Knowledge & Dataset Distillation
Knowledge Distillation Under Strict Black-Box Teacher Preconditions
Generated automatically from the limitations stated in 4 papers (ICLR, NeurIPS, EMNLP), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
Distillation from closed-source, proprietary API-hosted models into local, highly compressed student models without requiring access to weights, hidden states, or full output distributions.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Benchmark existing white-box, logit-based, and output-only distillation techniques across standard vision and language tasks under progressively restricted teacher access (full states $\rightarrow$ logits $\rightarrow$ top-k logits $\rightarrow$ hard labels/text), measuring downstream student task accuracy.
- 2
Develop a surrogate-guidance framework where a local, lightweight proxy model estimates pseudo-logits and synthetic intermediate targets from black-box teacher sample outputs, measuring distillation fidelity against true white-box baselines.
- 3
Formulate an active query-selection strategy that optimizes which inputs to submit to the black-box teacher API to maximize student learning efficiency under fixed query budgets, measured by student performance per API call.
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
If prompting or simple supervised fine-tuning on raw black-box teacher outputs already saturates student capacity limits, rendering specialized black-box distillation formulations redundant.
Evidence
Each paper's own statement of the limitation, verbatim.
- Taught Well Learned Ill: Towards Distillation-conditional Backdoor AttackNeurIPS 2025
Requires white-box control over the teacher model's training process (data, loss functions, and hyperparameters).
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language ModelsICLR 2025
Requires an accessible teacher's logits at the same inputs during training, so it is inapplicable when only teacher outputs or a black-box teacher API is available
- EMO: Embedding Model Distillation via Intra-Model Relation and Optimal Transport AlignmentsEMNLP 2025
Requires an additional learnable projection layer to bridge dimensionality gaps between teacher and student states.
- A Good Learner can Teach Better: Teacher-Student Collaborative Knowledge DistillationICLR 2024
Requires white-box access to both teacher and student logits and hidden representations, excluding black-box/decoder-only API teachers
Nearest existing work
- Multi-Level Logit DistillationCVPR 2023
- Does Knowledge Distillation Really Work?NeurIPS 2021
- On the Efficacy of Knowledge DistillationICCV 2019
- Knowledge Distillation with Refined LogitsICCV 2025
- Knowledge Distillation Detection for Open-weights ModelsNeurIPS 2025
- Knowledge Distillation for Large Language Models through Residual LearningICLR 2026
- What Makes a Good Dataset for Knowledge Distillation?CVPR 2025
- Knowledge Distillation With the Reused Teacher ClassifierCVPR 2022
- What Knowledge Gets Distilled in Knowledge Distillation?NeurIPS 2023
- Analyzing the Confidentiality of Undistillable Teachers in Knowledge DistillationNeurIPS 2021
- Tailoring Instructions to Student’s Learning Levels Boosts Knowledge DistillationACL 2023
- Shadow Knowledge Distillation: Bridging Offline and Online Knowledge TransferNeurIPS 2022
- Scaled Decoupled DistillationCVPR 2024
- Knowledge Distillation with Auxiliary VariableICML 2024
- Gap Preserving Distillation by Building Bidirectional Mappings with A Dynamic TeacherICLR 2025
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
Distillation Under Zero-Data and Unlabeled-Stream Preconditions
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.
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.