Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Learning to Learn with Contrastive Meta-Objective
Authors: Shiguang Wu, Yaqing Wang, Yatao Bian, Quanming Yao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We provide empirical studies to understand the effect of Con ML on synthetic data, which shows that learning to learn with Con ML brings generalizable alignment and discrimination abilities. Code is avaliable at https://github.com/LARS-research/Con ML. 4.1 Few-Shot Image Classification Performance To show Con ML brings learner-agnostic improvement, we integrate Con ML into various meta-learners and evaluate the meta-learning performance on few-shot image classification problem follow existing works [46, 15, 4]. We use two few-shot image classification benchmarks: mini Image Net [46] and tiered Image Net [35], evaluating on 5-way 1-shot and 5-way 5-shot tasks. |
| Researcher Affiliation | Academia | Shiguang Wu1 , Yaqing Wang2 , Yatao Bian3, Quanming Yao1,4 1Department of Electronic Engineering, Tsinghua University 2Beijing Institute of Mathematical Sciences and Applications 3 Department of Computer Science, National University of Singapore 4State Key laboratory of Space Network and Communications, Tsinghua University |
| Pseudocode | Yes | Algorithm 1 Mini-Batch Episodic Training (with Validation Loss). while Not converged do Sample a batch of tasks b p B(τ). for All τ b do Get task-specific model hτ = g(Dtr τ; θ); Get validation loss L(Dval τ ; hτ); end for Get episodic loss Le = 1 B P τ b L(Dval τ ; g(Dtr τ; θ)); Update θ by θ θ θLe. end while |
| Open Source Code | Yes | Code is avaliable at https://github.com/LARS-research/Con ML. |
| Open Datasets | Yes | We use two few-shot image classification benchmarks: mini Image Net [46] and tiered Image Net [35], evaluating on 5-way 1-shot and 5-way 5-shot tasks. |
| Dataset Splits | Yes | Splitting each Dτ into a training set Dtr τ = {(xτ,i, yτ,i)}n i=1 and a validation set Dval τ = {(xτ,i, yτ,i)}m i=n+1, the meta-training objective is minimizing Eτ p(τ)L(Dval τ ; g(Dtr τ; θ)). |
| Hardware Specification | No | Table 4 has shown the detailed relative consumption comparing the proposed method and standard method. |
| Software Dependencies | No | We train the GPT-2 [34]-like transformer for each function with ICL and ICL w/ Con ML respectively and compare the inference (meta-testing) performance. |
| Experiment Setup | Yes | We uses a common configuration for Con ML s hyperparameter for all meta-learners: task batch size B = 32, inner-task sampling K = 1, and πκ(Dtr τ Dval τ ) = Dtr τ, ϕ(a, b) = 1 a b/ a b and λ = 0.1. |