design, implementation of the proposed method, data construction or processing, qualitative assessment, or interpretation of experimental results. All AI-assisted outputs were subsequently inspected and revised by the authors. Any modified code was validated within the experimental workflow, and all generated or edited figures were checked for consistency with the corresponding technical descriptions and results. The authors independently developed and verified the scientific ideas, methods, experiments, analyses, and conclusions, and remain fully responsible for the accuracy and integrity of the final manuscript. References Ang et al. (2026) S. Ang, Y. Yang, C. Chen, and Y. Wang CLOVER: closed-loop value estimation & ranking for end-to-end autonomous driving planning. arXiv preprint arXiv:2605.15120. External Links: Link Cited by: §1, Table 1, Table 2. Cao et al. (2025) W. Cao, M. Hallgarten, T. Li, D. Dauner, X. Gu, C. Wang, Y. Miron, M. Aiello, H. Li, I. Gilitschenski, B. Ivanovic, M. Pavone, A. Geiger, and K. Chitta Pseudo-simulation for autonomous driving. In Conference on Robot Learning, External Links: Link Cited by: Table 2. Chi et al. (2025) H. Chi, H. Gao, Z. Liu, J. Liu, C. Liu, J. Li, K. Yang, Y. Yu, Z. Wang, W. Li, L. Wang, X. Hu, H. Sun, H. Zhao, and H. Zhao Impromptu vla: open weights and open data for driving vision-language-action models. In Advances in Neural Information Processing Systems, Vol. 38, p. 84771–84806. External Links: Document Cited by: §2. Chitta et al. (2023) K. Chitta, A. Prakash, B. Jaeger, Z. Yu, K. Renz, and A. Geiger TransFuser: imitation with transformer-based sensor fusion for autonomous driving. IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (11), p. 12878–12895. External Links: Document, Link Cited by: Table 1, Table 2. Dauner et al. (2023) D. Dauner, M. Hallgarten, A. Geiger, and K. Chitta Parting with misconceptions about learning-based vehicle motion planning. In Proceedings of the 7th Conference on Robot Learning, Proceedings of Machine Learning Research, Vol. 229, p. 1268–1281. External Links: Link Cited by: Table 1. Dauner et al. (2024) D. Dauner, M. Hallgarten, T. Li, X. Weng, Z. Huang, Z. Yang, H. Li, I. Gilitschenski, B. Ivanovic, M. Pavone, A. Geiger, and K. Chitta NAVSIM: data-driven non-reactive autonomous vehicle simulation and benchmarking. In Advances in Neural Information Processing Systems, Vol. 37, p. 28706–28719. External Links: Document Cited by: Figure 1, 3rd item, §1, §2, §3.2, §4.1, Table 1. Feng et al. (2026) L. Feng, Y. Gao, E. Zablocki, Q. Li, W. Li, S. Liu, M. Cord, and A. Alahi RAP: 3d rasterization augmented end-to-end planning. In International Conference on Learning Representations, Cited by: Table 1. Guan et al. (2021) Y. Guan, S. E. Li, J. Duan, J. Li, Y. Ren, Q. Sun, and B. Cheng Direct and indirect reinforcement learning. International Journal of Intelligent Systems 36 (8), p. 4439–4467. External Links: Document Cited by: §3.3. Guan et al. (2023) Y. Guan, Y. Ren, Q. Sun, S. E. Li, H. Ma, J. Duan, Y. Dai, and B. Cheng Integrated decision and control: toward interpretable and computationally efficient driving intelligence. IEEE Transactions on Cybernetics 53 (2), p. 859–873. External Links: Document Cited by: §1. Guan et al. (2026) Y. Guan, L. Tang, Y. Lyu, S. E. Li, Y. Yang, K. Sheng, B. Zhang, L. He, and K. Li Enhanced integrated decision and control for high-level automated vehicles and its experiment verification. IEEE Transactions on Automation Science and Engineering 23, p. 11578–11595. External Links: Document Cited by: §1. Guo et al. (2025) K. Guo, H. Liu, X. Wu, J. Pan, and C. Lv iPad: iterative proposal-centric end-to-end autonomous driving. arXiv preprint arXiv:2505.15111. Cited by: §1, §2, §2, Table 1. Hu et al. (2023) Y. Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang, L. Lu, X. Jia, Q. Liu, J. Dai, Y. Qiao, and H. Li Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 17853–17862. External Links: Document Cited by: §1, Table 1. Hwang et al. (2025) J. Hwang, R. Xu, H. Lin, W. Hung, J. Ji, K. Choi, D. Huang, T. He, P. Covington, B. Sapp, Y. Zhou, J. Guo, D. Anguelov, and M. Tan EMMA: end-to-end multimodal model for autonomous driving. Transactions on Machine Learning Research. Cited by: §2. Jia et al. (2024) X. Jia, Z. Yang, Q. Li, Z. Zhang, and J. Yan Bench2Drive: towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. In Advances in Neural Information Processing Systems, Vol. 37, p. 819–844. Cited by: §2. Jiang et al. (2026) B. Jiang, S. Chen, H. Gao, B. Liao, Q. Zhang, W. Liu, and X. Wang VADv2: end-to-end vectorized autonomous driving via probabilistic planning. In International Conference on Learning Representations, Cited by: §2. Jiang et al. (2025) S. Jiang, Z. Huang, K. Qian, Z. Luo, T. Zhu, Y. Zhong, Y. Tang, M. Kong, Y. Wang, S. Jiao, H. Ye, Z. Sheng, X. Zhao, T. Wen, Z. Fu, S. Chen, K. Jiang, D. Yang, S. Choi, and L. Sun A survey on vision-language-action models for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, p. 4583–4595. External Links: Document Cited by: §1, §2. Kirby et al. (2026) E. Kirby, A. Boulch, Y. Xu, Y. Yin, G. Puy, É. Zablocki, A. Bursuc, S. Gidaris, R. Marlet, F. Bartoccioni, A. Cao, N. Samet, T. Vu, and M. Cord Driving on registers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 32058–32069. Cited by: §C.2, Figure 1, §2, §2, §3.1, §4.1, Table 1. Li et al. (2026a) J. Li, J. Wu, D. Hu, X. Huang, B. Sun, Z. Hao, X. Lang, X. Zhu, and L. Zhang SGDrive: scene-to-goal hierarchical world cognition for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Cited by: Table 1. Li et al. (2025a) K. Li, Z. Li, S. Lan, Y. Xie, Z. Zhang, J. Liu, Z. Wu, Z. Yu, and J. M. Alvarez Hydra-MDP++: advancing end-to-end driving via expert-guided hydra-distillation. arXiv preprint arXiv:2503.12820. External Links: Link Cited by: Table 2. Li (2023) S. E. Li Reinforcement learning for sequential decision and optimal control. Springer Nature Singapore, Singapore. External Links: Document, ISBN 978-981-19-7784-8 Cited by: §3.3. Li et al. (2026b) Y. Li, S. Shang, W. Liu, B. Zhan, H. Wang, Y. Wang, Y. Chen, X. Wang, Y. An, C. Tang, L. Hou, L. Fan, and Z. Zhang DriveVLA-W0: world models amplify data scaling law in autonomous driving. In International Conference on Learning Representations, External Links: Link Cited by: Table 1, Table 2. Li et al. (2026c) Y. Li, K. Xiong, X. Guo, F. Li, S. Yan, G. Xu, L. Zhou, L. Chen, H. Sun, B. Wang, K. Ma, G. Chen, H. Ye, W. Liu, and X. Wang ReCogDrive: a reinforced cognitive framework for end-to-end autonomous driving. In International Conference on Learning Representations, Cited by: §2, §3.3, §4.1, Table 1, Table 2. Li