Non-productive loop determination method based on information discovery rate adaptive gating

Abstract

The present application relates to the technical field of information retrieval, and discloses a non-productive cycle determination method based on information discovery rate adaptive gating, which comprises the following steps: first, receiving an event stream called in the running process of an automated research assistant to obtain a standardized research event sequence arranged in ascending order of time; obtaining a comprehensive determination signal and outputting the same; calculating the difference file total number of the work area path set of two adjacent events according to the work area path set corresponding to the two adjacent events and the comprehensive determination signal, and outputting a real change indication signal as an invalid state; performing final running state determination according to the comprehensive determination signal, the work area real change indication signal and a marginal benefit diminishing alarm signal, and matching a corresponding recovery action according to a preset priority rule and outputting the same according to the final running state and the specific triggered alarm mode. The present application can output accurate cycle determination results and targeted recovery actions.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, specifically to a non-productive loop determination method based on adaptive gating of information discovery rate. Background Technology

[0002] The automated research assistant is an intelligent system built on a large language model, capable of autonomously performing complex tasks such as information retrieval, literature reading, data extraction, logical reasoning, and research report generation. In scenarios such as academic research, industry surveys, and policy analysis, the research assistant interacts with the environment by repeatedly calling tools such as search engines, academic databases, and document parsers to gradually collect information and synthesize it into structured reports.

[0003] However, in actual operation, research assistants frequently fall into unproductive cycles, that is, repeatedly performing similar search and browsing actions for extended periods without generating effective research progress. This manifests as repeatedly using the same keywords to search for the same topic, switching back and forth between different databases only to obtain duplicate literature, and performing multiple rounds of information retrieval without producing substantial analytical conclusions. This problem leads to a sharp increase in API call costs, meaninglessly prolonged research cycles, and a severe decline in user experience.

[0004] Existing technologies for detecting loops in research assistants mainly fall into three categories: The first category is surface-based detection methods based on text similarity. These methods calculate the edit distance or semantic similarity between adjacent search results and document abstracts, identifying loops when the similarity exceeds a threshold. Their drawback is that they only focus on the textual surface and cannot distinguish between "effective duplication" (such as deep cross-validation of core literature) and "ineffective duplication" (such as repeatedly searching for the same keyword to obtain the same results), resulting in a very high false positive rate in research assistant scenarios. The second category is hard-termination methods based on a fixed number of iterations or API call limits. For example, setting a maximum of 20 search rounds or a maximum of 50 document downloads. Their drawback is that the threshold is rigid and cannot adapt to the varying complexity of different research topics. Simple topics may only require 5 rounds, while complex interdisciplinary topics may require more than 50 rounds; a one-size-fits-all threshold either prematurely terminates effective exploration or fails to prevent ineffective loops in time. The third category is natural language judgment methods based on self-reflection of large language models. These methods allow the research assistant to determine whether it needs to continue searching. Its drawbacks lie in the lack of quantitative evidence, the judgment results being heavily influenced by model illusions, and the model tending to overconfidently continue searching to "ensure comprehensiveness," which in fact exacerbates the cycle.

[0005] The fundamental flaw in the aforementioned existing technologies lies in their modeling of the "should we stop?" question as a textual representation or subjective judgment problem, rather than an information gain problem. The core output of a research assistant is "newly discovered knowledge entities," including new literature, new concepts, and new data relationships. Existing methods do not measure, from an information theory perspective, whether the rate of new knowledge discovery per unit time is sufficient to support the costs of continued operation. Furthermore, existing methods cannot identify "pseudo-progress," i.e., situations where retrieval logs are frequently updated but the number of reported files in the workspace does not substantially increase.

[0006] Therefore, to address the above issues, a non-productive cyclic decision method based on information discovery rate adaptive gating is needed. Summary of the Invention

[0007] The purpose of this invention is to provide accurate loop determination results and targeted recovery actions.

