C³L anchors saliency to semantic concepts, breaking explanation drift in continual learning
kbag01 · hf · 2026-10-28
Tel Aviv University researchers identify a structural flaw in explanation-aware continual learning: pixel-level saliency regularization derives supervision from the model's own frozen saliency maps, creating a self-referential drift loop that reinforces spurious attention. Their C³L method anchors regularization to offline-discovered semantic concepts, suppressing spurious correlations via a relevance crowding-out effect.
More from Research
- VybBench: open benchmark scores image models' design fidelity with numbers, not votes — thetripathi58 · 2026-10-28
- Physical AI exec predicts dexterous manipulation breakthrough in 2-3 quarters as robots self-code — BradPorter_ · 2026-10-28
- NVIDIA open-sources Flora, a model suite that auto-labels robot videos for training GR00T — qinzytech · 2026-10-28
- Multimodal GraphRAG models missing and contradictory evidence for storm damage assessment — JeremyCMorgan · 2026-10-28
- SIMIT: VLMs self-improve at test time by imagining practice problems, +7.2% on 17 benchmarks — delliott · 2026-10-28
- Mistral sponsors Open Math OS with zero data retention, powering a Kimi K3/DeepSeek math platform — GuillaumeLample · 2026-10-28