Research map / Continual learning / Continual fine-tuning of LLMs
The Blind Spot of Adaptation: Quantifying and Mitigating Forgetting in Fine-tuned Driving Models
Runhao Mao, Hanshi Wang, Yixiang Yang, Qianli Ma, Jingmeng Zhou, Zhipeng Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Prompt Adapter on top of TAEM (ablation) | DEA on Qwen2.5VL-3B · Fidelity Driving Bench | KRR: 74.4% -> 79.0% | Improved |
| add Prompt Adapter on top of TAEM (ablation) | DEA on Qwen2.5VL-3B · Fidelity Driving Bench | Traffic-QA: 35.0 -> 41.0 | Improved |
| use TAEM mixture of LoRA experts instead of single LoRA (ablation) | Qwen2.5VL-3B · Fidelity Driving Bench | KRR: 64.6% -> 74.4% | Improved |
| use TAEM instead of single LoRA (ablation) | Qwen2.5VL-3B · Fidelity Driving Bench | Traffic-QA: 22.3 -> 35.0 | Improved |
| apply DEA to ImpromptuVLA-3B | ImpromptuVLA-3B · Fidelity Driving Bench | KRR: 68.4% -> 76.0% | Improved |
| DEA versus full fine-tuning | Qwen2.5VL-3B · Fidelity Driving Bench | KRR: 64.3% -> 79.0% | Improved |
| DEA versus base model on NoPR | Qwen2.5VL-3B · Fidelity Driving Bench | NoPR: 36.8 -> 29.0 | Hurt |
| LoRA fine-tuning on diverse driving corpus | six base VLMs · Fidelity Driving Bench | NoPR: InternVL3-2B 50.8 -> 30.8; InternVL2-4B 37.1 -> 16.0 | Hurt |
| LoRA fine-tuning on Traffic-QA | six base VLMs · Fidelity Driving Bench | Traffic-QA: LLAVA-OV-7B 35.9 -> 47.0; InternVL3-2B 31.5 -> 25.1 | Mixed |
| multi-source rather than single-source fine-tuning data | VLM driving models · Fidelity Driving Bench | KRR: RecogDrive-8B 74% vs safeAuto-BDDX 31.5% | Improved |
| full fine-tuning on driving data | InternVL3-8B (RecogDrive-8B) · Fidelity Driving Bench | NoPR: 55.2 -> 41.3 | Hurt |
Closest papers by meaning
- SGDrive: Scene-to-Goal Hierarchical World Cognition for Autonomous Driving (CVPR 2026 · similarity 0.89)
- DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous Driving (NeurIPS 2026 · similarity 0.89)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.89)
- DriveHierarchy: A Benchmark for Diagnosing VLM Driving Capabilities from Open-Loop Understanding to Closed-Loop Execution (NeurIPS 2026 · similarity 0.88)
- Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.88)
- Continuously Learning, Adapting, and Improving: A Dual-Process Approach to Autonomous Driving (NeurIPS 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
