cs.AIOct 4, 2026
SaveHallucination Across the Reasoning Lifecycle: Interface Visibility, Causal Evidence, and Release Control in Large Reasoning Models
Organizations: Zhejiang University, China · Binjiang Institute of Zhejiang University, China · Guangzhou University, China · Rensselaer Polytechnic Institute, USA · The University of Hong Kong, Hong Kong SAR, China · The Hong Kong University of Science and Technology (Guangzhou), China · The Hong Kong Polytechnic University, Hong Kong SAR, China · Research Center for Scientific Data Hub, Zhejiang Lab, China
Abstract
Reasoning errors can propagate into later decisions and memory. This survey synthesizes 312 papers and first-party reports on text-based reasoning hallucinations around three questions: what evidence is observable, what study designs establish, and which corrective actions the evidence supports. UIPCA records unsupported premises (U), invalid inferences (I), dependent reuse (P), visible answer-trace consistency (C), and action-policy failures (A). Across 58 reviewed sources, no comparison establishes that a specified intervention improves reasoning while reducing factual reliability under matched conditions. The synthesis connects diagnosis to verification, repair, selective release, and persistent-state control across memory, tools, and training feedback.
Figures & tables
Figure 1. Research river of reasoning hallucinations: 60 representative works grouped by primary analytical role. The regional headings name research themes, not formal UIPCA fields. In particular, “Propagation & Monitoring” groups related literature; records dependent reuse of a confirmed error, which repetition or a monitoring signal alone cannot establish. Likewise, the consistency region includes process-faithfulness work, while checks only visible answer–trace consistency. The separate action region informs the policy audit. Placement denotes neither chronology nor causal order, and groups are not exclusive. S7 provides canonical titles and links. A winding river runs from an input marker to a release gate. Sixty paper cards with years and institutional marks are grouped under evidence and grounding, inference origins, propagation and monitoring, answer–trace consistency, and selective action. These headings organize research themes rather than define formal UIPCA fields. In particular, P requires dependent reuse of a confirmed error; C compares the answer with the visible trace. The separate action region informs policy auditing. The river is not a temporal or causal sequence.
Figure 2. A guide to the review. Click section indices to open the six topics. In the evidence panel, denotes reasoning capability and factual reliability. The five checks concern the joint adverse-effect criterion, not study classes. Dashed links among H1–H5 are schematic, not causal claims. Six panels with section indices summarize the review. Diagrams show interface visibility, the UIPCA diagnostic record and separate policy audit, joint adverse-effect checks and hypotheses, detection signals and action selection, benchmark reporting, and checks before memory, tool, or feedback writes.
| Survey / Focus | Interface and Evidence Design | Persistence and Action | What This Review Adds |
|---|---|---|---|
| Broad and domain hallucination surveys ( Ji et al., 2023 ; Huang et al., 2025 ; Zhang et al., 2025 ; Sahoo et al., 2024 ; Lee et al., 2025 ) | Organize definitions, causes, benchmarks, mitigation, or domain-specific errors; interface disclosure is not the main axis | Controls are usually attached to an output, model stage, or domain failure | Retain their error types and checks, then record visibility, evidence design, dependency reuse, and action |
| Ho et al. ( Ho et al., 2026 ) | Multi-level definition and model lifecycle connect text and vision; causal matching is not the main organizing idea | Causes and mitigations span model families, but released claims are not tracked through dependencies or later writes | Add interface-specific unknown labels, study-design classes, and claim reuse |
| Song et al. ( Song et al., 2026 ) | Crosses reasoning types with fundamental, application, and robustness failures | Identifies causes and mitigation directions across informal, formal, and embodied reasoning | Add error location and separate process diagnosis from release-policy failure |
| Long- and image-grounded CoT surveys ( Chen et al., 2026c ; Dong et al., 2026 ) | Compare trace length, training method, task demand, or interleaved states | Emphasize capability and method trade-offs rather than later unsupported writes | Use their method distinctions while checking support, visibility, and action |
| Liu et al. ( Liu et al., 2026 ) | Evaluates reasoning hallucination in an educational setting and derives a task-specific classification | Tests uncertainty-based abstention and model-to-model correction within that application | Supply the broader, source-level evidence synthesis and interface-aware diagnostic record |
| Mirages of Logic ( Anonymous ACL submission, 2026 ) | Types visible premise, operation, logic, and conclusion failures | Supports local process diagnosis; dependency reuse, interface opacity, and release action are outside its core type labels | Retain error type, then add origin/reuse, unknown-by-interface, answer–trace consistency, and policy action |
Table 1 . What neighboring surveys organize, and which joint gap this review addresses.
