PaperlayerView exact revision on arXiv

Source-linked AI summary

AERA: Adaptive Evidence Residual Allocation for Efficient Test-Time Reasoning

Ziming Wang, Ivor Tsang, Hangwei Qian

arXiv:2608.27964v1cs.AI

TL;DR

Uniform test-time budgets waste computation, while common stopping signals describe current evidence rather than the value of further reasoning. AERA predicts residual utility from response prefixes; among 792 adjacent GPQA checkpoint transitions, 42 were recoveries and 14 were collapses.

  • Problem

    Adaptive inference seeks to identify when another reasoning unit is useful, but confidence, entropy, and agreement describe current responses rather than unperformed computation.

    Source: Introduction · p. 1

  • Method

    AERA predicts whether additional computation can improve an answer enough to justify its cost from response-prefix evidence, using future correctness only for offline supervision.

    Source: Introduction · p. 2

  • Results

    Across 792 adjacent GPQA checkpoint transitions, 42 were recoveries and 14 were collapses; conventional evidence sometimes strengthened during collapse and weakened during recovery.

    Source: Introduction · pp. 1–2

  • Takeaways & Limitations

    Residual-utility prediction can complement confidence-based stopping when checkpoint correctness evolves non-monotonically.

    Source: Conclusion · p. 7

    Online studies use one generation seed and research code and do not establish latency, energy, throughput, or production-overhead savings.

    Source: Limitations and Broader Impact · p. 7

Abstract

from arXiv · show

Test-time scaling improves language-model reasoning by generating additional candidate solutions, but allocating the same inference budget to every problem is computationally wasteful. Existing adaptive stopping methods commonly rely on confidence, agreement, or answer stability, implicitly assuming that stronger current evidence indicates that further computation is unnecessary. We show that this assumption can fail: checkpoint-level correctness evolves non-monotonically, and observable evidence may strengthen before an answer collapses or weaken before it recovers. Motivated by this mismatch, we introduce Adaptive Evidence Residual Allocation (AERA), a sequential controller that learns whether additional computation is likely to recover a better answer from checkpoint-observable evidence. AERA characterizes cumulative response prefixes using answer-distribution, temporal, re-solving, semantic, and compute features, and repeatedly decides whether to stop or allocate the next response block. Future checkpoint correctness is used only to construct offline supervision and is never available to the controller at inference time. Across GSM8K and GPQA Diamond, AERA identifies question-specific residual opportunities while substantially reducing inference computation. In a frozen-threshold incremental-generation evaluation on 300 untouched GSM8K questions, AERA achieves 92.61% accuracy versus 93.01% with 128 responses while reducing completion tokens by 95.99%. These results suggest that adaptive reasoning should estimate the future value of computation rather than equating present confidence with correctness.

Introduction

Uniform test-time scaling wastes computation because questions differ in how much reasoning they need, while current evidence can diverge from correctness and future recovery. AERA instead estimates the residual utility of more computation from the observed response prefix, using future correctness only for offline supervision.

Introduction

  • Prior adaptive methods vary sampled responses using answer stability or estimated difficulty (Li et al. 2024; Wang et al. 2025), or monitor intermediate trace signals (Yang et al. 2026; Qiao et al. 2025).These approaches leave unresolved what stopping should estimate when reasoning quality evolves non-monotonically.Source: Introduction · p. 1
  • Across 792 adjacent GPQA checkpoint transitions, 42 recoveries and 14 collapses show that correctness can change non-monotonically as reasoning proceeds.Conventional evidence signals can strengthen during collapse and weaken during recovery; Figure 1 illustrates a trajectory that recovers and then collapses.Source: Introduction · pp. 1–2, Figure 1 · p. 1
  • AERA summarizes each observed prefix with answer, temporal, re-solving, semantic, and compute features, then gates allocation of the next response block at each checkpoint.The gate is reevaluated as responses accumulate; future correctness supplies retrospective supervision, not an inference-time input.Source: Introduction · p. 2
  • The paper evaluates residual-utility allocation with full offline frontiers, nested calibration, and frozen-threshold generation, reporting a near-lossless GSM8K operating point selected without test access.Its stated contributions also include quantifying mismatches between observable evidence and checkpoint correctness, including recovery, collapse, oscillation, and confidence changes in the wrong direction.Source: Introduction · p. 2

Related Work

Related work spans pre-inference reasoning-efficiency changes, inference-time stopping and trace controllers, and studies of non-monotonic reasoning. AERA instead allocates sampled responses from observed aggregate states to estimate future recoverability.

