Research map / Label-efficient and robust learning / Knowledge distillation
Uncertainty-Aware Knowledge Distillation for Multimodal Large Language Models
Jingchen Sun, Shaobo Han, Deep Patel, Wataru Kohno, Can Jin, Changyou Chen
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 task-level Beta-KD weighting over manual weights (ablation) | CE + FKL · ScienceQA test | VQA accuracy: 48.6 -> 49.3 | Improved |
| add instance-level Beta-KD weighting (ablation) | CE + FKL · ScienceQA test | VQA accuracy: 48.2 -> 51.8 | Improved |
| instance-level Beta-KD on three-loss objectives (ablation) | CE + KD + FD objectives · ScienceQA test | VQA accuracy: +5.2 (FKL+FD) to -2.8 (TVD+FD) | Mixed |
| task-level Beta-KD on RKL+FD (ablation) | CE + RKL + FD · ScienceQA test | VQA accuracy: 49.8 -> 49.5 | No clear effect |
| use instance-level rather than task-level weighting (ablation) | Beta-KD · ScienceQA test | VQA accuracy: e.g. 49.3 -> 51.8 (FKL); 53.1 -> 51.0 (SFKL) | Mixed |
| match teacher in probability space with cosine distance instead of logit space (ablation) | CE distillation · ScienceQA test | VQA accuracy: 4.0 -> 47.2 | Improved |
| use Cosine-Probs instead of FKL as distillation loss (ablation) | CE distillation · ScienceQA test | VQA accuracy: 45.7 -> 47.2 | Improved |
| add Beta-KD (Instance) to Align-KD | Align-KD · 6-benchmark average | average: 63.4 -> 64.6 | Improved |
| add Beta-KD (Instance) to Cosine KD | Cosine KD · 6-benchmark average | average: 63.4 -> 65.5 | Improved |
| add Beta-KD (Instance) to LLaVA-KD | LLaVA-KD Qwen2.5-0.5B · TextVQA | accuracy: 52.0 -> 54.9 | Improved |
| add Beta-KD on POPE | Align-KD · POPE | accuracy: 85.4 -> 85.4 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TAID — TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Distilling What Matters: Confidence-Aware Selective Distillation for Large Language Models (NeurIPS 2026 · similarity 0.92)
- Bayesian Knowledge Distillation: A Bayesian Perspective of Distillation with Uncertainty Quantification (ICML 2024 · similarity 0.90)
- ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence (ICML 2025 · similarity 0.90)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.89)
- DistiLLM: Towards Streamlined Distillation for Large Language Models (ICML 2024 · similarity 0.88)
- PLD: A Choice-Theoretic List-Wise Knowledge Distillation (NeurIPS 2025 · 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.
