Reflection or Re-Generation? Why LLM Revision Fails Where Human Revision Succeeds

# Reflection or Re-Generation? Why LLM Revision Fails Where Human Revision Succeeds
Source: [https://arxiv.org/html/2607.28908](https://arxiv.org/html/2607.28908)
Yefan Tao, Gerald Friedland, Madhusudhanan Chandrasekaran, Luyang Kong Amazon \{tayefan, gfriedla, chandmad, luyankon\}@amazon\.com

###### Abstract

Reflection, the ability to revisit and revise prior reasoning, is central to how humans improve their own answers\. Large language models \(LLMs\) are increasingly prompted to “reflect,” yet whether this process resembles human revision remains unclear\. We introduce the Human–LLM Reflection Framework \(HRF\), a controlled two\-pass protocol that compares human and LLM revision behavior under identical conditions across self\-revision, peer\-revision, and cross\-agent settings\. Using an information\-theoretic analysis grounded in per\-iteration cross\-entropy reduction, we identify two distinct failure modes of LLM reflection\. On objective reasoning tasks with finite answer spaces, reflection yields near\-zero information gain \(Δ​I≈0\\Delta I\\approx 0\), behaving as neutral re\-generation indistinguishable from independent re\-sampling\. On subjective evaluation tasks, reflection produces statistically significant negative information gain \(Δ​I<0\\Delta I<0\), consistently moving predictions away from the target\. Human revision, by contrast, yields positive information gain across both settings\. Cross\-agent experiments localize the failure to the revision step rather than input quality: LLMs degrade even high\-quality human first\-pass responses\. Diagnostic analyses—revision conditioned on first\-pass correctness, and oracle\-guided revision against a random\-reshuffle baseline—show that which sub\-step dominates the failure varies by task and by model rather than reducing to a single mechanism: self\-error detection is present on objective multiple\-choice tasks but weak on subjective ones, and recovery under an oracle error signal exceeds a random\-reshuffle baseline for some models and falls below it for others\. The unifying account is structural: without external information, self\-conditioned revision cannot reduce uncertainty about the target, so LLM reflection is better understood as conditioned re\-generation than as genuine error\-driven revision\.

## 1Introduction

The ability to reconsider and improve prior reasoning is a hallmark of human cognition\. In educational psychology, this capacity is studied under the framework of metacognition—the monitoring and regulation of one’s own cognitive processesDunlosky and Metcalfe \([2008](https://arxiv.org/html/2607.28908#bib.bib1)\)\. A key insight from this literature is that effective revision involves two separable components:*error detection*\(recognizing that something is wrong\) and*error correction*\(producing a better response\)Nelson \([1990](https://arxiv.org/html/2607.28908#bib.bib2)\); Yeung and Summerfield \([2012](https://arxiv.org/html/2607.28908#bib.bib3)\)\. This distinction matters because it implies that the ability to generate correct answers does not, by itself, guarantee the ability to identify and fix errors in existing ones\.

Large language models \(LLMs\) have been prompted to engage in analogous behavior\. Reflection\-based methods instruct a model to revisit a previously generated response and decide whether to retain or revise itShinnet al\.\([2023](https://arxiv.org/html/2607.28908#bib.bib4)\); Jiet al\.\([2023](https://arxiv.org/html/2607.28908#bib.bib5)\); Renze and Guven \([2024](https://arxiv.org/html/2607.28908#bib.bib6)\)\. While some studies report improvements from such prompting, particularly on reasoning tasks, others find that LLMs frequently fail to self\-correct and may even degrade their own outputs without external feedbackHuanget al\.\([2023](https://arxiv.org/html/2607.28908#bib.bib7)\); Liet al\.\([2024b](https://arxiv.org/html/2607.28908#bib.bib8)\); Kamoiet al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib9)\)\. This tension raises a question that existing evaluations have not resolved: when LLMs are prompted to reflect, are they engaging in genuine error correction, or merely re\-generating responses conditioned on their prior output?

We argue that resolving this question requires two ingredients that prior work has largely lacked\. First, a*controlled comparison with human revision under identical conditions*, to establish an empirical reference for what effective reflection looks like\. Existing human–LLM comparisons have examined single\-pass responses across domains such as moral judgmentYaxet al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib10)\), theory of mindJoneset al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib11)\), and factual reasoningKamoiet al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib9)\), but have not studied the*dynamics*of revision—how answers change from a first attempt to a second\. Second, an*information\-theoretic characterization of revision dynamics*, to move beyond aggregate accuracy comparisons and ask whether each revision step accumulates task\-relevant information or merely reshuffles the model’s output distribution\.

To this end, we introduce the Human–LLM Reflection Framework \(HRF\): a controlled two\-pass protocol in which both human annotators and LLMs produce an initial response \(Pass 1\) and then revise it with access to the earlier output \(Pass 2\)\. HRF supports three revision settings—self\-revision, peer\-revision among agents of the same type, and cross\-agent revision in which humans revise LLM outputs and vice versa—enabling us to isolate whether failures stem from the revision mechanism itself or from the quality of the input being revised\.

We instantiate HRF on two tasks with discrete, finite answer spaces that enable rigorous information\-theoretic analysis: MalAlgoQASonkaret al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib12)\), a 4\-class mathematical reasoning benchmark, and IMDb\-Rating, a 10\-class sentiment evaluation task\. We complement these with TISERBazagaet al\.\([2025](https://arxiv.org/html/2607.28908#bib.bib13)\), an open\-ended temporal reasoning benchmark used for behavioral generalization\. We evaluate five representative LLMs spanning multiple families—Llama\-3\.1\-405B, Claude\-3\.5\-Sonnet, Mistral\-Large\-0724, GPT\-4o\-20240806, and DeepSeek\-R1—alongside non\-expert human annotators, under identical prompts, inputs, and evaluation criteria\. Two additional models \(GPT\-5\-mini and GPT\-5\.2\) are evaluated in Appendix[G](https://arxiv.org/html/2607.28908#A7)to assess generalization across model generations\.

Our central finding is that LLM reflection exhibits two distinct failure modes, unified by a common information\-theoretic explanation\. Our contributions are:

1. 1\.HRF, a controlled two\-pass protocol enabling matched comparison of human and LLM revision behavior across self\-revision, peer\-revision, and cross\-agent settings\.
2. 2\.A characterization of LLM self\-revision as conditioned re\-generation\.Under matched conditions, self\-revision does not accumulate information about the target: information gain is neutral on objective tasks \(Δ​I≈0\\Delta I\\approx 0\) and negative on the subjective task \(Δ​I<0\\Delta I<0\), while human revision is positive under identical conditions\. We give a unifying information\-theoretic account whose central content is a falsifiable prediction: without external information, self\-conditioning cannot reduceH​\(Y∣⋅\)H\(Y\\mid\\cdot\)\.
3. 3\.Diagnostic analyses of where revision fails\.Cross\-agent matrices localize the failure to the revision step rather than input quality\. Revision conditioned on first\-pass correctness and oracle\-guided revision against a random\-reshuffle baseline show that the dominant sub\-step—detecting which answers need revision versus producing a better one—varies by task and by model; we report this heterogeneity rather than attributing the failure to a single mechanism\.

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/hrf_only_v2.png)Figure 1:The Human–LLM Reflection Framework \(HRF\)\. Human and LLM annotators independently produce first\-pass responses; a reflection prompt then presents a prior response for optional revision in the second pass\.
## 2Framework and Experimental Setup

We introduce the Human–LLM Reflection Framework \(HRF\), a controlled evaluation protocol for studying reflective revision in humans and LLMs under identical conditions\. The overall pipeline is illustrated in Figure[1](https://arxiv.org/html/2607.28908#S1.F1)\.

### 2\.1Two\-Pass Protocol and Revision Settings

HRF follows a two\-pass design\. In the first pass, an annotator—human or LLM—produces an initial response based solely on the task input\. In the second pass, the same task is presented together with a previously generated first\-pass response, and the annotator decides whether to retain or revise it\. No additional information beyond the original input and the supplied prior response is provided, ensuring that any performance change is attributable to the revision process itself\.