Figure 3. UIPCA diagnostic record. Panel (a) locates unsupported premises or invalid inferences ( ) and dependent reuse ( ). Panel (b) compares the answer with the visible trace ( ), not hidden computation. Panel (c) audits the observed action against a fixed policy ( ). Panel (d) shows interface limits: unavailable evidence remains , meaning unknown. The fields answer different diagnostic questions, not five sequential stages. Four panels separate origin and dependent reuse, visible trace–answer consistency, action-policy compliance, and interface limits. U and I are alternative origin labels. P connects an origin claim to a dependent later claim. C compares trace and answer. A is a separate policy audit. Full, summary, and hidden interfaces permit different checks.
| Evaluation Target | Operational Question | Typical Unit / Evidence | Interface Boundary |
|---|---|---|---|
| Factuality / source grounding | Is a claim true at evaluation time, or licensed by the designated source? | Atomic claim or answer; fact label, source entailment, execution result | Available under all interfaces when the required external evidence exists |
| Process faithfulness | Does the displayed reasoning reflect what causally determines the answer? | Trace–answer pairs with controlled perturbations ( Lanham et al., 2023 ; Tutek et al., 2025 ) | Causal attribution requires intervention; textual consistency alone is insufficient |
| Uncertainty calibration | Do reported probabilities or scores match empirical correctness? | Answer set; Brier, ECE, calibration curve ( Damani et al., 2025 ; Khanmohammadi et al., 2026 ) | Behavioral confidence can be evaluated under all interfaces; internal signals require provider access |
| Corrigibility | Does valid feedback change the dependent trace and outcome? | Edited step or trajectory ( Lu et al., 2025 ) | Requires before/after evidence; local acceptance is not outcome recovery |
| Selective decision quality | Does a fixed policy choose answer, retrieve, verify, clarify, abstain, or escalate appropriately? | Item and action; selective risk, coverage, action counts | The action is observable; its justification depends on declared policy inputs and evidence |
Table 2 . What each evaluation target measures and what the interface permits an evaluator to establish.
Figure 4. Publication-year distribution of the 312-record analytical corpus. The solid blue series reaches 89 records observed through 30 August 2026; 263 records (84.3%) are dated 2023–2026. The 2026 count covers only part of the year. These counts describe the curated collection, not exhaustive field-wide publication totals. A solid blue line charts annual analytical-corpus records from 2016 through the partial 2026 observation. Counts remain below ten through 2021, rise to 25 in 2022, then to 49, 54, and 71 in 2023 through 2025. The solid blue segment reaches 89 observed records through 30 August 2026. A callout states that 263 of 312 records (84.3 percent) are dated 2023 through 2026. No full-year estimate is shown.
Figure 5. Two held-out FINE-CoT cases coded with U/I/P/C/A ( Shen et al., 2026 ) . Case A reuses an invalid Step 3 reduction ( ); its answer follows the wrong trace, so . Dashed arrows abbreviate paths through omitted steps. Case B contains a separate Step 8 sufficiency error and a released option that conflicts with valid earlier results ( ). The crossed line marks this contradiction, not dependence on Step 8. With no later reuse, . Case A shows the task, an invalid Step 3 residue, reuse at Steps 8 and 12, and a wrong but trace-consistent answer. Case B separates valid earlier results from an invalid Step 8 sufficiency judgment. The released option conflicts with those earlier results; no dependency from Step 8 to the answer is asserted. Both cases keep the action label unknown because no release policy is declared.
