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

Tianzhu Zhang

cs.CLSep 23, 2026

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

Authors: Tianzhu Zhang

Organizations: Nokia Bell Labs, Massy, France

Abstract

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.

Figures & tables

GSM8K check	OLMo-3-7B	Phi-4-plus
Matched items	1,319	1,319
Invalid initial / revised	18 / 13	2 / 4
Both answers parseable	1,301	1,314
Correct initial / revised	533 / 533	1,220 / 1,216
Helpful / harmful flips	0 / 0	27 / 31

Table 1: Checks recomputed from paired records, retaining invalid outputs in the accuracy denominator. Phi-4-plus denotes Phi-4-reasoning-plus. These two checks do not certify the complete model panel.

Figure 1: Initial and revised accuracy from the same 82-row summary source used by the complete tables in Appendix F . Each segment joins one setting’s two accuracies; blank positions indicate missing settings. DS-R1 abbreviates DeepSeek-R1-Distill, and instruction suffixes are shortened. The earlier diagnostic plots use different run summaries and are shown separately in Appendix E .

Model	Task	$\mathrm{C}\!\to\!\mathrm{C}$	$\mathrm{C}\!\to\!\mathrm{W}$	$\mathrm{W}\!\to\!\mathrm{C}$	$\mathrm{W}\!\to\!\mathrm{W}$	Net $\Delta$	Harm	Rec.	Takeaway
		Transition counts				Rates / delta (%)
Olmo-3.1-32B-Instruct	BoolQ	2767	40	96	367	+1.7	1.4	20.7	low-harm gain
Llama-3.1-70B-Instruct	BoolQ	2762	159	122	227	-1.1	5.4	35.0	net harmful
Llama-3.1-8B-Instruct	GSM8K	186	44	381	708	+25.5	19.1	35.0	high-gain/high-harm
DeepSeek-R1-Distill-Qwen-14B	GSM8K	1005	44	130	140	+6.5	4.2	48.1	gain/lower-harm
DeepSeek-R1-Distill-Qwen-7B	Corr2Cause	1446	50	301	449	+11.2	3.3	40.1	broad recovery

Table 2: Representative transition-count summaries. Counts sum to the nominal task split sizes; Harm and Rec. are conditional rates defined in Section 2 . These summaries and the gate-evaluable accuracy panel have different provenance and are not treated as item-matched gate records.

Figure 2: BoolQ prompt comparison across 11 models. Bars show mean accuracy change. The harmful and helpful curves show joint transition fractions, $\#\mathrm{C}\!\to\!\mathrm{W}/N$ and $\#\mathrm{W}\!\to\!\mathrm{C}/N$ , rather than conditional HarmRate and RecoveryRate. Their difference gives the accuracy change. No prompt-level uncertainty intervals are available in these plotted summaries.

Figure 3: Qwen2.5-7B on matched BoolQ initial answers. Left: accuracy change and joint harmful/helpful fractions of all evaluated examples. Right: the distribution of changes in diagnostic gold-answer log confidence; negative values indicate reduced confidence in the correct answer. Gold-answer confidence is an analysis diagnostic, not a deployable gate input.

Dataset	Model	Initial	Revised	Gate	Selected	Policy
Gate exceeds both unconditional baselines
GSM8K	Llama-3.1-8B	17.44	42.99	43.59	43.59	gate
Corr2Cause	Qwen2-57B-A14B	71.24	79.52	79.70	79.70	gate
BoolQ	Qwen2.5-7B	80.83	72.32	86.02	86.02	gate
BoolQ	gpt-oss-20b	82.45	84.13	85.08	85.08	gate
Recorded selector chooses an unconditional policy

Table 3: Representative outcomes generated from the complete summary panel. Bold marks the largest recorded accuracy in each row, not a statistically established improvement. DS-R1 abbreviates DeepSeek-R1-Distill.

Figure 4: Accepted transitions in the 15 available gate-detail summaries. Each equal-sized marker is one model–task setting; shape and color identify the task. Above the diagonal, accepted recoveries exceed accepted harmful flips, implying a gain over Stage 0 for a matched population. This does not establish a gain over unconditional Stage 1.

Dataset	Settings	S1 $<$ S0	Selected $\geq$ S0	Percentage	S0/S1/gate	Median saved
BoolQ	28	20	25	89.3	1/3/24	56.5
GSM8K	28	10	25	89.3	6/11/11	30.0
Corr2Cause	26	9	22	84.6	3/8/15	2.0

Table 4: Aggregates recomputed from the same 82 rows as Appendix F . The policy mix counts recorded choices; selected accuracy is not the maximum of the three evaluation accuracies. Savings are percentages of additional Stage 1 generated tokens, with medians computed from the source’s whole-percentage values.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Evaluation	Size	Development
BoolQ	validation	3,270	train: 9,427
GSM8K	test	1,319	train: 7,473
Corr2Cause	test	2,246	dev: 2,246

Appendix

Table 5: Nominal dataset splits. These sizes do not certify the denominator of a summary-only gate result.