Based on the source of the prior response, HRF distinguishes three settings\. In*self\-revision*, the annotator revises their own first\-pass output\. In*peer\-revision*, the prior response is produced by a different annotator of the same type, with first\-pass responses randomly reassigned\. In*cross\-agent revision*, humans revise LLM outputs and vice versa\. The cross\-agent setting serves a diagnostic purpose \(Section[5\.2](https://arxiv.org/html/2607.28908#S5.SS2)\): if reflection failures stem from poor first\-pass quality, providing a higher\-quality source should help; if the revision mechanism itself is deficient, performance should degrade regardless of source\.

### 2\.2Tasks and Datasets

We evaluate three tasks; the first two have discrete, finite answer spaces enabling the information\-theoretic analysis in Section[4](https://arxiv.org/html/2607.28908#S4)\.

#### IMDb\-Rating\.

A sentiment evaluation task pairing movie reviews with ground\-truth ratings on a discrete 1–10 scale \(908 items\)\.

#### MalAlgoQA\(Sonkaret al\.,[2024](https://arxiv.org/html/2607.28908#bib.bib12)\)\.

A 4\-choice mathematical reasoning benchmark \(709 questions, shuffled options\)\.

#### TISER\(Bazagaet al\.,[2025](https://arxiv.org/html/2607.28908#bib.bib13)\)\.

A temporal reasoning benchmark requiring consistency under evolving constraints \(250 samples\); because responses are free\-form, we use TISER for behavioral analysis only, with correctness assessed by a fixed LLM judge\(Alexandruet al\.,[2025](https://arxiv.org/html/2607.28908#bib.bib14)\)\.

All datasets were released after the training\-data cutoff dates of the models evaluated, minimizing contamination risk\.

### 2\.3Annotators and Evaluation

#### LLMs\.

We evaluate five models: Llama\-3\.1\-405B, Claude\-3\.5\-Sonnet, Mistral\-Large\-0724, GPT\-4o\-20240806, and DeepSeek\-R1\. All use identical prompts and strict JSON output constraints; sampling parameters are fixed across conditions\. Two additional models \(GPT\-5\-mini and GPT\-5\.2\) are evaluated in Appendix[G](https://arxiv.org/html/2607.28908#A7)\.

#### Humans\.

Annotations are collected via Amazon SageMaker Ground Truth, targeting 5 annotators per item\. Inter\-annotator agreement \(Krippendorff’sα\\alpha\) is reported in Appendix[A](https://arxiv.org/html/2607.28908#A1)\(Table[3](https://arxiv.org/html/2607.28908#A1.T3)\)\.

#### Metrics\.

IMDb\-Rating uses MSE \(lower is better\); MalAlgoQA and TISER use accuracy \(higher is better\)\. Reflection gain \(RG\) quantifies the relative change from first to second pass:

RG=ϕ​\(y^\(2\)\)−ϕ​\(y^\(1\)\)\|ϕ​\(y^\(1\)\)\|×100%,\\text\{RG\}=\\frac\{\\phi\(\\hat\{y\}^\{\(2\)\}\)\-\\phi\(\\hat\{y\}^\{\(1\)\}\)\}\{\|\\phi\(\\hat\{y\}^\{\(1\)\}\)\|\}\\times 100\\%,\(1\)whereϕ\\phiis the task metric, defined so that higher values indicate better performance\. For group\-level analysis, we aggregate via majority voting\.

## 3Empirical Overview

### 3\.1Reflection Gain Across Tasks

Table[1](https://arxiv.org/html/2607.28908#S3.T1)reports first\-pass performance, second\-pass performance, and reflection gain for all models and human annotators across the three tasks\. Three patterns emerge from this table\.

#### Humans consistently improve; LLMs do not\.

Across all three tasks, human annotators exhibit positive reflection gains:\+17\.8%\+17\.8\\%on IMDb\-Rating,\+6\.0%\+6\.0\\%on MalAlgoQA, and\+4\.7%\+4\.7\\%on TISER \(averages across annotators\)\. In contrast, no LLM achieves consistent positive gains across tasks\. Aggregating multiple LLM responses via majority voting partially stabilizes outcomes but does not close the gap to human performance\.

#### Reflection is actively harmful on subjective tasks\.

On IMDb\-Rating,*every*LLM exhibits negative reflection gain, with degradation ranging from−5\.1%\-5\.1\\%\(Mistral\) to−29\.2%\-29\.2\\%\(Claude\)\. On objective tasks \(MalAlgoQA, TISER\), reflection effects are smaller in magnitude and directionally mixed, with most models showing near\-zero or slightly negative gains\. The qualitative difference between subjective and objective tasks is striking: IMDb\-Rating is the only setting in which reflection*systematically*degrades all models, suggesting a failure mode distinct from the neutral re\-generation observed on objective tasks\.

Table 1:Reflection performance across tasks and passes\. Reflection gain \(RG\) denotes relative improvement from the first to the second pass\. For IMDb\-Rating the metric is MSE \(lower is better\); for MalAlgoQA and TISER it is accuracy \(higher is better\)\.
#### Model strength does not resolve the problem\.

DeepSeek\-R1, the strongest model in our evaluation \(MalAlgoQA first\-pass accuracy0\.9480\.948\), achieves only\+0\.6%\+0\.6\\%reflection gain on MalAlgoQA and−3\.0%\-3\.0\\%on TISER\. The failure of reflection is not confined to weaker models: even the highest\-performing LLM cannot reliably convert reflection into improvement\. Additional experiments with more recent models \(GPT\-5 series\) confirm this pattern \(Appendix[G](https://arxiv.org/html/2607.28908#A7)\)\.

### 3\.2Revision Frequency and Precision

Beyond aggregate accuracy, we examine how often annotators revise and how often those revisions are beneficial\. Detailed self\- and peer\-revision breakdowns are reported in Appendix[C](https://arxiv.org/html/2607.28908#A3)\.

#### Humans revise less but more precisely\.

Humans initiate revisions less frequently than LLMs, but a substantially larger fraction of human revisions are beneficial\. In contrast, LLMs revise more frequently, yet beneficial and harmful revisions occur at comparable rates, so increased revision frequency does not translate into effective reflection\.

#### LLMs show stronger self–peer asymmetry\.

Both humans and LLMs revise peer responses more frequently than their own, but this asymmetry is substantially larger for LLMs\. For LLMs, peer\-revision produces larger but less consistent changes, whereas human revisions remain comparatively stable across both settings\.

These patterns establish that LLM reflection fails, but do not explain why it fails, nor why the failure differs across task types\. The remainder of the paper addresses these questions: Section[4](https://arxiv.org/html/2607.28908#S4)introduces the information\-theoretic framework, Section[5\.1](https://arxiv.org/html/2607.28908#S5.SS1)identifies two distinct failure modes, and Section[5\.2](https://arxiv.org/html/2607.28908#S5.SS2)diagnoses their cause\.

## 4Information\-Theoretic Framework

To explain why LLM reflection fails—and why it fails differently across task types—we model reflection as an information acquisition process and ask whether each revision step reduces uncertainty about the target answer\.

Letxxdenote the task input,y∗y^\{\*\}the target answer, andy^\(1\)\\hat\{y\}^\{\(1\)\}the initial response\. A reflection step produces a revised response by conditioning on the prior answer:

y^\(2\)∼pθ​\(y∣x,y^\(1\)\)\.\\hat\{y\}^\{\(2\)\}\\sim p\_\{\\theta\}\(y\\mid x,\\hat\{y\}^\{\(1\)\}\)\.\(2\)From an information\-theoretic perspective, reflection is effective only if this conditioning reduces uncertainty abouty∗y^\{\*\}\. We formalize this via the per\-iteration information gain at revision stepkk:

Δ​I\(k\)=I​\(y∗;y^\(k\)∣x,y^\(k−1\)\),\\Delta I^\{\(k\)\}=I\(y^\{\*\};\\,\\hat\{y\}^\{\(k\)\}\\mid x,\\,\\hat\{y\}^\{\(k\-1\)\}\),\(3\)where positive values indicate that revision brings the output closer to the target and negative values indicate that it moves further away\. Effective reflection corresponds to sustained positiveΔ​I\(k\)\\Delta I^\{\(k\)\}across iterations\.

