Error Reflection Prompting (ERP)

• ERP is a prompting paradigm that integrates explicit negative examples and error analysis to guide models toward correct reasoning.
• It enhances calibration, safety, and overall performance by contrasting incorrect reasoning with annotated corrections.
• ERP methods scale through automation and have demonstrated measurable gains in tasks like arithmetic and commonsense reasoning.

Searching arXiv for the cited papers to ground the response in current arXiv records. Error Reflection Prompting (ERP) is a prompting paradigm for LLMs in which the model is exposed not only to a correct reasoning chain, but also to an explicitly incorrect one together with an explanation of the error. In its canonical form, ERP is presented as a method “comprised of an incorrect answer, error recognition, and a correct answer,” extending Chain-of-Thought (CoT) prompting with negative examples and explicit correction signals (Li et al., 22 Aug 2025). Within the broader literature on reflection, ERP denotes a family of techniques that aim to improve reasoning, calibration, safety, prompt quality, or policy learning by making errors legible and actionable rather than leaving them implicit.

1. Definition and canonical formulation

ERP was introduced as a direct response to a limitation of CoT prompting: CoT can supply step-by-step procedures, but it “lacks the ability of reflection and error correction,” so a model may perpetuate earlier mistakes through the rest of the reasoning trace (Li et al., 22 Aug 2025). The central design choice is to place erroneous reasoning inside the prompt in a structured way, so that the model learns both what line of reasoning is wrong and why it is wrong.

In the formulation described for the original ERP method, each example contains four elements: the question, an incorrect answer, an explicit error analysis, and a correct answer. The incorrect answer is intended to be plausible rather than arbitrary, typically reflecting common model failure modes such as misinterpretation, missed steps, or calculation error. The error analysis then explains those failures in natural language, after which the prompt supplies the correct step-by-step solution (Li et al., 22 Aug 2025).

The paper formalizes the error space as a set $E$ of relevant error types, with one or more errors $e_1, \ldots, e_n \subset E$ associated with a question $Q$. In that framing, ERP does not merely show a positive derivation; it pairs the derivation with annotated counterexamples. This suggests that ERP functions as a form of contrastive reasoning supervision at the prompt level, where “what not to do” is made explicit alongside “what to do” (Li et al., 22 Aug 2025).

A second defining feature is scalability through automatic construction. The reported automation pipeline generates plausible errors, constructs an incorrect solution that instantiates them, and then assembles the full ERP prompt by pairing that incorrect solution with error explanations and the correct solution. This allows ERP examples to be produced in batch rather than entirely by manual authoring (Li et al., 22 Aug 2025).

2. Reflection architectures beyond the canonical ERP template

Subsequent work expanded the reflection idea along several axes: when reflection occurs, what object is reflected upon, how many perspectives are used, and whether reflection is static or adaptive. These methods are closely related to ERP, but they reorganize the role of error analysis.

A useful way to situate the literature is by the structure of the reflective loop.

Method	Reflection target	Distinctive mechanism
ERP	Incorrect answer and its error types	Incorrect answer + error recognition + correct answer
PR-CoT	Initial CoT	Multi-perspective reflection over logic, completeness, ethics, alternatives
FaR	Retrieved or recalled facts	Fact elicitation followed by reflection for calibrated answering
MAPS	Failed mathematical solution	Adaptive auto-generated reflection prompts across multiple layers
REAP	Full problem-solving process	Reflection embedded before, during, and after deconstruction

MyGO Poly-Reflective Chain-of-Thought (PR-CoT) replaces single-dimensional critique with a structured set of predefined reflection prompts applied after an initial CoT. Its four stated perspectives are logical consistency, information completeness, bias/ethical consideration, and alternative solutions. After generating $CoT_{init}$, the model produces $R_i = LLM(CoT_{init}, Q, P_{ref,i})$ for each perspective and then synthesizes a revised $CoT_{final}$ and answer (Costa et al., 12 Jan 2026). Relative to canonical ERP, this moves from a single annotated error case to a modular critique stack over multiple failure dimensions.

Fact-and-Reflection (FaR) separates fact recollection from reasoning. Its sequence is fact elicitation, reflection over the listed facts, and optional confidence extraction. The paper emphasizes calibration rather than post-hoc correction, but the method still belongs to the reflection family because it uses explicit intermediate self-monitoring to reduce overconfidence (Zhao et al., 2024). A plausible implication is that ERP-style error understanding and FaR-style uncertainty expression address different failure modes: the former targets erroneous procedures, while the latter targets miscalibrated confidence.