[0008] This invention is implemented as follows: First, the event stream invoked during the operation of the automated research assistant is received, and each event is parsed in a structured manner. Then, all parsed events are sorted by timestamp and abnormal events are removed to obtain a standardized research event sequence arranged in ascending order of time. Based on the output standardized research event sequence, output the cumulative knowledge entity set and the total number of entities it contains, and the near-window entity set and the total number of entities it contains; Based on the cumulative knowledge entity set and the near window entity set, the number of entities in the near window entity set that are not included in the cumulative knowledge entity set is counted as the number of new knowledge entities discovered within the near window; the number of new knowledge entities within the near window is divided by the total number of cumulative knowledge entities to obtain the information discovery rate index and output it. Based on the standardized research event sequence, non-productive mode detection is performed in parallel, and independent alarm signals are output respectively; if the number of occurrences of any query keyword exceeds the preset duplicate query threshold, a duplicate query alarm is triggered, otherwise it is not triggered. Based on the output information discovery rate index and alarm signals, a preset information discovery rate gating threshold is set; adaptive gating processing is performed on retrieval chain alarms, analysis action ratio alarms, and duplicate query alarms to obtain a comprehensive judgment signal and output it; Based on the workspace path set and comprehensive judgment signal corresponding to two adjacent events, calculate the total number of files with differences in the workspace path set of two adjacent events, and output a real change indication signal as invalid. Based on the comprehensive judgment signal, the actual change indication signal of the working area, and the diminishing marginal returns alarm signal, the final operating status is determined. According to the final operating status and the specific alarm mode triggered, the corresponding recovery action is matched and output according to the preset priority rules.

[0009] Furthermore, the structured parsing of each event includes extracting seven core fields: event category tags, new knowledge entity set, new research artifact set, new document URL set, new query keyword set, estimated research report completion progress, and complete research workspace path set, to construct a single standardized research event. The event category tags include two types: retrieval and analysis. Retrieval corresponds to information acquisition actions, including search engine queries, academic database searches, and full-text document downloads. Analysis includes knowledge processing actions, including document abstract extraction, viewpoint comparison and synthesis, and report paragraph generation. The output of the cumulative knowledge entity set and the total number of entities contained therein, and the near window entity set and the total number of entities contained therein, includes taking the union of the new knowledge entity set corresponding to all events and the new research workpiece set, deduplicating them, constructing and updating the cumulative knowledge entity set; and setting a preset sliding window size, taking the union of the new knowledge entity set corresponding to all events within the window coverage area and the new research workpiece set, deduplicating them, constructing and updating the near window entity set.

[0010] Furthermore, the information discovery rate index ranges from 0 to 1; when the index approaches 1, it indicates that the efficiency of new knowledge discovery within the sliding window is high and the information gain is significant; when the index approaches 0, it indicates that the research action mainly involves repeatedly accessing known content and the information gain approaches zero; the gating threshold is dynamically adjusted according to the total amount of knowledge in the research topic.

[0011] Furthermore, the non-productive mode detection includes: Mode 1 is retrieval chain mode detection, which counts the length of the longest continuous retrieval event chain in the event sequence. If the length is greater than or equal to a preset continuous retrieval threshold, a retrieval chain alarm is triggered; otherwise, it is not triggered. Mode 2 is the action ratio analysis mode detection: the proportion of the number of analysis-type events within the sliding window range to the total number of events in the window is counted. If the proportion is less than the preset analysis ratio threshold, the action ratio analysis alarm is triggered; otherwise, it is not triggered. Mode 3 is the detection of diminishing marginal returns: the total number of new knowledge entities and new research artifacts corresponding to each event in the sliding window is counted in turn to form a quantity sequence arranged by time; if the quantity sequence is strictly monotonically decreasing and the total number corresponding to the last event in the sequence is less than the preset diminishing marginal return threshold, then the diminishing marginal return alarm is triggered; otherwise, it is not triggered. Mode 4 is the duplicate query mode detection: count the number of times each query keyword appears within the sliding window range; the four non-productive mode detections are executed independently of each other, and each outputs a binary alarm signal; among them, the retrieval chain mode is used to detect pure retrieval behavior without continuous analysis, the analysis action ratio mode is used to detect behavior with insufficient research depth and too low knowledge processing ratio, the diminishing marginal returns mode is used to detect behavior with continuous degradation of information discovery ability, and the duplicate query mode is used to detect behavior with solidified and locked retrieval strategies.

