External cognitive honesty decision-making engine for the reasoning stage and a differentiated conversation solidification and re-interrogation pattern mapping method

By introducing a lightweight domain classifier and an external decision engine during the reasoning stage of a large language model, the model's internal hesitation is quantified and linked with external validation results to execute three-level decisions. This solves the cognitive honesty problem of large language models when generating answers and achieves efficient, auditable decision-making and differentiated processing of repaired inputs.

CN123117348APending Publication Date: 2026-11-17SHANGHAI QIYUANCHENG BUSINESS SERVICE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611385482.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing large language models suffer from several problems when generating answers, including a lack of quantification of internal uncertainty, a lack of linkage decision-making between internal uncertainty and external knowledge verification, a lack of domain-adaptive thresholds, insufficient differentiated repair after warnings and rejections, and a lack of standardized interfaces with downstream intelligent retry and scheduling modules, resulting in insufficient guarantees of cognitive honesty.

Method used

By setting an adaptive threshold using a lightweight domain classifier, the model's internal hesitation is quantified and linked with external validation results. Three-level decision-making is performed to distinguish the repair data boundaries between CAUTION and ABSTAIN, and reusable re-interrogation mode instructions are output, maintaining clear boundaries with other modules.

Benefits of technology

It enables forced cognitive honesty without modifying model parameters, quantifies internal hesitation and links it with external validation results, avoids erroneous release in high-risk hesitation scenarios, distinguishes the repair data boundaries between CAUTION and ABSTAIN, and reduces the latency and computing power consumption of downstream repeated retrieval and validation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure FT_1
    Figure FT_1
  • Figure FT_2
    Figure FT_2
Patent Text Reader

Abstract

This invention discloses an external cognitive honesty decision-making engine for the inference stage and a differentiated conversation solidification and re-interrogation pattern mapping method, belonging to the field of artificial intelligence and natural language processing technology. This engine does not modify the parameters of the underlying large language model, but instead reads or receives the probability distribution of the Top-5 candidate tokens from the output layer at the inference service API layer, calculates the probability entropy H and probability concentration C to mark model hesitation; receives upstream bidirectional fact anchoring verification results and optional four-layer classification detection evidence; integrates the internal uncertainty marker with the external verification pass rate and reverse coverage rate, and executes three-level decisions according to the priorities of OUTPUT, ABSTAIN, and CAUTION; and generates differentiated conversation solidification instructions and re-interrogation pattern mapping instructions for CAUTION and ABSTAIN. In the CAUTION scenario, the original response text is retained for pattern correction; in the ABSTAIN scenario, the original response text is prohibited as reconstruction input, and reconstruction is only based on the key fact list and the omission list. This invention outputs decision signals, solidification instructions, and pattern mapping instructions, without performing prompt word construction, model regeneration, lightweight verification, or optimized response output, and without updating the scheduling strategy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and natural language processing technology, specifically relating to a method and system for adding an external cognitive honesty decision-making layer during the reasoning stage of a large language model. More particularly, it involves an engine architecture that integrates the quantification of uncertainty in the model output layer with external knowledge verification results, executes hierarchical decisions, and achieves closed-loop input repair through differentiated session solidification and re-interrogation pattern mapping. This invention does not perform specific regeneration and prompt word construction for intelligent retries, does not retry verification, and does not update the scheduling strategy; it only outputs DecisionSignal, SessionFixInstruction, and RetryInstruction, and maintains interface coordination with PA1 declaration flow, PA2 fact anchoring verification, PA4 classification and evidence detection, PA5 intelligent retry execution, and PA6 scheduling parameters. Background Technology

[0002] The frequent "illusion" phenomenon in large language models when generating answers is partly due to the structured incentives inherent in some training paradigms. Training methods, such as reinforcement learning based on human feedback, often reward "confident, complete, and fluent" answers more highly, while giving lower rewards or not considering "acknowledging uncertainty, refusing to answer, or declaring unverifiable responses" as preferred samples. This can easily lead the model to produce definitive statements after training, even if its internal state does not support such a conclusion. While training-phase intervention methods, such as constitutional AI, direct preference optimization, or similar alignment techniques, can alleviate this tendency, they usually require adjusting model parameters, reward models, or preference data, making them difficult to directly apply to deployed third-party models. Furthermore, for closed-source models or third-party API models already in service online, business users typically cannot modify their training process or internal weights.

