Teaching VLMs What to Say, Not How to Reason: Rethinking Counterfactual Reasoning in Autonomous Driving
Organizations: Department of Electrical Engineering, Korea Advanced Institute of Science and Technology
Abstract
End-to-end autonomous driving systems have been shown to suffer from shortcut learning and causal confusion, limiting their ability to generalize robustly under distribution shift. To mitigate these limitations, recent works based on Vision-Language Models (VLMs) have explored incorporating linguistic reasoning capabilities into driving decision-making, reporting improvements in driving performance. In particular, counterfactual (CF) reasoning has been proposed to enable self-reflective, high-level causal understanding. However, it remains unclear whether the language-based reasoning mitigates shortcut learning or merely shifts it from visual representations to linguistic patterns, creating an illusion of causal understanding. Moreover, existing approaches often conflate intervention-based reasoning with CF reasoning, leaving the effects of genuine CF reasoning unexplored. In this work, we examine what CF reasoning actually induces in VLM-based autonomous driving and observe that it shifts the action distribution toward more conservative behaviors while increasing uncertainty in reasoning and decision-making. This effect persists under reinforcement fine-tuning, suggesting that it is not merely a transient artifact. Furthermore, visual attention analysis shows that CF reasoning does not alter visual grounding, while causal intervention experiments reveal that masking causally relevant and non-causal factors in accident scenarios results in comparable decision change rates. Overall, these results suggest that CF reasoning does not induce genuine causal reasoning grounded in visual evidence. Instead, it acts as a linguistic shortcut that reshapes the action distribution, teaching models what to say but not how to reason, highlighting a gap between language-based reasoning and causal understanding in autonomous driving.
Figures & tables
| Model | Metric | Accident Recog. | Ego Inclusion | Decision | ||||
|---|---|---|---|---|---|---|---|---|
| Yes | No | Involv. | Not In. | Stop | Decel. | Keep | ||
| InternVL2-1B | P@1 w/o | 47.5 | 64.4 | 45.8 | 64.4 | 20.4 | 29.2 | 78.0 |
| P@1 w/ | 60.3 | 50.6 | 60.3 | 50.6 | 26.2 | 24.9 | 67.2 | |
| +12.8 | -13.8 | +14.5 | -13.8 | +5.8 | -4.3 | -10.8 | ||
| P@16 w/o | 99.4 | 99.7 | 99.4 | 99.7 | 87.2 | 90.3 | 99.8 | |
| P@16 w/ | 99.9 | 99.9 | 99.9 | 99.9 | 93.3 | 93.6 | 100.0 | |
| Model | Metric | Accident Recog. | Ego Inclusion | Decision | ||||
|---|---|---|---|---|---|---|---|---|
| Yes | No | Involv. | Not In. | Stop | Decel. | Keep | ||
| InternVL2.5-1B (GRPO) | P@1 w/o | 36.6 | 79.9 | 36.0 | 80.4 | 16.9 | 16.7 | 87.2 |
| P@1 w/ | 39.6 | 73.5 | 39.6 | 73.5 | 24.6 | 17.9 | 80.3 | |
