When a language model has to make a decision under uncertainty, such as which crop to plant or which stock to buy, one effective recipe is to have the model rank possible outcomes and then compute the expected-utility-maximizing action outside the model. Spending more inference compute on more rankings helps. Rankings have a limit, though: they say which outcome is preferred, not by how much. Expected utility needs exactly that magnitude, because a forecast of 30% chance of drought must be weighed against how bad drought actually is. This note proposes spending the same inference budget on a different kind of question, borrowed from revealed-preference economics. Each outcome is shown with a synthetic "price," and the model is asked which outcome it would take net of price. Every answer of that kind yields a linear inequality on utility differences. Across a schedule of prices, the inequalities pin down the size of the gaps, not just their order.
The second payoff is a built-in consistency check for domains that have no answer key. A set of priced choices either admits some utility function that explains all of them or it does not. When it does not, the minimal "slack" needed to restore consistency is a computable, label-free measure of where the model's preferences contradict themselves. We use that signal to decide where the next unit of inference compute goes. The core claim is that, at equal token budgets, priced-choice elicitation followed by a small linear program produces more accurate cardinal utilities, and therefore better decisions, than pairwise ranking fitted with a Bradley–Terry model. We also claim that the inconsistency signal outperforms uniform resampling as a rule for allocating extra queries.
Where This Sits
The method this proposal modifies is DeLLMa. It breaks decisions into four steps: list the uncertain states, forecast them, elicit utilities over state–action pairs through minibatch rankings fitted to a Bradley–Terry model (a logistic model that turns win/loss comparisons into strength scores), and take the expected-utility argmax. It reports that more forecast samples and denser ranking overlap give linear gains in decision accuracy. We keep its first, second, and fourth steps unchanged and replace only the third. Bradley–Terry strengths are identified only up to the logistic link the model assumes. Treating them as cardinal von Neumann–Morgenstern utilities is a modeling convenience, not something the comparisons establish. The file's description of DeLLMa likewise notes that it never uses prices, budgets, or demand queries. Test-Time Compute Games is the other explicit bridge between economics and inference compute, but it treats compute as a billed commodity in a provider market with fixed user demand. It does not use compute to infer anyone's preferences, so it is adjacent rather than overlapping.
The method comes from the demand-query line of work. Optimization with Demand Oracles. shows that asking "what would you choose at these prices" carries far more structure than asking for values, and recovers it through an LP whose dual is separated by demand queries. Restificar et al. supply the intuition that choosing against a certain alternative reveals cardinal parameters. Two papers constrain the design. LLM answers will be noisy and occasionally irrational, so we cannot assume exact optimality. Badanidiyuru et al. show that low-regret choices still yield valid linear bounds, and Learning from Stochastically Revealed Preference shows how to handle the "empty intersection" case, where no single utility explains every choice, with a margin objective. That paper also warns that accuracy drops from 87–99.9% at ten or fewer items to 55–68% at 25 items when utilities are sampled by MCMC in high dimension. We therefore keep the per-instance utility dimension small (the DeLLMa state–action tables) and solve an LP rather than sampling. The rule of reflecting only when triggered follows Scaling Test-time Compute for LLM Agents, which found that indiscriminate reflection degrades agents and that list-wise comparison beats scalar scoring. The latter finding is a warning against the simplest cardinal alternative, which is simply asking for numbers.
Core Claim
Let $u_k$ be the unknown utility of state–action outcome $k$. A priced-choice query shows a small menu $M$ with prices $p_j$ and records the chosen item $c$. Under quasilinear choice with slack $\xi \ge 0$, each answer imposes
$$u_c - p_c \;\ge\; u_j - p_j - \xi_{M} \quad \text{for all } j \in M.$$
We fit utilities by minimizing total slack $\sum_M \xi_M$ subject to all such constraints, plus a normalization that fixes the worst outcome at 0 and the best at 1. Pairwise rankings correspond to the special case where every price is zero. In that case each constraint gives only the sign of $u_c - u_j$. Nonzero prices bound the size of the gap, and the bound tightens as the price grid becomes finer. The minimized total slack is the label-free inconsistency score. Its per-item contributions identify which outcomes the model is confused about.