[0003] Another approach employs a prompting engineering method, using system prompts, few-shot examples, or output format constraints to guide the model in expressing uncertainty when faced with uncertainty. This approach is simple to implement, but the constraint strength is limited: the model may still ignore the constraints due to decoding randomness, insufficient prompt coverage, or contextual competition, continuing to generate deterministic or even fabricated content; for high-risk scenarios, prompts alone cannot form an auditable and mandatory basis for decision-making.

[0004] In addition, existing reasoning phase schemes also have the following common defects.

[0005] Defect 1: Lack of a quantifiable internal uncertainty mechanism. Some systems can only judge whether the model is "confident" based on the surface wording of the generated text, and cannot use internal states such as the probability distribution of candidate tokens in the output layer to objectively characterize whether the model hesitates at key generation positions; while the model's text expression is often overconfident, and surface confidence cannot be equated with internal uncertainty.

[0006] Defect 2: Lack of coordinated decision-making between internal uncertainty and external knowledge verification. The probability distribution of the model output layer reflects "whether the model hesitates internally," while external knowledge base verification reflects "whether the answer is supported by evidence." These are different information sources; without fusion rules and priority adjudication, the system either relies excessively on model self-evaluation or excessively on post-event comparison, making it difficult to execute consistent and auditable rejections or warnings when there is uncertainty and verification fails.

[0007] Deficiency 3: Lack of domain-adaptive thresholds. Different domains have significantly different tolerances for uncertainty: casual conversation scenarios can be moderately lenient, professional academic scenarios require standardized verification, while medical, legal, and financial scenarios require stricter handling for any critical uncertainty. A uniform threshold can easily lead to overly conservative approaches in low-risk scenarios or missed detections in high-risk scenarios.

[0008] Defect 4: Insufficient differentiated repair after warnings and rejections. When the system outputs a CAUTION with a warning, the original response may still contain verified correct information. It's suitable to revise the question statement and supplement any missing facts while retaining the correct parts. However, when the system outputs an ABSTAIN indicating an active rejection, the original response may contain serious factual errors, high-risk omissions, or unverifiable content, making it unsuitable as a basis for subsequent generation. If the same retry strategy is used for both scenarios, it's easy for correct information in the CAUTION scenario to be wasted, while incorrect information in the ABSTAIN scenario to be inherited.

[0009] Defect 5: Lack of standardized interface with downstream intelligent retry and scheduling modules. If the decision engine only outputs natural language conclusions without outputting decision type, key fact list, omission list, problem statement list, verification metadata, and pattern mapping instructions, downstream modules cannot reuse verified results. This causes the entire retrieval and verification process to be repeated for each retry, increasing latency and computing power consumption.

[0010] Therefore, existing technologies urgently need an external decision-making mechanism for ensuring cognitive honesty in the LLM inference stage. This mechanism should not modify the underlying model parameters, but should be able to read or receive uncertainty signals from the output layer, integrate external verification results, execute hierarchical decisions according to safety baselines, and generate differentiated session solidification and re-interrogation pattern mappings for warning / refusal scenarios. This invention does not address model training alignment itself, nor does it replace external knowledge base retrieval; its focus is on the decision diversion and repair orchestration of inputs at the API layer during the inference stage. Summary of the Invention

[0011] The problems to be solved by the present invention are: how to enforce cognitive honesty without modifying model parameters; how to quantify internal hesitation of the model and link it with external verification results; how to prevent high-risk hesitation scenarios from being erroneously released according to priority; how to distinguish the repaired data boundary between CAUTION and ABSTAIN; and how to output reusable instructions to the intelligent retry module, while maintaining clear boundaries with PA1 / PA2 / PA4 / PA6.