Model	Acc0	Applied acc.	Candidate acc.	KEEP count	CHANGE count	Failed-change count
Llama-3.1-8B-Instruct	17.5	43.0	44.8	568	731	20
Qwen2.5-32B-Instruct	37.9	37.1	0.0	586	69	664
Olmo-3.1-32B-Instruct	47.5	48.6	26.8	1225	26	68

Appendix

Table 6: Answer-selection sensitivity on representative GSM8K verifier-style runs. “Applied acc.” follows the answer-selection rule, whereas “Candidate acc.” scores the parsed Stage 1 candidate directly when present. This is a protocol diagnostic, not evidence that parser or action-format artifacts are semantic ISC failures. Denominators may differ slightly across columns because each rule has its own parseability condition.

Feature group	Examples	Role in the learned gate
Parser and final-answer format	parse OK, final present, final count, final-line length	Captures whether the initial answer satisfies the task contract and whether refinement may mainly repair formatting or extraction risk.
Prompt and generation metadata	prompt length, prompt truncation, generation length, post-final truncation	Records length and truncation conditions visible after Stage 0 that may signal unreliable or incomplete initial outputs.
Generation log-probability summaries	mean/min/last log-probability, log-probability standard deviation	Measures confidence and instability in the emitted Stage 0 text when token log-probabilities are available.
Binary-answer confidence features	yes/no confidence, entropy, logit margin, probability gap	Used for BoolQ and Corr2Cause to estimate whether the initial yes/no answer is uncertain enough to justify refinement.
Integer-answer candidate features	integer confidence, entropy, logit margin, number of candidates	Used for GSM8K when candidate scores are available to capture uncertainty over numeric final answers.
Parsed-answer consistency features	parsed/scored answer match, signed margin, parsed integer is best	Checks whether the parsed answer agrees with the model’s scored candidate preference under the same Stage 0 prompt.

Appendix

Table 7: Feature groups used by the reported learned gates. Examples are shortened aliases for the corresponding implementation fields. All features are pre-Stage 1 signals extracted from the initial generation or its scoring pass. The reported gates do not use post-Stage 1 agreement, refined-answer confidence, trace completion, or gold-label diagnostics.

Dataset	Model	Gated accuracy	Accepted $\mathrm{W}\!\to\!\mathrm{C}$	Accepted $\mathrm{C}\!\to\!\mathrm{W}$
BoolQ	Qwen2.5-7B	86.0	276	106
BoolQ	gpt-oss-20b	85.1	121	35
BoolQ	Olmo-3.1-32B	87.8	81	17
BoolQ	Phi-4-mini-reasoning	83.6	80	38
BoolQ	Phi-4-reasoning-plus	88.2	8	2
BoolQ	Qwen2.5-72B	89.0	5	3

Appendix

Table 8: Separate accepted-transition summaries underlying Figure 4 . Gated accuracy is a percentage. Coverage is withheld because the supplied summary contains inconsistent endpoints; source values and automatic checks are retained in the accompanying CSV files.

Dataset	Feature group	Runs	$>$ S0	$>$ both	Gain	Saved
BoolQ	conf/logprob	17	12	10	+0.4	69.2
BoolQ	parser/format	23	19	15	+0.4	54.8
BoolQ	trace/protocol	13	9	4	+0.1	76.4
BoolQ	full gate	25	21	19	+0.5	57.2
GSM8K	conf/logprob	14	9	4	+0.1	50.5
GSM8K	parser/format	18	14	2	+0.6	0.1

Appendix

Table 9: Feature-restricted baseline comparisons. Panels differ in coverage and predictor class, so these are not controlled ablations. Gain is the median gated accuracy improvement over Stage 0 on the reported evaluation data, in percentage points. Saved is the median replay estimate of generated Stage 1 tokens avoided, relative to always using refinement. Feature-only rows use simple one-dimensional thresholds selected from the training split, while “full gate” reports the selected tabular gate suite.

Model	Task	Net benefit 95% CI	Harm rate 95% CI
Olmo-3.1-32B-Instruct	BoolQ	[+1.0, +2.4]	[1.0, 1.9]
Llama-3.1-70B-Instruct	BoolQ	[-2.1, -0.1]	[4.6, 6.3]
Llama-3.1-8B-Instruct	GSM8K	[+22.9, +28.4]	[14.2, 24.3]
DeepSeek-R1-Distill-Qwen-14B	GSM8K	[+4.6, +8.4]	[3.0, 5.5]
DeepSeek-R1-Distill-Qwen-7B	Corr2Cause	[+9.6, +12.7]	[2.5, 4.3]
Phi-4-reasoning-plus	Corr2Cause	[+5.9, +8.7]	[2.5, 4.4]

Appendix

Table 10: Bootstrap intervals for the representative transition rows. Net benefit and harm rates are expressed in percentage points. Intervals are resampled from observed transition counts and do not include run-selection uncertainty.

Figure 5: Original BoolQ diagnostic panel, using its separate summary populations. This is not the 28-setting retained policy panel.