et al. (2025b) Z. Li, W. Yao, Z. Wang, X. Sun, J. Chen, N. Chang, M. Shen, Z. Wu, S. Lan, and J. M. Alvarez Generalized trajectory scoring for end-to-end multimodal planning. arXiv preprint arXiv:2506.06664. External Links: 2506.06664 Cited by: §1, §2. Liao et al. (2025) B. Liao, S. Chen, H. Yin, B. Jiang, C. Wang, S. Yan, X. Zhang, X. Li, Y. Zhang, Q. Zhang, and X. Wang DiffusionDrive: truncated diffusion model for end-to-end autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 12037–12047. External Links: Document, Link Cited by: §1, §2, Table 1, Table 2. Liu et al. (2026) S. Liu, Z. He, G. Zhan, L. Tao, Z. Zheng, J. Wu, Y. Wang, Y. Guan, K. Sheng, B. Zhang, et al. Stapo: stabilizing reinforcement learning for llms by silencing rare spurious tokens. arXiv preprint arXiv:2602.15620. Cited by: §3.3. Luo et al. (2026a) Y. Luo, F. Li, Q. Chen, S. Xu, J. Liu, Z. Song, Z. Yang, and F. Wen Unleashing vla potentials in autonomous driving via explicit learning from failures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 24833–24842. Cited by: Table 1. Luo et al. (2026b) Y. Luo, F. Li, S. Xu, Y. Ji, Z. Zhang, B. Wang, Y. Shen, J. Cui, L. Chen, G. Chen, H. Ye, Z. Yang, and F. Wen LaST-VLA: thinking in latent spatio-temporal space for vision-language-action in autonomous driving. arXiv preprint arXiv:2603.01928. External Links: Link Cited by: Table 2. Luo et al. (2026c) Y. Luo, F. Li, S. Xu, Z. Lai, L. Yang, Q. Chen, Z. Luo, Z. Xie, S. Jiang, J. Liu, L. Chen, B. Wang, and Z. Yang AdaThinkDrive: adaptive thinking via reinforcement learning for autonomous driving. In Proceedings of the IEEE International Conference on Robotics and Automation, External Links: Link Cited by: Table 1. Ren et al. (2026) Z. Ren, Y. Lyu, C. Li, L. Zhang, and S. E. Li Momentum as residual-driven multiplier correction for deep learning optimization. arXiv preprint arXiv:2608.12925. Cited by: §4.1. Shi et al. (2025) C. Shi, S. Shi, K. Sheng, B. Zhang, and L. Jiang DriveX: omni scene modeling for learning generalizable world knowledge in autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 28599–28609. External Links: Link Cited by: Table 1. Tian et al. (2026) H. Tian, T. Li, H. Liu, J. Yang, Y. Qiu, G. Li, J. Wang, Y. Gao, Z. Zhang, L. Wang, H. Ye, L. Chen, and H. Li SimScale: learning to drive via real-world simulation at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 36365–36374. Cited by: §4.1. Wang et al. (2026) X. Wang, X. Wang, T. Zhou, G. Chen, X. Gui, Z. Xu, X. Wu, F. Tan, H. Zhou, and M. Yang ChainFlow-VLA: causal flow planning with vision-language models. arXiv preprint arXiv:2605.23270. Cited by: Table 1. Wang et al. (2025) Y. Wang, X. Li, W. Wang, J. Zhang, Y. Li, Y. Chen, X. Wang, and Z. Zhang Unified vision-language-action model. arXiv preprint arXiv:2506.19850. Cited by: Table 1. Weng et al. (2024) X. Weng, B. Ivanovic, Y. Wang, Y. Wang, and M. Pavone PARA-Drive: parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 15449–15458. External Links: Link Cited by: Table 1. Xie et al. (2026) C. Xie, C. Sima, T. Li, B. Sun, J. Wu, Z. Hao, and H. Li FLARE: learning future-aware latent representations from vision-language models for autonomous driving. arXiv preprint arXiv:2601.05611. Cited by: Table 1. Xing et al. (2026) Y. Xing, Z. Ke, Z. Liu, Y. Jiang, W. Yu, and J. Wang CLEAR: cognition