PaperlayerView exact revision on arXiv

Source-linked AI summary

When Should LLMs Trust Their Own Revisions? A Risk-Aware Study of Intrinsic Self-Correction

Tianzhu Zhang

arXiv:2609.35832v1cs.CLcs.AI

TL;DR

Intrinsic self-correction can repair errors but can also overturn correct answers, making the decision to revise consequential. The paper tracks transitions across 29 open-weight models and three benchmarks, then compares revision policies and gating. Outcomes vary by setting, and a large net gain can coexist with substantial harm to initially correct answers.

  • Problem

    Because revision without external evidence can repair errors or disturb correct answers, the practical question is when its cost and risk justify replacing the initial response.

    Source: §1 · p. 1

  • Method

    The study tracks correctness transitions across 29 open-weight models and three benchmarks, compares refinement instructions on matched BoolQ answers, and evaluates runtime policies using initial-answer information.

    Source: §1 · p. 2, §4.1 · p. 4

  • Results

    Llama-3.1-8B gains 25.5 points on GSM8K while refinement overturns 19.1% of its initially correct answers, illustrating that net gains can coexist with substantial harm.

    Source: §3.2 · p. 4

  • Takeaways & Limitations

    Across evaluated settings, selective gating is useful in some cases, while keeping the initial answer or always accepting revision is better in others.

    Source: §7 · p. 9

    Unresolved provenance for historical gate and selector summaries prevents treating selector gains as verified unbiased estimates.

    Source: §6 · p. 9

Abstract

from arXiv · show

Intrinsic self-correction asks a language model to revise its own answer without receiving new external evidence. A second pass can recover mistakes, but it can also overturn answers that were already correct. We study this trade-off across 29 open-weight LLMs on BoolQ, GSM8K, and Corr2Cause by tracking correctness transitions between initial and revised answers. Aggregate accuracy can conceal substantially different revision behavior: for example, Llama-3.1-8B improves by 25.5 percentage points on GSM8K, while refinement changes 19.1% of initially correct answers into wrong ones. A controlled BoolQ study further shows that refinement prompts shift the balance between recovery and harm. We then compare three runtime choices: keeping the initial answer, always accepting the revision, and selectively invoking revision using signals available after the initial response. The comparison identifies settings where learned gating is useful and others where a simpler unconditional policy performs better. These results suggest treating intrinsic self-correction as a revision policy rather than as a uniformly beneficial second pass, and evaluating it through both the corrections it recovers and the errors it introduces.

1 Introduction

The paper frames intrinsic self-correction as a decision about whether revision is worth its cost and risk, not merely whether a second answer is more accurate. It studies correctness transitions and runtime policies across models, tasks, and prompts.

1 Introduction

  • Because intrinsic revision adds no external evidence, it can repair an error or disturb an answer that was already correct, while also incurring another generation cost.The runtime policy determines which answer reaches the user.Source: §1 · p. 1
  • A useful instance-level gate must predict revision consequences from pre-revision features and be compared against both keeping every initial answer and revising every answer.Otherwise, improvements from the revision procedure itself may be misattributed to selection.Source: §1 · pp. 1–2
  • Across 29 open-weight models and three benchmarks, the study tracks correctness transitions and compares keeping, revising, and selectively gating answers.It also compares five refinement instructions on matched BoolQ initial answers.Source: §1 · p. 2
  • Llama-3.1-8B rises from 17.4% to 43.0% with unconditional GSM8K revision and 43.6% with gating, while Qwen2.5-7B BoolQ falls from 80.8% to 72.3% with revision but reaches 86.0% with gating.These examples distinguish gains from revision from gains attributable to selection.Source: §1 · p. 2
  • The contribution combines a transition-level risk audit, protocol-validity checks, and comparisons of setting-level policy choice with instance gating.It examines whether initial-answer information can support selective computation for a fixed revision procedure.Source: §1 · p. 2

2 Revision Protocol and Evaluation

The evaluation generates an initial response and a protocol-dependent revision without external evidence, extracts and scores both answers, and classifies their correctness transitions. It also reports record-level checks and limits on what those checks establish.

