Why Reasoning Models Hallucinate: A Lifecycle Framework for Diagnosis, Evidence, and Release Control

1 Introduction

Reasoning models can err from unsupported premises or from arithmetic/logic mistakes on correct premises. Fixing these requires different tools. Answer accuracy alone cannot select the repair; a correct answer may still rest on unsupported explanations. Reasoning systems vary widely: prompt-generated steps, multi-path sampling with aggregation, post-training reasoning, tool use, or hidden compute budgets. Training recipes and inference settings cannot be conflated. Vendors may show full traces, summaries, or nothing; fluent explanations are generated text, not proof of internal computation. Unlike surveys organized by error type, model lifecycle, or trace length, this work links local errors to downstream use, visibility, and action. Contributions include the UIPCA audit record, a research-design review of reasoning vs. factual reliability, and cross-interface verification, repair, selective release, and persistent-write control.

Figure 1 | Survey roadmap: from scope and observability into diagnosis, research evidence, detection/intervention, evaluation, and persistent-state control. Reasoning capability and factual reliability are independent measurement targets; dashed lines between hypotheses are not established causal relations. Source: original paper Figure 2, PDF p.3.

The paper centers on four research questions: how to diagnose and repair; what current evidence says about reasoning and factual reliability; what detection and intervention each interface permits; and how benchmarks and release strategies assess critical outputs and persistent state. The endpoint is not a higher detection score but better-grounded operational decisions.

2 Scope, Evidence Model, and Survey Method

2.1 Concepts and Evaluation Units

Large reasoning models are defined as systems that go beyond direct generation to allocate or supervise multi-step computation, covering visible drafts, hidden thinking, multi-path search, tool use, and post-training reasoning. Reinforcement learning, supervised fine-tuning, and distillation often change data, reward, optimization, and trace retention simultaneously; labels alone cannot isolate a single cause. "Hallucination" is judged against the task's declared reference frame: factual QA asks whether a claim is true; source faithfulness asks whether specified material supports it. An externally true fact may still lack task-required support when the task demands grounding in a given document. Three further independent targets are process faithfulness (does displayed reasoning reflect the actual decision process, requiring causal intervention, not text matching), calibration (do probabilities/scores match actual accuracy), and correctability (does valid feedback fix downstream reasoning and results). None substitutes for action quality: knowing uncertainty does not guarantee appropriate verification or abstention.

2.2 Boundary Between Hallucination and Deception

Wrong, persistent, or self-contradictory statements do not prove deceptive intent. Deception requires goal-driven concealment or strategic misrepresentation of evidence, e.g., controlled experiments on incentive and information access. UIPCA records observable support, dependency, consistency, and policy failures without inferring mental states. The same record may stem from knowledge gaps, inference slips, or reward pressure; even factually correct outputs can constitute action-policy failures if they violate pre-specified release rules.

2.3 Observable Interfaces and Evidence Status

Full traces expose ordered steps; summaries expose selected statements; hidden interfaces expose no process record. Comparable elements — question, specified evidence, answer, recorded actions — should not be ignored because the process is hidden. Evidence status should be classified as supported, unsupported, or unknown. Missing required grounding in specified documents confirms source-grounding failure; open-world retrieval finding no source does not prove the claim false. Missing process remains unknown; it is not upgraded to "no error" because of model confidence, sample consistency, or correct outcome. Moreover, showing different views of the same run differs from changing how the model generates or compresses reasoning; disclosure itself may alter behavior, so interface visibility must not be conflated with generation mechanism effects.

2.4 Search, Screening, and Evidence Strength

The corpus comprises 312 papers and first-party reports available by 30 Aug 2026, collected via OpenAlex, arXiv, citation tracing, and targeted checks for under-represented topics. The authors explicitly call this a curated collection, not an exhaustive census; year counts do not equal total publication volume. The dedicated reasoning–reliability review covers 34 in-corpus sources plus 24 external sources with full text obtained, totalling 58 sources — not 58 additional independent experiments. Supplementary materials archive retrieval, screening, evidence review, coding, and figures. The authors distinguish direct single-study evidence, convergent support from compatible studies, limited/mixed evidence, and proposed tests with no results yet. Language models assisted drafting, but protocol, source selection, and conclusions remain the authors' responsibility.