Becausey∗y^\{\*\}is not directly observable, we estimateΔ​I\(k\)\\Delta I^\{\(k\)\}using cross\-entropy reduction\. For each revision stepkk, we compute the cross\-entropyH​\(y∗∣p\(k\)\)H\(y^\{\*\}\\mid p^\{\(k\)\}\)between the model’s predicted answer distributionp\(k\)p^\{\(k\)\}and the one\-hot ground\-truth labely∗y^\{\*\}, wherep\(k\)p^\{\(k\)\}is estimated fromK=10K=10independent runs under identical decoding settings\. The empirical information gain is then

Δ​Iemp\(k\)=H​\(y∗∣p\(k−1\)\)−H​\(y∗∣p\(k\)\)\.\\Delta I^\{\(k\)\}\_\{\\mathrm\{emp\}\}=H\(y^\{\*\}\\mid p^\{\(k\-1\)\}\)\-H\(y^\{\*\}\\mid p^\{\(k\)\}\)\.\(4\)
We are explicit about what this estimator does and does not provide\. With a point\-mass target,H​\(y∗∣p\(k\)\)H\(y^\{\*\}\\mid p^\{\(k\)\}\)reduces to−log2⁡p^\(k\)​\(y∗\)\-\\log\_\{2\}\\hat\{p\}^\{\(k\)\}\(y^\{\*\}\), so on finite\-label tasksΔ​Iemp\(k\)\\Delta I^\{\(k\)\}\_\{\\mathrm\{emp\}\}is a monotone reframing of the change in the empirical probability of the correct answer, and on the ordinal rating task the distance\-aware analogue12​log2⁡\(MSE1/MSE2\)\\tfrac\{1\}\{2\}\\log\_\{2\}\(\\mathrm\{MSE\}\_\{1\}/\\mathrm\{MSE\}\_\{2\}\)is one\-to\-one with the MSE ratio\. It is therefore not an independent source of evidence beyond accuracy and MSE, and we refer to it as*predictive cross\-entropy reduction*rather than claiming a new measure\. Its role is twofold: it places tasks scored by accuracy and by MSE on one bits axis so the two failure modes can be compared, and it yields a falsifiable prediction—without external information, self\-conditioning cannot reduceH​\(Y∣⋅\)H\(Y\\mid\\cdot\)—that the self\-, peer\-, cross\-agent, oracle, and human experiments are designed to test\.

We apply this analysis to the two tasks with discrete answer spaces—MalAlgoQA \(4\-class\) and IMDb\-Rating \(10\-class\)—using matched items throughout\. TISER is excluded because its free\-form answers lack a finite label space, making entropy\-based measures ill\-defined; we rely on behavioral metrics for that task\.

## 5Failure Modes and Diagnosis of LLM Reflection

### 5\.1Two Failure Modes

Figure[2](https://arxiv.org/html/2607.28908#S5.F2)reports per\-model information gainΔ​Iemp\(1\)\\Delta I^\{\(1\)\}\_\{\\mathrm\{emp\}\}on the two tasks with discrete answer spaces, estimated via cross\-entropy reduction overK=10K=10repeated runs\.

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/info_gain_comparison.png)Figure 2:Per\-model information gainΔ​I\\Delta Ion MalAlgoQA \(left, 4\-class\) and IMDb\-Rating \(right, 10\-class\)\. All bars are at or below zero\. On MalAlgoQA,Δ​I\\Delta Iis weakly negative and largely indistinguishable from zero \(neutral re\-generation\)\. On IMDb\-Rating,Δ​I\\Delta Iis substantially negative across all models \(distributional flattening\)\. Error bars show 95% confidence intervals overK=10K=10runs\.Two distinct patterns emerge:

#### Objective tasks: neutral re\-generation \(Δ​I≈0\\Delta I\\approx 0\)\.

On MalAlgoQA, all five models produceΔ​I\\Delta Ivalues that are weakly negative and close to zero\. Confidence intervals either include zero or sit just below it\. Repeated\-run pairedtt\-tests \(K=10K=10\) confirm that self\-revision does not yield statistically significant improvement over stochastic re\-generation for four of five models; the one exception \(Claude\-3\.5\) shows a small but significant*decrease*\(Appendix[D](https://arxiv.org/html/2607.28908#A4)\)\. On TISER, the same statistical analysis reveals a similarly stable pattern: no model achieves significant positive gain \(Appendix[D](https://arxiv.org/html/2607.28908#A4)\)\. Taken together, these results indicate that on objective tasks, conditioning on the first\-pass response introduces no new information abouty∗y^\{\*\}—revision is statistically indistinguishable from independent re\-sampling\.

#### Subjective tasks: distributional flattening \(Δ​I<0\\Delta I<0\)\.

On IMDb\-Rating, the picture is qualitatively different\. All five models exhibit negativeΔ​I\\Delta I, and the effect is substantially larger in magnitude than on MalAlgoQA\. Pairedtt\-tests over repeated runs confirm that all five models show statistically significant MSE increases under reflection compared to re\-generation \(Appendix[D](https://arxiv.org/html/2607.28908#A4)\)\. Unlike the near\-neutral pattern on objective tasks, reflection on IMDb\-Rating moves predictions away from the target: second\-pass predictions are systematically*further*from the ground\-truth rating than first\-pass predictions\. We refer to this pattern as distributional flattening, and Appendix[E](https://arxiv.org/html/2607.28908#A5)provides geometric evidence consistent with predictions drifting toward a model\-internal prior; we note that this is one mechanism consistent with the data \(others, such as variance reduction under conditioning or anchoring on the prior text, are not ruled out\), whereas the negativeΔ​I\\Delta Iitself is a direct measurement\.

#### Interpretation\.

Both failure modes stem from the same limitation: without an external error signal, conditioning on the first\-pass response cannot introduce new information abouty∗y^\{\*\}\. On objective tasks with a small answer space, this produces neutral re\-sampling \(Δ​I≈0\\Delta I\\approx 0\)\. On subjective tasks, where the model’s prior concentrates on a narrower range than the ground truth, uninformative conditioning triggers regression toward default predictions, overwriting first\-pass outputs that had captured informative distributional features \(Δ​I<0\\Delta I<0\)\. Geometric analysis in embedding space corroborates this: LLM second\-pass responses drift toward the model’s distributional prior, whereas human revisions remain localized \(Appendix[E](https://arxiv.org/html/2607.28908#A5)\)\.

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/imdb_reflection_heatmap.png)

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/malalgoqa_reflection_heatmap.png)

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/tiser_reflection_heatmap.png)

Figure 3:Cross\-model reflection matrices across the three evaluation tasks: IMDb\-Rating \(left\), MalAlgoQA \(center\), and TISER \(right\)\. Rows denote reflectors \(second pass\), columns denote first\-pass sources\. Each cell reports the net improvement rate \(beneficial minus harmful revision rate\)\.

### 5\.2The Failure Lies in the Revision Step, Not the Input

The preceding analysis shows that LLM reflection fails to accumulate information\. Two explanations are possible: first\-pass responses may be too poor to build upon, or the revision mechanism may lack the ability to identify which responses need correction\. We examine cross\-agent revision outcomes from the HRF protocol \(Section[2](https://arxiv.org/html/2607.28908#S2)\) to distinguish between these hypotheses\.

Figure[3](https://arxiv.org/html/2607.28908#S5.F3)reports the cross\-model reflection matrices for each task, where rows denote second\-pass revisers and columns denote first\-pass sources\. Each cell reports the net improvement rate: the difference between beneficial and harmful revisions\.

Two patterns emerge\. First,human revisers produce non\-negative net improvement across nearly all source types and tasks\.The human row is consistently non\-negative regardless of whether the first\-pass response comes from a weak model, a strong model, or another human\. This holds even when the source output is substantially stronger than the human’s own first\-pass performance \(e\.g\., LLM outputs on IMDb\-Rating, where LLMs achieve lower MSE than humans in the first pass\)\.

Second,LLM revision outcomes are source\-dependent and dominated by distributional attraction rather than error correction\.On IMDb\-Rating, the LLM–LLM cells are almost universally negative: LLMs degrade each other’s outputs despite those outputs being reasonably strong\. The LLM–Human cells are an apparent exception, showing mostly positive values—but this reflects the fact that human first\-pass predictions on IMDb are weaker than LLM predictions \(Table[1](https://arxiv.org/html/2607.28908#S3.T1)\), so pulling toward the model’s prior happens to reduce error\. The improvement is incidental rather than diagnostic: it tracks distributional proximity, not targeted correction\. On MalAlgoQA, the pattern is similar—only the strongest model \(DeepSeek\-R1\) acts as a broadly effective reviser, and its effectiveness tracks the strength gap with the source rather than error\-specific reasoning\.