Controlling reasoning before inference

  • Prior efficiency methods change, compress, or budget reasoning before observing intermediate evidence, whereas AERA leaves the reasoner unchanged and allocates sampled responses after observing aggregate state.Related approaches include post-training for switching, budget-aware policies, concise traces, structural or parallel reasoning, and training-free token limits or reasoning bypasses.Source: Related Work · p. 2

Adaptive stopping during inference

  • Unlike stability-based stopping (Aggarwal et al. 2023; Li et al. 2024) and difficulty-adaptive budgets (Wang et al. 2025), AERA predicts recoverability from cumulative populations at aligned checkpoints.Future canonical aggregate correctness supervises continuation, while the population is re-observed after each response block; unlike present-convergence signals, AERA estimates whether evidence may recover when convergence and correctness diverge.Source: Related Work · p. 2

Non-monotonic reasoning and residual utility

  • Reasoning trajectories can be non-monotonic, including inverted-U accuracy with chain-of-thought length (Wu et al. 2026), motivating attention to residual utility.Related work studies revisiting unproductive paths (Wang et al. 2026), useful intermediate states within negative final trajectories (Tian et al. 2026), and failures and recoveries within long reasoning processes (Xu et al. 2026).Source: Related Work · pp. 2–3

Empirical Motivation: Evidence Is Not Correctness

Correctness changes non-monotonically across response checkpoints, and observable evidence can move in the opposite direction from correctness. Individual evidence coordinates are informative but cannot be treated as correctness labels.

Empirical Motivation: Evidence Is Not Correctness

  • Across checkpoints 4, 8, 16, 32, 64, and 128, the authors label Re2-selected answers correct or wrong and call trajectories oscillatory after at least two correctness changes.Correctness is used to characterize the phenomenon, not as a controller input.Source: Empirical Motivation: Evidence Is Not… · p. 3

Correctness Evolves Non-Monotonically

  • GPQA has 42 recovery versus 14 collapse transitions from checkpoints 4–64, while GSM8K has 33 monotonic recoveries, 11 oscillatory recovered cases, one monotonic collapse, and six oscillatory failures.The GPQA counts cover the range with aligned AEC states; the GSM8K trajectories span six checkpoints across 1,319 questions.Source: Correctness Evolves Non-Monotonically · p. 3

Evidence Can Move Against Correctness

  • Among 14 GPQA collapses, majority ratio rose in six, entropy fell in seven, and semantic consensus rose in five; one improved all three while becoming incorrect.Conversely, four recoveries had falling majority ratio, six had rising entropy, and 33 had falling semantic consensus; Figure 3 shows substantial overlap between evidence distributions across correctness outcomes, not a causal relationship.Source: Evidence Can Move Against Correctness · p. 3, Figure 3 · p. 3

AERA: Adaptive Residual Compute Allocation

AERA uses checkpoint-observable evidence to decide whether to stop or allocate another response block while leaving response generation unchanged. Future checkpoint correctness supplies offline supervision only, not an inference-time signal.

Method Overview

  • A shared gate controls a frozen checkpointed reasoner without changing response generation and is reevaluated after each allocated response block.It advances sequentially rather than jumping to a predicted final budget, allowing newly generated evidence to be observed.Source: Method Overview · p. 3, Residual Gate Learning · p. 4

Residual-Utility Objective

  • Under the evaluated binary assumptions, AERA’s residual-utility target is exactly recoverability: continue when currently wrong and a later checkpoint can be correct, then stop at earliest recovery.The experiments validate this binary special case, not the full graded-utility formulation.Source: Residual-Utility Objective · pp. 3–4

Training Labels and Information Boundary

  • Ground truth and later checkpoint outcomes construct offline labels only; inference uses the question, observed responses, checkpoint, cumulative cost, and schedule, not correctness estimates.Canonical correctness is computed with the frozen answer-extraction, normalization, and Re2 aggregation rule for label construction, but never enters the evidence state.Source: Training Labels and Information Boundary · p. 4

Observable Evidence Characterization

  • AEC describes each observed prefix with compute, answer-distribution, temporal, re-solving, and semantic features without estimating correctness.Its semantic features include similarity, dispersion, centroid movement, block novelty, and a 32-dimensional projection.Source: Observable Evidence Characterization · p. 4