and latent evaluation for adaptive routing in end-to-end autonomous driving. arXiv preprint arXiv:2606.06219. Cited by: §1, Table 1. Xiong et al. (2025) Z. Xiong, J. K. Eappen, and S. Jagannathan FLoRA: a framework for learning scoring rules in autonomous driving planning systems. IEEE Robotics and Automation Letters 10 (4), p. 4101–4108. External Links: Document Cited by: §2. Xu et al. (2026) Y. Xu, E. Zablocki, Y. Yin, E. Ramzi, E. Kirby, A. Boulch, and M. Cord Test-time trajectory optimization for autonomous driving. arXiv preprint arXiv:2606.07170. Cited by: Table 1. Yao et al. (2026) W. Yao, Z. Li, S. Lan, Z. Wang, X. Sun, J. M. Alvarez, and Z. Wu DriveSuprim: towards precise trajectory selection for end-to-end planning. Proceedings of the AAAI Conference on Artificial Intelligence 40 (14), p. 11910–11918. External Links: Document Cited by: §1. Yuan et al. (2024) C. Yuan, Z. Zhang, J. Sun, S. Sun, Z. Huang, C. D. W. Lee, D. Li, Y. Han, A. Wong, K. P. Tee, and M. H. Ang DRAMA: an efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601. External Links: Link Cited by: Table 1. Zeng et al. (2025) S. Zeng, X. Chang, M. Xie, X. Liu, Y. Bai, Z. Pan, M. Xu, and X. Wei FutureSightDrive: thinking visually with spatio-temporal cot for autonomous driving. In Advances in Neural Information Processing Systems, Vol. 38. External Links: Document Cited by: Table 1. Zheng et al. (2024) W. Zheng, R. Song, X. Guo, C. Zhang, and L. Chen GenAD: generative end-to-end autonomous driving. In European Conference on Computer Vision, p. 87–104. External Links: Document Cited by: §2. Zheng et al. (2025a) Y. Zheng, R. Liang, K. Zheng, J. Zheng, L. Mao, J. Li, W. Gu, R. Ai, S. E. Li, X. Zhan, and J. Liu Diffusion-based planning for autonomous driving with flexible guidance. In International Conference on Learning Representations, Cited by: §3.3. Zheng et al. (2025b) Y. Zheng, P. Yang, Z. Xing, Q. Zhang, Y. Zheng, Y. Gao, P. Li, T. Zhang, Z. Xia, P. Jia, X. Lang, and D. Zhao World4Drive: end-to-end autonomous driving via intention-aware physical latent world model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, p. 28632–28642. External Links: Link Cited by: Table 1. Zhou et al. (2026a) X. Zhou, X. Han, F. Yang, Y. Ma, V. Tresp, and A. Knoll OpenDriveVLA: towards end-to-end autonomous driving with large vision language action model. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, p. 13782–13790. External Links: Document Cited by: §1, §2. Zhou et al. (2025) Z. Zhou, T. Cai, S. Z. Zhao, Y. Zhang, Z. Huang, B. Zhou, and J. Ma AutoVLA: a vision-language-action model for end-to-end autonomous driving with adaptive reasoning and reinforcement fine-tuning. In Advances in Neural Information Processing Systems, Vol. 38, p. 31725–31761. External Links: Document Cited by: §2, Table 1. Zhou et al. (2026b) Z. Zhou, R. Yang, X. Qi, Y. Guo, S. X. Chen, T. Feng, K. Pistunova, Y. Shen, L. Su, and J. Ma SpanVLA: efficient action bridging and learning from negative-recovery samples for vision-language-action model. arXiv preprint arXiv:2604.19710. Cited by: Table 1. Appendix A Theoretical Analysis of Candidate Competition To provide a theoretical explanation for the trajectory evaluation bottleneck discussed in Sec. 3.1, we consider a stylized setting in which an oracle trajectory competes with multiple near-optimal