Source-linked AI summary
Conformal Thinking: Risk Control for Reasoning on a Compute Budget
Xi Wang, Anushri Suresh, Alvin Zhang, Rishi More, William Jurayj, Benjamin Van Durme, Mehrdad Farajtabar, Daniel Khashabi, Eric Nalisnick
TL;DR
Reasoning LLMs improve with additional test-time computation, but selecting fixed budgets or adaptive thresholds requires balancing error risk against compute. Conformal Thinking frames stopping as distribution-free risk control with upper and lower thresholds, and reports efficiency gains while meeting the specified risk target.
Problem
Adaptive reasoning requires setting thresholds or budgets that balance error risk against computational cost, while threshold values depend on the uncertainty signal.
Method
Conformal Thinking uses distribution-free risk control to select an upper confidence threshold and a parametric lower threshold from a user-specified risk and validation data.
Results
Simultaneous upper and lower thresholds consistently yield efficiency gains at the same target risk across the evaluated stopping approaches.
Takeaways & Limitations
The method enables reasoning to terminate both after sufficient confidence and when further progress appears unlikely, improving computational efficiency within the specified risk target.
Takeaways & Limitations
The approach assumes monotonic risk functions and unaffected upper-loss risk after adding the lower threshold, and evaluation is limited mainly to scientific and math reasoning tasks.
Abstract
from arXiv · showhide
Reasoning Large Language Models (LLMs) enable test-time scaling, with dataset-level accuracy improving as the token budget increases, motivating adaptive reasoning -- spending tokens when they improve reliability and stopping early when additional computation is unlikely to help. However, setting the token budget, as well as the threshold for adaptive reasoning, is a practical challenge that entails a fundamental risk-accuracy trade-off. We re-frame the budget setting problem as risk control, limiting the error rate while minimizing compute. Our framework introduces an upper threshold that stops reasoning when the model is confident (risking incorrect output) and a novel parametric lower threshold that preemptively stops unsolvable instances (risking premature stoppage). Given a target risk and a validation set, we use distribution-free risk control to optimally specify these stopping mechanisms. For scenarios with multiple budget controlling criteria, we incorporate an efficiency loss to select the most computationally efficient exiting mechanism. Empirical results across diverse reasoning tasks and models demonstrate the effectiveness of our risk control approach, demonstrating computational efficiency gains from the lower threshold and ensemble stopping mechanisms while adhering to the user-specified risk target.
1. Introduction
Adaptive reasoning can reduce waste by stopping when confidence is sufficient, but threshold selection remains difficult because thresholds depend on the signal and target risk. Conformal Thinking reframes stopping-rule selection as distribution-free risk control and adds complementary upper and lower mechanisms.
- Motivation: Adaptive reasoning adjusts token use to instance difficulty by stopping when uncertainty falls below a threshold.This supports instance-dependent budgets rather than applying one fixed budget to every problem.
- Motivation: Threshold selection is difficult because uncertainty signals can have arbitrary, uninterpretable ranges and require signal-dependent calibration.Figure 1 indicates that required thresholds vary across signals and target test risks.
- Framework: Conformal Thinking maps a user-specified error risk to stopping criteria using distribution-free risk control.The framework treats early termination as introducing error risk and uses validation data to select corresponding criteria.
- Framework: The framework distinguishes upper-threshold false-positive risk from lower-threshold false-negative risk, targeting different sources of inefficiency.The upper threshold limits tokens after effective convergence, while the lower threshold avoids continued reasoning when progress is insufficient.
- Contributions: At the same target risk, different stopping approaches have different efficiencies, while simultaneous upper and lower thresholds consistently yield efficiency gains.The paper presents this as a central contribution and reports that no single method is uniformly most efficient across target-risk levels.
2. Background: Adaptive Early Stopping via Confidence (Upper) Threshold
Reasoning models often spend more tokens than necessary, motivating confidence-based upper-threshold exits that allocate computation adaptively across instances. These exits stop when a stabilized confidence signal crosses a selected threshold, but signal noise and threshold interpretation remain practical concerns.
- Overthinking: Reasoning traces can greatly exceed the length needed for many instances, creating unnecessary inference cost and making reasoning length a crucial hyperparameter.Models may continue reasoning even after a correct answer could already be elicited.
- Confidence monitoring: Adaptive early stopping monitors a scalar confidence or uncertainty signal computed from the partial reasoning trajectory.Common signals derive from entropy or confidence measured after the reasoning segment.
- Signal stabilization: Per-time-point signals may be noisy or inconveniently scaled, so smoothing, reciprocal, or normalization transformations can improve signal quality.The transformation is application-dependent and produces a stabilized signal for stopping.
- Upper-threshold exit: The upper-threshold exit mechanism halts reasoning at the earliest time the transformed score exceeds λ and emits the answer.It unifies stop-when-confident approaches while avoiding generation of later reasoning tokens.
- Efficiency: Adaptive reasoning can achieve the same dataset-level accuracy with fewer total tokens than assigning a fixed budget to every question.Different instances receive different token counts according to how long confidence takes to reach the target threshold.