[0012] Furthermore, in the adaptive gating process, if the rate index is higher than the gating threshold, the corresponding alarm is suppressed and forcibly determined to be in an untriggered state; if the rate index is lower than or equal to the gating threshold, the original triggering state of the corresponding alarm is retained; alarms with diminishing marginal returns are not subject to gating suppression and always retain their original triggering state; a logical OR operation is performed on all alarm signals after gating to obtain a comprehensive judgment signal and output it. When the information discovery rate index is higher than the gating threshold, the retrieval chain alarm, analysis action ratio alarm, and duplicate query alarm are all forcibly set to an untriggered state; when the information discovery rate index is lower than or equal to the gating threshold, the retrieval chain alarm, analysis action ratio alarm, and duplicate query alarm all maintain their original triggered state; the diminishing marginal returns alarm always maintains its original triggered state and is not affected by the gating threshold; the four types of gating alarm signals are logically ORed, and if any alarm is in a triggered state, the overall signal is determined to be in a triggered state.

[0013] Furthermore, the total number of differing files in the adjacent event work area path set includes the sum of the number of newly added, deleted, or modified files; if the total number of differing files is greater than the preset work area change threshold, it is determined that there is a real change in the work area, and a real change indication signal is output as a valid state; otherwise, it is determined as a false progress state, and a real change indication signal is output as an invalid state. The result is obtained by comparing the complete set of working area paths of the current event with that of the previous event, and summing the number of paths that do not contain each other in the two sets; when the total number of difference files is less than or equal to the working area change threshold, it is judged as pseudo progress, that is, frequent retrieval actions but no substantial increase in research results.

[0014] Furthermore, regarding the determination of the comprehensive judgment signal, if the comprehensive judgment signal is in a triggered state and the actual change indication signal of the working area is in an invalid state, or the diminishing marginal returns alarm is in a triggered state, then it is determined to be a non-productive cycle state; otherwise, it is determined to be a valid operating state. The priority of the recovery actions, from high to low, includes the action corresponding to the diminishing marginal returns alarm, the action corresponding to the repeated query alarm, the action corresponding to the retrieval chain alarm, and the action corresponding to the analysis action ratio alarm. When multiple modes are triggered simultaneously, the recovery action with the highest priority is executed.

[0015] Furthermore, the recovery actions corresponding to multiple alarm modes include the action corresponding to the diminishing marginal returns alarm: terminating the current research branch and outputting the research report for the current progress; The corresponding action for duplicate query alerts is to randomly adjust the retrieval strategy and skip recent historical retrieval contexts. The corresponding action for the retrieval chain alarm is: force a switch to analysis mode, and retrieval can only be restored after at least one knowledge processing action is completed; The actions analyzed compared to the alarms were: expanding the search space, changing the query keywords, or switching the data source.

[0016] Furthermore, the estimated progress value of the research report is between 0 and 1, and serves as the basis for the completeness of the output report content when the recovery action of triggering the termination branch and outputting the report is executed.