candidates under imperfect evaluation. The analysis characterizes how increasing candidate competition affects oracle recovery when evaluation errors are present. Proposition A.1 (Candidate Competition under Imperfect Evaluation). Let ri=ℛ(i)r_i=R( τ_i) denote the reference planning score of candidate i τ_i. Consider a candidate set containing one oracle trajectory i⋆i with score r⋆r and K−1K-1 competing trajectories with score r⋆−γr -γ, where γ>0γ>0. Suppose the evaluator predicts si=ri+ϵi,s_i=r_i+ _i, (16) where ϵii=1K\ _i\_i=1^K are i.i.d. evaluation errors drawn from a continuous distribution with cumulative distribution function F. Let i^=argmaxisi i= _is_i denote the selected candidate and define the evaluation gap as Δeval(K)=r⋆−ri _eval(K)=r -r_ i. Then the probability of recovering the oracle trajectory is PK=Pr(i^=i⋆)=ϵ⋆[F(ϵ⋆+γ)K−1],P_K= ( i=i )=E_ε [F(ε +γ)^K-1 ], (17) where ϵ⋆ε is the evaluation error associated with the oracle trajectory. The expected evaluation gap is therefore [Δeval(K)]=γ(1−PK).E [ _eval(K) ]=γ(1-P_K). (18) Moreover, if Pr(0<F(ϵ⋆+γ)<1)>0 \! (0<F(ε +γ)<1 )>0, then PKP_K strictly decreases with K, while [Δeval(K)]E[ _eval(K)] strictly increases with K. Proof. Condition on the oracle evaluation error ϵ⋆=eε =e. A competing trajectory is ranked below the oracle trajectory if r⋆−γ+ϵi<r⋆+e,r -γ+ _i<r +e, which is equivalent to ϵi<e+γ _i<e+γ. This event occurs with probability F(e+γ)F(e+γ). Since the K−1K-1 competing trajectories have independent evaluation errors, the conditional probability that all of them are ranked below the oracle is Pr(i^=i⋆∣ϵ⋆=e)=F(e+γ)K−1. ( i=i ε =e)=F(e+γ)^K-1. (19) Taking the expectation over ϵ⋆ε yields equation 17. Under the assumed score structure, the evaluation gap is zero when the oracle trajectory is selected and equals γ otherwise. Hence, [Δeval(K)]=γPr(i^≠i⋆)=γ(1−PK),E [ _eval(K) ]=γ ( i≠ i )=γ(1-P_K), (20) which gives equation 18. To establish monotonicity, define X=F(ϵ⋆+γ)∈[0,1].X=F(ε +γ)∈[0,1]. Then PK−PK+1=[XK−1(1−X)].P_K-P_K+1=E [X^K-1(1-X) ]. (21) Since XK−1(1−X)≥0X^K-1(1-X)≥ 0, we have PK+1≤PKP_K+1≤ P_K. Furthermore, if Pr(0<X<1)>0 (0<X<1)>0, then [XK−1(1−X)]>0E[X^K-1(1-X)]>0, and therefore PK+1<PKP_K+1<P_K. Combining this result with equation 18 immediately gives [Δeval(K+1)]>[Δeval(K)].E[ _eval(K+1)]>E[ _eval(K)]. Thus, under imperfect evaluation, increasing the number of competing trajectories reduces the probability of recovering the oracle and enlarges the expected evaluation gap. ∎ This result provides a simplified explanation for the generation–evaluation asymmetry observed in multi-candidate planning. Increasing the number of trajectory hypotheses can improve candidate coverage and oracle potential, but it also exposes the evaluator to more competing alternatives. Consequently, when evaluation remains imperfect, stronger candidate generation alone does not guarantee a comparable improvement in final trajectory selection. Appendix B Interpretation of Candidate Space Refinement We provide a simple interpretation of the exponential weighting used in Stage 2. Let =(,⋆) ξ=( x, τ ) denote a supervised training example drawn from the empirical demonstration distribution p()p_D( ξ), and let R()=R(,⋆)R( ξ)=R( x, τ ) denote its downstream planning reward. We consider a reward-tilted demonstration distribution that favors