Residual Gate Learning

  • The learned gate is shared across checkpoints, with checkpoint context encoded by next_checkpoint and budget fraction; its sequential policy does not commit to a one-shot budget prediction.Feature medians, means, and standard deviations are fitted on the training partition and frozen.Source: Residual Gate Learning · p. 4

Sequential Allocation and Calibration

  • Inference visits checkpoints 4, 8, 16, 32, 64, and 128 in order; a separate calibration set selects a threshold under an allowed accuracy-loss constraint, defaulting to full compute if none is feasible.Checkpoint 128 is terminal regardless of the gate output.Source: Sequential Allocation and Calibration · p. 5

Experiments · Experimental Setup · Metrics and Evaluation Protocols

The experiments assess whether AERA improves the accuracy–compute trade-off over fixed budgets and adaptive controls, including when a threshold is selected without test access. They combine offline frontiers, nested calibration, and incremental generation with a frozen threshold.

Experiments

  • AERA is evaluated against fixed budgets and adaptive controls using offline frontiers, nested calibration, and frozen-threshold incremental generation.The protocols test whether its accuracy–compute advantage persists without test access during threshold selection.Source: Experiments · p. 5

Experimental Setup

  • The evaluation uses a frozen Qwen2.5-7B Re2 reasoner with temperature 0.6, top-p 0.95, up to 128 responses, and checkpoints at 4, 8, 16, 32, 64, and 128.Responses have a maximum generation length of 16,384 tokens; benchmarks are GSM8K (1,319 test questions) and GPQA Diamond (198 questions).Source: Experimental Setup · p. 5

Metrics and Evaluation Protocols

  • Reported measures are Re2 accuracy, mean responses per question, and response saving relative to 128 responses, with complete frontiers for thresholds 0.1 through 0.9.Response saving is 100(1−Eq[cq]/128); mean responses measures sampling count, not latency or FLOPs, and completion-token savings are reported separately when available.Source: Metrics and Evaluation Protocols · p. 5
  • Frozen-threshold tests select a threshold using calibration questions only, freeze it before accessing a disjoint test set, and compare test stopping with a separately generated paired Fixed-128 pool.Calibration questions are generated to 128 responses.Source: Metrics and Evaluation Protocols · p. 5
  • Question-level folds keep all checkpoints for a question together; nested evaluation uses separate model-training, inner-validation, and calibration partitions before one outer-fold evaluation.Inner validation selects training duration, calibration selects the operating threshold, and the frozen model–threshold pair is evaluated once on the outer fold; ordinary five-fold frontiers are descriptive, not untouched outer-test estimates.Source: Metrics and Evaluation Protocols · p. 5

Baselines · Implementation Details

The section defines fixed, adaptive, randomized, one-shot, oracle, and checkpoint-aligned response-sampling controls for comparison with AERA. It also specifies the gate’s training settings and the observable-feature constraints of the final controller.

Baselines

  • Fixed-budget controls stop at 4, 8, 16, 32, 64, or 128 responses, while evidence heuristics stop at thresholds on majority ratio, semantic consensus, or normalized entropy.The heuristic results are shown as complete post-hoc sweeps.Source: Baselines · p. 5
  • Matched random shuffling preserves AERA’s empirical budget distribution, whereas one-shot allocation predicts a terminal checkpoint from checkpoint 4 and cannot revise it.Oracle utility selects checkpoints using future correctness and is a non-deployable upper reference.Source: Baselines · p. 5
  • Checkpoint-aligned ESC applies window unanimity with default w = 5, and ASC applies its Beta-posterior stopping criterion with default C = 0.95 at shared checkpoints.Both variants preserve their original decision statistics while sharing AERA’s action space and evaluator.Source: Baselines · p. 5
  • The response-only proxy is not labeled DSC because faithful DSC also requires question-difficulty ranking and budget preallocation.Unless stated otherwise, heuristic sweeps and oracle comparisons are diagnostic references rather than independently calibrated policies.Source: Baselines · p. 5, Implementation Details · p. 6

Implementation Details

  • The gate uses λ = 0.10, five question-level folds, Adam with learning rate 10−3 and weight decay 10−4, up to 300 epochs, and patience 25.Semantic features come from a frozen lightweight sentence encoder and are compressed to a 32-dimensional projection; encoder and preprocessing details are in the supplement.Source: Implementation Details · pp. 5–6
  • The final controller uses 55 numerical inputs restricted to predefined checkpoint-observable features, excluding outcome-like and future-derived fields.The supplement provides the complete feature inventory and a provenance audit of excluded columns.Source: Implementation Details · p. 6

Offline Accuracy–Compute Frontiers

Figure 4 summarizes the offline accuracy–response frontiers for fixed budgets, ESC, ASC, and AERA; Table 2 gives retrospective representative trade-offs, not independently calibrated operating points.

