Knowledge & Dataset Distillation
Distillation Under Zero-Data and Unlabeled-Stream Preconditions
Generated automatically from the limitations stated in 4 papers (ICLR, ECCV, ICCV), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Current distillation frameworks depend on strict data preconditions, including access to original downstream training sets, in-domain unlabeled datasets, out-of-distribution reference data, or sample-level identity annotations. When proprietary, privacy, or security constraints prevent the release of both the underlying training set and external proxy data, these methods cannot execute at all. Furthermore, when teacher outputs arrive as an uncurated stream without identity labels, existing feature-bank and prototype-matching methods break down. Resolving these preconditions allows model compression and dataset synthesis in strictly zero-data or privacy-restricted deployments.
Why it matters
Knowledge transfer and student model compression become possible in privacy-critical domains where raw datasets, domain-matched surrogate data, and sample identity labels cannot be shared or stored.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Inverted-generator teacher distillation: Train a generator to synthesize inputs directly from the teacher's activation statistics and logits without reference data, measuring downstream classification accuracy and synthetic sample fidelity across standard vision benchmarks.
- 2
Label-free stream distillation: Formulate teacher-to-student feature transfer using contrastive clustering or optimal transport over streaming mini-batches rather than indexed identity banks, measuring performance degradation relative to full-supervision baselines.
- 3
Sensitivity analysis across missing data assumptions: Benchmark existing data-free, surrogate-data, and stream-based distillation techniques across varying degrees of proxy data mismatch to quantify the performance cliff when data preconditions are violated.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
Generative data-free inversion methods may suffer severe mode collapse on large-scale label spaces, failing to recover sufficient feature diversity to match even simple baseline models trained on small public proxies.
Evidence
Each paper's own statement of the limitation, verbatim.
- DisTaC: Conditioning Task Vectors via Distillation for Robust Model MergingICLR 2026
Requires access to unlabeled task-specific data for distillation, which may be restricted by privacy or security constraints.
- Adapt without Forgetting: Distill Proximity from Dual Teachers in Vision-Language ModelsECCV 2024
Requires reference/wild data for distillation; performance depends on this external data distribution, which is not characterized
- ICD-Face: Intra-class Compactness Distillation for Face RecognitionICCV 2023
Requires identity labels at distillation time (feature banks are indexed per identity), so it cannot distill from an unlabeled teacher-output stream or when training identities are unknown
- Bridging the Gap between Decision and Logits in Decision-based Knowledge Distillation for Pre-trained Language ModelsACL 2023
Requires access to the downstream task training dataset; cannot distill when no training data is publicly available
Nearest existing work
- Zero-Shot Knowledge Distillation in Deep NetworksICML 2019
- Knowledge Distillation Detection for Open-weights ModelsNeurIPS 2025
- What Makes a Good Dataset for Knowledge Distillation?CVPR 2025
- Zero-shot Knowledge Transfer via Adversarial Belief MatchingNeurIPS 2019
- Flow Map Distillation Without DataCVPR 2026
- Learning Student Networks in the WildCVPR 2021
- Coupling the Generator with Teacher for Effective Data-Free Knowledge DistillationICCV 2025
- Analyzing the Confidentiality of Undistillable Teachers in Knowledge DistillationNeurIPS 2021
- Learnability-Guided Diffusion for Dataset DistillationCVPR 2026
- Weighted Distillation with Unlabeled ExamplesNeurIPS 2022
- Discovering and Overcoming Limitations of Noise-engineered Data-free Knowledge DistillationNeurIPS 2022
- Shadow Knowledge Distillation: Bridging Offline and Online Knowledge TransferNeurIPS 2022
- Continual Distillation of Teachers from Different DomainsCVPR 2026
- Random Conditioning for Diffusion Model Compression with DistillationCVPR 2025
- Data-Free Knowledge Distillation for Image Super-ResolutionCVPR 2021
Related open problems
Knowledge & Dataset Distillation
Cross-Modal 3D Distillation Without Paired and Synchronized LiDAR Streams
Existing cross-modal distillation frameworks for 3D perception strictly require synchronized, spatially calibrated LiDAR point clouds and pre-trained LiDAR teachers alongside camera or radar feeds during training. This strict precondition limits distillation to expensive research vehicles with high-end sensor rigs, preventing its application to the vast majority of production fleet datasets where LiDAR is absent, uncalibrated, or asynchronously logged. As a result, low-cost sensor suites cannot leverage pre-trained LiDAR representations unless rigid multimodal collection setups are maintained during training.
Knowledge & Dataset Distillation
Knowledge Distillation Under Strict Black-Box Teacher Preconditions
Current high-performance distillation techniques rely on white-box access to intermediate representations, full logit vectors, or direct control over the teacher's training dynamics. When the teacher is a proprietary API or closed-source system that only returns discrete text or top-1 predictions, these methods cannot run. Downstream practitioners are structurally blocked from transferring capabilities from frontier closed models into compact architectures using state-of-the-art distillation losses.
Knowledge & Dataset Distillation
Knowledge Distillation and Rollout Transfer from Black-Box Teachers Without Logit Access
High-performing knowledge distillation and policy transfer techniques rely on white-box teacher access to evaluate conditional entropy, top-$k$ predictive distributions, and log-probabilities during rollouts. This requirement structurally prevents distillation from state-of-the-art proprietary or API-only models that output only discrete decisions or text tokens without internal logits. Furthermore, existing workarounds often rely on fragile parametric assumptions, such as Gaussianity of teacher logits under data augmentation, which fail on complex, multi-modal output distributions.