Taken together, these results indicate that the failure lies in the revision mechanism, not in source quality\. Human revisers consistently improve outputs regardless of source strength\. LLM revisers, by contrast, shift outputs toward their own distributional prior—a process that helps when the source is weaker but harms when the source is comparable or stronger\. This locates the failure in the revision step itself; it does not yet tell us which sub\-step of revision is responsible\. We turn to that question next, measuring error detection and error correction directly, and find that neither is uniformly the bottleneck: the dominant sub\-step varies by task and by model\.

### 5\.3Error Detection Is Present but Task\-Dependent

To ask whether models detect their own errors, we measure the per\-run revision rate conditioned on first\-pass correctness \(K=10\), i\.e\.P​\(revise∣wrong\)P\(\\text\{revise\}\\mid\\text\{wrong\}\)versusP​\(revise∣correct\)P\(\\text\{revise\}\\mid\\text\{correct\}\)\. A model that detects its errors should revise wrong answers more often than correct ones\. On the objective task MalAlgoQA this ratio is large \(roughly33–21×21\\timesacross models\), so a detection signal clearly exists; on the subjective ordinal task IMDb it is weak \(roughly0\.90\.9–1\.6×1\.6\\times\)\. Detection is therefore task\-dependent, not uniformly absent\. Crucially, where the signal exists it does not translate into improvement: reflection gain on MalAlgoQA remains near zero across all five models \(Table[1](https://arxiv.org/html/2607.28908#S3.T1)\) despite wrong answers being revised far more often\. The model flags candidates for revision but does not reliably fix them—which motivates isolating the correction step directly\.

### 5\.4Error Correction Is Limited and Model\-Dependent

To isolate correction from detection, we introduce oracle\-guided revision: a diagnostic experiment in which the model receives a minimal external signal indicating whether its first\-pass answer was correct, without revealing the correct answer itself\.

For each question, the oracle routes the model into one of two conditions based on first\-pass correctness:

- •Oracle\-wrong: “Your previous answer is incorrect\. Please reconsider the question and provide a revised answer\.”
- •Oracle\-correct: “Your previous answer is correct\. Please confirm or reconsider your answer\.”

This design isolates*error detection*—knowing which answers to revise—from*error correction*—knowing what to revise them to\. The oracle provides exactly one bit of information per item and never reveals the correct answer\.

Table[2](https://arxiv.org/html/2607.28908#S5.T2)reports results on MalAlgoQA averaged over 10 runs\. We report*recovery rate*\(fraction of incorrect first\-pass answers corrected\) and*retention rate*\(fraction of correct first\-pass answers preserved\)\. Because MalAlgoQA is a 4\-choice task, a model that excludes its original answer and guesses uniformly achieves a recovery rate of33\.3%33\.3\\%\.

Table 2:Oracle\-guided revision on MalAlgoQA\. Recovery rate: fraction of incorrect first\-pass answers corrected\. Retention rate: fraction of correct first\-pass answers preserved\. Random re\-guess baseline:0\.3330\.333\. All values averaged over 10 runs\.Oracle\-guided revision improves accuracy over self\-reflection for all five models, with gains ranging from\+1\.1\+1\.1points \(DeepSeek\-R1\) to\+7\.4\+7\.4\(GPT\-4o\)\. But the recovery rate—how often a flagged\-wrong answer is actually corrected—is*model\-dependent*and does not support a uniform “correction is intact” reading\. Recovery exceeds the33\.3%33\.3\\%random\-reshuffle baseline only for the stronger models \(DeepSeek\-R1:41\.3%41\.3\\%, Claude:37\.2%37\.2\\%\), sits at chance for GPT\-4o \(33\.4%33\.4\\%\), and falls well below it for the weaker models \(Mistral:17\.1%17\.1\\%, Llama:16\.8%16\.8\\%\)\. If correction capacity were intact and only detection missing, a flagged\-wrong answer should recover at least at the reshuffle baseline; for most models it does not\. What is consistent across models is*retention*: the oracle reliably prevents correct→\\rightarrowwrong changes \(9797–99\.9%99\.9\\%\), which is where much of its accuracy gain comes from, especially for weaker models\.

Taken with the detection measure above, the picture is one of heterogeneity rather than a single bottleneck\. Detection is present on objective tasks but weak on subjective ones; correction, as unlocked by an oracle signal, is genuine for the strongest models but limited for weaker ones, whose main benefit from the oracle is retention\. We therefore do not localize the failure to a single sub\-step\. The consistent finding is structural: a binary external signal is enough to change behavior at all \(unlike self\-conditioning\), but often insufficient to overwrite a confident wrong prior\. While a perfect oracle is unavailable in practice, approximate error signals are readily obtainable through tool\-based verification, verifier models, or lightweight human feedback\(Cobbeet al\.,[2021](https://arxiv.org/html/2607.28908#bib.bib20); Lightmanet al\.,[2023](https://arxiv.org/html/2607.28908#bib.bib21)\), suggesting that such mechanisms are a more promising investment than increasing reflection depth or prompt complexity\.

Multi\-iteration experiments \(up to 5 passes\) using iterative self\-revision, self\-ensemble, and cross\-model strategies further confirm that incremental information gain collapses to near zero after the first revision step across all configurations \(Appendix[F](https://arxiv.org/html/2607.28908#A6)\)\.

## 6Related Work

### 6\.1Metacognition and Human Revision

The study of how humans monitor and revise their reasoning has a long history in cognitive psychology\.Flavell \([1979](https://arxiv.org/html/2607.28908#bib.bib15)\)introduced the concept of metacognition, andNelson \([1990](https://arxiv.org/html/2607.28908#bib.bib2)\)formalized it as a two\-level system in which a meta\-level process monitors and controls object\-level cognition\. Subsequent work has shown that effective revision depends on monitoring accuracy—the ability to distinguish correct from incorrect responses—rather than on revision volume\(Dunlosky and Metcalfe,[2008](https://arxiv.org/html/2607.28908#bib.bib1); Veenmanet al\.,[2006](https://arxiv.org/html/2607.28908#bib.bib16)\)\. Error monitoring studies further demonstrate that humans detect and signal their own mistakes during task performance, driving adaptive learning\(Yeung and Summerfield,[2012](https://arxiv.org/html/2607.28908#bib.bib3)\)\.

We use human annotators as a behavioral reference for what effective revision looks like under our protocol, not as a claim about human metacognition: our non\-expert crowdworker sample is not designed to measure monitoring capacity in the sense ofNelson \([1990](https://arxiv.org/html/2607.28908#bib.bib2)\)\. At the behavioral level, human annotators in our study produce sparse, targeted revisions \(Section[3\.2](https://arxiv.org/html/2607.28908#S3.SS2)\) with positive information gain across both objective and subjective tasks \(Sections[5\.1](https://arxiv.org/html/2607.28908#S5.SS1)\), whereas LLM revision does not improve under identical conditions\. For LLMs, our detection and oracle experiments \(Sections[5\.3](https://arxiv.org/html/2607.28908#S5.SS3),[5\.4](https://arxiv.org/html/2607.28908#S5.SS4)\) show that which sub\-step of revision is deficient varies by task and by model\.

### 6\.2Reflection and Self\-Correction in LLMs

Reflexion\(Shinnet al\.,[2023](https://arxiv.org/html/2607.28908#bib.bib4)\)uses verbal self\-reflection as a reinforcement signal for sequential decision\-making; subsequent work applies similar ideas to hallucination mitigation\(Jiet al\.,[2023](https://arxiv.org/html/2607.28908#bib.bib5)\)and general problem solving\(Renze and Guven,[2024](https://arxiv.org/html/2607.28908#bib.bib6)\)\. However, the reliability of intrinsic self\-correction remains contested:Huanget al\.\([2023](https://arxiv.org/html/2607.28908#bib.bib7)\)show that LLMs cannot self\-correct reasoning without external feedback, confirmed byKamoiet al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib9)\)andLiet al\.\([2024a](https://arxiv.org/html/2607.28908#bib.bib17)\)\. Our work adds a matched human baseline, characterizes two distinct failure modes \(neutral re\-generation on objective tasks, distributional flattening on subjective tasks\), and shows via direct detection and oracle experiments \(Sections[5\.3](https://arxiv.org/html/2607.28908#S5.SS3),[5\.4](https://arxiv.org/html/2607.28908#S5.SS4)\) that the dominant sub\-step of the failure varies by task and by model rather than reducing to a single bottleneck\. Our use of “revision” follows this prompting literature rather than the formal belief\-revision tradition \(AGM;Alchourrónet al\.,[1985](https://arxiv.org/html/2607.28908#bib.bib22); Darwiche and Pearl,[1997](https://arxiv.org/html/2607.28908#bib.bib23)\), where revision requires triggering information that conflicts with the agent’s prior state; self\-conditioned re\-prompting supplies no such information, which is consistent with our finding that it behaves as re\-generation\.