[0012] S0 Domain adaptive threshold adjustment. A lightweight domain classifier is used to identify the domain to which a question belongs, and H_threshold, C_threshold, T_pass1, R_cov1, T_pass2, and R_cov2 are set; the thresholds are configurable parameters, and preferred examples are given in the embodiments. The PA6 scheduling module may issue threshold versions or scheduling parameters, but PA3 shall adopt them only within the allowable range of the safety baseline, and shall not lower the high-risk threshold due to user portraits or scheduling preferences.

[0013] S1 Input. Receive user query Q, LLM generated token stream T, and optional domain label D.

[0014] S2 Uncertainty quantification. When the output layer probabilities are available, read the Top-5 candidate token probability distribution {p1,...,p5}, and calculate H=-Σpi ln pi and C=p1 / (p2+p3+p4+p5). When the denominator is 0, C is regarded as infinity; when H>H_threshold or C<C_threshold, the generation position is marked as hesitation U. If the third-party API does not expose the output layer probabilities, U=unknown; the threshold shall not be relaxed due to unknown in high-risk scenarios, and higher verification depth or safety rejection may be triggered, and the unknown event shall be recorded.

[0015] S3 External result reception. Receive the VerificationResult of PA2 two-way dynamic fact anchoring verification through a standardized API, including the positive pass rate F, negative coverage rate V, key fact list K, omission list M, problem statement list P, evidence pointer, threshold version and knowledge base version; optionally receive the PA4 DetectionResult as evidence supplement, but the intervention suggestion of PA4 does not replace the final decision of this engine. This engine does not take the retry result of PA5 as the basis for the first decision, and only reads whether the retry is successful in the subsequent audit link.

[0016] S4 Three-level decision making. Execute in order of priority: First priority OUTPUT, when F≥T_pass1, V≥R_cov1 and no hesitation U exists, directly output the final response A; Second priority ABSTAIN, when the OUTPUT condition is not satisfied, F<T_pass2, V<R_cov2 and U exists, return the standard refusal response script; Third priority CAUTION, when neither the OUTPUT nor ABSTAIN conditions are met, output a response with a warning attached. Security coverage rule: This decision takes precedence over the PA2 classification reference and the PA4 detection layer intervention suggestion, but it is not allowed to change a numerical statement that has been judged as contradictory by PA2 to pass without new evidence; it is not allowed to delete or tamper with the evidence pointer; in high-risk fields, U=unknown only allows increasing the verification depth or performing a security refusal response, and does not allow lowering the threshold.

[0017] S5 Differentiated session solidification and pattern mapping. When CAUTION is triggered, generate the first solidification instruction SessionFixInstruction-caution, solidify K, M, P, A_original and verification metadata, and send RetryInstruction-corrective to PA5; when ABSTAIN is triggered, generate the second solidification instruction SessionFixInstruction-abstain, solidify K, M and verification metadata, prohibit using A_original or its derived text as reconstruction input, only save the irreversible audit fingerprint original_audit_hash, and send RetryInstruction-reconstructive to PA5. The solidified data is associated with the session ID, timestamp, query semantic feature vector and TTL, and is hard deleted upon expiration; behavior records must be authorized, anonymized, closable, and deletable.

[0018] The present invention further provides an engine, comprising a domain adaptation threshold module, an uncertainty quantification module, a two-way verification result receiving module, a three-level decision making module, and a differentiated session solidification and pattern mapping module; outputting DecisionSignal, SessionFixInstruction and RetryInstruction.

[0019] 3.3 Coordination relationship of peripheral modules Relationship with PA1: PA1 only provides ClaimStream; this engine does not generate atomic statements, does not modify spans, and does not perform atomic splitting.

[0020] Relationship with PA2: PA2 provides VerificationResult and classification reference; this engine takes its own fusion decision as the final shunting basis, but shall not overwrite the contradictory evidence confirmed by PA2.

[0021] Relationship with PA4: PA4 provides DetectionResult and differentiated intervention suggestions; this engine can write it into DecisionSignal as evidence pointers and triggering reasons, but does not directly equate PA4 priority with end-user traffic allocation.

[0022] Relationship with PA5: PA5 receives the RetryInstruction from this engine and performs corrections / refactoring; this engine does not construct prompt words, does not call LLM regeneration, does not perform lightweight verification, and does not output A′.