| Setting | Premise Object | Early Check | Later Check / Action | Unresolved Boundary |
|---|---|---|---|---|
| Source-grounded text | Passage, database record, or time-indexed fact | Claim extraction and source entailment ( Chern et al., 2023 ; Hu et al., 2024 ) | Verify, retrieve, revise, abstain | Hidden evidence use and open-world completeness |
| Image or video | Object, region binding, frame, or temporal segment | Coverage and grounding ( Li et al., 2023a ; Guan et al., 2024 ; Lv et al., 2026 ; Ma, 2026 ) | Reasoning check, targeted retrieval, abstention | Whether the rationale causally used the grounded evidence |
| Function or repository code | Requirement, API, dependency, program state, repository snapshot | Type/contract/static check ( Tian et al., 2024 ; Erfanian et al., 2026 ) | Sandbox, properties, hidden tests, patch review ( Liu et al., 2023 ; Jimenez et al., 2023 ) | Untested paths, security, and unwritten intent |
Table 3 . Shared lifecycle with domain-specific evidence and release consequences.
| Comparison | Outcome | Outcome | Study Design | Unmet Criterion |
|---|---|---|---|---|
| Post-training recipes ( Yao et al., 2025 ) | No independent threshold | Factual QA rises or falls by recipe | Comparison of 12 separately trained systems | Data, reward, and optimizer vary together |
| System comparisons ( Chen et al., 2025a ) | System type, not matched | Higher hallucination for two reasoning systems | Reasoning versus non-reasoning systems | Checkpoint and output detail differ |
| Natural trace groups ( Gekhman et al., 2026 ) | Recall expansion reported separately | Lower accuracy with an incorrect extracted fact | Observed groups on the same questions | Facts are grouped, not experimentally manipulated |
| Qwen3 thinking mode ( Wang et al., 2026b ) | No independent | Aggregate gain with some correct-to-incorrect flips | Fixed checkpoint and items; mode varies | Prompts/decoding differ; beneficial flips coexist |
| Test-time budget ( Zhao et al., 2025 ) | Tokens, not independent capability | Non-monotonic accuracy and hallucination | Model/task fixed in many budget comparisons | Joint effect unresolved; answer willingness mediates outcomes |
| Factuality-aware RL ( Li and Ng, 2025 ) | Math reasoning accuracy | Factuality/hallucination measured separately | Training-pipeline comparison | Reasoning and factuality training vary together |
Table 4 . Representative S3 candidates. : separately defined reasoning capability; : factual reliability. Study design and unmet joint-effect criteria are reported separately.
Figure 6. Five hypotheses, their proposed tests, and current evidence. Each row pairs a change with an outcome to measure; dashed arrows do not denote established causal effects. H4 has direct evidence from one seeded memory-attack setting, not an estimate of ordinary error reuse. The tests require the matching controls described in the text. Table 5 specifies the remaining decisive experiments. Five aligned rows map H1 to generation and verification, H2 to reward arms, H3 to relevance and support, H4 to seeded premises, and H5 to uncertainty signals. Each row shows the outcome to measure and its current evidence grade. H1 and H3 have limited or mixed evidence. H2 and H5 have supported synthesis with bounded scope, while H4 has direct evidence in one seeded memory attack.
| Hypothesis | Evidence grade | Representative evidence | Decisive test still missing |
|---|---|---|---|
| H1: Unchecked claims accumulate | Limited or mixed | Faulty traces and complexity-related errors ( Lu et al., 2025 ; Heyman and Zylberberg, 2025 ) | Fix the initial trace; vary unchecked claims, verification, and stopping |
| H2: Rewards favor answering | Supported synthesis; bounded reward interventions | Reward granularity and calibrated reflection ( Palandye et al., 2026 ; Damani et al., 2025 ) | Randomize reward arms from one base checkpoint; fix data, budget, optimization, and evaluation; use multiple seeds |
| H3: Relevance outweighs evidence | Limited or mixed | Wrong guidance and sensitivity to supplied evidence ( Lu et al., 2025 ; Shi et al., 2023 ) | Vary relevance and support for one task using matched wording templates |
| H4: Errors are reused | Direct, single-setting seeded attack | FARMA: forged reasoning stored and reintroduced ( Karamchandani et al., 2026 ) | Seed supported, unsupported, and refuted premises; vary independent repair and measure actual reuse |
| H5: Uncertainty changes action | Supported synthesis; limited transfer | Uncertainty-guided control and the monitoring–control gap ( Yan et al., 2026 ; Yu et al., 2026b ) | Change only the signal; measure action and useful coverage, then test distribution shift |
Table 5. H1–H5: claims, evidence, and decisive tests. Representative studies support the indicated part of each hypothesis, not every link. The last column gives the remaining experiment.
