The Retry Is Not the Lesson: How ReTeach Learns from Failed Answers

TL;DR for operators

A model generates an incorrect answer. An automated verifier can reject it but cannot explain the mistake. ReTeach¹ shows how that failure can become training material without a worked solution or a larger external teacher.

The model’s teacher revisits unsuccessful attempts, reflects on possible errors, and retries. But the successful retry is not simply copied into the student’s training targets. Instead, the teacher uses its additional reasoning context to produce better-informed token predictions along the student’s original response. The student learns from those predictions without needing that context itself. On SciKnowEval-Physics, explicit reflection achieves 64.24% accuracy, versus 62.91% with retries but no reflection (Section 4.3, Figure 4).

Across six benchmarks, ReTeach achieves 65.50% average accuracy, compared with 64.11% for GRPO and 64.15% for CEPO. For teams training models on verifiable workflows, the opportunity is to extract more learning value from failures. The trade-offs are additional training computation, dependence on reliable verification and stable teacher predictions, and modest gains established on only one model backbone without formal significance testing.

A failed answer can become training material

Suppose an AI assistant generates a tool call with the wrong arguments. A validator rejects the call, but its only feedback is failure. The developer has no corrected example to add to the training set, and the validator provides no explanation of which argument was wrong.

Ordinary reward-based training can use that failure as a negative outcome. It cannot directly identify which token-level decisions caused the mistake. Retrying may eventually produce a valid call, but repeated attempts do not necessarily improve the model’s future first response.

ReTeach investigates whether the model can construct its own teaching information from this situation.

The distinction between retrying and explicitly reflecting is measurable. In the paper’s SciKnowEval-Physics ablation, accuracy reaches 64.24% with reflection, compared with 62.91% when retries remain but explicit reflection is removed (Section 4.3, Figure 4).

That 1.33-percentage-point difference suggests that examining unsuccessful attempts contributes something beyond generating another candidate. The more consequential design choice is how this information reaches the student.

The teacher learns from failures without receiving a reference solution

ReTeach begins with a student model and a teacher initialized from the same Qwen3-8B checkpoint. The student generates an original response, which a binary verifier judges correct or incorrect.

For an incorrect response, the teacher follows a repeated procedure:

1. Examine the original answer and any subsequent unsuccessful attempts.
2. Generate an explicit reflection about the previous attempts.
3. Produce another response conditioned on that reflection.
4. Check the new response and stop upon verified success or exhaustion of the retry budget.

The reported default permits two reflection-and-retry rounds.

The teacher has access to the question, its own attempt history, and binary verification outcomes. It does not receive reference answers, reference reasoning traces, or external explanations of the failure. Earlier reflections are not preserved as separate historical entries; the teacher generates a new reflection using the accumulated unsuccessful attempts.

When the procedure ends, the final teaching context contains the unsuccessful attempts preceding the last retry and the final reflection. Importantly, the successful retry itself is not simply copied into the student as a supervised answer.

Instead, ReTeach compares two sets of token predictions along the student’s original response.

The teacher predicts tokens while seeing the additional deliberative context. The student predicts tokens from its original prompt and response prefix, without that context. Training minimizes the Jensen-Shannon divergence between their predictive distributions, averaged across the original response’s tokens (Section 3.2, Equation 7).

This is on-policy self-distillation: the student generates the attempts used for training, while a differently informed version of the same model supplies richer token-level targets.

The teacher’s advantage comes from its reasoning procedure and additional context, rather than from a larger pretrained model or privileged reference solution. Successful correction indicates that this context can support a valid answer, although it does not establish that every reflection is accurate.

Not every failed attempt deserves the same training weight

A reflection process can finish in three distinct states. ReTeach treats them differently rather than assuming that every generated teacher context is equally reliable.

Source: Appendix A, Table 2; outcome categories defined in Section 3.2.

The objective averages losses separately within each category, then applies the prescribed weights. This prevents the number of examples in a category from automatically determining its contribution. An absent category contributes zero, without redistributing its weight.

Category B captures the central learning opportunity: a failed response for which self-generated reflection eventually supports a verified correction. Category A helps preserve initially successful behavior.

Category C is more ambiguous. In mathematics and science, increasing its weight generally damages performance. In Tool-Use, a modest weight of 0.2 helps.

The authors suggest that strict tool-call verification can reject partially useful responses. An unresolved teacher attempt may therefore contain worthwhile information alongside errors. This is a plausible explanation of the task difference, not evidence that all partially correct tool calls should be retained.