2.1 Generation and scoring

  • The same model generates a revision from the question, previous answer, and refinement instruction without retrieval, tool output, or oracle feedback; extraction and applied KEEP/CHANGE rules determine the scored answer.Direct-final, verifier, and reasoning variants use different revision payloads and task-specific extraction.Source: §2.1 · p. 2
  • Across 82 summaries of 29 checkpoints, the panel covers BoolQ validation, GSM8K test, and Corr2Cause test; missing answers are invalid and scored incorrect in the accuracy denominator.The reported splits contain 3,270 BoolQ, 1,319 GSM8K, and 2,246 Corr2Cause examples; gate-summary denominators require their own matched records.Source: §2.1 · p. 2

2.2 Record-level protocol checks

  • For Phi-4-reasoning-plus, correctness fell from 1,220 to 1,216 answers, with 27 recoveries and 31 harmful flips, among budgeted traces followed by answer probes—not unrestricted completed reasoning traces.Source: §2.2 · p. 3

2.3 Correctness transitions

  • The four transitions distinguish preserved correctness, persistent errors, harmful flips, and helpful recoveries; only W→C and C→W change accuracy.HarmRate and RecoveryRate condition on initially correct and initially wrong answers, respectively, so their net effect depends on initial accuracy.Source: §2.3 · p. 3

3 How Revision Changes Correctness

Revision outcomes vary across model–task settings and refinement instructions: revision can improve accuracy while also overturning correct answers. Net accuracy therefore does not fully describe the transition trade-off or establish whether instance gating helps.

3.1 Revision gains vary across settings

  • Unconditional revision reduces accuracy in 20 of 28 BoolQ, 10 of 28 GSM8K, and 9 of 26 Corr2Cause settings, while gains reach 25.55 points for Llama-3.1-8B on GSM8K and 27.64 points for Qwen3-Coder-30B-A3B on Corr2Cause.Between-setting variation motivates policy comparisons by model, task, and protocol, but does not by itself show that instance gating is useful.Source: §3.1 · pp. 3–4

3.2 Net gains can conceal harmful revisions

  • 381 recoveries outweigh 44 harmful flips for Llama-3.1-8B on GSM8K, yielding a 25.5-point gain, yet 19.1% of its initially correct pool is overturned.The count-based 95% bootstrap intervals are [22.9, 28.4] points for gain and [14.2, 24.3]% for conditional harm; they exclude run-selection uncertainty.Source: §3.2 · p. 4
  • Llama-3.1-70B on BoolQ recovers 122 errors but overturns 159 correct answers, so its net loss does not mean it lacks correction capability.Whether a gate can retain recoveries while blocking damage depends on information available before the second pass.Source: §3.2 · p. 4

3.3 The refinement instruction changes the trade-off

  • In the matched BoolQ experiment, all five refinement-instruction families have negative mean accuracy changes, with neutral wording showing the smallest loss; forceful switching instructions increase harmful transitions more than helpful ones.For Qwen2.5-7B, prompt wording shifts both accuracy change and the mix of revisions, but these observations do not establish that beneficial changes can be selected in advance.Source: §3.3 · p. 4

4 Choosing When to Revise

The section compares keeping an initial answer, always revising, and using an initial-response gate, while distinguishing revision gains from the added value of selection. Its reported comparisons are exploratory, and the cost analysis covers only accuracy-aligned preferences.

4.1 Policies and runtime decisions

  • A gate uses the initial response and its generation features to choose whether to invoke the fixed revision procedure; acceptance policies generate the revision before choosing an answer.The gate’s gain over the initial answer differs from its incremental gain over unconditional revision, and it can underperform revision if it blocks too many recoveries.Source: §4.1 · p. 4
  • Figure 1 joins initial and revised accuracies for each setting using the same 82-row source as the complete Appendix F tables; blank positions mark missing settings.Earlier diagnostic plots use different run summaries and are shown separately.Source: Figure 1 · p. 5

4.2 Features, fitting, and result provenance

  • The gate uses initial-response confidence, parser and answer-format state, generation metadata, and candidate-answer agreement; the recovered trainer selects configurations and thresholds with out-of-fold training predictions.The reported gate and setting-level selector outcomes remain exploratory because saved configurations and selected-policy artifacts are not matched to every summary row, preventing retrospective confirmation of evaluation independence.Source: §4.2 · pp. 5–6

4.3 Incremental value and failure modes

  • 17.44% to 42.99% with unconditional revision and 43.59% with gating for Llama-3.1-8B on GSM8K; Qwen2.5-7B on BoolQ falls from 80.83% to 72.32% with revision but reaches 86.02% with gating.Other settings show selection can forgo revision gains, and selected accuracy falls below the initial baseline in three BoolQ, three GSM8K, and four Corr2Cause settings.Source: §4.3 · p. 6

4.4 Harm and inference-cost preferences

  • The experiments evaluate only the accuracy-aligned case λ = 1, η = 0, and token-savings estimates exclude the initial response, gate overhead, and prompt-prefill cost.A harm–cost sweep addressing other deployment preferences is not part of the reported experiment.Source: §4.4 · p. 7

5 Related Work

Prior work studies iterative feedback, verification, learned correction, and self-reflection, while this paper holds the correction operation fixed and examines when to invoke it. It evaluates whether pre-revision signals can support that decision and whether a learned gate improves on both unconditional policies.

