Agency in Collaborative Learning 2026
2026-04-13..2026-04-17
Berkeley, United States
who it's for
machine learning researchers, mechanism design and algorithmic economics researchers, multi-agent systems researchers, data platform and federated learning engineers, language model fine-tuning engineers, ai policy and regulation researchers, product managers for data marketplaces
topics
incentive-aware collaboration platforms; payment-based peer prediction; statistical evaluability of generative models; multilingual scaling and data mixing; human-verified synthetic data filtering; learning preferences with latent state; retrieval of influential LLM training examples; aggregation effects on model diversity; competition-based alignment guarantees; delegation and complementarity in multi-agent models
| title | author | length |
|---|---|---|
| Spotlight Talks | Chido Onyeze, Kumar Kshitij Patel | 03h 50m |
Shows a simple artificial‑money mechanism that provably approximately implements the core at equilibrium in repeated allocation with packing constraints… | ||
| Rethinking AI Alignment and Evaluation under Heterogeneous Preferences | Kunhe Yang | 42m |
Shows that fitting one reward to heterogeneous users is information-theoretically doomed: mixtures of Bradley–Terry utilities are non-identifiable and pairwise… | ||
| Emergent Alignment via Competition | Natalie Collina | 48m |
Claim: competition among multiple misaligned providers can yield outcomes as good as a perfectly aligned model when the user utility lies in the convex hull of… | ||
| Platforms for Efficient and Incentive-Aware Collaboration | Mingda Qiao | 44m |
Shows collaboration as a graph where each node’s payoff needs one unit from its closed neighborhood, and a central platform knows the hidden node permutation… | ||
| Complementarity and Delegation in Multi-Agent Models | Jon Kleinberg | 59m |
Builds Maia by retraining Leela’s head on billions of Lichess games so Maia mimics human moves (Maia 1100 becomes the site’s most-played bot and its argmax… | ||
| Can Humans’ Verification of Synthetic Data Save Models from Collapse? | Haifeng Xu | 46m |
Shows that retraining only on human-verified synthetic examples can avoid the known collapse from unfiltered self-generated data by proving in a Gaussian… | ||
| Mechanisms for Honest Data Sharing under Competition | Florian Dorner | 52m |
NeurIPS 2023 work proposes a concrete mechanism: players send empirical means (k samples each), attackers add Gaussian noise with variance α, and a payment… | ||
| A Theoretical Framework for Statistical Evaluability of Generative Models | Han Shao | 44m |
Statistical evaluability is formalized: some metrics are impossible to test from finite i.i.d. samples while others are. Concrete findings: KL and Rényi are… | ||
| Learning Preferences with Latent State | Ramesh Johari | 48m |
Claims that preferences are recoverable when the human uses an unknown finite-state machine, gives a concrete estimator and proofs of consistency. Method fits… | ||
| AI Ecosystems: Structure, Strategy, and Regulation | Ben Laufer | 44m |
A game-theoretic fine-tuning model formalizes generalist upstream investment, domain specialist downstream adaptation, and a revenue-sharing parameter delta… | ||
| Can Users Fix Algorithms? A Game-Theoretic Analysis of Collective Content Amplification in... | Chara Podimata | 54m |
Paper models recommendation as a game using matrix completion and game theory and proves that coordinated uprating by a subset of majority users can force a… | ||
| Scaling Data Beyond English: Multilingual Scaling and Data Mixing Laws | Shayne Longpre | 36m |
Paper runs 800 pretraining runs from 10M–8B parameters on MADLAD-400 and derives a cross-lingual transfer matrix and a bilingual transfer score that quantify… | ||
| Peer-Predictive Self-Training for Language Model Reasoning | Yiling Chen | 52m |
Label-free fine-tuning called Peer-Predictive Self-Training has models generate answers sequentially, treat the final aggregated answer as a hard training… | ||
| One AI, two AI, red AI, blue AI: Impacts of Aggregation on Model Diversity and Consumer Utility | Kate Donahue | 41m |
Analyzes how noisy aggregation of multiple models changes user welfare and firm incentives, formalizes aggregation with a Bradley-Terry/softmax choice model… | ||
| Increasing Participant Agency by Understanding the Impact of Training Data | Andrew Ilyas | 46m |
Shows a scalable method to compute the meta-gradient—a gradient of final model behavior with respect to continuous training-data parameters—by… | ||
| Preference Alignment from Choices and Response Times | Alireza Fallah | 45m |
Combines choice and response-time data to recover latent preferences using drift-diffusion models (DDM) and log-normal race models, with a provable one-over-n… | ||
| Proper Scoring Rules for Text and Super-alignment | Jason Hartline | 49m |
Proposes an algorithm that converts instructor reviews into a numeric rubric by calling summarization and question-answering oracles (domain-knowledge-free… | ||
| Spotlight Talks | Rashida Hakim, Chhavi Yadav | 41m |
Proposes a fair exchange contract that pairs agents one-for-one and proves the induced game is supermodular, so pure Nash equilibria exist, form a lattice, and… | ||
| Incentivizing Truthful Data Sharing in Collaborative Machine Learning and Data Marketplaces | Kirthevasan Kandasamy | 48m |
Mechanisms claim that truthful data submission is a (approximate) Nash equilibrium with truthful reporting and that, for normal mean estimation, the social… | ||
| Power and Limitations of Aggregation in Compound AI Systems | Meena Jagadeesan | 46m |
Presents a principal–agent model with conic feasibility constraints, feature-weights matrix α and per-agent budget levels, and proves aggregation expands the… | ||
| The Agentic Data Science Multiverse | Steven Wu | 48m |
Built autonomous AI analysts on LLMs that run complete Python analysis pipelines and an LLM auditor that checks methodological validity, generating thousands… | ||
| Efficient Retrieval of Influential LLM Training Examples | Roger Grosse | 51m |
Argues that training-data attribution splits into an Estimation Problem (accurately estimating average causal effect) and a Retrieval Problem (finding top… | ||
| nothing matches | ||