Research map / Large language models / Language, safety and interpretability

Incentivizing Quality Text Generation via Statistical Contracts

NeurIPS 2024 · Large language models › Language, safety and interpretability · approach: Zero-shot machine-text detection · 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 cost-robust pay-for-performance contracts for LLM text generation under unknown inference costs, proving they correspond to minimax composite hypothesis tests, and evaluates them on code and MT-Bench benchmarks.

Problem

Pay-per-token pricing creates moral hazard: an LLM-serving agent can secretly use a cheaper model, and the principal does not know the agent's internal inference costs.

Approach

Define cost-robust contracts that incentivize the target action for any cost vector within a known spread bound, and show that optimal such contracts (min-budget and min-pay) are optimal minimax composite hypothesis tests scaled by a risk-dependent constant, generalizing a two-action result to arbitrary action counts and objectives.

Key results

• On MT-Bench, cost-robust monotone min-pay contract has expected payment 4.82 vs 4.19 cost-aware (+15%).
• On MT-Bench, cost-robust min-budget contract has budget 6.63 vs 6.16 cost-aware (+1.7%).
• Cost uncertainty raises optimal payment by about 14.7-14.9% on both MBPP and HumanEval code-generation contracts.

Measured effects reported in the paper

Surprising findings

• Easier task (MBPP, higher pass rates) yields more expensive contracts, because the target model is harder to distinguish from alternatives when pass-rate gaps are small.
• Optimal monotone cost-robust contracts have an intuitive three-level structure: zero pay for lowest scores, base pay for intermediate, extra pay for highest.

Limitations

• Empirical outcome distributions come from benchmarks and do not reflect the richer real-world prompt distribution.
• Analysis assumes known cost-uncertainty bounds and a single target action; strictly intermediate targets need extra implementability assumptions.
• Cost estimates are order-of-magnitude, based on electricity energy consumption data.

Experimental setup

Datasets: MT-Bench MBPP HumanEval

Benchmarks: MT-Bench MBPP HumanEval

Metrics: expected payment (min-pay) budget (min-budget) stdev of payment (min-variance) pass@1

Baselines: cost-aware optimal contracts threshold contracts

Builds on: Saig et al. (NeurIPS 2023) two-action min-budget contract / hypothesis testing c Neyman-Pearson lemma Karlin-Rubin theorem LLM-as-a-judge (MT-Bench)

Closest papers by meaning

• A Theory of Training Profit-Optimal LLMs (NeurIPS 2026 · similarity 0.84)
• Fairshare Data Pricing via Data Valuation for Large Language Models (NeurIPS 2025 · similarity 0.83)
• Pricing and Competition for Generative AI (NeurIPS 2024 · similarity 0.83)
• Test-Time Compute Games (NeurIPS 2026 · similarity 0.82)
• Cost-of-Pass: An Economic Framework for Evaluating Language Models (ICLR 2026 · similarity 0.82)
• When Should LLMs Be Less Specific? Selective Abstraction for Reliable Long-Form Text Generation (NeurIPS 2026 · similarity 0.82)