Outcome analyses of intrinsic correction

  • Huang et al. (2024) document failures of reasoning self-correction without external feedback, while Kamoi et al. (2024) distinguish evaluation settings often grouped under self-correction and Zhang et al. (2025) examine prompt bias and unstable revisions.Tyen et al. (2024) separate locating a mistake from correcting a known one; Yang et al. (2025) distinguish preserving correct answers from correcting errors.Source: §5 · p. 7

Feedback, verification, and search

Learned correction and within-generation reflection

  • Trained correctors and self-reflective methods improve the correction operation or generation process, whereas this study holds that operation fixed and examines its invocation (Welleck et al., 2023; Kumar et al., 2025; Mu et al., 2026).Reasoning-tuned checkpoints test additional explicit revision after existing deliberation, not a causal effect of reasoning training.Source: §5 · p. 8

6 Scope and Limitations

The evaluation is limited in model, benchmark, and prompt coverage, and unresolved result provenance constrains policy comparisons. Protocol choices and uncertainty measures further limit what can be inferred about deployment costs and transfer.

Scope

  • The panel covers open-weight checkpoints up to 123B parameters and three benchmarks, excluding commercial frontier models, larger open-weight systems, and a harder modern benchmark; the prompt experiment covers only BoolQ.The comparisons do not establish cross-domain transfer or a causal effect of model training.Source: §6 · p. 8

Result provenance

  • Unmatched historical gate and selector summaries lack links to frozen configurations and disjoint IDs, so their comparisons are descriptive rather than verified unbiased estimates.Recovered paired records validate two GSM8K settings, not the full panel; inconsistent coverage fields are withheld pending matched-record reconstruction.Source: §6 · pp. 8–9

Protocol and deployment limits

  • Protocol choices can change measured transitions, count-based intervals condition on a fixed run and parser, and the study does not evaluate end-to-end latency or energy.Gate fitting requires labeled development data; transfer, calibration burden, and explicit harm–cost preferences remain unevaluated.Source: §6 · p. 9

7 Conclusion

Across evaluated settings, revision both repairs errors and overturns correct answers, with the balance varying by model, task, and prompt. The conclusion is to treat self-correction as a policy choice: the best option may be gating, keeping the initial answer, or accepting the revision.

Conclusion

  • Revision both repairs errors and overturns correct answers, and their balance varies by model, task, and refinement prompt.Aggregate accuracy therefore captures only part of what revision changes.Source: §7 · p. 9
  • An instance-level gate preserves useful corrections in some settings, while keeping the initial answer or accepting every revision performs better in others.The conclusion evaluates the decision using both recovered and introduced errors under the intended revision protocol.Source: §7 · p. 9
  • The value of selective intrinsic self-correction lies in deciding when additional inference is worth using, not in revision alone..Source: §7 · p. 9

Ethics Statement

The work evaluates existing open-weight models on public benchmarks and creates no new datasets. The stated scope is inference-time policy evaluation, with deployment contexts beyond the benchmarks requiring validation.

Ethics Statement

  • The work uses public benchmarks and does not involve human subjects, personal data, or new dataset creation.Source: Ethics Statement · p. 9
  • The studied gating policies aim to reduce harmful answer changes during inference, not to add model capabilities or enable adversarial applications.The authors advise validating transition rates in use cases beyond the evaluated benchmarks.Source: Ethics Statement · p. 9

A Reproducibility Details

The appendix documents the retained benchmark splits, inference setup, prompt and parsing protocols, and implementation details for the learned gates. It also records checks needed to interpret protocol-dependent outputs and gate results.

A Reproducibility Details

  • Table 5 reports nominal dataset split sizes, but those sizes do not establish or certify the denominator used for a summary-only gate result.The comparison is an 82-row panel; smaller checkpoints and StarCoder2 appear only in separate diagnostic panels, which are not pooled with that comparison.Source: Table 5 · p. 10, A Reproducibility Details · pp. 10–11

A.1 Compute environment

  • Large-model inference used a four-GPU A100 80GB Ray node, while smaller models used one- or two-GPU actors; this setup is not a throughput benchmark.The pipeline records package versions, model and tokenizer identifiers, prompt settings, and memory caps for each run.Source: §A.1 · p. 11