| +3.0 | -6.4 | +3.6 | -6.9 | +7.7 | +1.2 | -6.9 | ||
| P@16 w/o | 92.9 | 99.7 | 92.9 | 99.7 | 73.8 | 75.1 | 100.0 | |
| P@16 w/ | 96.3 | 99.9 | 96.3 | 99.9 | 82.7 | 84.4 | 99.8 | |
| Ablation | perception | decision | counterfactual | |||||||||
| Type | CF | RFT | ||||||||||
| InternVL2.5-1B | SFT | ✗ | — | 9.56 | 10.73 | 18.67 | 7.81 | 8.60 | 14.91 | 7.93 | 8.83 | 15.14 |
| SFT | ✓ | — | 9.54 | 10.91 | 19.04 | 8.62 | 9.67 | 16.54 | 8.73 | 9.71 | 16.72 | |
| RFT | ✗ | GRPO | 9.39 | 10.85 | 18.85 | 8.06 | 8.91 | 15.21 | 8.46 | 9.34 | 15.91 | |
| RFT | ✓ | GRPO | 10.39 | 11.46 | 19.74 | 9.22 | 9.79 | 16.66 | 8.70 | 9.39 | 16.00 | |
| Qwen2.5-VL-3B | SFT | ✗ | — | 16.42 | 20.04 | 32.50 | 17.38 | 19.75 | 32.05 | 16.12 | 18.33 | 29.96 |
| InternVL2-1B | InternVL2-2B | InternVL2.5-1B | Qwen2.5-VL-3B | VideoLLaMA2-7B | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SFT | SFT | SFT | RFT (GRPO) | SFT | RFT (GRPO) | RFT (GDPO) | SFT | |||||||||
| ✗ | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | ✓ | ✗ | ✓ | |
| Causal | 0.455 | 0.494 | 0.408 | 0.451 | 0.425 | 0.382 | 0.307 | 0.340 | 0.457 | 0.524 | 0.404 | 0.520 | 0.417 | 0.484 | 0.434 | 0.508 |
| Non-causal | 0.463 | 0.498 | 0.367 | 0.404 | 0.416 | 0.373 | 0.310 | 0.346 | 0.457 | 0.546 | 0.416 | 0.504 | 0.399 | 0.548 | 0.396 | 0.519 |
| Random | 0.475 | 0.508 | 0.385 | 0.416 | 0.452 | 0.363 | 0.295 | 0.355 | 0.458 | 0.519 | 0.411 | 0.524 | 0.388 | 0.531 | 0.414 | 0.513 |
| (m) | Subset | ||||
|---|---|---|---|---|---|
| 0.3m | 5,119 | 8.85% | 3.07% | pp | |
| 2,286 | 17.94% | 5.95% | pp | ||
| 874 | 29.52% | 11.21% | pp | ||
| 1,412 | 10.76% | 2.69% | pp | ||
| 0.5m | 5,119 | 6.58% | 1.93% | pp | |
| 1,676 | 18.85% | 5.55% |
| Metric | Qwen2.5-VL-3B (SFT) | Qwen2.5-VL-3B (GRPO) | Qwen2.5-VL-3B (GDPO) | |||
|---|---|---|---|---|---|---|
| No-CF | CF | No-CF | CF | No-CF | CF | |
| Baseline FPR | 25.00% | 69.00% | 21.58% | 66.96% | 12.35% | 61.01% |
| Adversarial FPR | 33.25% | 72.25% | 31.21% | 63.24% | 22.92% | 60.20% |
| Neutral FPR | 28.25% | 78.75% | 29.65% | 77.46% | 22.21% | 70.32% |
| Hazard Specificity (Adv. - Neu.) | +5.00 pp | 6.50 pp | +1.56 pp | 14.21 pp | +0.71 pp | 10.12 pp |
Appendix figures & tables19 assets
Supplementary material from the paper’s appendix.
Appendix
| Model | w/ CF | w/o CF |
|---|---|---|
| InternVL2.5-1B (GRPO) | 0.365 | 0.499 |
| Qwen2.5-VL-3B (GRPO) | 0.392 | 0.461 |
| Qwen2.5-VL-3B (GDPO) | 0.370 | 0.439 |
| Dataset | Year | Type | Perspect. | #Frames | Acc. Recog. | Desc. | CF QA | Decision |
|---|---|---|---|---|---|---|---|---|
| VIENA 2 [ 55 ] | 2019 | synth | vehicle | 2,250,000 | ✓ | ✗ | ✗ | ✗ |
| GTACrash [ 56 ] | 2019 | synth | vehicle | 751,610 | ✓ | ✗ | ✗ | ✗ |
| Argus [ 57 ] | 2021 | real | roadside | 120,000 | ✓ | ✗ | ✗ | ✗ |
| YoutubeCrash [ 58 ] | 2021 | real | vehicle | 7,720 | ✓ | ✗ | ✗ | ✗ |