3. Related Works
Prior work mainly uses upper-threshold or stability-based exits and often relies on hand-tuned criteria. Conformal Thinking adds a lower-confidence stopping rule and distribution-free threshold calibration, distinguishing its approach from related risk-controlled reasoning methods.
- Existing early exits: Most existing early-exit methods monitor confidence, uncertainty, or intermediate-answer stability and stop when the model appears sufficiently certain or converged.These methods include entropy, trial-answer confidence, and answer-consistency heuristics.
- Conformal Thinking: Conformal Thinking formalizes stopping when reasoning is not becoming sufficiently more confident and combines this lower threshold with an upper threshold.The dual rule captures both confident success and confident failure within one trajectory.
- Risk calibration: Unlike prior hand-tuned cutoffs, sweeps, and heuristics, the framework uses distribution-free risk control to select thresholds with finite-sample guarantees.This calibration targets user-specified metrics rather than relying only on manually chosen signal values.
- Risk-controlled reasoning: Thought Calibration calibrates a probe-based stopping threshold but uses a consistency signal, whereas Conformal Thinking targets generic signals.Thought Calibration can terminate unsolvable instances, but its mechanism is tied to consistency-based probing.
- Risk-controlled reasoning: PAC Reasoning controls whether to use a reasoning-chain output rather than controlling the chain’s length itself.This differs from Conformal Thinking’s focus on stopping the reasoning process.
- Overthinking: Evidence that longer reasoning is not always beneficial provides complementary motivation for adaptive stopping, while Conformal Thinking does not select among completed trajectories.Related approaches instead use multi-sample selection, structured decoding, or broader reasoning-system pipelines.
4. Method: Conformal Thinking
Conformal Thinking frames adaptive reasoning as risk-controlled early stopping, using upper and lower thresholds to balance correctness against wasted computation. A finite-sample calibration procedure selects feasible stopping rules and, when several satisfy the risk target, chooses the most efficient one.
- Framework: The method models each reasoning step with an intermediate prediction and scalar signal, then uses these quantities to define adaptive stopping policies.The signal can be confidence or another scalar derived from the partial reasoning trajectory.
- Two-threshold policy: Two thresholds control complementary exits: the upper threshold stops when confidence indicates a correct answer, while the lower threshold stops when confidence is insufficient to justify continued reasoning.The thresholds are ordered λ+ > λ−, so an instance exits through only one threshold.
- Risk and efficiency losses: The framework defines correctness losses for upper-threshold false positives and lower-threshold false negatives, alongside efficiency losses that quantify wasted reasoning relative to the total budget.The efficiency losses help distinguish thresholds with comparable correctness risk by measuring unnecessary computation.
- Parametric lower threshold: The lower threshold uses a parametric sigmoid schedule whose slope controls how rapidly confidence must increase for reasoning to continue.The parameterization can represent linear, exponential, logarithmic, or nearly constant threshold shapes.
- Calibration: Risk control uses finite-sample correction on a validation set to select thresholds that respect a user-specified tolerance, rather than relying on potentially overfit empirical risk.Naive calibration can exceed the target test risk, whereas corrected calibration is reported to keep test risk below the tolerance across resamplings.
- Calibration: When multiple signal-threshold candidates satisfy the risk constraint, the procedure selects the candidate with the smallest estimated efficiency loss.The final routine enumerates candidates, enforces the risk budget, and then minimizes efficiency loss among feasible options.
5. Empirical validation
Experiments show that finite-sample risk control keeps test risk below the user-specified target, while signal ensembling and dual thresholds improve token efficiency. Lower thresholds are especially useful when unsolvable instances are common, and UCB becomes more valuable with smaller validation sets or distribution shift.
- Risk control: UCB keeps unseen-test risk below the user-specified tolerance, whereas naive calibration often exceeds the target.This pattern holds across early-stopping signals and is driven by finite-sample correction for validation uncertainty.
- Ensembling signals: The signal ensemble selects the most efficient signal–threshold pair on validation data and transfers efficiency gains to the test set.On Qwen3-8B, the ensemble consistently selects the probe signal across target tolerances.
- Dual thresholds: When solvable instances dominate, upper-only stopping captures most savings and lower-only stopping adds little.The complementarity of the thresholds therefore depends on the solvable:unsolvable composition of the evaluation set.
- Dual thresholds: When unsolvable instances are common, adding the lower threshold shifts the accuracy–token curve left relative to upper-only stopping.Upper thresholds mainly save tokens on solvable instances, while lower thresholds halt unsolvable runs that would otherwise consume the full budget.
- Robustness: The advantage of UCB over naive calibration becomes more pronounced as validation-set size decreases.Additional robustness experiments also show that short-to-long length shift particularly challenges lower-threshold risk, while UCB remains below the target under dataset shift.
6. Conclusions, Limitations, and Future Work
Conformal Thinking adds a lower progress-based threshold to confidence-based stopping, allowing reasoning to terminate both after confidence is reached and when further progress appears unlikely. The approach assumes monotonic risks and is evaluated mainly on scientific and mathematical reasoning tasks.