MAPS, or Multi-Layered Self-Reflection with Auto-Prompting, makes reflection iterative and adaptive. It begins with a CoT solution, checks correctness, and if the answer is wrong generates a problem-specific reflection prompt from a meta-prompt. The model then re-solves the task, potentially repeating the cycle for several layers. The key distinction from fixed reflection prompts is that each new prompt is tailored to the observed failure and prior reasoning steps (Loureiro et al., 30 Jun 2025).

REAP embeds reflection throughout a single prompt rather than confining it to an after-the-fact repair stage. Its reported components are Reflection, Explicit Problem Deconstruction, and Advanced Prompting; reflection is used at multiple stages, including literal interpretation, key-insight checking, uncertainty handling, and a final contradiction review (Lingo et al., 2024). This suggests a broader conception of ERP in which reflection is not only corrective but also preventive.

3. Empirical performance, evaluation protocols, and observed gains

The empirical evidence for ERP is mixed but nontrivial. In the original ERP study, arithmetic reasoning and commonsense reasoning were evaluated on GSM8K, AQUA, MATH, StrategyQA, and CSQA using GPT-3.5 and GPT-4 with 4-shot or 5-shot prompting (Li et al., 22 Aug 2025). Reported arithmetic results show consistent gains over CoT in several settings: for GPT-3.5, GSM8K improved from 74.6% to 77.8%, AQUA from 54.3% to 58.7%, and MATH from 40.2% to 41.1%; for GPT-4, GSM8K improved from 95.4% to 95.7%, AQUA from 75.9% to 76.7%, and MATH from 54.1% to 54.8% (Li et al., 22 Aug 2025). In commonsense reasoning, GPT-3.5 improved from 65.3% to 66.1% on StrategyQA and from 75.8% to 77.3% on CSQA, while GPT-4 improved from 79.9% to 80.0% on StrategyQA and from 85.3% to 86.3% on CSQA (Li et al., 22 Aug 2025).

The same study reports that ERP reduced certain error frequencies, including algebraic or calculation, misinterpretation, and logical errors, while also shifting the model’s error profile so that some categories, such as “wrong selection of answer,” increased (Li et al., 22 Aug 2025). That pattern is important: ERP does not simply lower all error rates uniformly; it can redistribute the failure surface.

Other reflection frameworks report stronger gains when the reflective structure is more elaborate. PR-CoT evaluates logical consistency (LC), error correction rate (ECR), and human judgments on arithmetic, commonsense reasoning, ethical decision-making, and logical puzzles. In the reported table, PR-CoT achieves the highest LC and ECR across all listed tasks, for example improving LC on arithmetic from 85% with CoT to 94%, and ECR from 15% with MCOT to 17%; on ethical decision-making, LC rises from 74% with CoT to 84% and ECR from 18% to 21% (Costa et al., 12 Jan 2026). Human evaluation rates average overall reasoning quality as 4.4 for PR-CoT, 4.0 for MCOT, and 3.4 for CoT (Costa et al., 12 Jan 2026).

FaR evaluates reflection with calibration metrics rather than answer accuracy alone. On StrategyQA and WebQ, it reports a 23.5% reduction in Expected Calibration Error and a 13.9% reduction in MacroCE relative to strong baselines, with ECE dropping from 30.3 for Standard and 26.0 for the best advanced prompting baseline to 22.8 for FaR final (Zhao et al., 2024). The method also increases explicit expressions of concern, with 8.8% versus 3.9% of examples compared to CoT, and when such concerns are expressed the average confidence score drops by 13.2% (Zhao et al., 2024).

At the same time, the literature explicitly warns against assuming that reflection is uniformly beneficial. A separate study on self-reflection finds that outcomes are highly sensitive to prompt wording and that self-reflection “cannot improve upon the state-of-the-art chain-of-thought (CoT) prompting” when CoT is already strong (Liu et al., 2024). That paper reports only marginal gains in some reasoning settings but substantial changes in safety-related behavior, including a 75.8% reduction in toxic responses while preserving 97.8% non-toxic ones, a 77% reduction in gender biased responses while preserving 94.3% unbiased ones, and a 100% reduction in partisan leaning response while preserving 87.7% non-partisan ones (Liu et al., 2024). The broader empirical picture is therefore heterogeneous: ERP-style reflection may improve accuracy, calibration, safety, or error localization, but the magnitude and direction of change depend strongly on prompt design and task structure.