2.5 Pre-specified Paired Joint-Adverse-Effect Test

To prove an intervention improves reasoning while harming factual reliability, one must measure both in the same comparison, pre-specify meaningful gain/loss thresholds, and require a joint confidence region supporting both directions. Observing two scores move oppositely is insufficient. Inference-time experiments must fix checkpoints and non-target settings; training experiments must start from a common base model and control non-target training choices; both must match questions and evaluators. Longer traces or the label "reasoning model" cannot substitute for capability measurement. Failure to meet the standard yields "unresolved," not "no harm."

3 From Evidence to Action: The Lifecycle

3.1 Operational Definitions

UIPCA's five fields are not internal model mechanisms nor mandatory stages in every run:

Unsupported Premise : input facts/conditions lack required grounding; check source first.

Invalid Local Inference : premises may be correct but arithmetic, logic, or state transition fails; fix that step first.

Dependency Reuse : later statements depend on an already-confirmed error; repair must propagate along affected branches; mere repetition or two wrong steps do not prove dependency.

Visible Answer–Trace Consistency : checks whether the final answer contradicts the visible reasoning; does not evaluate hidden computation, nor does consistency prove correctness.

Action-Policy Failure : compares actual action against a pre-fixed policy, e.g., requiring independent verification but releasing without it.

The paper uses 1 for confirmed failure, 0 for excluded failure, and an unknown symbol for insufficient evidence. The earliest-error marker locates the earliest confirmed premise or inference failure; when earlier steps are undecided, it does not equal the true first error.

Figure 2 | Premise and inference locate error onset; dependency field records downstream reuse; consistency field checks visible contradictions; action field audits fixed policy separately. Full, summary, and hidden interfaces permit different checks; evidence unavailable remains unknown. Source: original paper Figure 3, PDF p.4.

3.2 Two Running Examples

Example 1 computes (1024 × 1047 × 1050 × 1053) mod 33. Step 3 writes the remainder of 1024÷33 as 29 (correct is 1); subsequent wrong remainders are reused in the product, yielding output 27 vs. reference 3. Thus local inference error and dependency reuse exist, but the answer follows the erroneous trace, so visible consistency passes. Consistency pass ≠ correct answer.

Example 2 judges which conditions determine a two-digit number. Earlier steps correctly find two condition sets each sufficient for 43; step 8 incorrectly treats a mere restatement of "tens minus units equals 1" as another sufficient condition. This error has no downstream reuse; a separate issue is that the final option contradicts the earlier valid result.

Figure 3 | Both original cases answer incorrectly but require different repairs: top case needs recomputation of the wrong remainder and dependent products; bottom case needs checking sufficiency conditions and option mapping. Neither provides a release policy, so action fields remain unknown, not compliant. Source: original paper Figure 5, PDF p.9.

Answer scoring cannot select these repair points. These are diagnostic case studies, not frequency estimates; the authors also do not retroactively attribute hypothetical policies to the original cases.

3.3 Textual Evidence and Cross-Domain Transfer Checks

Text needs source and timestamp tags; visual claims need object, region, or frame references; code can use contracts, static analysis, sandboxes, and tests. The diagnostic questions are shared, but a single undifferentiated verifier cannot serve all domains. A visual error may be a missed object or a seen-but-unused object; continuing in pure text cannot verify unobserved image details. Passing code tests only proves covered properties, not all requirements, security, or unwritten user intent. Repository tasks must also record build environment, snapshots, and patch interactions.

3.4 Lifecycle Components and Repair Points

Premise issues can be addressed by retrieval, source verification, or input correction; local inference issues need rule, math, or execution checks. If a graph model hallucinates an edge, downstream deduction may be perfectly consistent on the wrong input — fix the input, not just the reasoning rule. Reuse issues require proving downstream dependency and checking branches; independent evidence can stop error propagation, while self-check may re-adopt the same faulty premise. Visible consistency needs trace alignment; process faithfulness still requires changing premises/steps and observing the answer. Action audit separately measures erroneous release, unnecessary refusal, missed clarification, and missed escalation, combined with cost and evidence requirements. Verbal confidence is not a release policy.