Figure 7. Complementary detection mechanisms and action control. Panels (a)–(e) illustrate evidence checks, sampling, execution, monitoring, and model-based judging. The controller chooses an action; audits policy compliance. Table 6 compares the methods. Five panels depict external evidence checking, semantic agreement across samples, executable checks with independent specifications, monitoring of visible steps and optional internal states, and model-based assessment of claim support and visible answer–trace consistency. A bottom strip connects risk signals, evidence status, and task policy to action selection. A separate A audit compares the observed action with the predefined policy.
| Signal / Examples | Unit / Covered Component | Interface / Access | Feasible Action | Boundary and Cost |
|---|---|---|---|---|
| External evidence: CLATTER, RefChecker ( Eliav et al., 2025 ; Hu et al., 2024 ) | Claim; U and source-backed I | All interfaces; source required | Accept, retrieve, revise, abstain | Strong support; extraction and checking cost grows with the number of claims |
| Uncertainty or learned scoring: Semantic Entropy Probes, FaithLens, UPAIR ( Kossen et al., 2024 ; Si et al., 2025 ; Yan et al., 2026 ) | Answer/trace; routes checks rather than establishing support or C | White-box probe or learned detector | Verify, resample, reject | Checkpoint and training-domain dependence |
| Verification and calibrated action: VeriFY, CURE, Latent Critic ( Altinisik et al., 2026 ; Liu and Wang, 2026 ; Vijayvargiya and Lokesh, 2026 ) | Claim/answer; support checks and action selection | Black-box model, labels, or internal critic | Answer, verify, abstain | Coverage depends on boundary estimation and detector transfer |
| White-box trajectory: TRACED, HARP, Reasoning Denoiser ( Jiang et al., 2026 ; Hu et al., 2025a ; Fang et al., 2026 ) | Step/trajectory; I and P routing | Internal states or trained features | Stop, inspect, escalate | Early but checkpoint-specific; no external factual evidence |
| Online search/repair: HaluSearch, Token-Guard ( Cheng et al., 2025b ; Zhu et al., 2026 ) | Prefix/segment; I and P | Generation-time score and access | Redirect, prune, regenerate | Detection and corrective policy effects are coupled |
| Executable and grounded checks ( Ling et al., 2023 ; Pan et al., 2023 ; Ma, 2026 ; Jin et al., 2026 ) | Constraint, region, program state; U/I | Evidence, tools, or sandbox | Bind, execute, repair, abstain | The check covers only declared evidence and tested properties |
Table 6 . Detection-to-action comparison. Component mappings identify possible uses of a signal, not a claim that every listed method directly annotates UIPCA.
| Supervision or verifier | Target | Useful intervention | Characteristic failure |
|---|---|---|---|
| Evidence-backed premise check | Reject or retrieve before a premise enters | Open-world evidence gaps; judge shares generator bias | |
| Rule, math, or execution check | Reject an invalid local transition | Incomplete checker; valid alternative paths scored as wrong | |
| Dependency-aware reward | Stop reuse or invalidate a branch | Missing edges; rewards favor style or late correction rather than preventing reuse | |
| Answer–trace or action verifier | Compare the answer with the trace; revise, verify, abstain, or escalate | Late detection; exploitable scores, unnecessary refusal, or an unrepaired trace |
Table 7. Process supervision and verifier placement by lifecycle target. No verifier type covers all components.