Offline Accuracy–Compute Frontiers

  • +0.24 points on GSM8K (95% CI [−0.03, +0.49]) and −0.21 on GPQA (95% CI [−3.34, +2.84]) were AERA’s paired accuracy differences versus Fixed-128 at τ = 0.3; neither interval establishes a 5% difference.The intervals come from question-level paired bootstrap comparisons with 20,000 resamples.Source: Offline Accuracy–Compute Frontiers · p. 6
  • ASC at its published 0.95 threshold reaches 94.43% with 9.40 responses on GSM8K and 38.99% with 81.82 responses on GPQA; matched-compute AERA accuracy is comparable, not uniformly higher.Against checkpoint-aligned ESC, AERA reaches a similar accuracy range with markedly fewer exposed responses: ESC’s five-response unanimity window uses 20.30 responses on GSM8K for 94.31% and 115.80 on GPQA for 38.85%. Against nondominated ASC mixtures using no more computation than AERA at τ = 0.3, paired differences are −0.09 points on GSM8K (95% CI [−0.30, +0.09]) and −0.93 on GPQA (95% CI [−3.45, +1.46]).Source: Figure 4 · p. 6, Offline Accuracy–Compute Frontiers · p. 6
  • At τ = 0.3, AERA exceeds compute-matched random routing by 1.18 accuracy points on GSM8K (94.35% versus 93.17% at 8.67 responses) and 4.50 on GPQA (38.44% versus 33.94% at 34.02 responses), with both p < 10−4.Permuting AERA’s stopping checkpoints within each evaluation fold preserves the checkpoint distribution and mean response count while removing question-budget dependence; the gains therefore reflect question-specific routing.Source: Offline Accuracy–Compute Frontiers · p. 6
  • Nested calibrated replay yields 94.28% accuracy at 7.88 responses on GSM8K versus 94.12% for Fixed-128, a paired difference of +0.17 points (95% CI [−0.17, +0.50]).Under the same two-point accuracy-constrained calibration rule, AERA and ASC are indistinguishable on GSM8K (93.72% versus 93.70%, at 4.48 versus 4.22 responses); the aggregate GPQA comparison favors ASC, indicating less stability for AERA’s learned gate under small-sample calibration shift.Source: Offline Accuracy–Compute Frontiers · p. 6

When Does Residual Gating Help?

Residual gating shows its clearest advantage on collapse questions, while hindsight diagnostics find that AERA’s continuation decisions are enriched for questions with later residual opportunity. These findings are conditional, and the hindsight grouping is retrospective.

When Does Residual Gating Help?

  • Among questions AERA continued beyond checkpoint 4 at τ = 0.3, 42.4% on GSM8K and 47.0% on GPQA had a higher-scoring later checkpoint, versus 0.9% and 33.3% among questions stopped there.Mean best-future gains were 14.14 versus 0.09 points on GSM8K and 24.58 versus 1.78 points on GPQA; this retrospective grouping does not mean either policy observed future outcomes.Source: When Does Residual Gating Help? · p. 6
  • On 16 GPQA collapse questions, the gate obtains 69.24% Re2 accuracy at 17.25 responses versus ASC’s 36.98% at 110.0 responses, a +32.25-point difference.The 95% bootstrap interval for the paired difference is [+15.89, +49.59].Source: When Does Residual Gating Help? · p. 7
  • On 16 GSM8K collapse questions, the gate’s paired difference is +2.91 points, with a 95% interval of [−0.53, +7.91], at 82.5 versus 92.5 responses.The small subsets do not establish global superiority; the results support a narrower claim that prospective supervision is useful when agreement precedes degradation.Source: When Does Residual Gating Help? · p. 7

Frozen-Threshold Incremental Generation · Ablation Study

The frozen-threshold evaluation reports AERA’s accuracy and token use on untouched GSM8K questions, alongside question-level routing behavior and limits on fixed-budget comparisons. The ablation findings show that a one-shot predictor struggles under severe target imbalance, while sequential re-observation retains value.

