A reasoning model that thinks step by step, pauses, reconsiders, and revises is paying for each extra round in latency and tokens. Recent early-exit methods cut this cost by stopping as soon as the model looks confident. They treat every wrong answer as equally bad. In clinical triage, code execution, or any other safety-critical setting, that assumption is exactly wrong. Stopping early and missing a myocardial infarction is not comparable to stopping early and confusing two benign conditions. This proposal builds a stopping rule for iterative refinement, meaning any loop in which a model drafts an answer and then keeps revising it. The rule decides whether one more round of thinking is worth its latency by asking how much it would reduce cost-weighted risk, not how confident the current answer is. It is also allowed to fall back to an earlier round's answer if later thinking drifts toward a cheaper-looking but more dangerous conclusion.
In two sentences: a confidence threshold tells you whether the model is sure, but under asymmetric costs what matters is whether the probability the model still assigns to the catastrophic alternatives is small enough to stop paying for more thought. We estimate that quantity at every pause in a reasoning chain, stop when the expected cost reduction from continuing falls below the price of a round, and output the answer from whichever round carried the lowest estimated cost. We then test on medical differential diagnosis whether this beats simple class-specific thresholds at matched token budgets.
Core Claim
The claim has three parts. First, the right stopping signal under a cost matrix is the minimum expected cost of committing now. Let $p_t(i)$ be a calibrated probability that the true class is $i$ after round $t$, and let $C_{ij}$ be the cost of answering $j$ when the truth is $i$. The cost of committing at round $t$ is
$$R_t = \min_j \sum_i p_t(i)\, C_{ij}.$$
The confidence of the top answer is a different quantity, and it can be high while $R_t$ is still large. This happens because a small residual probability on a high-cost class dominates the sum.
Second, the decision to continue should compare the expected drop in this risk against a per-round latency price $\beta$. With a stopping time $\tau$ and an output $\hat y_\tau$, the goal is to minimize
$$\mathbb{E}\big[C_{y,\hat y_\tau}\big] + \beta\, \mathbb{E}[\tau].$$
Third, refinement drift is itself asymmetric. The "overthinking" documented by Dynamic Early Exit in Reasoning Models overwrites correct early deductions. When the overwritten answer was the dangerous-but-correct one, the damage is severe. So the output rule should be allowed to recall the round with the lowest $R_s$ for $s \le \tau$, rather than being forced to take the last draft.
What would be new is a stopping policy for autoregressive reasoning that is driven by a cost matrix and allowed to recall earlier answers, with its severe-miss rate controlled on held-out data. Existing cost-aware stopping for deliberation uses heuristic energy thresholds over multi-agent debate. Existing optimal stopping with recall is symmetric and has only been tested on VGG networks.
Relation to Prior Work
The closest work is When Agreement Becomes Unsafe: Loss-Aware Energy Control for Diagnostic Deliberation (D-CEM). It monitors a two-agent clinical debate and computes a "diagnostic-risk energy" from three terms: disagreement between the agents, expected harm under a clinical loss matrix, and the margin between the top two candidates. It then picks among four actions: continue, steer the agents, stop and decide, or stop and escalate. D-CEM motivates this proposal, because it shows that cost-blind deliberation converges on harmful consensus. It also supplies one of our tools, its "confusion-induced posteriors," which map a model's stated answer through its historical confusion matrix to a probability over the true condition. We reuse that mapping to calibrate $p_t$.
We differ from D-CEM in three ways. We target a single reasoning chain, which is the cheaper and far more common deployment. We replace the heuristic energy thresholds, which the D-CEM authors themselves list as a limitation, with a stopping rule derived from the objective above. And we add recall of earlier rounds, which D-CEM does not have.
Two early-exit lines supply the rest of the machinery, and both are symmetric. Dynamic Early Exit in Reasoning Models (DEER) is a training-free method for large reasoning models. It pauses at "action transition points" such as "Wait," elicits a trial answer through a parallel branch that shares the model's key-value cache, and halts when a single confidence threshold $\lambda$ is exceeded. We adopt DEER's pause-and-probe mechanism wholesale as our sensor. What we replace is its cost-blind global $\lambda$.
Hold That Exit: Near Optimal Early-Exit Inference via Recall proves that no stopping rule that forbids returning to earlier outputs can approximate the optimal accuracy-latency trade-off. It solves stopping with recall by dynamic programming, but only on VGG image classifiers, with a symmetric loss and an approximately 8% accuracy penalty from discretizing loss trajectories. We carry its recall formulation into reasoning chains under a cost matrix. Because DEER's pause points give only a handful of decision points per chain, a coarse state space should keep the discretization cost small; we measure this.