higher-reward examples while remaining close to the original training distribution: q⋆=argmaxq[∼q[R()]−1βDKL(q()∥p())].q = _q [E_ ξ q[R( ξ)]- 1βD_KL (q( ξ)\|p_D( ξ) ) ]. (22) Introducing a Lagrange multiplier for the normalization constraint ∫q()=1 q( ξ)\,d ξ=1, the stationary condition is R()−1β(logq⋆()p()+1)+λ=0,R( ξ)- 1β ( q ( ξ)p_D( ξ)+1 )+λ=0, (23) which yields the Gibbs-form solution q⋆()=p()exp(βR())Zβ,Zβ=∼p[exp(βR())].q ( ξ)= p_D( ξ) (β R( ξ) )Z_β, Z_β=E_ ξ p_D [ (β R( ξ) ) ]. (24) Therefore, Eq. equation 22 corresponds to exponentially tilting the original demonstration distribution according to downstream planning quality. Training the candidate generator under the tilted distribution in Eq. equation 24 gives ∼q⋆[ℒtraj(,ϕ)]=1Zβ∼p[exp(βR())ℒtraj(,ϕ)].E_ ξ q [L_traj( ξ;φ) ]= 1Z_βE_ ξ p_D [ (β R( ξ) )L_traj( ξ;φ) ]. (25) Since ZβZ_β is independent of the candidate-generator parameters ϕφ, minimizing Eq. equation 25 is equivalent, up to a positive constant factor, to the reward-weighted objective in Eq. equation 11. Importantly, our supervision is deterministic: each scene x is associated with a single expert trajectory ⋆=f⋆() τ =f ( x), such that p(,)=p()δ(−f⋆()).p_D( x, τ)=p_D( x)δ ( τ-f ( x) ). (26) Under this setting, the exponential tilting in Eq. equation 24 does not change the supervision target within an individual scene. Instead, it reweights the relative contributions of different demonstrations, assigning greater optimization weight to scenes whose expert trajectories achieve higher downstream planning rewards. Consequently, Stage 2 remains a fully supervised refinement procedure, while the shared candidate generator is preferentially optimized toward trajectory patterns associated with higher planning quality. Appendix C Experiment Details C.1 Evaluation Metrics We employ benchmark-specific metrics to comprehensively evaluate planning performance. For closed-loop planning, we report the Predictive Driver Model Score (PDMS) on NAVSIM v1 and the Extended Predictive Driver Model Score (EPDMS) on NAVSIM v2. PDMS on NAVSIM v1. PDMS integrates five sub-metrics: No At-Fault Collision (NC), Drivable Area Compliance (DAC), Time-to-Collision (TTC), Comfort (C), and Ego Progress (EP). It is computed as PDMS=NC×DAC×(5×EP+5×TTC+2×C12).PDMS=NC×DAC× ( 5×EP+5×TTC+2×C12 ). (27) EPDMS on NAVSIM v2. EPDMS extends PDMS with additional components that provide a more comprehensive evaluation of safety, traffic-rule compliance, driving progress, lane keeping, and comfort. It includes No At-Fault Collision (NC), Drivable Area Compliance (DAC), Driving Direction Compliance (DDC), Traffic Light Compliance (TLC), Ego Progress (EP), Lane Keeping (LK), History Comfort (HC), Time-to-Collision (TTC), and Extended Comfort (EC). It is computed as EPDMS=NC×DAC×DDC×TLC×(5×EP+2×LK+2×HC+5×TTC+2×EC16).EPDMS=NC×DAC×DDC×TLC× ( 5×EP+2×LK+2×HC+5×TTC+2×EC16 ). (28) C.2 Implementation Details The trajectory generator follows the standard configuration of DrivoR (Kirby et al., 2026). It takes 4 camera views and the ego state as input. We use DINOv2 ViT-S as the visual encoder and adapt it using LoRA, producing 16 visual registers for each view. The trajectory decoder and Scorer each consist of 4 layers. The Safety Head is implemented as a 2-layer MLP that maps candidate features to 2 binary risk predictions: collision and departure from the drivable area. For semantic modulation, we use a frozen InternVL3-2B initialized with the pretrained weights from ReCogDrive. Its hidden features condition an 8-block DiT Modulator. The main model and training configurations are summarized in Table 4. Table 4: Model and training configurations. Configuration Value Candidate number K 64 Future planning steps T 8 Planning time interval 0.5 s Visual registers per view 16 Trajectory decoder layers 4 Scorer layers 4 DiT blocks 8 DiT sampling steps 5 Global batch size 64 Safety threshold δ 0.95 Evaluator loss weight λeval _eval 1 Reward coefficient β 1 Profiles per group G 8 Top-k candidates k 16 Lower clipping bound ϵlow _low 0.2 Upper clipping bound ϵhigh _high 0.28 Appendix D Supplementary Experimental Results D.1 Computational Time and Memory Analysis We profile the computational and memory overhead of iDriveVLA on an A800 GPU with a single-element batch. The complete pipeline requires 297.20 ms per sample, only 186.10 ms more than DrivoR. This sub-second latency satisfies the real-time inference requirements of practical driving applications. Table 5: Computational and memory overhead of the iDriveVLA inference pipeline. Module Time (ms) Parameter memory (MiB) Vision Encoder 97.73 90.53 Action Decoder 1.25 41.74 Scorer 3.40 7.15 Safety Head 0.20 2.02 Modulator 194.62 3999.51 Total 297.20 4140.95 D.2 Qualitative Analysis of Trajectory Selection Selection failures of DrivoR. Figure 6: Representative catastrophic selection failures of DrivoR. Yellow curves denote the candidate trajectories, while red and green curves denote the selected and oracle trajectories, respectively. Despite the presence of a safe oracle trajectory, the evaluator selects an unsafe candidate with zero PDMS. Figure 6 shows representative failures at intersections and in interactions with vehicles and pedestrians. The candidate overlays demonstrate that the generator already provides multiple plausible alternatives, including a safe oracle trajectory. Nevertheless, DrivoR selects an unsafe candidate with zero PDMS. Although the selected and oracle trajectories differ only subtly in curvature or lateral offset, these differences determine collision risk and road compliance, exposing the evaluator’s limited sensitivity to fine-grained safety risks and interaction semantics. Selection results of iDriveVLA. Figure 7: Qualitative comparison between the trajectory selected by iDriveVLA and the oracle trajectory. The selected trajectories closely follow or overlap with the oracle trajectories, demonstrating reliable identification of high-quality candidates. We further compare the trajectories selected by iDriveVLA with the oracle trajectories in Figure 7. The oracle trajectory is defined as the candidate achieving the highest ground-truth PDMS within the generated candidate set. In Figure 7(a), the selected trajectory closely follows the oracle trajectory, with only a minor geometric deviation, demonstrating a near-oracle selection. In Figure 7(b), iDriveVLA successfully identifies the oracle candidate, as indicated by the overlap between the selected and oracle trajectories. These examples show that the proposed evaluator can distinguish high-quality candidates from a diverse trajectory set and select optimal or near-optimal driving behaviors.