### 6\.3Human–LLM Comparative Studies

Joneset al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib11)\)evaluate LLMs on theory\-of\-mind tasks calibrated against human performance\.Yaxet al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib10)\)compare human and LLM reasoning on cognitive tasks\.Kamoiet al\.\([2024](https://arxiv.org/html/2607.28908#bib.bib9)\)contrast human and LLM error patterns in self\-correction settings\. These studies evaluate single\-pass responses rather than revision dynamics\. To our knowledge, no prior work has \(i\) compared human and LLM revision under a unified multi\-pass protocol, \(ii\) included cross\-agent revision, or \(iii\) provided an information\-theoretic characterization of revision dynamics\. HRF addresses all three gaps\.

## 7Discussion and Conclusion

#### Reflection as conditioned re\-generation\.

Our information\-theoretic analysis reveals that LLM “reflection” is better understood as conditioned re\-generation: without an external error signal, self\-conditioned revision cannot reduce uncertainty about the target, yielding neutral re\-sampling on objective tasks \(Δ​I≈0\\Delta I\\approx 0\) and distributional flattening on subjective tasks \(Δ​I<0\\Delta I<0\)\.

#### Where revision fails varies by task and model\.

The monitoring\-and\-control framework from cognitive science\(Nelson,[1990](https://arxiv.org/html/2607.28908#bib.bib2)\)distinguishes error detection from error correction\. Our direct measurements show that neither is uniformly the bottleneck\. Self\-error detection is present on objective multiple\-choice tasks \(wrong answers are revised several times more often than correct ones\) but weak on subjective ordinal judgments\. Correction, as unlocked by an oracle error signal, exceeds a random\-reshuffle baseline for the strongest models but falls below it for weaker ones, whose oracle benefit comes mainly from retention\. Rather than a single dissociation, the failure is heterogeneous across tasks and models, unified only by the structural point below\.

#### Implications for system design\.

Our findings suggest three design principles for effective revision systems:

1. 1\.*External verification is necessary\.*Self\-conditioned revision without new information cannot accumulate task\-relevant signal\. Effective reflection requires an independent source of error detection—whether from a separate verifier model, tool\-based checking, or human feedback\.
2. 2\.*Weak–strong asymmetry can be exploited\.*Improvements are more likely when a stronger agent revises a weaker response \(Section[5\.2](https://arxiv.org/html/2607.28908#S5.SS2)\), suggesting that multi\-agent architectures with heterogeneous capabilities may be more effective than single\-model self\-revision\.
3. 3\.*Iteration depth should be bounded\.*Diminishing returns across iterations \(Appendix[F](https://arxiv.org/html/2607.28908#A6)\) imply that computational resources are better allocated to independent sampling or external verification than to repeated self\-conditioning\.

#### Limitations\.

Our information\-theoretic analysis is restricted to tasks with finite answer spaces and reliable ground truth \(MalAlgoQA and IMDb\-Rating\)\. Extending these measures to open\-ended generation remains an open challenge\. Human annotators are non\-expert crowdworkers; expert annotators might exhibit different revision patterns, though the structural differences we observe \(sparse, directional revision versus high\-variance re\-sampling\) are unlikely to depend on domain expertise alone\. Our second\-pass prompt asks the model to agree or disagree with the prior answer, which may itself encourage revision\. As a control, we re\-ran the IMDb second pass with a neutral prompt that asks the model to revise only upon identifying a specific error; this sharply reduces the revision rate and attenuates, though does not eliminate, the negative reflection gain, indicating that part of the IMDb effect is prompt\-induced disagreement while part reflects genuine drift away from the target\. This also means our revision\-rate magnitudes are prompt\-dependent and should be read as behavior under an agree\-or\-disagree instruction rather than as intrinsic constants\. Finally, our oracle experiments use a binary error signal; more nuanced forms of feedback \(e\.g\., identifying the type of error\) may yield additional insights\.

#### Conclusion\.

We introduced the Human–LLM Reflection Framework \(HRF\) and identified two failure modes of LLM reflection through per\-iteration information\-theoretic analysis: neutral re\-generation on objective tasks \(Δ​I≈0\\Delta I\\approx 0\) and distributional flattening on subjective tasks \(Δ​I<0\\Delta I<0\)\. Human revision, by contrast, yields positive information gain across both settings\. Diagnostic experiments localize the failure to the revision step rather than input quality, and show that which sub\-step dominates—detecting which answers to revise versus producing a better one—varies by task and by model rather than reducing to a single mechanism\. The unifying point is structural: without external information, self\-conditioned revision cannot reduce uncertainty about the target\. These results reframe LLM reflection as conditioned re\-generation and point toward revision architectures grounded in external verification rather than self\-conditioned prompting\.

## LLM Usage Disclosure

Large language models are central to this work in two capacities\. First, as experimental subjects: five models \(Llama\-3\.1\-405B, Claude\-3\.5\-Sonnet, Mistral\-Large\-0724, GPT\-4o\-20240806, and DeepSeek\-R1, with GPT\-5\-mini and GPT\-5\.2 in supplementary experiments\) were evaluated under the HRF protocol to study reflection behavior\. Second, as an evaluation tool: a fixed LLM judge\(Alexandruet al\.,[2025](https://arxiv.org/html/2607.28908#bib.bib14)\)was used to assess correctness of free\-form responses on the TISER benchmark, as described in Section[2](https://arxiv.org/html/2607.28908#S2)\.