Beyond Greedy Exits (UAT) adapts one global exit threshold at test time using a bandit, with a learned reliability score. It is again symmetric, and serves us as a baseline.
Two papers constrain the design. Learning to Offload with Low Regret under Asymmetric Misclassification Costs shows that under asymmetric costs a single confidence cutoff should become class-specific thresholds bracketing an "ambiguous region." Its guarantees hold only for binary, one-shot classification with a perfect fallback server. The simplest multiclass generalization of its idea, per-class thresholds, is our strongest simple baseline. Online Learnability of Chain-of-Thought Verifiers formalizes that accepting a wrong reasoning step (a soundness error) is far worse than rejecting a correct one, which merely costs more search. That is the same asymmetry our stopping rule exploits, but its results are theoretical and assume the realizable setting.
Finally, LEC gives linear-expectation constraints for controlling risk conditional on a selection decision. We apply this idea to the stopping decision, as a calibration layer that caps the severe-miss rate.
Research Program
We build the system in four pieces.
The sensor is DEER's trial-answer probe, run at each action transition point. It yields a distribution over candidate diagnoses at round $t$, formed from answer-token probabilities pooled over a few prompt phrasings as in DEER-PRo.
The calibrator converts this raw distribution into $p_t(i)$. It does so by fitting a round-indexed confusion-induced posterior, in the style of D-CEM, on training trajectories. The round index matters because calibration drifts as the chain lengthens.
The continuation model is a small regressor trained on logged full-length trajectories. From features of the current state, it predicts the expected minimum cost reachable within the remaining budget. The features are $R_t$, $R_{\min} = \min_{s\le t} R_s$, the posterior mass on classes whose cost exceeds a set level, how often the answer has flipped between rounds, and the remaining budget. Training on full-length logs means we observe the counterfactual of continuing for free. The rule stops when the predicted improvement in cost falls below $\beta$. At the stopping round it outputs the answer from the round $s^* = \arg\min_{s \le \tau} R_s$, with the committed answer chosen by the cost-weighted rule inside that round.
The certifier tunes a single scalar offset on the stopping boundary using held-out data. It enforces, in the style of LEC, that the rate of missed high-severity diagnoses stays below a target $\alpha$.
The main application is differential diagnosis, which is the domain where the file supplies both benchmarks and prior cost-aware deliberation results. The cost matrix follows D-CEM's asymmetric clinical loss structure over high-risk conditions such as anaphylaxis and myocardial infarction versus benign ones. For robustness we also sweep a family of matrices by scaling the severe-miss penalty.
Baselines
The naive reference points bracket the latency range. At one end is a single-pass answer with no refinement. At the other is the full-budget chain that always runs to natural termination. Both commit using the cost-weighted Bayes rule, so that differences in cost come from stopping and not from the output rule. Any stopping policy must lie at or below the line joining these two points on the cost-versus-tokens plot.
The strongest existing reasoning-chain method is DEER with its global threshold $\lambda$. We tune $\lambda$ on the same validation split and budget as our method and report its whole curve, not one operating point. The file contains no reported result on Lacuna for DEER on DDXPlus or MedQA, so we run it ourselves with the released procedure on identical backbones.
The strongest simple competitor is a multiclass version of the HI-LCB-2D idea, which we call cost-thresholded DEER. It uses DEER's probe, the same calibrated $p_t$, and the cost-weighted output rule, and stops when $R_t$ falls below a per-class-tuned threshold. It shares every component with our method except the continuation model and recall. Because of that, it is simultaneously the most dangerous competitor and the key ablation. We tune it with the same number of validation evaluations as our method.
The strongest existing cost-aware deliberation method is D-CEM. We run it in its native two-agent form with the same backbone for both agents, and compare at matched total tokens. The reported figures available are that D-CEM reduced harmful consensus "to near zero, an order-of-magnitude improvement over standard pooling methods," and stopped easy cases in "2–8 rounds" compared with Free Discussion baselines that used their full 22-round budget (When Agreement Becomes Unsafe).
We also include UAT's bandit-tuned global threshold, and RECALL's symmetric dynamic program fed with the same trial answers. The latter isolates what the cost matrix adds on top of recall.
Beating these baselines means one of two things on DDXPlus: a lower expected cost at every matched mean token budget than cost-thresholded DEER, or a comparable harmful-error rate to D-CEM at substantially fewer tokens. In either case the certified severe-miss rate must stay within $\alpha$ on test data.
