Research map / Optimization / Optimizers
The Role of Learning Algorithms in Collective Action
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that the learning algorithm (DRO re-weighting, JTT/LfF, iterative DRO with validation control, and SGD's simplicity bias) substantially changes the success of algorithmic collective action, extending Bayes-optimal bounds with effective collective size and surrogate-distribution bounds.
Problem
Prior theory on algorithmic collective action assumes Bayes (sub-)optimal classifiers, ignoring how the choice of learning algorithm (distributionally robust optimization, simplicity-biased SGD) affects whether a collective can plant a signal successfully.
Approach
Defines an effective collective size under sample re-weighting, derives success bounds for two-stage and iterative DRO re-weighting (including validation-set control), and proposes algorithm-aware signal design that exploits simplicity bias by planting signals in simple features or by using surrogate distributions that discard complex features.
Key results
- For small collective size (around 0.1-0.2), JTT and LfF give the collective higher success than ERM on the 2D dataset and CIFAR-10; this reverses for larger collectives.
- With DRO on Waterbirds (alpha=0.3), success oscillates sharply across epochs and depends heavily on the collective fraction in the validation set, whereas ERM is nearly unaffected by it.
- On k-strips, collective success increases with the number of strips k, achieving higher success with a smaller collective size.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use JTT or LfF instead of ERM for a small collective | Collective signal planting · Synthetic 2D and CIFAR-10 | collective success | Helped |
| Use JTT or LfF instead of ERM for a large collective | Collective signal planting · Synthetic 2D and CIFAR-10 | collective success | Hurt |
| Increase the collective proportion in the validation set | DRO · CIFAR-10 and Waterbirds | collective success | Helped |
| Use maximum validation accuracy with a collective-free validation set for stopping | DRO · Waterbirds | collective success: success is 0 | Hurt |
| Increase the collective proportion in the validation set rather than the training set | DRO · CIFAR-10 and Waterbirds | collective success | Helped |
| Increase training-set collective proportion rather than validation-set proportion | ERM · CIFAR-10 and Waterbirds | collective success | Helped |
| Reduce training-set collective proportion from 0.1 to 0.001, even with full validation-set control | ERM · CIFAR-10 and Waterbirds | collective success: 0.1: almost full success; 0.001: does not go over 0.4 | Hurt |
| Increase the complexity of the x2 feature by increasing k | SGD · LMS-k | collective success | Helped |
| Decrease MNIST-label correlation | SGD · MNIST-CIFAR | collective success | Helped |
| Increase the number of horizontal strips and plant the signal in the simpler x1 feature | SGD with early stopping · k-strips | collective success | Helped |
Surprising findings
- Maximum validation accuracy as a stopping criterion for DRO yields zero collective success when the validation set contains no collective signal.
- Under a constructed example, the surrogate-distribution bound gives a success guarantee of about 1 - 9/N, exceeding the Bayes-optimal bound of 0.1.
Limitations
- Focuses only on the feature-label strategy for planting signals; other goals such as erasing signals are left for future work.
- Theoretical bounds are for idealized algorithms (e.g., ideal continuous reweighting) rather than practical implementations.
- Experiments are on small synthetic datasets and CIFAR-10/Waterbirds; the paper notes results depend on the algorithm and may not transfer to other algorithms.
Experimental setup
Datasets: CIFAR-10 Waterbirds MNIST-CIFAR LMS-k k-strips synthetic 2D uniform dataset
Benchmarks: CIFAR-10 Waterbirds synthetic 2D dataset MNIST-CIFAR LMS-k k-strips
Metrics: collective success
Baselines: ERM JTT LfF DRO
Builds on: algorithmic collective action (Hu et al. 2023) Just Train Twice (JTT) Learning from Failure (LfF) Group DRO / CVaR-DRO LMS-k simplicity bias dataset
Closest papers by meaning
- A Statistical Framework for Algorithmic Collective Action with Multiple Collectives (NeurIPS 2026 · similarity 0.90)
- Statistical Collusion by Collectives on Learning Platforms (ICML 2025 · similarity 0.86)
- Look-Ahead Reasoning on Learning Platforms (NeurIPS 2025 · similarity 0.83)
- Manipulation-Robust Selection of Citizens’ Assemblies (AAAI 2024 · similarity 0.82)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.82)
- MixMax: Distributional Robustness in Function Space via Optimal Data Mixtures (ICLR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.