4. Prompt optimization, memory, and dataset-level error diagnosis

A major branch of ERP research treats reflection not as an inference-time reasoning aid but as a mechanism for improving prompts over many examples. In this line of work, the object of correction is the prompt itself.

Exemplar-Guided Reflection with Memory (ERM) is framed explicitly against feedback-driven prompt optimizers such as ERP or ProTeGi. It argues that prior systems underuse historical feedback and retrieve exemplars only by semantic similarity. ERM adds exemplar-guided feedback generation, a feedback memory with priority scores and forgetting, and an exemplar factory that retrieves examples using both priority and semantic relevance (Yan et al., 2024). Empirically, it improves F1 by 10.1 on the LIAR dataset and reduces half of the optimization steps on ProTeGi (Yan et al., 2024). In this setting, “reflection” refers to diagnosing prompt deficiencies from erroneous cases and reusing that diagnostic information across optimization rounds.

Reflective Prompt Tuning (RPT) generalizes this idea with function-calling. A diagnostic function evaluates the target model on an entire optimization set, summarizes recurring failure modes, and returns a structured diagnostic report; an optimizer LLM then revises the prompt using both the current report and an accumulated memory of previous reports. The framework also supports confidence-aware optimization via calibration signals and final prompt selection. Across three reasoning tasks, RPT improves over initial prompts by up to 12.9 points and improves confidence calibration (Bayat et al., 20 May 2026). The emphasis here is not single-example self-correction but systematic, validation-driven prompt repair.

Contrastive Reflection moves even closer to a debugging workflow. It identifies error-heavy behavioral slices, augments them with nearby successful examples from the same region, and asks a Teacher LLM for a targeted prompt edit. Candidate edits are accepted only if validation performance improves, optionally under regression constraints. On HotpotQA retrieval-augmented QA, one tree-selected contrastive repair improves held-out exact-match accuracy from 51.4% to 60.4%; failure-only and random-evidence variants improve less and break more previously correct examples (Koh et al., 29 Jun 2026). The method thereby treats reflection as contrastive diagnosis over structured traces rather than as an isolated critique of one answer.

MetaReflection shifts the same logic into agent memory. It collects self-reflections over failed trajectories, then performs a meta-reflection step that distills them into reusable instructions stored as semantic memory. Across logical reasoning, biomedical semantic similarity, open-world question answering, and vulnerability threat detection, it reports gains from 4% to 16.82% over the raw GPT-4 baseline while requiring fewer LLM calls (Gupta et al., 2024). This suggests that ERP can be operationalized as durable instruction learning rather than repeated per-instance critique.

5. From prompting to training: learning reflection as a capability

Another strand of research treats reflection not as a prompt-time intervention but as a trainable ability. In these works, ERP-like behavior is internalized into the model through supervised fine-tuning, preference optimization, or reinforcement learning.

ReflectEvo proposes an iterative self-generated reflection pipeline for small LLMs. It constructs ReflectEvo-460k, described as a dataset with 460,000+ self-generated reflection examples spanning 17 datasets and 10 domains, and uses SFT and DPO to train improved reflection behavior. Reported BIG-bench results show Llama-3 improving from 52.4% to 71.2% and Mistral from 44.4% to 71.1%, with further gains under multi-turn reflection (Li et al., 22 May 2025). The paper emphasizes error localization and correction rather than merely producing more verbose reasoning.

“Teaching Large Reasoning Models Effective Reflection” distinguishes between effective and superficial reflection. It defines the Effective Reflection Ratio as

$$\mathrm{ERR} = \frac{\sum_{i=1}^n \mathrm{Accuracy}(y_i)\cdot R(y_i)}{\sum_{i=1}^n R(y_i)},$$

where $R(y_i)$ is the number of reflections in response $y_i$ (Wang et al., 19 Jan 2026). Its Self-Critique Fine-Tuning (SCFT) framework uses self-generated critiques filtered by rejection sampling, and RLERR adds reinforcement learning with reflection-quality rewards organized around five principles: Truthfulness, Constructiveness, Specificity, Substantiveness, and InfoGain. On AIME2024 and AIME2025, SCFT and RLERR improve both Pass@1 and ERR; for DeepSeek-R1-Distill-Qwen-7B, AIME2024 Pass@1 rises from 55.6% to 60.2% under SCFT and to 63.1% under RLERR, while ERR increases from 0.26 to 0.38 and then 0.50 (Wang et al., 19 Jan 2026).