[0017] Furthermore, the present invention provides a non-productive loop judgment system based on information discovery rate adaptive gating, comprising: a research event stream parsing module, which receives the original tool call event stream, performs structured parsing and temporal denoising, and outputs a standardized research event sequence; The entity set maintenance module maintains the cumulative knowledge entity set and the sliding window near window entity set, and outputs the two sets and the total number of entities they contain; The rate calculation module counts the number of new knowledge entities within the near window and calculates the information discovery rate index. The multi-mode evaluation module performs four non-productive mode detections in parallel and outputs alarm signals corresponding to each mode. The adaptive gating module performs gating suppression processing on alarm signals based on the information detection rate and outputs a comprehensive judgment signal; The workspace verification module compares the differences in workspace paths between adjacent events and outputs a signal indicating the actual change. The loop determination and recovery module comprehensively determines the running status and matches the corresponding recovery action; each module is connected sequentially according to the data flow direction, and the output of the previous module serves as the input of the subsequent module; The operation log recording module records the information discovery rate, comprehensive judgment signal, actual change signal of the working area, running status, and recovery action five-tuple data for each time step, which is used for analysis and threshold parameter tuning.

[0018] Furthermore, the present invention provides a computer storage medium storing a computer program that can run on a processor, wherein the processor executes the computer program to implement any one of the above-described embodiments.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention is based on information theory and quantification, and its decision-making logic is highly interpretable. This invention transforms the question of "whether to terminate a research branch" from a text similarity comparison or subjective judgment problem into an information gain quantification decision problem. It uses information discovery rate as the core quantification indicator, which has clear physical meaning and is easy for researchers to understand, debug, and optimize parameters. By dynamically suppressing surface loop alarms through information discovery rate, it allows the research assistant to continue running while it is still in the stage of rapidly discovering new literature and concepts, avoiding the problem of prematurely terminating effective exploration in traditional fixed threshold methods, and balancing the recall and precision of loop detection.

[0020] 2. This invention employs a four-mode parallel evaluation to cover all degradation scenarios. The four modes—retrieval chain, analysis action ratio, diminishing marginal returns, and repeated queries—correspond to different types of unproductive behavior by research assistants. The parallel detection mechanism ensures comprehensive coverage of all ineffective loop patterns. Through workspace path set difference analysis, it can accurately identify false activity phenomena such as "frequent updates to retrieval logs but no substantial increase in report files," further improving the accuracy of loop detection and avoiding being misled by superficial actions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the method of the present invention. Figure 2 This is a system structure diagram of the present invention; Figure 3 This is a structural diagram of the storage medium of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to describe selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example Please see Figure 1 This embodiment uses the task of "reviewing the research progress on non-radiative recombination loss of perovskite solar cells in the past five years" as a practical example of the application scenario of an automated research assistant. It performs real-time monitoring and cyclic determination of the research assistant's operating status and details the specific implementation process of the invention:

[0024] In this embodiment, the study of event stream reception and structured parsing is conducted first: The research assistant generates a continuous stream of tool call events during its operation. At time step 15, the research assistant performs a Google Scholar search, returning a list of 8 articles. The system performs structured parsing of this event, extracting seven core fields to construct a standardized research event: in: For the purpose of "retrieval," academic database queries fall under the category of information retrieval. Five new knowledge entities have been added: {“Perovskite”, “Non-radiative recombination”, “Defect passivation”, “Shockley-Read-Hall”, and “Carrier lifetime”}. No new report paragraphs or research artifacts were generated for this event; A total of 8 new document URLs were found. {“Perovskite defect passivation method”} is the search keyword used in this embodiment; The current research report is estimated to be 12% complete. { / workspace / report.md, / workspace / refs / } represents the set of file paths in the current research workspace.

[0025] Specifically, in this embodiment, the above fields are encapsulated as standardized research events. Added to the standardized research event sequence by timestamp middle.

[0026] In this embodiment, the knowledge entity set maintenance and information discovery rate calculation are performed. Set the sliding window size W=10, and the total number of events T=20. The cumulative knowledge entity set is the union of all new knowledge entities from all events and new research artifacts, calculated as follows: through , It includes 52 document-related entities and 10 report artifacts.

[0027] The near-window entity set is the union of new knowledge entities and new research artifacts from events 11 to 20, expressed as: According to statistics It contains 35 entities.