3.5 Applying the Framework Without Overclaiming

A single sentence may suit source verification yet contain multiple arithmetic operations needing separation; a token-level alert may not indicate which branch to restart. Localization granularity has practical value only when it changes actionable repair. Fields record location and relation, not severity. An unused faulty premise may be critical; a discarded long erroneous branch may be harmless; severity and recovery difficulty need separate measurement. Displayed dependencies and contradictions do not directly represent hidden computation.

3.6 Operational Sensitivity to Segmentation, Dependency, and Unknowns

Merging correct and incorrect operations inflates the first-error span; treating all adjacent steps as dependencies fabricates propagation. Absent preconditions or multiple derivations mean undetected dependency does not automatically exclude reuse. The authors conducted independent dual-annotation on 48 full-trace cases; on 334 shared source steps, Cohen's κ for premise, inference, and reuse were 0.653, 0.730, 0.858 respectively; first-error location agreed in 41/48 cases. All action labels were unknown because no policy was provided. These results support label consistency under fixed segmentation, not comprehensive validation of candidate dependency graphs or deployed repair efficacy.

4 What the Evidence Actually Proves

4.1 Candidate Evidence for Paired Joint Adverse Effect

Among the 34 in-corpus and 24 external sources, none satisfied the paper's pre-specified paired joint-adverse-effect standard; even relaxing thresholds and uncertainty-reporting requirements yielded no positive sources. This is a source-level design-and-reporting conclusion, not a field-level null effect. Candidate studies variously discuss full-pipeline differences, natural-trace correlations, single-item regressions amid overall gains, or monitorability declines; they do not answer the identical question.

4.2 Reading Evidence by Research Design, Not Title

Pipeline comparisons : different post-training recipes may raise or lower factual QA, but data, reward, and optimization change together, so reasoning cannot be isolated as the sole cause. Natural-trace grouping : error-fact traces correlate with lower accuracy, but difficulty, confidence, and other attributes confound; this does not prove the causal effect of inserting that fact. Local interventions : changing steps, tool availability, or prompts can test specific mechanisms; acceptance of correct local edits and final-outcome recovery must use separate denominators. Deployment and evaluation reports : when document sets, graders, and protocols shift, scores cannot be stitched into a continuous leaderboard. Factual access vs. compositional ability, monitorability vs. factuality must be measured independently.

4.3 Five Mechanism Hypotheses

Figure 4 | Five hypotheses each state what to vary, what to measure, and flag current evidence strength. Direct evidence for error reuse comes from a planted-memory attack scenario, not everyday reuse rates; arrows describe proposed experiments, not established causal effects. Source: original paper Figure 6, PDF p.13.

Unverified Statement Accumulation : longer traces increase error opportunities but also correction opportunities; experiments should fix the starting trace, vary verification and stopping rules, and count both new statements and corrections.

Reward Bias Toward Answering : sparse outcome rewards may encourage plausible-looking answers over calibrated abstention; should randomly assign reward schemes from the same model, controlling data, budget, and optimization.

Correlation Over Evidence : misleading guidance, sycophancy, and context neglect are compatible with this but do not directly prove it; need to cross-vary correlation strength and support level while matching wording.

Error Reuse : FARMA demonstrates an attack path where fabricated reasoning written to memory re-enters context; everyday rates still need controls for supported, unsupported, and refuted premises, plus varied independent correction opportunities.

Uncertainty Changes Action : detection signals must actually change verification, abstention, or escalation while maintaining useful coverage; acknowledging contradiction without resolving it does not count as protection.

4.4 Boundary Conditions of the Synthesis Conclusion

Output detail determines verifiable claim count; willingness to answer changes the answered subset; evaluator permissions limit available checks; checkpoint and training changes may overwhelm "thinking mode" effects. Therefore long vs. short answers, different budgets, and different post-trained models cannot be directly merged. The survey demands every result state its source, denominator, and supported relation. Absence of strict positive results leaves the causal question open; it does not cancel observed harms in specific systems.

5 From Detection to Intervention

5.1 Signals, Units, and Actions

Repeated sampling, semantic entropy, and self-consistency provide risk signals, but shared faulty premises make multiple outputs agree. Hidden-state and sparse feature probes may warn earlier but depend on model and training domain; they cannot conjure missing facts. FActScore operates at the fact level; MiniCheck checks document support for statements; FacTool and SAFE seek external evidence; step-level and streaming detectors locate process issues during generation. The unit determines repair power: sentence scores may not pinpoint an operation; internal anomalies may not yield a reliable source.

