Check What Breaks: Turning Reasoning Errors into Inference-Time Controls

TL;DR for operators

A generic instruction such as “pay more attention” is not a reliable reasoning control. In this study, that prompt barely changes performance for one model, substantially improves another, and makes a third slightly worse. A prompt that instead tells the model which failure points to verify reduces incorrect answers across all three tested model families.

The intervention comes from failure analysis rather than prompt intuition. The researchers manually examine 583 incorrect mathematical responses, classify them into 21 fine-grained error types, select eight frequent and actionable failures, and translate those into five verification questions covering conditions, assumptions, theorem prerequisites, logical consistency, and step validity.

For teams operating reasoning-heavy LLM systems, the relevant workflow is therefore: classify recurring failures, identify error classes that can be checked during inference, convert them into explicit controls, and measure whether those controls improve outcomes before investing in retraining. The evidence supports that workflow on the MATH benchmark across three open-weight models between 27B and 70B parameters. It does not establish that the same taxonomy or gains will transfer unchanged to other domains or model classes.

A plausible reasoning trace can still miss the decisive error

A multi-step model answer can look coherent while failing for a narrow reason: one condition was ignored, a theorem was invoked outside its prerequisites, or an implication did not actually follow from the preceding step. The difficulty for an operator is that “reasoning failure” is too broad a diagnosis to determine what should change.

The intuitive response is often to add an instruction such as “think carefully” or “check your work.” Yoshida, Nishida, and Nishida examine whether more specific guidance works better in Improving Mathematical Reasoning Capabilities in Large Language Models via Reasoning Process Error Classification.¹

Their experiment matters because it separates two different interventions. One merely asks the model to devote more attention to correctness. The other specifies what kinds of mistakes the model should actively look for.

On a 4,300-problem MATH test split, the second approach produces fewer mean incorrect answers than the default prompt for all three evaluated models. The generic-attention instruction does not show the same consistency.

The verification questions come from 583 observed failures

The researchers first generate zero-shot chain-of-thought solutions with Llama-3.3-70B-Instruct and isolate 583 genuinely incorrect responses after equivalence checking and manual verification.

Those failures are manually organized into five broad categories containing 21 fine-grained classes. Because a response can receive more than one error label, the counts are not mutually exclusive.

Logical and reasoning errors are the largest broad category with 230 instances, followed by problem-understanding errors with 185 and arithmetic or algebraic manipulation errors with 135. At the finer level, several failure modes recur often:

• applying an inappropriate theorem: 77 instances;
• misreading figures, tables, charts, or graphs: 69;
• incorrect numerical calculations: 61;
• incorrect logical implications: 57;
• ignoring stated problem conditions: 53.

The taxonomy is therefore more than a list of incorrect answers. It identifies where the reasoning process tends to become unreliable.

The authors then select eight frequent errors, concentrated mainly in problem understanding and logical reasoning, and convert them into five verification questions. These checks ask the model to examine whether it captured all conditions, introduced unsupported assumptions, satisfied theorem or formula prerequisites, maintained consistency with stated requirements, and preserved valid operations and implications through the reasoning process.

This distinction is operationally significant. “Be careful” does not identify a failure surface. “Check whether every theorem’s prerequisites are satisfied” does.

Specific checks outperform generic attention across all three models

The evaluation compares four prompt conditions under common settings: the default prompt, a generic “Pay Attention” prompt, a coarse prompt referring to broad error categories, and the proposed fine-grained verification prompt.

Mean number of incorrect answers on the 4,300-problem test split over 10 runs. Lower is better.

The generic instruction behaves differently across model families. It changes Llama only modestly, improves Qwen substantially, and produces more errors than the default prompt for Gemma. The fine-grained prompt produces the lowest mean incorrect-answer count in all three cases.

The coarse prompt is also informative. Merely naming broad categories such as reasoning or problem-understanding errors is not as effective as specifying concrete things to verify. Across all three models, the fine-grained prompt has fewer incorrect answers than the coarse version.

The paper also reports Mann-Whitney U comparisons and rank-biserial correlations between the proposed prompt and each baseline. Those tests support the reported within-benchmark differences. They should not, however, be read as evidence that every model will respond similarly to the intervention.

The numerical magnitude is also heterogeneous. The reduction from default is much larger for Qwen3-32B than for Llama or Gemma. “Works across three models” therefore describes the direction of the result, not a uniform treatment magnitude.

The experiment supports error-detection assistance, but does not measure it directly

The authors’ proposed mechanism is that a model can sometimes possess enough mathematical knowledge to produce a correct solution yet fail to notice an error-prone step while generating it. Explicit verification questions may make those points more salient during inference.

The observed results are consistent with that explanation. Fine-grained checks outperform both generic attention and broad-category reminders.

But the experiment does not manually reclassify the remaining failures after the intervention. It therefore cannot show, for example, that ignored-condition errors specifically fell after adding the corresponding check. The study demonstrates improved aggregate answer accuracy, not direct mediation through individual error classes.

The same distinction applies to cross-model generalization. Because the original taxonomy is built from Llama-3.3-70B-Instruct failures, improvements on Qwen and Gemma provide indirect evidence that at least some targeted weaknesses may recur across models. They do not establish that the three models have the same error-frequency distribution.

Cognaptus inference: failure analysis can become a lightweight control layer

For an LLM product team, the transferable idea is not necessarily these 21 mathematical error labels. It is the process used to create the intervention.

A production system already generates evidence about where it fails: rejected outputs, human corrections, evaluation traces, escalations, and benchmark regressions. Where those failures recur and can be expressed as explicit checks, the taxonomy can become an inference-time control layer.

The final step is an extension beyond what this paper measures, but it is important for production use. Aggregate accuracy can improve while the intended failure class remains unchanged. A deployment team interested in control rather than benchmark gains would want both outcome improvement and evidence that the targeted error mode actually declined.

This approach is most attractive when failures are recurrent, diagnosable, and expressible as verification criteria. It is less informative when errors arise from missing knowledge, inaccessible information, severe distribution shift, or failure modes that cannot be recognized from the model’s available context.

The boundary is the benchmark, not the workflow idea

The study is confined to MATH and three open-weight models ranging from 27B to 70B parameters. The taxonomy itself originates from one model. No evidence here establishes equivalent gains for smaller models, proprietary systems, non-mathematical reasoning, tool-using agents, or domain-specific workflows.

That limits claims about transfer, but it does not erase the operational pattern demonstrated inside the tested setting. The paper shows a complete path from observed failure traces to a structured taxonomy, from that taxonomy to explicit inference-time checks, and from those checks to measurable benchmark improvement.

The resulting decision rule is narrower than “prompt engineering works.” When a model repeatedly fails in recognizable ways, first determine whether those failures can be turned into concrete verification questions. Test that control against both the existing prompt and a generic instruction to be more careful. Retraining becomes one option in the escalation path rather than the automatic first response.

Cognaptus: Automate the Present, Incubate the Future.