Reasoning Has an Opportunity Cost: Allocate Compute Before You Spend It

TL;DR for operators

When several subtasks share one hard inference budget, spending more on the current task leaves less capacity for everything downstream. Under a 512-token budget, untreated Qwen2.5-1.5B-Instruct scored 0.33 on Medium and 0.15 on Hard three-question papers. A version trained to anticipate reasoning cost and allocate effort across the sequence scored 0.97 and 0.93.

ROI-Reasoning¹ addresses that allocation problem in two stages. Meta-Cognitive Fine-Tuning (MFT) teaches the model to predict a coarse cost level before detailed reasoning and to abstain on persistently low-return problems. Rationality-Aware Reinforcement Learning (RARL) then trains those decisions across the complete exam, where early expenditure reduces what can be done later.

The result is not evidence that shorter reasoning is generally better, nor that a 1.5B model is broadly stronger than larger models. It is evidence that selective reasoning can raise cumulative performance when multiple tasks compete for scarce capacity. For production systems, that suggests pre-execution routing among direct answer, deeper reasoning, escalation, deferral, or abstention—but the paper does not test those choices outside mathematics.

A fixed budget turns reasoning depth into an allocation decision

Inference-time scaling is often treated as a per-query setting: harder task, more reasoning. That changes when one workflow has several tasks left and only one nonrenewable computation budget. An expensive attempt on the first item can be locally sensible and globally poor if it prevents later items from being solved.

The paper formalizes this as an Ordered Stochastic Multiple-Choice Knapsack Problem: ordered problems are visible upfront, reasoning actions have uncertain cost and reward, earlier actions cannot be revised, and cumulative cost cannot exceed budget $B$.

The system is therefore not choosing reasoning depth for one problem in isolation. It is choosing how much shared capacity to consume now while preserving options for later tasks.

Cost prediction has to happen before expensive reasoning

MFT trains the model to emit a predicted reasoning-cost level before detailed reasoning. Rejection-sampled supervision teaches those levels first on single problems and then on three-problem sequences. Refusal examples add an NA option for problems that remain unsolved or appear to offer poor return on further computation.

The timing matters: a cost estimate produced after long reasoning cannot govern whether that reasoning should have been attempted.

Prompting alone does not reproduce the behavior in this benchmark. On Hard/512, Plan-and-Solve scores 0.05 and Least-to-Most 0.13, versus 0.81 for MFT. That supports a narrow conclusion: learned pre-computation judgment works better here than simply instructing the model to plan or skip.

Sequence-level training captures the opportunity cost

MFT improves cost awareness, but the remaining issue is whether spending on the current problem is worthwhile given tasks still ahead. RARL trains on the complete ordered exam under a hard budget. A problem earns reward only when its answer is correct and its predicted cost level matches realized token consumption; exam reward sums those calibrated correctness rewards across all three problems.

Adding RARL to MFT raises total Score in every main same-backbone condition: by 0.15 on Medium/1024, 0.15 on Medium/512, 0.17 on Hard/1024, and 0.12 on Hard/512. MFT+RARL is the highest-scoring Qwen2.5-1.5B variant in all four.

Appendix D is mechanism-supporting evidence rather than a second thesis. On Hard/512, P3 accuracy rises from 0.38 with MFT to 0.43 with MFT+RARL, while total Score rises from 0.81 to 0.93. The behavioral analysis similarly reports more shortening or abstention on costly problems. Both patterns are consistent with preserving capacity for later questions, though they do not isolate that behavior as the sole cause of the score gain.

Tight budgets are where allocation changes the comparison

Under 512 tokens, MFT+RARL scores 0.97 on Medium and 0.93 on Hard, versus 0.33 and 0.15 for the untreated backbone. It also exceeds the listed DeepSeek-V3.2 results of 0.88 and 0.49 and GPT-4o-mini results of 0.65 and 0.32.

That cross-scale result does not hold generally. At 1024 tokens, GPT-4o-mini scores 1.41 on Medium and 1.27 on Hard, above MFT+RARL’s 1.13 and 1.12. DeepSeek-V3.2 also leads on Medium/1024, 1.33 to 1.13.

Approximate ordering regret tells a related but qualified story. Relative to the untreated backbone, MFT+RARL reduces regret from 2.73 to 0.16 on Hard/512 and sharply lowers it in the other conditions. Yet it is not best among every ablation: MFT reports lower regret on Medium/512, and MFT+Greedy reports lower regret on Hard/512. The strongest empirical claim is higher cumulative Score, not universal dominance on every auxiliary metric.

What production systems can borrow—and what remains untested

The paper directly studies generated tokens on three-question mathematics exams. The business extension is therefore an inference.

This is most relevant when tasks genuinely compete for one constraint: a latency objective, per-session cost ceiling, tool-call budget, or throughput cap. If each request has an independent budget and no downstream dependency, the paper’s opportunity-cost mechanism matters much less.

Several boundaries block direct production claims. The 10,000 evaluation papers all contain exactly three mathematical problems in a fixed structure; the principal learned-method comparison uses one 1.5B backbone; difficulty uses coarse levels; and no confidence intervals or statistical significance tests are reported. Token count also omits major components of real inference cost.

One reporting inconsistency should remain visible: Appendix Table 2 prints Qwen2.5-7B-Instruct Medium/1024 position accuracies of 0.45, 0.50, and 0.32 but an overall Score of 0.27, inconsistent with those components and Table 1’s 1.27. The aggregate comparison here therefore follows Table 1.

Allocate reasoning before scaling it

ROI-Reasoning shifts inference control from one problem to a sequence sharing scarce capacity. Its strongest evidence is not that less reasoning is inherently efficient or that smaller models are generally superior. It is that where computation is spent can materially change how much total work is completed when the budget binds.

For production AI, the testable design hypothesis is clear: estimate cost before execution, account for the downstream cost of spending now, and train allocation against full-workflow outcomes. The paper establishes that logic in a controlled mathematical setting; richer cost signals and broader validation are still required before treating it as a general orchestration policy.

Cognaptus: Automate the Present, Incubate the Future.