[0023] Relationship with PA6: PA6 can issue P_sched, threshold version, security baseline, and mode preference; this engine is only used above the security baseline, does not send back profiles, and does not update scheduling policies.

[0024] Relationship with RAG / Orchestration Layer: This engine can be used as a plugin for the inference service API layer, outputting DecisionSignal upstream and fixed instructions to the session state management module, without directly modifying the underlying LLM architecture.

[0025] 3.4 Data Objects and Interface Fields DecisionSignal fields include: decision_id, decision_type∈{OUTPUT,CAUTION,ABSTAIN}, F, V, U∈{true,false,unknown}, domain, threshold_version, safety_floor_id, trigger_reason, evidence_pointer, trace_id, created_at. trigger_reason at least distinguishes: forward_contradiction, numeric_forced_contradiction, reverse_missing_high_priority, uncertainty_H, uncertainty_C, unknown_probability.

[0026] SessionFixInstruction-caution fields include: session_id, decision_type, K, M, P, A_original, metadata{F,V,threshold_version,kb_version,model_version,evidence_pointer}, ttl, query_vector_id.

[0027] The SessionFixInstruction-abstain field includes: session_id, decision_type, K, M, metadata, ttl, query_vector_id, and original_audit_hash; it may not contain the A_original field or may only contain the audit_hash field.

[0028] RetryInstruction-corrective fields include: session_id, mode=corrective, allowed_inputs=[K,M,P,A_original], forbidden_inputs=[], validation_scope=incremental_only.

[0029] RetryInstruction-reconstructive fields include: session_id, mode=reconstructive, allowed_inputs=[K,M], forbidden_inputs=[A_original], validation_scope=new_claims_only.

[0030] The interface supports idempotent keys, timeouts, rate limiting, error codes, and trace_id. Error codes include at least: PROBABILITY_UNAVAILABLE, EVIDENCE_POINTER_INVALID, SAFETY_FLOOR_VIOLATION, FIX_INSTRUCTION_SCHEMA_INVALID, RATE_LIMITED, and RETRY_MODULE_UNAVAILABLE.

[0031] 3.5 Status, Anomalies, and Degradation When the output layer probability is unavailable, U = unknown. If the domain is casual / general, a conservative strategy can continue. If the domain is professional / high-risk, the threshold is not relaxed by default, and higher verification depth or ABSTAIN / safe rejection is triggered first. If PA2 VerificationResult is missing, it must not enter OUTPUT; CAUTION or ABSTAIN is output according to the domain's safety baseline, and missing_verification_result is recorded. If PA5 is unavailable, this engine still outputs DecisionSignal and fixed instructions, and RetryInstruction enters the waiting queue or is switched to a backup retry link by the upstream orchestration layer; this engine does not modify the decision causality because of this. If A_original is empty but the decision is CAUTION, it is downgraded to only sending K / M / P and correction instructions, and a_original_missing=true is marked in DecisionSignal.

[0032] (1) Forced splitting of the inference stage can be implemented without modifying the model. (2) The output layer H / C quantifies "inner hesitation" to avoid relying solely on superficial confidence in the text. (3) The priority mechanism ensures that ABSTAIN takes precedence when there is serious uncertainty and external verification fails. (4) CAUTION preserves correct information, while ABSTAIN prohibits the inheritance of errors and corrects input differences. (5) Standardized interfaces reduce downstream duplicate retrieval and verification. (6) The boundaries with PA1 / PA2 / PA4 / PA5 / PA6 are clear, allowing for plug-in collaboration and auditable evidence chains. Attached Figure Description

[0033] Figure 1 The flowchart for the external cognitive honesty decision engine and re-interrogation connection in the reasoning stage shows the steps of S0 threshold adjustment, S1 input, S2 uncertainty quantification, S3 three-level decision, CAUTION / ABSTAIN session solidification, S5 waiting for re-interrogation, S6 repeated interrogation detection, S7 correction / reconstruction mode mapping, and downstream PA5 execution boundary.