Three further ablations each remove one ingredient. The first removes recall, so the output is always the last round. The second replaces $R_t$ with a symmetric error probability inside the continuation model. The third removes the round-indexed calibration.
Why not just…?
Why not just use per-class thresholds on DEER's trial answers? This is the real threat, and the proposal is designed to lose cleanly if it holds. Thresholding $R_t$ decides whether the current answer is safe enough, but it cannot tell apart a chain that is still making progress on a dangerous differential from one that has stalled. It also cannot undo a late flip away from a correct severe diagnosis. The DEER paper's pearl-reasoning observation (the critical deduction arrives early in about 75% of solved problems) suggests such late flips are common enough to matter. If cost-thresholded DEER matches our method on the cost-latency curve and on severe-miss rate, we will report that the continuation model and recall add nothing and that the contribution reduces to calibrated cost-weighted thresholds.
Why not just run D-CEM? D-CEM needs at least two agents deliberating over many rounds, and its thresholds are heuristic. Our question is whether a single chain with a principled stopping rule reaches D-CEM's safety at a fraction of its tokens. The matched-token comparison answers this directly. If D-CEM dominates even at equal tokens, the lesson is that disagreement between agents carries risk information that one chain cannot produce, which would itself be useful.
Why not always run the full chain whenever a dangerous class is anywhere in the differential? In differential diagnosis a high-risk condition is plausible for many cases, so this rule degenerates toward the full-budget baseline. It also inherits the overthinking harm that DEER documents. We include it as an extra operating point; its token cost and severe-miss rate will show whether it is merely expensive or actively worse.
Why not fine-tune the model with a severity-weighted loss, as PAMS does for pathology slides, so that a single threshold suffices? That changes what the model believes but not when it should stop. It also requires retraining a large reasoning model and does not transfer when the cost matrix changes. Our stopping rule works on a frozen model given only answer-token probabilities, and can be re-tuned for a new matrix from logged trajectories. The sweep over cost matrices tests the value of that re-tuning.
Proposed Experiments
The primary benchmark is DDXPlus: about 1.3 million synthetic vignettes over 49 pathologies with doctor-vetted differentials. Its original paper reports that training the AARLC model on differentials raised differential diagnosis recall from 21.96% to 97.73%, which shows the differentials carry the multi-candidate structure a cost matrix needs.
The secondary benchmark is MedQA-USMLE, recast as specialty triage following D-CEM. As a symmetric sanity check we use MATH-500 and GPQA Diamond from the DEER evaluation. With a uniform cost matrix our rule should collapse to roughly DEER's behavior, and if it underperforms DEER there, the method is miscalibrated.
We use two open reasoning backbones of different scale, because DEER reports that the best $\lambda$ shifts across models. We ask whether our method's single certified offset transfers better across scales.
We measure four things: expected cost per case under the cost matrix, severe-miss rate against the target $\alpha$, mean and 95th-percentile tokens and wall-clock time, and a severity-aware confidence measure following PAMS's asymmetric misclassification confidence (AsMC). We report full cost-versus-tokens curves over $\beta$.
Three checks decide the outcome. The first is the matched-budget comparison against cost-thresholded DEER, which tests whether the continuation model and recall earn their place. The second is a flip analysis. It counts how often later rounds move a correct high-severity answer to an incorrect low-severity one, and asks whether recall recovers exactly those cases; this checks the asymmetric-drift premise directly. The third is a shift test: calibrate on one subset of DDXPlus pathologies and test on held-out ones, and check whether the certified severe-miss rate holds. Failure here would confirm the miscalibration-under-shift limitation the file flags, and would bound what the certifier can promise.
References
- Dynamic Early Exit in Reasoning Models
- When Agreement Becomes Unsafe: Loss-Aware Energy Control for Diagnostic Deliberation
- Hold That Exit: Near Optimal Early-Exit Inference via Recall
- Beyond Greedy Exits: Improved Early Exit Decisions for Risk Control and Reliability
- Learning to Offload with Low Regret under Asymmetric Misclassification Costs
- Online Learnability of Chain-of-Thought Verifiers: Soundness and Completeness Trade-offs
- LEC: Linear Expectation Constraints for Selection-Conditioned Risk Control in Selective Prediction and Routing Systems
- Every Error has Its Magnitude: Asymmetric Mistake Severity Training for Multiclass Multiple Instance Learning
- DDXPlus: A New Dataset For Automatic Medical Diagnosis
- What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Exams