## References

- On the logic of theory change: partial meet contraction and revision functions\.The Journal of Symbolic Logic50\(2\),pp\. 510–530\.Cited by:[§6\.2](https://arxiv.org/html/2607.28908#S6.SS2.p1.1)\.
- A\. Alexandru, A\. Calvi, H\. Broomfield, J\. Golden, K\. Dai, M\. Leys, M\. Burger, M\. Bartolo, R\. Engeler, S\. Pisupati,et al\.\(2025\)Atla selene mini: a general purpose evaluation model\.arXiv preprint arXiv:2501\.17195\.Cited by:[§2\.2](https://arxiv.org/html/2607.28908#S2.SS2.SSS0.Px3.p1.1),[LLM Usage Disclosure](https://arxiv.org/html/2607.28908#Sx1.p1.1)\.
- A\. Bazaga, R\. Blloshmi, B\. Byrne, and A\. de Gispert \(2025\)Learning to reason over time: timeline self\-reflection for improved temporal reasoning in language models\.arXiv preprint arXiv:2504\.05258\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p5.1),[§2\.2](https://arxiv.org/html/2607.28908#S2.SS2.SSS0.Px3)\.
- R\. R\. Bouckaert and E\. Frank \(2004\)Evaluating the replicability of significance tests for comparing learning algorithms\.InPacific\-Asia conference on knowledge discovery and data mining,pp\. 3–12\.Cited by:[Appendix D](https://arxiv.org/html/2607.28908#A4.p1.7)\.
- K\. Cobbe, V\. Kosaraju, M\. Bavarian, M\. Chen, H\. Jun, L\. Kaiser, M\. Plappert, J\. Tworek, J\. Hilton, R\. Nakano,et al\.\(2021\)Training verifiers to solve math word problems\.arXiv preprint arXiv:2110\.14168\.Cited by:[§5\.4](https://arxiv.org/html/2607.28908#S5.SS4.p6.1)\.
- A\. Darwiche and J\. Pearl \(1997\)On the logic of iterated belief revision\.Artificial Intelligence89\(1\-2\),pp\. 1–29\.Cited by:[§6\.2](https://arxiv.org/html/2607.28908#S6.SS2.p1.1)\.
- T\. G\. Dietterich \(1998\)Approximate statistical tests for comparing supervised classification learning algorithms\.Neural computation10\(7\),pp\. 1895–1923\.Cited by:[Appendix D](https://arxiv.org/html/2607.28908#A4.p1.7)\.
- J\. Dunlosky and J\. Metcalfe \(2008\)Metacognition: a textbook for cognitive, educational, life span & applied psychology\.\.SAGE Publications Ltd \(CA\)\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p1.1),[§6\.1](https://arxiv.org/html/2607.28908#S6.SS1.p1.1)\.
- J\. H\. Flavell \(1979\)Metacognition and cognitive monitoring: a new area of cognitive–developmental inquiry\.\.American psychologist34\(10\),pp\. 906\.Cited by:[§6\.1](https://arxiv.org/html/2607.28908#S6.SS1.p1.1)\.
- J\. Huang, X\. Chen, S\. Mishra, H\. S\. Zheng, A\. W\. Yu, X\. Song, and D\. Zhou \(2023\)Large language models cannot self\-correct reasoning yet\.arXiv preprint arXiv:2310\.01798\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p2.1),[§6\.2](https://arxiv.org/html/2607.28908#S6.SS2.p1.1)\.
- Z\. Ji, T\. Yu, Y\. Xu, N\. Lee, E\. Ishii, and P\. Fung \(2023\)Towards mitigating llm hallucination via self reflection\.InFindings of the Association for Computational Linguistics: EMNLP 2023,pp\. 1827–1843\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p2.1),[§6\.2](https://arxiv.org/html/2607.28908#S6.SS2.p1.1)\.
- C\. R\. Jones, S\. Trott, and B\. Bergen \(2024\)Comparing humans and large language models on an experimental protocol inventory for theory of mind evaluation \(epitome\)\.Transactions of the Association for Computational Linguistics12,pp\. 803–819\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p3.1),[§6\.3](https://arxiv.org/html/2607.28908#S6.SS3.p1.1)\.
- R\. Kamoi, Y\. Zhang, N\. Zhang, J\. Han, and R\. Zhang \(2024\)When can llms actually correct their own mistakes? a critical survey of self\-correction of llms\.Transactions of the Association for Computational Linguistics12,pp\. 1417–1440\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p2.1),[§1](https://arxiv.org/html/2607.28908#S1.p3.1),[§6\.2](https://arxiv.org/html/2607.28908#S6.SS2.p1.1),[§6\.3](https://arxiv.org/html/2607.28908#S6.SS3.p1.1)\.
- M\. Li, L\. Chen, J\. Chen, S\. He, J\. Gu, and T\. Zhou \(2024a\)Selective reflection\-tuning: student\-selected data recycling for llm instruction\-tuning\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 16189–16211\.Cited by:[§6\.2](https://arxiv.org/html/2607.28908#S6.SS2.p1.1)\.
- Y\. Li, C\. Yang, and A\. Ettinger \(2024b\)When hindsight is not 20/20: testing limits on reflective thinking in large language models\.arXiv preprint arXiv:2404\.09129\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p2.1)\.
- H\. Lightman, V\. Kosaraju, Y\. Burda, H\. Edwards, B\. Baker, T\. Lee, J\. Leike, J\. Schulman, I\. Sutskever, and K\. Cobbe \(2023\)Let’s verify step by step\.InThe twelfth international conference on learning representations,Cited by:[§5\.4](https://arxiv.org/html/2607.28908#S5.SS4.p6.1)\.
- T\. O\. Nelson \(1990\)Metamemory: a theoretical framework and new findings\.InPsychology of learning and motivation,Vol\.26,pp\. 125–173\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p1.1),[§6\.1](https://arxiv.org/html/2607.28908#S6.SS1.p1.1),[§6\.1](https://arxiv.org/html/2607.28908#S6.SS1.p2.1),[§7](https://arxiv.org/html/2607.28908#S7.SS0.SSS0.Px2.p1.1)\.
- M\. Renze and E\. Guven \(2024\)Self\-reflection in llm agents: effects on problem\-solving performance\.arXiv preprint arXiv:2405\.06682\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p2.1),[§6\.2](https://arxiv.org/html/2607.28908#S6.SS2.p1.1)\.
- N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao \(2023\)Reflexion: language agents with verbal reinforcement learning\.InProceedings of the 37th International Conference on Neural Information Processing Systems,pp\. 8634–8652\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p2.1),[§6\.2](https://arxiv.org/html/2607.28908#S6.SS2.p1.1)\.
- S\. Sonkar, N\. Liu, M\. Le, and R\. Baraniuk \(2024\)Malalgoqa: pedagogical evaluation of counterfactual reasoning in large language models and implications for ai in education\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 15554–15567\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p5.1),[§2\.2](https://arxiv.org/html/2607.28908#S2.SS2.SSS0.Px2)\.
- M\. V\. Veenman, B\. H\. Van Hout\-Wolters, and P\. Afflerbach \(2006\)Metacognition and learning: conceptual and methodological considerations\.Metacognition and learning1\(1\),pp\. 3–14\.Cited by:[§6\.1](https://arxiv.org/html/2607.28908#S6.SS1.p1.1)\.
- N\. Yax, H\. Anlló, and S\. Palminteri \(2024\)Studying and improving reasoning in humans and machines\.Communications Psychology2\(1\),pp\. 51\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p3.1),[§6\.3](https://arxiv.org/html/2607.28908#S6.SS3.p1.1)\.
- N\. Yeung and C\. Summerfield \(2012\)Metacognition in human decision\-making: confidence and error monitoring\.Philosophical Transactions of the Royal Society B: Biological Sciences367\(1594\),pp\. 1310–1321\.Cited by:[§1](https://arxiv.org/html/2607.28908#S1.p1.1),[§6\.1](https://arxiv.org/html/2607.28908#S6.SS1.p1.1)\.

## Appendix AAnnotation Details

All human annotators are English\-speaking, based in India, with at least high school–level education\. Across tasks, we employ approximately 10–15 unique annotators per dataset\. The design targets 5 annotators per item, with the final average ranging between 4 and 5 due to occasional missed or invalid submissions\. Quality control includes removal of incomplete submissions and validation of required output formats\.

We built custom annotation UIs using liquid templates in Amazon SageMaker Ground Truth\. Example interfaces are shown in Figures[4](https://arxiv.org/html/2607.28908#A1.F4)–[6](https://arxiv.org/html/2607.28908#A1.F6)\.

#### Inter\-annotator agreement\.

Table[3](https://arxiv.org/html/2607.28908#A1.T3)reports Krippendorff’sα\\alphafor each task and pass\. For IMDb\-Rating, we use the ordinal variant ofα\\alphato account for the ordered 1–10 scale; for MalAlgoQA and TISER, we use the nominal variant\. TISER responses are free\-form; we map them to a ternary label set \(correct/incorrect/unknown\) before computing agreement\. The second\-pass column includes both self\-revision and peer\-revision annotations, roughly doubling the average number of annotations per item\.

Agreement is moderate to high across all settings\. IMDb\-Rating shows the strongest agreement \(α=0\.819\\alpha=0\.819first pass,0\.8920\.892second pass\), consistent with the ordinal structure of the task\. MalAlgoQA exhibits substantial agreement \(α=0\.685→0\.731\\alpha=0\.685\\to 0\.731\) with near\-ceiling majority agreement \(92\.4%→96\.6%92\.4\\%\\to 96\.6\\%\), indicating that most items admit a clear consensus answer\. TISER yields the lowestα\\alphavalues \(0\.529→0\.5600\.529\\to 0\.560\), reflecting the inherent ambiguity of free\-form temporal reasoning even after ternary mapping; this is consistent with human first\-pass accuracy on TISER being the lowest across all three tasks \(Table[1](https://arxiv.org/html/2607.28908#S3.T1)\)\. Across all three tasks, bothα\\alphaand majority agreement increase from the first to the second pass, suggesting that revision produces mild convergence among annotators—consistent with the positive reflection gains reported in Section[3](https://arxiv.org/html/2607.28908#S3)\.