[0034] Figure 2 This is a diagram comparing the solidified content of differentiated sessions. It shows that in the CAUTION scenario, K / M / P / A_original and verification metadata are solidified and the correction mode is entered. In the ABSTAIN scenario, K / M and verification metadata are solidified, the original reply is prohibited as a reconstruction input, and the reconstruction mode is entered. Detailed Implementation

[0035] Example 1: Medical Strict ABSTAIN. A user inquires about aspirin dosage, categorized as medical / high-risk. Strict thresholds are given as follows: T_pass1=0.95, R_cov1=0.90, T_pass2=0.80, R_cov2=0.70, H_threshold=1.2, C_threshold=1.0. The LLM responds with "1000mg". PA2 determines the values ​​are contradictory and the reverse coverage is insufficient. S2 reads the Top-5 probabilities, obtaining H=1.30>1.2 and C=0.82<1.0, marking U=true. ABSTAIN is satisfied, and the standard dialogue is output. S5 generates a second fixed instruction, only fixing K / M / metadata, disabling A_original as reconstruction input, and sending a reconstruction mode instruction to PA5.

[0036] Example 2: CAUTION Correction. When a user queries GDP growth rate and a professional / academic standard threshold, PA2 determines the numbers are contradictory but the reverse coverage is complete. S2 marks hesitation but does not meet the ABSTAIN reverse low coverage condition, executes CAUTION, and outputs a warning. S5 generates the first fixed instruction, retaining A_original, P, K, and M, and sends a correction mode instruction to PA5. Upon the user's re-query, PA5 corrects the numbers based on A_original and supplements any omissions.

[0037] Example 3: Probability Unavailable. When the third-party API does not expose logprobs, U = unknown. In casual / professional scenarios, a conservative strategy can continue; in medical / legal / financial scenarios, the threshold must not be lowered, the verification depth should be increased, or a secure rejection should be returned, and the unknown event should be recorded. If the PA2 result is also missing, OUTPUT is not allowed; CAUTION or ABSTAIN should be output according to the safety baseline, and missing_verification_result should be recorded.

[0038] Embodiment 4: Interface sample. DecisionSignal: {decision_id:"d1",decision_type:"ABSTAIN",F:0.0,V:0.33,U:true,domain:"medical",threshold_version:"th-med-2026q1",safety_f loor_id:"sf-high-1",trigger_reason:["numeric_forced_contradiction","uncertainty_H"],evidence_pointer:["ptr-7","ptr-8"],trace_id:"tr-3"}. RetryInstruction-reconstructive: {session_id: "s1", mode: "reconstructive", allowed_inputs: ["K", "M"], forbidden_inputs: ["A_original"], validation_scope: "new_claims_only"}.

[0039] Example 5: Coordination Example. PA6 issues threshold_version="th-std-2026q1" and mode_pref="prefer_corrective"; PA3 only uses this if the version meets the safety baseline. PA4 returns citation trigger evidence, and PA3 records citation_unverified in trigger_reason, but ultimately executes ABSTAIN due to the contradiction in PA2's value and U=true, instead of directly allowing it because of PA4's orange flag.

[0040] Example 6: Simulation Comparison (Simulation settings, not actual testing). Among the same batch of high-risk queries, compared to the group with only text confidence hints, the group with only PA2 post-validation, and the group fused by this engine, this engine showed higher consistency in rejecting responses on the subset of "model hesitation and external validation failure," and reduced the number of cases where contradictory PA2 values ​​were mistakenly allowed. On the chatter subset, the CAUTION ratio decreased. This simulation description should be replaced with real logs or controlled experiments before formal submission.