Figure 8. From claim checks to release and persistent-state control. Claim support and visible-process diagnostics inform a controller governed by a fixed policy. Harm, reversibility, coverage, and cost affect action selection, not evidence status. audits the chosen action against that policy. State-changing operations also require evidence, write-authority, and version checks. Private or shared memory, tool or world state, and training feedback are parallel destinations. Dashed paths are conditional; retained output may become evidence in a later round, but reuse is not inevitable. Available records feed claim checking and, when visible information permits, process localization. Evidence and a separate fixed policy feed action selection: release, verify or revise, clarify or abstain, or escalate. Before any state-changing operation, a write gate binds evidence, preserves unknowns, checks write authority and versions, and logs provenance and repairs. Authorized writes may enter private or shared memory, tool or world state, or training feedback in parallel. Failed checks lead to blocking or repair; retained state may become input to a later round.
| Family | Primary Unit | Directly Reveals | Required Companion Report |
|---|---|---|---|
| Factual QA / grounding: SimpleQA, TriviaQA, FACTS ( Wei et al., 2024a ; Joshi et al., 2017 ; Cheng et al., 2025a ) | Answer or claim–source pair | Outcome factuality or source support | Evidence/evaluator version; extraction rule; answer rate and coverage |
| Trace / perturbation: FINE-CoT, C2-Faith, RACE ( Shen et al., 2026 ; Mittal and Arike, 2026 ; Wang et al., 2025a ) | Step, dependency, trace, answer | Selected process errors; coverage depends on suite labels | Segmentation, codebook, boundary cases, interface, unobserved steps |
| Calibration / answerability: RMCB, VeriFY, AbstentionBench ( Khanmohammadi et al., 2026 ; Altinisik et al., 2026 ; Kirichenko et al., 2025 ) | Confidence, answer, action | Ranking, calibration, answer/abstain behavior | Brier/ECE, selective risk, coverage, unnecessary refusal |
| Long context / changing evidence: RULER, Lost in the Middle, StreamingQA, HalluDial ( Hsieh et al., 2024 ; Liu et al., 2024 ; Liška et al., 2022 ; Luo et al., 2024 ) | Retrieved answer or turn | Outcome under available, positioned, or changing evidence | Retrieval/binding record, evidence time, first-error annotation |
| Multimodal: POPE, HallusionBench, AMBER ( Li et al., 2023a ; Guan et al., 2024 ; Wang et al., 2023 ) | Object, region, claim, answer | Perception/grounding and selected outcome | Coverage, binding, reasoning, answer, abstention as separate fields |
| Code / software: CodeHalu, CRUXEval, EvalPlus, SWE-bench ( Tian et al., 2024 ; Gu et al., 2024 ; Liu et al., 2023 ; Jimenez et al., 2023 ) | Program state, function, repair, patch | Selected execution, repair, or issue-resolution behavior | Environment, tests/coverage, analyzer, security, repository snapshot |
Table 8 . Benchmark families and reporting coverage needed for a release decision. Named resources illustrate each family; protocols and metrics are not interchangeable. Memory governance is separated from answer quality and semantic memory errors.
| System | Control point | Check or evidence | Remaining boundary |
|---|---|---|---|
| A-MAC ( Zhang et al., 2026 ) | Candidate admission and conflict-aware replacement | Weighted utility, overlap-based confidence, novelty, recency, and type prior | Admission precision/recall does not establish semantic truth; overlap is a support proxy |
| TrustMem ( Yang et al., 2026 ) | Learned write, revise, and prune behavior | Transition verifier scores coverage, preservation, and faithfulness for reward and preference training | Frozen-LLM judgments depend on the supplied chunk and prior memory; no independent truth guarantee |
| Governed Memory ( Taheri, 2026 ) | Typed extraction, entity-scoped retrieval, policy routing | Heuristic quality gates, schema checks, provenance metadata, and bounded retrieval | Self-evaluation; sequential conflict tests leave concurrent multi-agent writes unresolved |
| MemClaw / ArgusFleet ( Margalit et al., 2026 ) | Scoped access, supersession, and controlled propagation | Service traces test authorized visibility, conflict handling, and provenance reconstruction | One-service self-evaluation; admission can pre-empt correction; visibility is not dependent reasoning use |
Table 9. Persistent-memory controls and the evidence they provide. Systems are compared by control point, not by scores from different tasks. Evaluation resources in Table 8 serve a separate role.