| IITH Road [ 59 ] | 2022 | real | roadside | 127,138 | ✓ | ✗ | ✗ | ✗ |
| TAD [ 60 ] | 2022 | real | roadside | 24,810 | ✓ | ✗ | ✗ | ✗ |
| Decision Class | Train | Test | Total |
|---|---|---|---|
| stop | 1,026 (15.4%) | 336 (25.0%) | 1,362 |
| decelerate | 1,423 (21.4%) | 336 (25.0%) | 1,759 |
| keep | 4,212 (63.2%) | 672 (50.0%) | 4,884 |
| Total | 6,661 | 1,344 | 8,005 |
| Category | Count | Percentage (%) |
|---|---|---|
| Correct | 77 | 77.0% |
| Minor Errors (e.g., phrasing) | 9 | 9.0% |
| Major Errors (e.g., logic flaws) | 8 | 8.0% |
| Hallucinations (e.g., phantom objects) | 6 | 6.0% |
| Usable (Correct + Minor) | 86 | 86.0% |
| Model | Metric | Accident Recog. | Ego Inclusion | Decision | ||||
|---|---|---|---|---|---|---|---|---|
| Yes | No | Involv. | Not In. | Stop | Decel. | Keep | ||
| InternVL2-1B | P@1 w/o | 69.4 | 58.0 | 58.0 | 69.4 | 33.3 | 32.1 | 71.4 |
| P@1 w/ | 66.7 | 30.0 | 66.7 | 30.0 | 22.2 | 32.1 | 59.2 | |
| -2.7 | -28.0 | +8.7 | -39.4 | -11.1 | 0.0 | -12.2 | ||
| InternVL2-2B | P@1 w/o | 41.7 | 66.6 | 41.7 | 66.6 | 11.1 | 25.0 | 83.7 |
| P@1 w/ | 50.0 | 56.0 | 50.0 | 56.0 | 22.2 | 32.1 | 83.7 | |
| Metric | Accident Recog. | Ego Inclusion | Decision | ||||
|---|---|---|---|---|---|---|---|
| Yes | No | Involv. | Not In. | Stop | Decel. | Keep | |
| P@1 w/o CF | 50.6 | 57.3 | 46.7 | 60.7 | 19.3 | 32.3 | 65.8 |
| P@1 w/ CF | 64.2 | 44.8 | 61.5 | 46.0 | 37.0 | 28.5 | 53.2 |
| +13.6 | -12.5 | +14.8 | -14.7 | +17.7 | -3.8 | -12.6 | |
| P@16 w/o CF | 95.7 | 93.9 | 95.7 | 97.9 | 69.5 | 91.8 | 96.9 |
| P@16 w/ CF | 99.7 | 97.6 | 99.7 | 98.7 | 94.5 | 92.6 | 98.3 |
| Category | Rationale for Selection | Related Works |
|---|---|---|
| LLaMA & Mistral | Standard baseline in early VLA research. We adopt VideoLLaMA2-7B to represent this lineage. | DriveMLM [ 67 ] , RDA-Driver [ 68 ] , DriveLM [ 1 ] , DriveGPT4 [ 2 ] , VLAAD [ 69 ] , OccLLaMA [ 70 ] , OmniDrive [ 5 ] , LMDrive [ 71 ] , BEVDriver [ 72 ] , CoVLA-Agent [ 73 ] , DSDrive [ 74 ] |
| Qwen Family | Dominant in 2025 research. We adopt Qwen2.5-VL-3B to align with SOTA trends. | ImpromptuVLA [ 75 ] , DriveAgent-R1 [ 13 ] , FastDriveVLA [ 76 ] , AutoDrive-R 2 [ 28 ] , CoReVLA [ 77 ] , AutoVLA [ 27 ] , E3AD [ 78 ] , UniUGP [ 79 ] , Reasoning-VLA [ 80 ] |
| InternVL Family | Vision-Centric alternative. We adopt InternVL2 to verify scalability. | Drive-R1 [ 81 ] , SimLingo [ 10 ] , VDRive [ 82 ] , DriVLM [ 83 ] , Li et al. [ 84 ] |
| Setting | Model | Hyperparameter | Value |
| SFT | InternVL2-1B | Batch size | 4 |
| Precision | bf16 | ||
| Epoch | 5 | ||
| Learning rate | |||
| Weight decay | 0.05 | ||
| # of frames | 8 |
VLMs for Autonomous Driving
Vision-language models for autonomous driving connect camera and sensor data with language to interpret scenes, predict hazards, and plan actions.
See all 226 papers