- Conclusion: Conformal Thinking combines an upper confidence threshold with a lower threshold that enforces a schedule of confidence improvement.The dual-threshold rule targets both confident success and cases where the model appears unlikely to become correct.
- Limitations: The method assumes risks are monotonic in the hyperparameters and that upper-threshold risk is unaffected by adding the lower threshold.These assumptions can be rigorously verified only in large-sample regimes.
- Limitations: The evaluation is limited to scientific and mathematical reasoning because existing uncertainty signals do not support early stopping for less structured reasoning and output.
- Future work: Future work includes instance-wise threshold setting, applications to agentic systems, parallel reasoning chains, and improved uncertainty-signal calibration.
A. Extended experiment specifications.
The extended specifications define the uncertainty signals and probe-based alternatives used to monitor reasoning progress. Confidence uses forced-answer likelihood, whereas EAT uses next-token entropy, and probes predict stepwise correctness from hidden representations.
- Uncertainty signals: Confidence and EAT measure uncertainty at each thought chunk during reasoning generation.Confidence and EAT are the two primary uncertainty metrics evaluated.
- Confidence: Confidence appends a termination tag and forcing string, greedily rolls out an answer, and scores it with length-normalized log likelihood.
- EAT: EAT appends the thought termination tag and forcing string, then uses the entropy of the next-token distribution without generating a rollout.
- Probe signals: Probe signals use hidden representations of candidate-answer steps to train a two-layer MLP that predicts stepwise correctness.The probe is trained on AIME 1983–2024 trajectories with labels determined by correctness relative to the gold answer.
B. Risk control and finite-sample correction
The calibration procedure searches over signals and threshold parameters, retaining candidates whose adjusted risk satisfies the user’s budget and selecting the one with minimum efficiency loss. Distribution-free risk control uses validation estimates while accounting for finite-sample uncertainty.
- Calibration objective: The framework selects a signal–threshold pair that satisfies risk budget ϵ while minimizing efficiency loss.
- Calibration procedure: Among feasible candidates, the procedure returns the signal–threshold pair with minimum estimated efficiency loss.
- Calibration procedure: For each signal and threshold, the procedure estimates adjusted risk from validation losses and retains candidates with adjusted risk at most ϵ.Threshold parameters include the upper threshold λ+ and lower-threshold parameter c.
B.2. Why finite-sample correction is needed
Raw empirical risk can be over-optimistic because sampling noise and broad candidate searches may select configurations that only appear feasible on validation data. The framework therefore uses a conservative finite-sample-adjusted risk estimate to enforce the target constraint.
- Sampling noise can make a candidate appear feasible on validation data while violating the target risk after deployment.
- Scanning many signal-threshold candidates can exploit random downward fluctuations in empirical risk during selection.
- The method replaces raw empirical risk with a conservative finite-sample-adjusted quantity and requires it to satisfy the risk target.
B.3. Calibration methods
The calibration methods differ in how they enforce the risk constraint: Naive uses no finite-sample correction, whereas UCB applies a concentration-based upper confidence bound. The correction is most conservative for small validation sets and vanishes as sample size grows.
- Naive calibration enforces the risk constraint without finite-sample correction.
- The Naive baseline can work when n is large but offers no protection against validation overfitting.
- UCB uses a concentration-based upper confidence bound for losses bounded in [0, 1].
- UCB declares a candidate feasible only when its corrected risk is at most ϵ, with larger corrections for small n and vanishing corrections as n grows.
- For a finite candidate grid, a union bound can strengthen the fixed-candidate guarantee by replacing δ with δ/|G|.
C. Violation of false positive guarantee
The false-positive guarantee depends on how the lower threshold abstains from incorrect versus correct samples. It remains valid when the mechanism filters out at least as large a fraction of incorrect samples, but can fail under distribution shift that disproportionately removes correct samples.
- The false-positive rate is defined using the numbers of incorrectly and correctly answered queries.
- The lower threshold changes false-positive risk according to the fractions of incorrect and correct samples it abstains from.
- When the lower threshold abstains proportionally more from incorrect than correct samples, the false-positive guarantee remains valid.
- A significant calibration-to-test distribution shift can make the lower threshold generalize poorly and disproportionately abstain from correct samples.
- If the lower threshold functions as intended, the upper-threshold guarantee remains intact.
D. Ablation study results
The ablations examine validation-set size, length shift, and dataset shift using false-positive and false-negative risk. Principled risk control is strongest with small validation sets and under dataset shift, while lower-threshold control is less robust to short-to-long length shifts.
- Principled risk control shows better risk control than Naive cross-validation when validation sets are small.
- Short-to-long length shift creates greater risk-control challenges, especially for false-negative risk governed by the lower threshold.
- Under length shift, upper-threshold control alleviates excessive risk for most signals, except token-based signals.
- The lower threshold lacks robustness to length shift because its shape depends on the reasoning horizon.
- Under dataset shift, principled risk control again yields more controlled risk.