Research map / Learning theory and algorithms / Prediction and decision losses

A Statistical Framework for Algorithmic Collective Action with Multiple Collectives

NeurIPS 2026 · Learning theory and algorithms › Prediction and decision losses · approach: Adaptive testing and experiments · 0 citations (OpenAlex)

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.

Proposes a statistical framework for algorithmic collective action (ACA) with multiple collectives in classification, deriving per-collective lower bounds on signal planting and unplanting success under neutrality and contamination-suboptimal classifiers, with synthetic climate-adaptation simulations.

Problem

Existing ACA theory covers only a single collective acting on a learning system, whereas real collective actions are fragmented into multiple collectives with different sizes, strategies and goals that may conflict or align.

Approach

Generalizes the single-collective measure-theoretic ACA framework to M collectives acting on a shared training mixture; per-collective success lower bounds include a cross-collective conflict term, with computable approximations (best-case aggregated knowledge, worst-case mass bound, anonymity proxy) that let each collective evaluate its bound with partial information.

Key results

• For two collectives, per-collective lower bounds show a trade-off frontier: shifting mass toward one collective raises its bound at the other's expense, and reduced feature overlap skews the frontier.
• With M=4 collectives, global success is highest when mass-weighted label entropy is low; success drops sharply when both mass and label entropies are high, driving the guarantee near zero.
• Reduction to M=1 reproduces prior single-collective bounds, which the authors report as complete fidelity in their ablations.

Surprising findings

• Mass-weighted uniformity alone does not prevent interference; without label alignment, higher mass entropy is insufficient.
• For feature-only planting, known-interventions bounds are nearly indistinguishable from per-collective bounds.

Limitations

• Assumes collectives are disjoint and the feature space is finite.
• Relies on neutrality assumption and an epsilon-contamination-suboptimal classifier; bounds are conservative and use Hoeffding radii.
• Experiments are purely synthetic simulations with no real-world data.

Experimental setup

Datasets: simulated neighborhood adaptation-records dataset (3,000,000 synthetic instances)

Benchmarks: synthetic climate-adaptation simulation

Metrics: per-collective lower bound on success global success (min and mass-weighted average) minimum sample size for 10% success

Baselines: single-collective ACA bounds (Theorems 3.3/3.5/3.7 of prior work)

Builds on: algorithmic collective action (single collective framework) Hoeffding inequality data leverage

Closest papers by meaning

• The Role of Learning Algorithms in Collective Action (ICML 2024 · similarity 0.90)
• Statistical Collusion by Collectives on Learning Platforms (ICML 2025 · similarity 0.87)
• Collective Counterfactual Explanations: Balancing Individual Goals and Collective Dynamics (NeurIPS 2025 · similarity 0.85)
• Look-Ahead Reasoning on Learning Platforms (NeurIPS 2025 · similarity 0.85)
• Learning Decentralized Partially Observable Mean Field Control for Artificial Collective Behavior (ICLR 2024 · similarity 0.84)
• Position: Evolving AI Collectives Enhance Human Diversity and Enable Self-Regulation (ICML 2024 · similarity 0.84)