Claims

1. A reasoning-stage plug-in cognitive honesty decision-making engine, characterized in that, comprising: a domain adaptive threshold module, configured to identify the domain to which a question belongs and set an uncertainty threshold and a pass rate / coverage threshold; an uncertainty quantification module, configured to read the Top-5 candidate token probability distribution when output layer probabilities are available, calculate the probability entropy H=-Σpi ln(pi) and the probability concentration C=p1 / (p2+p3+p4+p5), and mark model hesitation U when H is greater than the uncertainty threshold or C is less than the concentration threshold; output U=unknown and perform processing according to the safety bottom line when the output layer probabilities are not available; a two-way verification result receiving module, configured to receive upstream two-way fact anchoring verification results, including the forward pass rate F, the reverse coverage V, the key fact list K, the omission list M, the question statement list P and evidence pointers; a three-level decision module, configured to execute decisions according to priority: the first priority is OUTPUT, when F≥the first pass rate threshold, V≥the first coverage threshold and there is no hesitation U, directly output the final reply; the second priority is ABSTAIN, when the OUTPUT condition is not satisfied, F<the second pass rate threshold, V<the second coverage threshold and there is hesitation U, return a standard refusal reply phrase; the third priority is CAUTION, when neither the OUTPUT nor ABSTAIN conditions are satisfied, output a reply with a warning; the three-level decision module is configured to perform execution with reference to the grading of the upstream fact verification layer in priority, but shall not change the numerical statement that has been judged as contradictory by the upstream to passed when no new evidence is added, and shall not delete or tamper with the evidence pointers; a differentiated session solidification and pattern mapping module, configured to generate a first solidification instruction including K, M, P, the original reply text A_original and verification metadata when in CAUTION, and send a correction mode instruction to the intelligent retry module; generate a second solidification instruction including K, M and verification metadata when in ABSTAIN, prohibit using A_original or its derived text as a reconstruction input, only save an irreversible audit fingerprint, and send a reconstruction mode instruction to the intelligent retry module; the solidified data is associated with a session ID, a timestamp, a semantic feature vector and TTL, and is hard deleted upon expiration.

2. The engine according to claim 1, characterized in that, The thresholds are configurable parameters; preferably, a loose threshold is adopted in the chat / common sense domain, a standard threshold is adopted in the professional / academic domain, and a strict threshold is adopted in the medical / legal / financial domain, where any uncertainty triggers more strict processing or a safety refusal reply.

3. The engine according to claim 1, characterized in that, The ABSTAIN condition is F<T_pass2 and V<R_cov2 and hesitation U exists; CAUTION is the remaining cases that do not satisfy the OUTPUT and ABSTAIN conditions; When U=unknown, the threshold shall not be lowered in high-risk domains.

4. The engine according to claim 1, characterized in that, The verification metadata includes the forward pass rate, the reverse coverage, the domain threshold, the uncertainty quantification indicator, the query semantic feature vector, the threshold version, the model version, the knowledge base version and the evidence pointers.

5. The engine according to claim 1, characterized in that, The TTL is set per domain; preferably, 300 seconds for chat, 600 seconds for professional use, and 900 seconds for medical / legal / financial use; user behavior records are authorized, anonymized, switchable off and deletable.

6. The engine according to claim 1, characterized in that, The three-level decision module may optionally receive four-layer classification detection evidence as a supplement to trigger_reason, but does not directly equate the detection layer intervention suggestions with the final diversion; the engine does not accept the retry results of the intelligent retry module as the basis for the initial decision.

7. The engine according to claim 1, characterized in that, The scheduling module can issue threshold versions, security baselines, and mode preferences; the engine only uses scheduling parameters above the security baseline, does not send back user profiles, and does not update the scheduling strategy.

8. A cognitive honesty decision-making method for the reasoning stage based on the engine described in any one of claims 1-7, characterized in that, The process includes S0 domain threshold adjustment, S1 input, S2 uncertainty quantification, S3 receiving external verification results, S4 three-level decision-making, and S5 differential session solidification and pattern mapping.

9. The method according to claim 8, characterized in that, The correction mode instruction instructs the downstream intelligent retry module to retain correct statements, correct questionable / contradictory statements in the problem statement list P, and supplement key facts in the omission list M; the reconstruction mode instruction instructs the downstream intelligent retry module to construct a complete response from scratch based on K and M, without inheriting the original response content.

10. The method according to claim 8, characterized in that, The method does not perform downstream intelligent retry prompt word construction, model regeneration, lightweight verification, and final optimization response output; the method only outputs decision signals, solidification instructions, and pattern mapping instructions.