Frozen-Threshold Incremental Generation

  • On 300 untouched GSM8K questions, AERA reaches 92.61% versus 93.01% for Fixed-128 while reducing completion tokens by 95.99%.The paired difference is −0.41 points (95% bootstrap interval [−1.79, +0.80]); AERA exposes 4.44 responses on average.Source: Frozen-Threshold Incremental Generation · p. 7
  • The pre-specified rule selects τ = 0.9 using 50 calibration questions; calibration favors AERA by 0.65 points, but its untouched-test accuracy is lower.The rule chooses the least-token threshold within two percentage points of Fixed-128 calibration accuracy, with no fallback if infeasible; the sign change cautions against treating calibration performance as final.Source: Frozen-Threshold Incremental Generation · p. 7
  • Against Fixed-4, AERA gains only +0.13 points while using 0.44 additional responses on average, and is statistically unresolved against every fixed budget in Table 3.The paired 95% confidence interval for the Fixed-4 difference is [−0.21, +0.58]; this evaluation validates the frozen protocol but does not establish superiority to the best fixed budget.Source: Frozen-Threshold Incremental Generation · p. 7
  • The online policy stops at four responses on 283 of 300 questions; its score exceeds Fixed-128 on 58, falls below it on 15, and matches it on 227.The passage attributes the small aggregate loss to a few routed cases rather than a broad decline.Source: Frozen-Threshold Incremental Generation · p. 7

Ablation Study

  • Under severe target imbalance, the one-shot predictor collapses toward checkpoint 4 and is substantially less accurate.Source: Ablation Study · p. 7
  • The ablation findings support sequential re-observation, while the hindsight oracle indicates remaining headroom, especially on GPQA.Source: Ablation Study · p. 7
  • Feature-group diagnostics, semantic-feature ablations, computation-cost sensitivity, and the complete leakage audit are reported in supplementary material.Source: Ablation Study · p. 7

Limitations and Broader Impact

The evidence is limited by replay-based evaluation, benchmark-specific calibration, and small or single-seed studies; the authors call for broader validation and safeguards against premature stopping.

Limitations and Broader Impact

  • Offline response counts and online completion tokens are proxies, and the online studies do not establish latency, energy, throughput, or production-overhead savings.Most results replay one frozen reasoner’s response pools; online completion-token counts omit prompt processing, semantic encoding, and controller overhead.Source: Limitations and Broader Impact · p. 7
  • Within-benchmark training, asymmetric source-to-target transfer, and small studies limit generalization and make modest accuracy differences uncertain.Nested folds provide the cleanest within-dataset estimate, while representative ordinary-replay thresholds are post hoc; only nested and online results use held-out calibration.Source: Limitations and Broader Impact · p. 7
  • A direct ASC comparison is competitive on GSM8K and stronger in aggregate GPQA accuracy-constrained evaluation, supporting a collapse-specific rather than uniform advantage over posterior stopping rules.A GPQA online stress test failed its calibration constraint, so the strict online-efficiency claim is restricted to GSM8K.Source: Limitations and Broader Impact · p. 7
  • The unsupervised semantic projection is benchmark dependent, motivating evaluation across reasoners and seeds, conservative calibration under distribution shift (Li et al. 2025), and end-to-end serving-cost measurement.Source: Limitations and Broader Impact · p. 7
  • Because premature stopping can disproportionately harm difficult or shifted queries, high-stakes systems should retain minimum budgets and support escalation.Source: Limitations and Broader Impact · p. 7

Conclusion

AERA estimates whether additional computation can improve an answer from observable response-prefix evidence, rather than treating confidence as correctness. The conclusion reports near-full-compute GSM8K accuracy at lower generation cost and supports question-specific allocation as a complement to confidence-based stopping.

  • AERA learns recoverability from observable response-prefix evidence and re-evaluates whether to continue after each response block.Source: Conclusion · p. 7
  • Trajectory analysis, nested replay, and frozen online generation show near-full-compute GSM8K accuracy with far lower generation cost.Compute-matched randomization further indicates gains come from question-specific allocation, not a marginal stopping distribution.Source: Conclusion · p. 7
  • Residual-utility prediction complements confidence-based stopping when checkpoint correctness evolves non-monotonically, with broader adoption still a future step.The conclusion characterizes AERA as task-dependent.Source: Conclusion · p. 7

Related papers

Loading 2608.27964v1…