B Protocol and Parsing Audits

  • Refinement prompts combine the question, previous answer, and instruction, with task-specific direct-answer, KEEP/CHANGE, or reasoning-plus-answer-probe variants.Five controlled BoolQ prompt families vary how strongly they favor retaining or changing the initial answer.Source: B Protocol and Parsing Audits · p. 11
  • Task-specific parsing maps explicit yes/no answers or normalized final integers to labels; missing valid answers count as incorrect in recovered paired records.Trace-plus-probe outputs are scored as answers but do not establish that budget-limited reasoning completed, and generation and candidate scoring must use the same serialized prefix for comparable confidence features.Source: B Protocol and Parsing Audits · p. 12
  • Verifier diagnostics distinguish the applied KEEP/CHANGE answer from the parsed revision candidate, so equal applied and initial accuracy need not mean the candidate was unchanged.These diagnostics use separate parsing conditions and should not replace retained-panel accuracy values.Source: B Protocol and Parsing Audits · p. 12

C Gate Features and Implementation

  • The recovered gate implementation trains separate help and harm models on correctness transitions, selects thresholds using out-of-fold predictions, then evaluates frozen thresholds.Its code uses evaluation labels for outcome metrics and does not enforce ID disjointness, so the runner must check the overlap report.Source: C Gate Features and Implementation · p. 12
  • Reported gates use initial-response or scoring-pass signals, not post-revision agreement, refined-answer confidence, trace completion, or gold-label diagnostics.The implementation supports logistic regression, HistGradientBoosting, and LightGBM; a saved configuration is needed to identify the path and parameters behind a historical row.Source: C Gate Features and Implementation · p. 12

C.1 Numerical consistency checks

The checks assess whether reported policy summaries are numerically consistent, clarify what transition intervals cover, and distinguish retained results from separate diagnostic panels. Several denominators and source linkages remain unverified.

C.1 Numerical consistency checks

  • All 15 accepted-transition count gains are compatible with the displayed one-decimal accuracies under the nominal task-denominator assumption, but this does not prove shared item IDs.For each complete-panel row, the audit checks that selected accuracy matches the recorded policy rather than selecting the highest evaluation accuracy after the fact.Source: §C.1 · p. 12
  • Three reported 100.0%-coverage rows fail the endpoint identity, and the Qwen-14B transition summaries do not establish whether the compared recoveries and harms share item IDs.The table omits coverage pending verification; the available summaries cannot confirm the item matching needed to infer blocked examples.Source: §C.1 · pp. 12–13

D Bootstrap Uncertainty

  • Bootstrap intervals condition on the chosen run and extraction rule, omitting model-sampling, prompt-choice, and run-selection uncertainty.They also cannot certify summary provenance or resolve incompatible gate-coverage values.Source: D Bootstrap Uncertainty · p. 13, Table 10 · p. 16

E Separate Diagnostic Panels

  • The original two-stage plots mix summary populations and checkpoints outside the retained panel, so their GSM8K Llama-3.1-70B drop cannot support the retained result.The diagnostic shows roughly 35.5% to 8.3%, whereas the retained summary reports 35.48% to 36.09%; scoring/run linkage is not established, and Figure 1 is the source-consistent comparison.Source: E Separate Diagnostic Panels · p. 13, Figure 7 · p. 17
  • Table 8 withholds coverage because its supplied accepted-transition summary has inconsistent endpoints.The diagnostic figures have distinct populations and must not be treated as the retained policy panel.Source: Figure 5 · p. 14, Table 8 · p. 15, Figure 6 · p. 16
  • Table 9 compares feature-restricted baselines with differing coverage and predictor classes, so its rows are not controlled ablations.Its gain is median gated-accuracy improvement over Stage 0, while saved is median replay-estimated Stage 1 tokens avoided relative to always revising.Source: Table 9 · p. 15

F Complete Per-Model Policy Results

The complete per-model policy results are organized as separate summary panels for BoolQ, GSM8K, and Corr2Cause. They report initial, revised, gated, and selected-policy accuracies, alongside token savings where available.

F Complete Per-Model Policy Results

  • The BoolQ complete summary panel reports percentages for accuracies and savings; the selected policy follows the recorded selection, and unavailable gate results retain the unconditional fallback.Saved is the percentage of additional revision output tokens avoided by the selected policy, rounded to whole percentages.Source: Table 11 · p. 18, F Complete Per-Model Policy Results · p. 18
  • The GSM8K complete summary panel reports percentages for accuracies and savings; item-level gate denominators are not linked to these summaries.N is therefore left as a dash, and nominal benchmark sizes must not be substituted.Source: F Complete Per-Model Policy Results · p. 18, Table 12 · p. 19
  • The Corr2Cause complete summary panel reports percentages for accuracies and savings, with selected accuracy following the recorded policy even when another policy scores higher.A reported tie remains a tie at the source’s whole-percentage precision.Source: F Complete Per-Model Policy Results · p. 18, Table 13 · p. 20
Loading 2609.35832v1…