Classical Supervised Learning: Trees, Ranking & Aggregation
Empirical Robustness and Scaling of Jacobian- and Pairwise-Based Aggregation Beyond BERT-Base
Generated automatically from the limitations stated in 4 papers (ICML, AISTATS, ACL), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Several classical supervised learning, ranking, and aggregation methods rely on per-example parameter Jacobians and all-pairs batch formulations, restricting their empirical verification to the linear readout layers of English BERT-base and ResNet-50. As a result, it is unknown whether the spectral variance assumptions and downstream performance of these algorithms hold on larger transformer architectures, decoder-only models, or multilingual representations. Practitioners currently have no evidence indicating whether these techniques fail gracefully or break down entirely when scaled beyond small, English-specific encoder backbones.
Why it matters
Establishes the empirical boundaries of Jacobian- and pairwise-based aggregation methods, determining whether these classical learning formulations can be reliably applied to modern transformer scales and non-English domains.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Conduct a cross-architecture benchmark of existing Jacobian- and pairwise-aggregation algorithms across larger transformer models (e.g., RoBERTa-Large, DeBERTa-v3, and small decoder-only LLMs) and multilingual backbones (e.g., XLM-RoBERTa); measure memory footprint, wall-clock scaling, and downstream ranking performance.
- 2
Evaluate the trade-offs of restricted layer Jacobians versus parameter-efficient Jacobian approximations (such as low-rank projection or adapter-based Jacobians); measure approximation error and downstream aggregation accuracy.
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 computing per-example parameter Jacobians proves fundamentally prohibitive on larger backbones without introducing approximation errors that render the aggregation method ineffective, or if frozen embedding baselines consistently match the method across all scale regimes.
Evidence
Each paper's own statement of the limitation, verbatim.
- Active Learning with Low-Rank Structure for Data SelectionICML 2026
Relies on pre-trained encoder embeddings (e.g., BERT) to capture the dominant spectral variance of the downstream task.
- A Bayesian Information-Theoretic Approach to Data AttributionAISTATS 2026
Requires computing per-example parameter Jacobians, which necessitates layer restrictions (e.g., linear layers only) on larger models like BERT.
- Understanding the Training Speedup from Sampling with Approximate LossesICML 2024
Evaluated only on BERT-base and a modified ResNet-50; not tested on larger transformer models
- Aggregating Multiple Heuristic Signals as Supervision for Unsupervised Automated Essay ScoringACL 2023
Requires ~25GB GPU memory for training (due to all-pairs within batch of 32); smaller batches trade off longer training; also English-only due to English-pretrained BERT
Nearest existing work
- Are Neural Rankers still Outperformed by Gradient Boosted Decision Trees?ICLR 2021
- What are the best Systems? New Perspectives on NLP BenchmarkingNeurIPS 2022
- LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative FusionACL 2023
- Better than Average: Paired Evaluation of NLP systemsACL 2021
- Erwin: A Tree-based Hierarchical Transformer for Large-scale Physical SystemsICML 2025
- Determine-Then-Ensemble: Necessity of Top-k Union for Large Language Model EnsemblingICLR 2025
- Applying BERT to Document Retrieval with BirchEMNLP 2019
- Rankformer: A Graph Transformer for Recommendation based on Ranking ObjectiveTheWebConference 2025
- Modularized Transfomer-based Ranking FrameworkEMNLP 2020
- Generalized Robust Bayesian Committee Machine for Large-scale Gaussian Process RegressionICML 2018
- Estimating the Hessian Matrix of Ranking Objectives for Stochastic Learning to Rank with Gradient Boosted TreesSIGIR 2024
- Precise Zero-Shot Pointwise Ranking with LLMs through Post-Aggregated Global Context InformationSIGIR 2025
- Learning-to-Rank with Partitioned Preference: Fast Estimation for the Plackett-Luce ModelAISTATS 2021
- MosaicBERT: A Bidirectional Encoder Optimized for Fast PretrainingNeurIPS 2023
- Recursive Tree-Structured Self-Attention for Answer Sentence SelectionACL 2021