Figure 5 | Complementary detection mechanisms: external sources, sampling uncertainty, executable checks, process monitoring, and model grading. Controller synthesizes evidence and policy to choose release, verify-revise, clarify-abstain, or escalate; action field re-audits actual action. Source: original paper Figure 7, PDF p.15.

Fast monitors add value by routing cases fit for strong checks, not just ranking hard cases first. Must measure downstream error release, false alarms, human burden, and cost. Detector comparison and full-system comparison are two different experiments.

5.2 Intervention Method Families

Increasing inference-time budget, paths, or debate may expand recall but can also share source errors; should report answer rate, selective accuracy, path diversity, and latency. Trace length alone proves neither repair nor waste. Retrieval, tool critique, and independent verification provide stronger grounding than pure re-thinking, but weak retrieval or unavailable tools introduce failures. Early checks can reject steps before reuse; post-generation critique-rewrite may improve the answer yet leave the original error trace; both need independent final checks. Process supervision moves checks into training: source support locates premises, rules and execution locate inferences, dependency-aware rewards must further prove downstream reference to a confirmed error. Fact verifiers may miss logic; step rewards may miss false premises; no single verifier covers all fields. Auto-labels can inherit grader bias; optimization may game wording or scores. Knowledge editing, activation steering, and decoding interventions should check associated facts, multi-hop answers, and unrelated capabilities; a single edited item improving does not prove the whole dependency chain is fixed. Multimodal and code interventions likewise need matched evidence units.

5.3 Action Mapping Across Interfaces

Showing Example 1 as full trace, schematic summary, and answer-only are three views of the same completed run, not three new generations. Full trace locates step 3 and downstream reuse; summary omitting arithmetic cannot localize; hidden view cannot check answer–trace consistency. Yet all three views can independently compute 3 from the question. If a pre-specified policy requires independent verification before numeric release, all three should block 27; incomplete verification → abstain. This demonstrates policy; the original case's action field stays unknown. More visibility improves localization but does not automatically guarantee benefits outweigh costs.

5.4 Calibrated Tiered Control Flow

First, decompose visible answer and reasoning statements; judge supported, unsupported, unknown; where localizable, do component-level repair; unknown routes to retrieval, clarification, abstention, or escalation — not counted as supported. Fast triage selects strong verification; factual disputes go to sources, arithmetic disputes to recomputation, high-stakes disputes to experts. Each layer's high score alone is insufficient: triage changes the sample the verifier sees; a conservative controller can manufacture high-reliability appearance via mass escalation. Full systems should jointly report repair success, selective risk, coverage, workload, and cost. Action audit only checks policy adherence; one compliant action does not retroactively rewrite evidence as sufficient.

6 Evaluation, Deployment, and Research Agenda

6.1 Benchmark Coverage and Joint Metrics

Factual QA and grounding benchmarks mainly test answers or statements; process benchmarks provide steps and perturbations; calibration and answerability benchmarks test probabilities and answering actions. Long context adds retrieval and evidence binding; vision needs region coverage; code needs environment and test coverage. Long-term memory benchmarks add cross-session recall, knowledge update, and selective forgetting. HaluMem separately evaluates retrieval, update, and memory QA; MemoryData compares components, retrieval, and cost; shared memory governance further tests authorization, visible versions, and conflict resolution. Operation labels or retrieval source IDs do not prove downstream reasoning depended on the error.

6.2 Critical Benchmark Gaps

First, endpoint scores hide paths; they cannot tell whether a correct answer came from an erroneous trace. Second, model graders are sensitive to prompts, few-shot exemplars, and model families; when generation and grading share biases, agreement ≠ independent confirmation; grader upgrades may change scores without changing system capability. Third, action coverage is lacking. Clarification, retrieval, tool verification, expert escalation, write refusal, and rollback are rarely measured together. Need external-verification labels, full action and cost logs, not just more error-answer counts.