Experiential Reinforcement Learning (ERL) incorporates an explicit experience-reflection-consolidation loop into RL training. Given an initial attempt $y^{(1)}$, the model receives environmental feedback, generates a reflection $e_1, \ldots, e_n \subset E$0, and produces a refined second attempt $e_1, \ldots, e_n \subset E$1; successful second attempts are then internalized into the base policy so that reflection is not needed at deployment. The paper reports gains of up to +81% in complex multi-step environments and up to +11% in tool-using reasoning tasks, with no additional inference-time overhead after training (Shi et al., 15 Feb 2026). Relative to prompt-time ERP, ERL treats reflection as a credit-assignment mechanism inside policy optimization.

The pre-training study “Rethinking Reflection in Pre-Training” pushes the origin of these abilities even earlier. By injecting deliberate errors into chains-of-thought, it evaluates whether models can recognize and correct mistakes through explicit or implicit reflection. The paper reports that self-correcting abilities appear in base OLMo-2-7B models trained on only about 198B tokens and continue to improve through 4 trillion tokens, with an OLMo2-7B model pre-trained on 4 trillion tokens displaying self-correction on six self-reflection tasks (AI et al., 5 Apr 2025). This suggests that ERP-style behavior is not exclusively an alignment or post-training phenomenon.

6. Limitations, misconceptions, and unresolved questions

The ERP literature repeatedly cautions against an overly broad claim that “more reflection” necessarily means “better reasoning.” One recurrent limitation is prompt sensitivity. The self-reflection study on Llama-3 8B shows that upward-biased prompts that aggressively search for mistakes can increase false positives and degrade accuracy, while downward-biased prompts that verify correctness can lower false positives but leave false negatives extremely high. Its Mixture of Prompts framework yields only slight accuracy gains, with MMLU moving from 64.1% to 65.2% at best (Liu et al., 2024). A common misconception, therefore, is that self-reflection is a drop-in replacement for stronger reasoning strategies; the reported evidence does not support that generalization.

A second limitation is superficial reflection. The SCFT/RLERR work identifies many reflections as “superficial,” meaning they restate, rubber-stamp, or perform tangential checks without materially improving the solution. Its entire training framework is motivated by the observation that not all reflections are beneficial and that unproductive reflective behavior adds computation overhead (Wang et al., 19 Jan 2026). PR-CoT reaches a similar conclusion indirectly through ablation: omitting any reflection type degrades performance, and removing the ethical consideration perspective yields the largest drop in relevant domains (Costa et al., 12 Jan 2026). This suggests that reflection quality and reflection diversity matter more than the mere presence of a reflective prompt.

A third issue is cost. Canonical ERP increases prompt length because it includes an incorrect solution, an error explanation, and a correct solution; the original paper notes increased token and computational cost as a practical limitation (Li et al., 22 Aug 2025). Related systems show the same trade-off in different forms: MAPS reports that deeper reflection layers improve accuracy but increase token usage and costs, and PR-CoT reports that increasing the number of perspectives improves logical consistency and error correction while also increasing inference time and token consumption (Loureiro et al., 30 Jun 2025, Costa et al., 12 Jan 2026).

Generalization is also unresolved. The original ERP results are stronger in arithmetic than in commonsense reasoning, and the paper explicitly notes diminishing returns on larger models and reduced gains on the MATH benchmark (Li et al., 22 Aug 2025). More broadly, reflection methods differ in what they optimize—accuracy, calibration, safety, neutrality, or interpretability—so cross-paper comparisons are not always commensurate. A plausible implication is that ERP should be treated less as a single algorithm than as a design space for converting failures into structured supervision signals.

Across this literature, the central question is no longer whether LLMs can reflect at all, but under what conditions reflections are faithful, specific, actionable, and worth their computational cost. In that sense, ERP has evolved from a prompt template into a general research program on error-aware reasoning, error-aware prompt repair, and error-aware policy learning (Li et al., 22 Aug 2025).