Table 3:Inter\-annotator agreement \(Krippendorff’sα\\alpha\) by task and pass\. IMDb\-Rating uses ordinalα\\alpha; MalAlgoQA and TISER use nominalα\\alpha\(TISER responses mapped to correct/incorrect/unknown\)\. Majority% is the fraction of items on which a strict majority of annotators agree\.![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/IMDB_UI.png)Figure 4:Example annotation UI in Amazon SageMaker Ground Truth for the IMDb\-Rating task\.![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/MalAlgoQA_UI.png)Figure 5:Example annotation UI in Amazon SageMaker Ground Truth for the MalAlgoQA task\.![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/TISER_UI.png)Figure 6:Example annotation UI in Amazon SageMaker Ground Truth for the TISER task\.

## Appendix BPrompt Templates

Each task uses a two\-pass prompt structure\. The first\-pass prompt contains the base task instructions\. The second\-pass prompt extends it with a reflection preamble \(shown inblue\) that presents the prior annotator’s response and asks the model to agree or disagree\. All prompts are used identically for human and LLM annotators\.

IMDb\-RatingYou are an experienced movie reviewer\.You will be given an IMDB movie review, along with another annotator’s review rating and explanation\. Carefully read the review, then consider the annotator’s rating and explanation\. Based on your own judgment, either agree or disagree with the annotator’s rating\.Read the provided IMDB movie review carefully and assign a sentiment rating from 1 to 10, where 1 means extremely negative and 10 means extremely positive\. Write a short comment \(1\-\-2 sentences\) explaining your reasoning\.Do not simply copy the annotator’s explanation\. Agreement is not required\.Respond in this JSON format:\{"rating": <integer\>, "comment": "<your explanation\>"\}

MalAlgoQAYou are an assistant for a math QA evaluation task\. You will be given a math question, four answer choices \(A\-\-D\), four rationales \(R1\-\-R4\), and a selected answer \(e\.g\., ‘B’\), and another annotator’s chosen rationale with explanation\.Your task is toread the question and selected answer yourself, then consider the annotator’s rationale and explanation\. Based on your own judgment, either agree or disagree\.Identify which rationale \(R1, R2, R3, or R4\) best matches the given selected answer, and provide a short comment \(1\-\-2 sentences\) explaining your reasoning\.Important:\-\- The selected answer may be incorrect\. Your task is to find the rationale that most likely supports the \*selected\* answer\.\-\- Rationales R1\-\-R4 are not necessarily aligned with answer options A\-\-D by position\. You must determine the correct match based on content\.\-\- Do not simply copy the previous explanation\. Agreement is not required\.Respond in this JSON format:\{"rationale": "<R1\|R2\|R3\|R4\>", "comment": "<brief explanation\>"\}

TISERYou are an assistant for a temporal reasoning evaluation task\. You will be given a context describing events and a question about the temporal order or relationship of those events, and another annotator’s answer with reasoning\.Your task:1\. Carefully read the context and understand the sequence of events\.2\. Review the annotator’s answer and reasoning\.3\. Based on your own judgment, either agree or disagree\. Provide your own answer and reasoning\. Do not simply copy the previous explanation\.Important:\-\- Your answer must be based only on the provided context\.\-\- Agreement with the previous annotator is not required\.\-\- Provide a short explanation \(1\-\-2 sentences\) that makes your reasoning clear\.\-\- If the question is not answerable with the context, put "unknown" in your answer\.Respond in this JSON format:\{"answer": "<your answer\>", "comment": "<your explanation\>"\}

## Appendix CSelf\- vs\. Peer\-Revision Plots

This appendix reports revision outcomes for all models \(including humans\) under two revision types:*self\-revision*\(revising one’s own first\-pass answer\) and*peer\-revision*\(revising an answer produced by another annotator\)\. In each subplot, the x\-axis enumerates models, and the y\-axis reports the fraction of revision outcomes falling into four categories: self\-revision leading to a beneficial change, peer\-revision leading to a beneficial change, self\-revision leading to a harmful change, and peer\-revision leading to a harmful change\.

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/imdb_revision_bar.png)

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/malalgoqa_revision_bar.png)

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/tiser_revision_bar.png)

Figure 7:Self\- and peer\-revision outcomes on IMDb\-Rating \(left\), MalAlgoQA \(center\), and TISER \(right\)\.
## Appendix DStatistical Testing Details

To assess whether reflection provides systematic improvements beyond stochastic decoding variability, we perform repeated\-run statistical testing using paired comparisons\. For each model–task pair, the full evaluation is repeatedK=10K=10times under two second\-pass conditions: \(i\)*re\-generation*, where the model answers the dataset again without access to its first\-pass output, and \(ii\)*reflection*, where the model conditions on its own first\-pass output following the reflection protocol \(Section[2](https://arxiv.org/html/2607.28908#S2)\)\. Each repetition yields a run\-level performance value \(e\.g\., MSE or accuracy\), producing two paired sets ofKKmeasurements\. We compute paired differences

dk=Metricrefl\(k\)−Metricre​\-​gen\(k\),d\_\{k\}=\\mathrm\{Metric\}^\{\(k\)\}\_\{\\mathrm\{refl\}\}\-\\mathrm\{Metric\}^\{\(k\)\}\_\{\\mathrm\{re\\text\{\-\}gen\}\},and apply a standard two\-sided pairedtt\-test\(Dietterich,[1998](https://arxiv.org/html/2607.28908#bib.bib18); Bouckaert and Frank,[2004](https://arxiv.org/html/2607.28908#bib.bib19)\)to evaluate whether conditioning on the first\-pass output produces a systematic performance change relative to re\-generation\. For each model–task pair, we report the mean paired changeΔ\\Delta, the associatedtt\-statistic, two\-sidedpp\-value, and a 95% confidence interval, along with the rejection decision atα=0\.05\\alpha=0\.05\.

Table[4](https://arxiv.org/html/2607.28908#A4.T4)reports the complete results\. Across all settings, we do not observe any case in which reflection yields a statistically significant*positive*improvement over stochastic re\-generation\.

ModelMetricre​\-​gen\\mathrm\{Metric\}\_\{\\mathrm\{re\\text\{\-\}gen\}\}Metricrefl\\mathrm\{Metric\}\_\{\\mathrm\{refl\}\}Δ\\Delta\(%\)ttpp\-valueC​I95%CI\_\{95\\%\}H0H\_\{0\}rej\.?IMDbMetric: MSE \(lower is better↓\\downarrow\)Mistral\-Large1\.7471\.884−\-7\.917\.862\.46×10−82\.46\\\!\\times\\\!10^\{\-8\}\(0\.120, 0\.155\)✓Llama\-3\.1\-405B1\.9672\.240−\-13\.918\.052\.24×10−82\.24\\\!\\times\\\!10^\{\-8\}\(0\.238, 0\.307\)✓Claude\-3\.51\.5922\.048−\-28\.754\.031\.28×10−121\.28\\\!\\times\\\!10^\{\-12\}\(0\.437, 0\.475\)✓GPT\-4o1\.7422\.013−\-15\.519\.111\.36×10−81\.36\\\!\\times\\\!10^\{\-8\}\(0\.239, 0\.303\)✓DeepSeek\-R11\.8141\.958−\-7\.98\.082\.05×10−52\.05\\\!\\times\\\!10^\{\-5\}\(0\.104, 0\.184\)✓MalAlgoQAMetric: Accuracy \(higher is better↑\\uparrow\)Mistral\-Large0\.7380\.742\+0\.62\.037\.26×10−27\.26\\\!\\times\\\!10^\{\-2\}\(−\-0\.001, 0\.009\)✗Llama\-3\.1\-405B0\.7670\.762−\-0\.6−\-0\.724\.87×10−14\.87\\\!\\times\\\!10^\{\-1\}\(−\-0\.020, 0\.010\)✗Claude\-3\.50\.8290\.809−\-2\.4−\-11\.838\.67×10−78\.67\\\!\\times\\\!10^\{\-7\}\(−\-0\.024,−\-0\.016\)✓GPT\-4o0\.7940\.792−\-0\.2−\-0\.436\.76×10−16\.76\\\!\\times\\\!10^\{\-1\}\(−\-0\.009, 0\.006\)✗DeepSeek\-R10\.9570\.954−\-0\.4−\-1\.751\.15×10−11\.15\\\!\\times\\\!10^\{\-1\}\(−\-0\.008, 0\.001\)✗TISERMetric: Accuracy \(higher is better↑\\uparrow\)Mistral\-Large0\.4710\.475\+0\.92\.185\.71×10−25\.71\\\!\\times\\\!10^\{\-2\}\(−\-0\.000, 0\.009\)✗Llama\-3\.1\-405B0\.4940\.490−\-0\.8−\-1\.372\.04×10−12\.04\\\!\\times\\\!10^\{\-1\}\(−\-0\.011, 0\.003\)✗Claude\-3\.50\.5300\.548\+3\.54\.361\.84×10−31\.84\\\!\\times\\\!10^\{\-3\}\(0\.009, 0\.028\)✓GPT\-4o0\.5260\.528\+0\.30\.496\.37×10−16\.37\\\!\\times\\\!10^\{\-1\}\(−\-0\.006, 0\.009\)✗DeepSeek\-R10\.5400\.536−\-0\.7−\-1\.879\.51×10−29\.51\\\!\\times\\\!10^\{\-2\}\(−\-0\.009, 0\.001\)✗Table 4:Paired run\-level statistical tests overK=10K=10repeated runs per model and task\.Δ\\Deltais the paired relative change in percentage points\.C​I95%CI\_\{95\\%\}reports the confidence interval on the raw metric difference\.H0H\_\{0\}rejected indicatesp<0\.05p<0\.05\.
## Appendix EGeometric Evidence for Distributional Attraction

