Research map / Learning theory and algorithms / Submodular and game algorithms
Autobidder's Dilemma: Why More Sophisticated Autobidders Lead to Worse Auction Efficiency
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.
Gives a fine-grained price-of-anarchy analysis of first-price auctions with slice-based non-uniform bid-scaling autobidders, showing efficiency worsens as bidders gain finer non-uniform control, parameterized by slice balancedness.
Problem
Autobidders with value-maximizing objectives and RoS constraints can bid non-uniformly across auction slices; prior PoA results (2 for non-uniform, optimal welfare for uniform bid-scaling) do not explain why more sophisticated non-uniform bidding can reduce welfare.
Approach
Parameterize autobidder power by the number and balancedness of slices, define an unbalancedness measure over the market, and prove a tight PoA bound of 1 - 1/(2t) for instances with unbalancedness at most t, so finer slicing and less balanced slices weakly lower efficiency.
Key results
- Theorem 1: for any t in [2/5, 1], the infimum PoA over instances with unbalancedness at most t equals 1 - 1/(2t).
- Unbalancedness is at least 2/5 whenever there are at least two slices, so PoA is at most 4/5 even with perfectly balanced slices.
- Per-auction optimal bidding (unbalancedness 1) yields PoA 1/2, matching prior first-price results, while uniform bidding (one slice) achieves full efficiency.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase non-uniform bidding freedom by refining the market slices | First-price auctions with value-maximizing autobidders · First-price auctions | PoA: 1 -> 1/2 | Hurt |
| Use multiple slices rather than uniform bidding over the entire market | First-price auctions with value-maximizing autobidders · First-price auctions with at least 2 slices | PoA: full efficiency -> at most 4/5 | Hurt |
| Improve a slice's balancedness without reducing the balancedness of other slices | First-price auctions with partially non-uniform bidding · First-price auctions | PoA | Helped |
| Subdivide a slice | First-price auctions with partially non-uniform bidding · First-price auctions | PoA | Hurt |
Surprising findings
- Even with perfectly balanced slices, any multi-slice market has efficiency loss of at least 1/5 (PoA at most 4/5), so efficiency loss is unavoidable once non-uniform bidding is allowed.
- More powerful individual autobidders, each weakly better off given others' bids, can collectively reduce social welfare, a prisoner's-dilemma-like effect.
Limitations
- Analysis is restricted to slice-based non-uniform bidding with a regularity assumption that bidders do not underbid unless they win all their rightful auctions.
- Results are purely theoretical; no experiments are run, and empirical validation relies on prior work by Deng et al. 2024.
- The PoA is defined per instance and worst-case over instances with given unbalancedness, not for specific real-world markets.
Experimental setup
Metrics: price of anarchy liquid welfare
Baselines: uniform bid-scaling (single slice) per-auction optimal bidding (PoA 1/2)
Builds on: Deng et al. 2024 slice-based non-uniform bidding model Balseiro et al. 2021 uniform bid-scaling optimal welfare Aggarwal et al. 2019 PoA of value maximizers Liaw et al. 2023 PoA of first-price auctions
Closest papers by meaning
- Efficiency of the First-Price Auction in the Autobidding World (NeurIPS 2024 · similarity 0.93)
- Sample Complexity of Automated Mechanism Design (NeurIPS 2016 · similarity 0.86)
- A Unified Framework for Uniform-Price Resource Allocation Mechanisms (NeurIPS 2026 · similarity 0.86)
- On the Power of Randomization for Obviously Strategy-Proof Mechanisms (AAAI 2025 · similarity 0.85)
- Optimal Auction Design for Mixed Bidders (AAAI 2025 · similarity 0.84)
- No-Regret Online Autobidding Algorithms in First-price Auctions (NeurIPS 2025 · similarity 0.84)
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.