Figure 6: Original Corr2Cause diagnostic panel. Run populations and model coverage differ from the retained results.

Figure 7: Original GSM8K diagnostic panel. Its large Llama-3.1-70B drop must not be equated with the retained run’s positive change.

Model	$N$	Initial	Revised	Gate	Selected policy	Selected accuracy	Saved tokens (%)
DS-R1-Llama-70B	—	87.65	87.50	87.20	gate	87.20	20
DS-R1-Llama-8B	—	85.37	83.62	83.62	S1	83.62	0
DS-R1-Qwen-14B	—	87.16	87.09	88.13	gate	88.13	56
DS-R1-Qwen-32B	—	88.90	88.07	89.33	gate	89.33	57
DS-R1-Qwen-7B	—	77.77	74.80	78.81	gate	78.81	52
Devstral-2-123B-2512	—	91.41	30.49	—	S0	91.41	100

Appendix

Table 11: BoolQ: complete summary panel. Accuracies and savings are percentages.

Model	$N$	Initial	Revised	Gate	Selected policy	Selected accuracy	Saved tokens (%)
DS-R1-Llama-70B	—	89.20	90.80	90.80	S1	90.80	0
DS-R1-Llama-8B	—	54.59	59.29	59.36	S1	59.29	0
DS-R1-Qwen-14B	—	79.53	86.05	85.90	S1	86.05	0
DS-R1-Qwen-32B	—	87.19	90.30	89.84	S1	90.30	0
DS-R1-Qwen-7B	—	72.25	77.79	77.56	S1	77.79	0
Devstral-2-123B-2512	—	39.20	39.20	39.20	gate	39.20	88

Appendix

Table 12: GSM8K: complete summary panel. Accuracies and savings are percentages.

Model	$N$	Initial	Revised	Gate	Selected policy	Selected accuracy	Saved tokens (%)
DS-R1-Llama-8B	—	82.63	83.25	83.56	gate	83.56	98
DS-R1-Qwen-14B	—	80.54	81.70	81.66	S1	81.70	0
DS-R1-Qwen-32B	—	80.94	81.12	81.12	S1	81.12	0
DS-R1-Qwen-7B	—	66.61	77.78	71.28	S1	77.78	0
Devstral-2-123B-2512	—	78.05	40.03	80.19	gate	80.19	74
Llama-3.1-70B	—	84.64	82.72	84.64	gate	84.64	100

Appendix

Table 13: Corr2Cause: complete summary panel. Accuracies and savings are percentages.

LLM Self-Correction

Large language model self-correction studies how language models detect and repair errors in their own answers, reasoning, and actions.

See all 159 papers

When Does Intrinsic Self-Correction Help? A Task-Sensitive Analysis
Jun 22, 2026Elroy Stav, Dvir Berlowitz, Maayan Orner +1LLM Self-Correction LLM Evaluation
Department of Computer Science and Artificial Intelligence, Bar-Ilan University · Independent Researcher
Self-Correction as Feedback Control: Error Dynamics, Stability Thresholds, and Prompt Interventions in LLMs
Apr 24, 2026Aofan Liu, Jingxiang MengLLM Self-Correction Language Model-Based Control
1Peking University. · University of Chicago.
The Self-Correction Illusion: LLMs Correct Others but Not Themselves
Jun 4, 2026Kuan-Yen Chen, Fang-Yi Su, Jung-Hsien ChiangLLM Self-Correction LLM Prompting
National Cheng Kung University

Jun 22, 2026J/K move · Enter open · S save

When Does Intrinsic Self-Correction Help? A Task-Sensitive Analysis

Elroy Stav, Dvir Berlowitz, Maayan Orner, Sarit Kraus

Department of Computer Science and Artificial Intelligence, Bar-Ilan University · Independent Researcher

LLM Self-Correction LLM Evaluation LLM Self-Refinement Metacognition in Language Models

Open paper AskPDF arXiv

Intrinsic self-correction (SC) aims to improve large language model outputs by prompting a model to revisit its own initial answer without external feedback. Recent studies have questioned the reliability of this approach, showing that models often struggle to judge whether their initial responses are correct. In this work, we take a task-sensitive view of SC. Rather than asking whether it works in general, we examine settings where SC may operate through different mechanisms: verifying explicit constraints, revisiting a complex reasoning process, or providing a second opinion over competing strategies in word-game tasks. Across multiple benchmarks and models, we find that SC can yield consistent performance gains when the underlying task structure facilitates these modes of revision. These results suggest that SC is best understood as a task-dependent inference-time strategy whose usefulness depends on the role the revision stage can play in a given task, rather than as a uniformly reliable method for improving initial model outputs.

Abstract

LLM Self-Correction

Explore similar work

When Does Intrinsic Self-Correction Help? A Task-Sensitive Analysis

Self-Correction as Feedback Control: Error Dynamics, Stability Thresholds, and Prompt Interventions in LLMs

The Self-Correction Illusion: LLMs Correct Others but Not Themselves