[0028] The number of entirely new knowledge entities within the near window is calculated as the cardinality of the difference between the near window entity set and the cumulative knowledge entity set. Calculated , Eight new knowledge entities that had not appeared before were discovered within the immediate window. For example, the "interface engineering" concept discovered in the 18th event did not appear in the previous cumulative entity set.

[0029] Specifically, calculate the information discovery rate metric: This value is below the preset threshold. This indicates that the rate of knowledge discovery within the recent window has slowed significantly, posing a risk of unproductive cycles.

[0030] In this embodiment, four modes are evaluated in parallel, and four modes of detection are performed in parallel for events 11 to 20: Search chain detection: Calculate the length of the longest consecutive search chain. Upon inspection, it was found that the category for six consecutive events, from event 15 to event 20, was "search." The research assistant repeatedly switched between Google Scholar, PubMed, and arXiv but no analytical paragraphs were generated. Therefore... Preset continuous search threshold , This triggers a search chain alarm. .

[0031] Specifically, the analysis compares the detection of actions and calculates the percentage of actions analyzed within the sliding window:

[0032] Of the last 10 events, only events 12 and 14 were analytical, involving a total of two knowledge processing actions. Preset analysis percentage threshold , No alarm was triggered. .

[0033] Specifically, the diminishing marginal returns detection calculates the total number of new knowledge entities and new research artifacts for each event in the last 10 steps. The resulting quantity sequence is [9,7,6,5,4,3,2,1,1,0]. The determination rule is: The sequence is strictly monotonically decreasing, and the last event... Less than the preset marginal decrease threshold This triggers a diminishing marginal returns warning. .

[0034] In this embodiment, duplicate query detection counts the occurrence frequency of each query keyword in the most recent 10 events. It was found that the keyword "perovskite nonradiative complex" appeared 4 times, exceeding the preset duplicate query threshold. This triggers a duplicate query alert. .

[0035] The final output is a set of four alarm signals: .

[0036] In this embodiment, adaptive gating and determination of actual changes are performed. Step 5 Adaptive Gating Calculation: Current Information Discovery Rate Gating does not suppress three types of alarms: retrieval chain, analysis action ratio, and duplicate query; diminishing marginal returns alarms are never suppressed by gating. Therefore, the outputs of each gating mechanism are: , , ,

[0037] The overall decision signal is obtained by performing a logical OR operation on the outputs of each gate: The risk of a non-productive cycle has been identified.

[0038] In this embodiment, the determination of actual changes in the workspace involves comparing the workspace path sets of the 20th event with those of the 19th event. and Calculate the symmetric difference cardinality:

[0039] Calculated This means the report file size has not increased and no new references have been added to the bibliography. (Preset workspace change threshold) , Therefore, the actual change indication signal in the work area It was determined to be a false progress status.

[0040] In this embodiment, the final determination and recovery action Final state determination: The determination rule is as follows: current and ,at the same time , The criteria for determining a non-productive cycle are met, therefore = "Non-productive cycle".

[0041] Recovery action selection: Multiple modes are triggered simultaneously, sorted by priority. The diminishing marginal returns alarm has the highest priority, so it is matched with the corresponding recovery action: Action="Terminate the current branch and output the current research report".

[0042] Upon receiving the instruction, the research assistant terminates the current review research branch and compiles the effective content accumulated from the 12% progress, including the abstracts of the 8 extracted documents and key findings, into a phased research report, thus avoiding further wasting API call quotas on meaningless duplicate searches.

