Knowledge & Dataset Distillation
Knowledge Distillation and Rollout Transfer from Black-Box Teachers Without Logit Access
Generated automatically from the limitations stated in 3 papers (ICML, ACL), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
Enables distilling state-of-the-art capabilities and uncertainty calibration from closed-source commercial APIs and black-box models into compact student architectures without internal weight or logit access.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Non-parametric uncertainty estimation via perturbation queries: Estimate teacher output entropy and top-$k$ rank statistics from empirical output frequencies across stochastic input perturbations, measuring distillation fidelity against white-box logit baselines on standard classification and text benchmarks.
- 2
Logit-free importance weighting for rollouts: Design sample-weighting heuristics based on student disagreement and rejection sampling over teacher-generated sequences rather than explicit teacher logit entropy, measuring student downstream accuracy and rollout memory overhead.
- 3
Distribution-free decision boundary alignment: Formulate an optimization objective that aligns student predictive distributions to teacher hard decisions using ranking loss constraints, measuring performance on black-box vision and language distillation tasks.
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
The query complexity required to reliably estimate teacher uncertainty from discrete outputs could become cost-prohibitive compared to standard supervised fine-tuning on raw teacher outputs.
Evidence
Each paper's own statement of the limitation, verbatim.
- Hybrid Policy Distillation for LLMsICML 2026
Requires white-box logit/log-probability access to the teacher model, making it inapplicable to black-box or API-only teacher models.
- Entropy-Aware On-Policy Distillation of Language ModelsICML 2026
Requires white-box logit access to the teacher model to evaluate conditional entropy and top-k distribution, precluding black-box API teachers.
Requires additional compute and memory during rollout processing to store and compute teacher entropy and top-k logit distributions.
- Bridging the Gap between Decision and Logits in Decision-based Knowledge Distillation for Pre-trained Language ModelsACL 2023
Relies on the assumption that teacher logits on augmented data follow a Gaussian distribution, which may not hold in practice
Nearest existing work
- Bridging the Gap between Decision and Logits in Decision-based Knowledge Distillation for Pre-trained Language ModelsACL 2023
- Multi-Level Logit DistillationCVPR 2023
- Knowledge Distillation with Auxiliary VariableICML 2024
- Knowledge Distillation with Refined LogitsICCV 2025
- Single Teacher, Multiple Perspectives: Teacher Knowledge Augmentation for Enhanced Knowledge DistillationICLR 2025
- Logit Standardization in Knowledge DistillationCVPR 2024
- Speculative Knowledge Distillation: Bridging the Teacher-Student Gap Through Interleaved SamplingICLR 2025
- Scaled Decoupled DistillationCVPR 2024
- A statistical perspective on distillationICML 2021
- EA-KD: Entropy-based Adaptive Knowledge DistillationICCV 2025
- Uncertainty-Aware Knowledge Distillation for Multimodal Large Language ModelsCVPR 2026
- Does Knowledge Distillation Really Work?NeurIPS 2021
- Entropy-Aware On-Policy Distillation of Language ModelsICML 2026
- Knowledge Distillation for Large Language Models through Residual LearningICLR 2026
- How to Train the Teacher Model for Effective Knowledge DistillationECCV 2024
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
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.