Research Program
We build an elicitation module that works as a drop-in for DeLLMa's ranking step. For each instance, it first runs a cheap ranking pass to get an ordering. It then issues priced menus of two to four outcomes, choosing prices adaptively so that each question roughly bisects the current feasible interval for the utility gaps in that menu. This is a myopic, information-seeking rule in the spirit of the information-seeking learner in Learning Human Contribution Preferences in Collaborative Human-Robot Tasks. Prices are presented in the same unit as the stated goal, such as a revenue penalty or a fixed cost of choosing an option, so the question stays natural language that the model can reason about. After each batch, the LP is re-solved. The remaining budget goes preferentially to outcomes that carry high slack or that sit in menus where the expected-utility argmax is close to a tie. Every query's prompt and response tokens are logged, so all comparisons are made at matched total elicitation tokens.
The analysis has three parts. The first measures cardinal accuracy directly: in both DeLLMa datasets, the realized revenue or return of each state–action pair is known, so we can compare elicited utilities to the true payoffs after an affine rescaling. The second measures decision accuracy as the number of candidate actions grows. The third is a sensitivity analysis that asks how often the final decision would change under monotone distortions of the Bradley–Terry utilities. That number tells us in advance how much room cardinal information has to matter.
Baselines
The naive reference point is DeLLMa's zero-shot and chain-of-thought prompting on the same base model. The file reports that the full DeLLMa pipeline beats these by up to 40% in decision accuracy. A frontier reasoning model answering directly is also included. On Agriculture, DeLLMa reports 33.3% accuracy for o1-preview.
The strongest competitor is DeLLMa-Pairs itself. DeLLMa reports 73.3% decision accuracy on Agriculture, and it scales with exactly the resource we are reallocating. We rerun it with the same base model, the same state enumeration and forecast samples, and the same total elicitation-token budgets. We sweep its minibatch overlap to trace its full compute curve, giving both methods the same hyperparameter tuning budget on a held-out split. Beating it means a higher decision-accuracy curve at matched tokens on both Agriculture and Stocks, with the gap widening as action sets grow. A tie at the low-budget end with a clear gain at the high end would also count, because that is where additional rankings should saturate.
Two ablations isolate the key ingredient. The first is priced choice with fixed uniform prices and no adaptive pricing or slack-targeted allocation; it tests whether prices alone carry the gain. The second keeps adaptive pricing but spends the extra budget uniformly at random rather than where slack is high; it tests the verifier claim.
Why not just…?
Why not just ask the model for numeric utilities directly? This is the cheapest source of cardinal information and the most serious threat. Scaling Test-time Compute for LLM Agents found that scalar scoring is worse than list-wise comparison, and DeLLMa documents a "sentiment trap" in direct generation. Neither result tested direct scores used as utilities against payoff ground truth. We therefore include a compute-matched baseline that averages many direct scores per outcome. If it matches priced choice on cardinal accuracy, the proposal reduces to "averaged scalar scores suffice," and we will report that.
Why not just spend the budget on more pairwise rankings? DeLLMa's own compute scaling is linear, so it might keep pace. The compute-matched curves answer this directly. The sensitivity analysis also predicts where rankings should fail: when decisions flip under monotone distortions of the Bradley–Terry utilities, ordinal information is not enough to choose correctly.
Why not just compute expected payoffs from the numbers in the reports with a code interpreter? The USDA and price data contain quantitative information, and Mei et al. show that execution grounding can sharply change scaling. This is a real threat on these two datasets specifically. We include a tool-augmented baseline. If it wins, the honest conclusion is that priced elicitation matters only where payoffs are not mechanically computable. We would then add a variant of the benchmark whose goals involve stated, non-monetary user preferences, such as risk aversion or a stated aversion to particular crops, where no calculator applies.
Decisive Checks
The proposal succeeds if three things hold at matched tokens. First, priced-choice utilities correlate more strongly with realized payoffs than Bradley–Terry utilities do. Second, this translates into higher decision accuracy than DeLLMa-Pairs on both 120-instance datasets, especially at the largest action sets. Third, slack-targeted allocation beats random allocation.
The proposal fails if the sensitivity analysis shows that decisions almost never depend on the cardinal scale. In that case the measured slack, though interpretable, would buy nothing. A separate failure mode is that the LLM's priced choices turn out so noisy that the minimized slack is large everywhere, which would mean the model does not behave like a consistent price-taker even approximately. Either outcome is informative, and both would be reported.
References
- DeLLMa: Decision Making Under Uncertainty with Large Language Models
- Test-Time Compute Games
- Optimization with demand oracles.
- Constructing utility models from observed negotiation actions
- Response Prediction for Low-Regret Agents.
- Learning from Stochastically Revealed Preference
- Scaling Test-time Compute for LLM Agents
- Learning Human Contribution Preferences in Collaborative Human-Robot Tasks
- a1: Steep Test-time Scaling Law via Environment Augmented Generation