For workflow developers, verifier design and training-data weighting are connected decisions. A strict pass/fail rule may be appropriate for deployment while concealing distinctions that matter during learning.

More reflection is not automatically better supervision

The paper’s component tests identify two further constraints on reflective self-teaching.

First, teacher updates need stability. If the teacher immediately tracks every student parameter change, its predictions become an unstable training target. The tracking configuration’s entropy falls from 0.129 to 0.036 nats, while its reported gradient norm spikes to 25.27. Training fails to improve upon the base model.

A frozen teacher is more stable but produces weaker results.

ReTeach instead updates teacher parameters through an exponential moving average (EMA), with a decay coefficient of 0.9999. This allows the teacher to incorporate student improvements gradually rather than immediately matching a changing student. The reported EMA gradient norms remain comparatively stable, approximately 0.08–0.14 (Section 4.3, Figures 4–5).

Second, additional retries have diminishing returns.

Source: Section 4.3, Figure 6.

Moving from one to two rounds adds 0.70 percentage points. A third adds only 0.14 points. The third round reduces across-seed variability but requires additional teacher inference.

These are mechanism and sensitivity tests, not independent demonstrations of broad superiority. They show that the quality, selection, and stability of teacher signals matter alongside the amount of reflection performed.

Six benchmarks show a measurable but modest advantage

The main comparison evaluates Qwen3-8B post-training methods across three mathematical reasoning benchmarks, two undergraduate science benchmarks, and structured tool calling.

Source: Section 4.2, Table 1. Unweighted averages of benchmark accuracy, with results reported across three random seeds.

ReTeach exceeds GRPO by 1.39 percentage points and CEPO by 1.35 percentage points. It achieves the highest mean among compared methods on five of the six benchmarks.

The exception matters. On SciKnowEval-Material, CEPO reaches 70.83%, ahead of ReTeach’s 69.17%. On ToolAlpaca, meanwhile, ReTeach reaches 62.72%, compared with 60.85% for GRPO and 61.86% for CEPO.

The aggregate also combines different evaluation settings. Mathematics uses thinking-enabled inference with avg@12, while science and tool use use thinking-disabled inference with avg@16. ReTeach therefore does not establish a uniform thinking-disabled deployment benefit across all six benchmarks.

The results support a comparative improvement within the tested setup. They do not establish statistical significance, because formal significance tests are not reported.

The deployment decision starts with verification and compute cost

For an organization training a model to generate structured API calls or perform executable workflows, ReTeach suggests a different use for failed training attempts.

If a validator can reliably determine success, unsuccessful generations can initiate reflection-based teacher construction. The resulting contexts may provide detailed training signals without requiring purchased reference reasoning traces or a separately trained larger teacher.

Cognaptus’s inference is that three decisions would determine whether this is worth implementing.

First, establish verifier reliability. A validator that rejects useful outputs or accepts incorrect ones changes both the reflection stopping condition and the reliability of the resulting training signals. The Tool-Use category-weight findings make this especially relevant.

Second, compare the deployed student, not just successful teacher retries. The intended benefit is improved student accuracy after distillation, without needing the complete teacher reflection procedure at ordinary inference. However, actual inference behavior must be measured under the task’s required thinking configuration.

Third, account for the entire training budget. Reflective retries, teacher generation, and token-distribution matching consume additional resources. The paper uses eight NVIDIA H20 GPUs for its experiments but does not establish an end-to-end, compute-normalized advantage over every baseline.

A useful internal comparison would hold the model, verifier, and evaluation protocol fixed while measuring deployed accuracy, training GPU-hours, generation volume, and inference latency against GRPO or another available distillation method.

Where the evidence stops

The reported experiments use one Qwen3-8B backbone with LoRA adaptation. Whether the same mechanism works equally well for smaller models, different model families, or tasks without objective correctness criteria remains unanswered.

The category weights are also selected from the reported ablations rather than independently validated. And while successful retries provide outcome-level evidence that a reflection context is informative, they do not certify the correctness of each intermediate diagnostic statement.

ReTeach demonstrates a specific capability: a model’s failed attempt can be converted into more informative training supervision through verified reflection, even without access to a reference explanation.

For model developers, the question is no longer limited to whether a failed answer can be corrected. It is whether the information generated during correction improves the deployed model enough to justify the additional training computation.

That is the distinction that makes the approach worth testing—and the measurement the paper has not yet completed.

Cognaptus: Automate the Present, Incubate the Future.