How precise should a training-contest leaderboard be?

What it found

In a winner-take-all contest ranked on a noisy holdout of m samples, equilibrium effort is V·a_n√m/s with a₂=a₃=1/(2√π) exactly, but the formula stops being an equilibrium once V/σ² passes a ceiling (12.35, 7.78, 6.11, 5.60, 5.80 for n=2, 3, 5, 10, 20) that sits below the zero-rent limit and falls to half of it at n=20. The organiser's optimal holdout is min(m_cap, (wVa_n/2sκ)²): when evaluation is cheap the ceiling, not cost, binds. Quadrature checked against Monte Carlo and brute-force best replies. Stylised.

A figure from its own output

Paper, code and results

• README and method
• White paper
• Experiments and results
• Code
• Tests

More in Paying contributors

• What does it cost to buy decentralised compute?Buying k units of decentralised compute by reverse auction: exact VCG payment k(k+1)/(n+1) is exactly 2× winners' cost, optimal reserve is exactly v/2, first-price equilibrium matches by revenue equivalence.
• What does it cost to make workers race?Paying the first k of n finishers makes speed a contest: π'(1)=(1−k/n)(H_n−H_{n−k}), equilibrium speed R(n−k)(H_n−H_{n−k})/(nc), exact round time nc/((n−k)R); paying all finishers buys no speed, a budget favours winner-take-all until an accuracy floor k_min binds, and recruiting more workers barely speeds the round.
• Who drops out of a race when workers have unequal costs?Paying the first k of n finishers when costs differ: proof of the π'(1) telescoping identity, exact exit rule (slow workers idle iff c_s/c_f ≥ n_f/(n_f−k), never for k ≥ n_f), cost ratio 4 excludes slow workers for k ≤ 3 of 4 fast, convex costs make headcount buy speed.
• How much accuracy can a market buy from contributors who value their privacy?Buying private data with a telescoping log-score market: payment equals information, best Gaussian-mechanism noise in closed form, participation iff κa<1 so precision is capped at λ/(cΔ²) for any number of contributors, and early arrivals over-collect (1–22% welfare loss).
• Paying for improvements, not for overfittingPaying for model improvements against a hidden holdout: telescoping budget, a payment-feedback overfitting attack (paid 0.15 while true loss worsens 0.37), and the Ladder fix with its measured cost.
• What does submitting copies earn under Shapley payments?Shapley and leave-one-out payments under replication: Shapley pays every extra copy even when worthless, saturating at stand-alone value ((n+2)/2× honest), naive evaluators hand over the pot, LOO pays copies 0 but distributes 1/(n+1), dedup underpays lookalikes by 45%.

All 8 Paying contributors projects →

Other research topics

Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1

Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.