UBD-Bench: Benchmarking and Analyzing Unimodal Branch Degradation in Multimodal Perception for Autonomous Driving
We introduce UBD-Bench, a systematic benchmark for studying unimodal branch degradation (UBD) in multimodal perception.
ProofPaper ↗
Key points
- Multimodal perception improves autonomous-driving performance by combining complementary sensor information.
- UBD-Bench uses a Paired Frozen-Branch Probe to compare the task-relevant representation quality of the same modality after independent unimodal training and multimodal joint training under matched decoding conditions.
- Across four autonomous-driving perception tasks and 10 models, we find that Camera branches consistently degrade after multimodal joint training, whereas LiDAR branches exhibit weaker and more model-dependent changes, revealing a pronounced modality asymmetry.
- Guided by this diagnosis, we introduce unimodal-to-multimodal distillation at the degraded representation.
Sources (1)
- [1]UBD-Bench: Benchmarking and Analyzing Unimodal Branch Degradation in Multimodal Perception for Autonomous DrivingarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 06:11 PM
We introduce UBD-Bench, a systematic benchmark for studying unimodal branch degradation (UBD) in multimodal perception.
Multimodal perception improves autonomous-driving performance by combining complementary sensor information.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 18, 2026Google Open-Sources Harness2, a Recursive Agent Harnessing Framework
- Oct 18, 2026Policy Is Not Capability: Some Safety Benchmarks Measure Which Definition of "Unsafe" You Trained On
- Oct 18, 2026MT Forcing: Towards Stable and Consistent Multi-Turn Image Editing with Autoregressive Diffusion
- Oct 17, 2026Nemotron 3.5 Bolt, Muse Spark Open, and a Fast LFM2.5 24B
- Oct 16, 2026Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning
- Oct 16, 2026OfficeTown: Generative Organizations for Evolving Agent Intelligence