To assess whether reflection produces targeted error correction or instead yields a conditioned variant of a model’s typical output, we analyze reflection as geometric drift in embedding space relative to an item\-conditioned task prior\.

For each questionqq, we obtain a set ofKKunconditioned second\-pass reasoning samples generated without access to the first\-pass response\. We take their mean embedding

μq=1K​∑k=1K𝐞q,regen\(k\)\\mu\_\{q\}=\\frac\{1\}\{K\}\\sum\_\{k=1\}^\{K\}\\mathbf\{e\}^\{\(k\)\}\_\{q,\\mathrm\{regen\}\}as a local approximation of the model’s intrinsic reasoning distribution for that item\. Given the first\-pass reasoning embedding𝐞q\(1\)\\mathbf\{e\}^\{\(1\)\}\_\{q\}and the second\-pass \(reflection\) embedding𝐞q\(2\)\\mathbf\{e\}^\{\(2\)\}\_\{q\}, we compute their cosine distances to the estimated prior,

dq\(1\)=dist​\(𝐞q\(1\),μq\),dq\(2\)=dist​\(𝐞q\(2\),μq\),d^\{\(1\)\}\_\{q\}=\\mathrm\{dist\}\(\\mathbf\{e\}^\{\(1\)\}\_\{q\},\\,\\mu\_\{q\}\),\\qquad d^\{\(2\)\}\_\{q\}=\\mathrm\{dist\}\(\\mathbf\{e\}^\{\(2\)\}\_\{q\},\\,\\mu\_\{q\}\),and define the per\-question drift asΔq=dq\(2\)−dq\(1\)\\Delta\_\{q\}=d^\{\(2\)\}\_\{q\}\-d^\{\(1\)\}\_\{q\}\. Negative values indicate that reflection moves the output closer to the model’s task prior; positive values indicate movement away from it\.

Figure[8](https://arxiv.org/html/2607.28908#A5.F8)visualizes the distributions ofdq\(1\)d^\{\(1\)\}\_\{q\}\(first pass\) anddq\(2\)d^\{\(2\)\}\_\{q\}\(reflection\) for all five LLMs and human annotators on MalAlgoQA\. Each panel reports the mean driftΔ¯\\bar\{\\Delta\}in its title\.

All five LLMs exhibit a consistent pattern: the reflection distribution \(orange\) is shifted toward smaller cosine distances relative to the first\-pass distribution \(blue\), with meanΔ¯\\bar\{\\Delta\}ranging from−0\.113\-0\.113\(GPT\-4o\) to−0\.178\-0\.178\(DeepSeek\-R1\)\. This indicates that second\-pass responses systematically move closer to the model’s own task prior, consistent with distributional attraction rather than targeted error correction\. Notably, DeepSeek\-R1—the strongest model in our evaluation and the one exhibiting the largest apparent reflection gains on MalAlgoQA—shows the most pronounced drift toward its prior, suggesting that even its improvements may reflect regression to a strong default rather than genuine error\-specific revision\.

In contrast, human annotators exhibit a slight*positive*mean drift \(Δ¯=\+0\.086\\bar\{\\Delta\}=\+0\.086\), with the first\-pass and reflection distributions largely overlapping\. Human revisions do not systematically converge toward a shared prior; instead, they remain localized and item\-specific, consistent with selective error correction rather than default\-mode re\-sampling\.

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/malalgoqa_dist_shift_all_models_overlap.png)Figure 8:Distributions of cosine distance to the estimated task prior for first\-pass \(blue\) and reflection \(orange\) responses on MalAlgoQA\. Each panel reports the mean per\-question driftΔ=dq\(2\)−dq\(1\)\\Delta=d^\{\(2\)\}\_\{q\}\-d^\{\(1\)\}\_\{q\}\. All five LLMs show negative drift \(reflection moves toward the prior\); human annotators show slight positive drift, indicating no systematic prior\-aligned movement\.
## Appendix FDiminishing Returns Across Iterations

The main text establishes that a single reflection step yields near\-zero or negative information gain\. A natural follow\-up question is whether*repeated*revision can overcome this ceiling by accumulating small gains across iterations\. We test this by measuring per\-iteration information gain across up to five revision passes on MalAlgoQA under three strategies:*iterative self\-revision*, in which the same model revises its own output repeatedly;*self\-ensemble revision*, in which independent first\-pass samples from the same model serve as prior responses; and*cross\-model peer revision*, in which a different model’s output serves as the prior response\. These strategies are illustrated in Figure[9](https://arxiv.org/html/2607.28908#A6.F9)\.

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/revision_strategy_only.png)Figure 9:Three multi\-iteration revision strategies evaluated in this appendix\.Left:Iterative self\-revision, where model M1 repeatedly revises its own prior output\.Center:Self\-ensemble revision, where M1 generates two independent first\-pass responses and revises by conditioning on both\.Right:Cross\-model peer revision, where M1 revises after seeing a first\-pass response from a different model M2\.Figure[10](https://arxiv.org/html/2607.28908#A6.F10)reports incremental information gain per iteration under both primary–auxiliary model assignments, obtained by swapping the roles of GPT\-4o and DeepSeek\-R1\.

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/info_gain_comparison_gpt.png)

\(a\) Primary: DeepSeek\-R1; auxiliary: GPT\-4o\.

![Refer to caption](https://arxiv.org/html/2607.28908v1/figures/info_gain_comparison_deepseek.png)

\(b\) Primary: GPT\-4o; auxiliary: DeepSeek\-R1\.

Figure 10:Incremental, label\-anchored information gain per revision iteration on MalAlgoQA for iterative self\-revision, self\-ensemble revision, and cross\-model peer revision\.Across all strategies and both assignments, incremental information gain is largely confined to a single revision step and collapses to near zero thereafter\. Iterative self\-revision saturates immediately\. Self\-ensemble and cross\-model revision occasionally produce larger first\-step updates with high variance but no persistence across iterations\. A transient gain appears only when the auxiliary model is stronger than the primary model, and only at the first step; reversing the assignment eliminates the effect, yielding near\-zero gain from the outset\. Despite the two models making different errors on approximately30%30\\%of items, such error diversity does not translate into sustained improvements under iterative revision\.

These results confirm that the information ceiling is structural: computational resources spent on iterative self\-revision yield diminishing returns after the first pass and are better allocated to independent sampling or external verification\.

## Appendix GAdditional Model Results

To verify that the reflection failures documented in the main text are not confined to specific model versions, we evaluate two recent GPT\-5 series models under the same HRF two\-pass protocol\. Identical prompts and evaluation procedures are used throughout\.

Table[5](https://arxiv.org/html/2607.28908#A7.T5)reports mean accuracy and standard deviation overK=5K\{=\}5repeated runs on MalAlgoQA\.

Across both models, second\-pass revision does not produce systematic improvements\. Instead, we observe small but consistent performance decreases, along with increased variance in the second pass\. These results confirm that the conditioned re\-generation behavior documented in the main text persists in more recent and stronger reasoning systems\.

Table 5:MalAlgoQA accuracy \(%\) for GPT\-5 series models under the HRF self\-revision protocol\.Δ\\Deltais the change from first to second pass in percentage points\.