6.3 Lifecycle-Oriented Benchmark Construction

Benchmarks should connect the same error, available evidence, and actual action; use different views of the same run to test visibility loss; then control premise or inference changes to verify whether safeguards actually prevent harm. Must also distinguish ranking, probability calibration, and decision: AUROC measures ranking, not probability accuracy; Brier score and ECE test different probability properties. Finding source support for a step does not prove it influenced the answer. VeriFY's model-relative recall includes unnecessary abstentions, depends on how the protocol judges base-model knowledge, and cannot be treated as general objective recall. Deployment value of error localization should be tested by improved repair and action, not just label agreement.

6.4 Release Policies Aware of Interface Limits

Release policy chooses among answer, search-evidence, revise, clarify, abstain, escalate. Total refusal avoids error release but yields zero coverage; total answer may exceed risk tolerance. Risk and coverage thresholds must be declared upfront. Example: 10,000 requests, 6,000 released, 30 errors → selective risk 0.5%, coverage 60%. This meets an illustrative constraint of risk ≤0.6% and coverage ≥55%, but fails a 0.2% risk ceiling. Numbers are illustrative, not deployment results, and must account for sampling uncertainty and distribution shift.

Figure 6 | Evidence and visible diagnosis enter a fixed-policy controller; before state change, check evidence, write permissions, and version. Memory, tool state, and training feedback are parallel destinations; stored content may be reused later, but reuse is not inevitable. Source: original paper Figure 8, PDF p.19.

Full traces support process checks; summaries only check disclosed content; missing process cannot be logged as zero errors. If current evidence and resources cannot simultaneously meet risk and coverage targets, the policy is infeasible. Identical alerts should trigger different actions for reversible writing errors vs. high-stakes advice; verbal uncertainty cannot substitute for risk control.

6.5 From Released Answer to Persistent State

Answers can enter private/shared memory, tool effects, and training feedback. Text edits, external action rollback, and parameter correction are distinct recovery operations; self-feedback and iteration do not themselves prove repair. A-MAC scores on utility, overlap confidence, etc., at admission; but literal overlap with dialogue snippets ≠ semantic entailment. TrustMem checks coverage, retention, and support for old-to-new memory transitions, still limited by evaluator and pre-existing erroneous memories. Admission and curation separately ask: can it enter? does the change preserve valid information?

Shared state adds writer permissions, concurrency, visible versions, and revocation. The paper recounts MemClaw/ArgusFleet single-service observations: near-duplicate gating may reject contradictory writes before async conflict detection, thereby blocking legitimate corrections. This shows control ordering needs verification; it does not extrapolate to a universal memory hallucination rate. Provenance records who wrote or derived, not truth; visibility to other agents does not prove dependency reuse. Blocking write or replacing version does not fix existing summaries, tool effects, or training feedback; affected consumers must be traced. The authors propose counting expected consumers who actually reuse an error within a fixed observation window before independent repair. Example: three consumers each with 50% selection probability, 20% pre-repair if selected → expected unrepaired uses = 1.2; raising repair probability to 60% → 0.6. Figures are hypothetical demonstrations, not measured diffusion rates or general propagation thresholds.

6.6 Experiments to Close Open Loops

First, use paired designs to separate verification gains from mere answer reduction, jointly measuring reasoning and factuality. Second, on the same cases compare error localization vs. answer-level warnings to test whether localization truly improves repair, not just annotation agreement. Third, evaluate transfer and recovery timing: translation may flip entities, negation, units; visual evidence and context position may change information use; code success depends on specific requirements and tests. Early premise correction, mid-branch interruption, and late answer revision should be compared under shared budget, including adversarial evidence, tool unavailability, and distribution shift.

7 Conclusion

Reliability cannot be summarized by a single aggregate score. UIPCA separates premises, inferences, error dependency reuse, visible consistency, and action-policy failures; unknown stays unknown, no inference of deception, and text traces are not automatically treated as true computational mechanisms. The dedicated review of 58 sources did not establish the paper's strictly defined "reasoning gain accompanied by factual reliability loss," but specific pipeline differences and error traces remain worth studying. The five hypotheses offer subsequent experimental directions, not a unified causal law. Practical systems should jointly measure answer quality, process support, confidence, action, coverage, and cost; critical statements entering memory, tools, and training feedback also need evidence, permission, and version checks. The governance goal is both reducing current error release and ensuring the next round inherits a grounded, correctable state.