[0043] Example 2 like Figure 2 This invention provides a non-productive loop judgment system based on information discovery rate adaptive gating, comprising: a research event stream parsing module, which receives the original tool call event stream, performs structured parsing and time-series denoising, and outputs a standardized research event sequence; The entity set maintenance module maintains the cumulative knowledge entity set and the sliding window near window entity set, and outputs the two sets and the total number of entities they contain; The rate calculation module counts the number of new knowledge entities within the near window and calculates the information discovery rate index. The multi-mode evaluation module performs four non-productive mode detections in parallel and outputs alarm signals corresponding to each mode. The adaptive gating module performs gating suppression processing on alarm signals based on the information detection rate and outputs a comprehensive judgment signal; The workspace verification module compares the differences in workspace paths between adjacent events and outputs a signal indicating the actual change. The loop determination and recovery module comprehensively determines the running status and matches the corresponding recovery action; each module is connected sequentially according to the data flow direction, and the output of the previous module serves as the input of the subsequent module; The operation log recording module records the information discovery rate, comprehensive judgment signal, actual change signal of the working area, running status, and recovery action five-tuple data for each time step, which is used for analysis and threshold parameter tuning.

[0044] Example 3 like Figure 3 In this embodiment, the present invention provides a computer storage medium storing a computer program that can run on a processor, wherein the processor executes the computer program to implement any one of the above-described embodiments.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations will be apparent to those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A non-productive cycle determination method based on adaptive gating of information discovery rate, characterized in that: include: First, receive the event stream invoked during the operation of the automated research assistant, and perform structured parsing on each event; All parsed events are sorted by timestamp and outliers are removed to obtain a standardized sequence of research events arranged in ascending order of time. Based on the output standardized research event sequence, output the cumulative knowledge entity set and the total number of entities it contains, and the near-window entity set and the total number of entities it contains; Based on the cumulative knowledge entity set and the near window entity set, count the number of entities in the near window entity set that are not included in the cumulative knowledge entity set; divide the number of new knowledge entities in the near window by the total number of cumulative knowledge entities to obtain the information discovery rate index and output it. Based on the standardized research event sequence, non-productive mode detection is performed in parallel, and independent alarm signals are output respectively; If the number of occurrences of any query keyword exceeds the preset duplicate query threshold, a duplicate query alarm will be triggered; otherwise, it will not be triggered. Based on the output information discovery rate index and alarm signals, a preset information discovery rate gating threshold is set; adaptive gating processing is performed on retrieval chain alarms, analysis action ratio alarms, and duplicate query alarms to obtain a comprehensive judgment signal and output it; Based on the workspace path set and comprehensive judgment signal corresponding to two adjacent events, calculate the total number of files with differences in the workspace path set of two adjacent events, and output the real change indication signal as invalid. Based on the comprehensive judgment signal, the actual change indication signal of the working area, and the diminishing marginal returns alarm signal, the final operating status is determined. According to the final operating status and the specific alarm mode triggered, the corresponding recovery action is matched and output according to the preset priority rules.

2. The non-productive cycle determination method based on information discovery rate adaptive gating according to claim 1, characterized in that, The structured parsing of each event involves extracting seven core fields: event category tags, new knowledge entity sets, new research artifact sets, new document URL sets, new query keyword sets, estimated research report completion progress, and complete research workspace path sets, to construct a single standardized research event. The event category tags include two types: retrieval and analysis. Retrieval corresponds to information acquisition actions, including search engine queries, academic database searches, and full-text document downloads. Analysis includes knowledge processing actions, such as document abstract extraction, viewpoint comparison and synthesis, and report paragraph generation. The output of the cumulative knowledge entity set and the total number of entities contained therein, and the near window entity set and the total number of entities contained therein, includes taking the union of the new knowledge entity set corresponding to all events and the new research workpiece set, deduplicating them, constructing and updating the cumulative knowledge entity set; and setting a preset sliding window size, taking the union of the new knowledge entity set corresponding to all events within the window coverage area and the new research workpiece set, deduplicating them, constructing and updating the near window entity set.

