Research map / Probabilistic and scientific modeling / Molecules and proteins
Coverage ≠ Exposure: Auditable Control of Same-Support Tail Failures under Multimodal Missingness
Ziteng Hong, Bingzhi Chen, Yishu Liu, Sudong Cai, Guangming Lu
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 closed-loop TailPressure control (H-CES) over static schedule | Static-TP · six vision-language routes | Worst-Head / Shift-Reweight: Worst-Head +2.1 to 3.2 points; Shift-Reweight +1.4 to 2.5 points | Improved |
| add closed-loop TailPressure control (H-CES) over static schedule | Static-TP · sentiment routes (MulT) | Worst-Head / Shift-Reweight: Worst-Head gains of 2.5 to 2.6 points; Shift-Reweight 2.2 to 2.7 points | Improved |
| use H-CES instead of LoRA with global weight decay | LoRA · MMBench-EN, LLaVA-NeXT | Worst-Head: 45.9 -> 53.1 | Improved |
| add H-CES | Full score · vision-language routes | Full: Full changes remain within 0.2 points relative to Static-TP | No clear effect |
| add closed-loop control (H-CES vs Static-TP), five seeds | Static-TP · MMBench-EN | Worst-Head / Shift-Reweight: 2.5 and 2.2 points | Improved |
| add closed-loop control (H-CES vs Static-TP), five seeds | Static-TP · MMMU | Worst-Head / Shift-Reweight: 2.1 and 1.6 points | Improved |
| add train-time gating plus eval-time gating together (full H-CES) (ablation) | H-CES · MMBench-EN | Worst-Head: 49.1 (train-gated only) vs 53.1; 46.7 (eval-gated only) vs 53.1 | Improved |
| keep sample-wise alignment of gates (vs shuffled gates) (ablation) | H-CES · MMBench-EN | Worst-Head: 46.2 (shuffle train) and 46.0 (shuffle eval) vs 53.1 | Improved |
| use correct controller-to-group mapping (vs permuted mapping) (ablation) | H-CES · MMBench-EN | Worst-Head: 46.9 -> 53.1 | Improved |
| use contract-defined groups (vs random grouping of same sizes) (ablation) | H-CES · MMBench-EN | Worst-Head: 48.2 -> 53.1 | Improved |
| use per-group TailPressure control (vs matched global WD schedule or uniform WD budget) (ablation) | H-CES · MMBench-EN | Worst-Head: 47.6 and 47.9 vs 53.1 | Improved |
| use weight-decay actuator (vs per-group LR scaling) (ablation) | H-CES · MMBench-EN | Worst-Head: 47.4 -> 53.1 | Improved |
Closest papers by meaning
- Calibration Is Not Control: Intervention Advantage for LLM-Agent Oversight (NeurIPS 2026 · similarity 0.83)
- Multi-Agent Coordination via Support-Preserving Distillation (NeurIPS 2026 · similarity 0.82)
- $h$-control: Training-Free Camera Control via Block-Conditional Gibbs Refinement (NeurIPS 2026 · similarity 0.81)
- LiSA: Lifelong Safety Adaptation via Conservative Policy Induction (NeurIPS 2026 · similarity 0.81)
- Combining Cost Constrained Runtime Monitors for AI Safety (NeurIPS 2025 · similarity 0.81)
- DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising (NeurIPS 2026 · similarity 0.81)
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.