3. The non-productive cycle determination method based on information discovery rate adaptive gating according to claim 1, characterized in that, in, The non-productive mode detection includes: Mode 1 is retrieval chain mode detection, which counts the length of the longest continuous retrieval event chain in the event sequence. If the length is greater than or equal to a preset continuous retrieval threshold, a retrieval chain alarm is triggered; otherwise, it is not triggered. Mode 2 is the action ratio analysis mode detection: the proportion of the number of analysis-type events within the sliding window range to the total number of events in the window is counted. If the proportion is less than the preset analysis ratio threshold, the action ratio analysis alarm is triggered; otherwise, it is not triggered. Mode 3 is the detection of diminishing marginal returns: the total number of new knowledge entities and new research artifacts corresponding to each event in the sliding window is counted in turn to form a quantity sequence arranged by time; if the quantity sequence is strictly monotonically decreasing and the total number corresponding to the last event in the sequence is less than the preset diminishing marginal return threshold, then the diminishing marginal return alarm is triggered; otherwise, it is not triggered. Mode 4 is the duplicate query mode detection: count the number of times each query keyword appears within the sliding window range; the four non-productive mode detections are executed independently of each other, and each outputs a binary alarm signal.

4. The non-productive cycle determination method based on information discovery rate adaptive gating according to claim 1, characterized in that, The adaptive gating process, if the rate index is higher than the gating threshold, suppresses the corresponding alarm and forces it to be in an untriggered state; if the rate index is lower than or equal to the gating threshold, retains the original triggering state of the corresponding alarm; alarms with diminishing marginal returns are not suppressed by gating and always retain their original triggering state; perform a logical OR operation on all gating alarm signals to obtain a comprehensive judgment signal and output it. When the information discovery rate index is higher than the gating threshold, the retrieval chain alarm, analysis action ratio alarm, and duplicate query alarm are all forcibly set to an untriggered state; when the information discovery rate index is lower than or equal to the gating threshold, the retrieval chain alarm, analysis action ratio alarm, and duplicate query alarm all maintain their original triggered state; the diminishing marginal returns alarm always maintains its original triggered state and is not affected by the gating threshold; the four types of gating alarm signals are logically ORed, and if any alarm is in a triggered state, the overall signal is determined to be in a triggered state.

5. The non-productive cycle determination method based on information discovery rate adaptive gating according to claim 1, characterized in that, The total number of differing files in the adjacent event work area path set includes the sum of the number of newly added, deleted, or modified files; if the total number of differing files is greater than the preset work area change threshold, it is determined that there is a real change in the work area, and a real change indication signal is output as a valid state; otherwise, it is determined as a false progress state, and a real change indication signal is output as an invalid state. The result is obtained by comparing the complete set of working area paths of the current event with that of the previous event, and summing the number of paths that do not contain each other in the two sets; when the total number of difference files is less than or equal to the working area change threshold, it is judged as pseudo progress, that is, frequent retrieval actions but no substantial increase in research results.

6. The non-productive cycle determination method based on information discovery rate adaptive gating according to claim 1, characterized in that, The determination of the comprehensive judgment signal is as follows: if the comprehensive judgment signal is in a triggered state and the actual change indication signal of the working area is in an invalid state, or the diminishing marginal returns alarm is in a triggered state, then it is determined to be a non-productive cycle state; otherwise, it is determined to be a valid operating state. The priority of the recovery actions, from high to low, includes the action corresponding to the diminishing marginal returns alarm, the action corresponding to the repeated query alarm, the action corresponding to the retrieval chain alarm, and the action corresponding to the analysis action ratio alarm. When multiple modes are triggered simultaneously, the recovery action with the highest priority is executed.

7. The non-productive cycle determination method based on information discovery rate adaptive gating according to claim 6, characterized in that, The recovery actions corresponding to multiple alarm modes include the action corresponding to the diminishing marginal returns alarm: terminating the current research branch and outputting the research report for the current progress; The corresponding action for duplicate query alerts is to randomly adjust the retrieval strategy and skip recent historical retrieval contexts. The corresponding action for the retrieval chain alarm is: force a switch to analysis mode, and retrieval can only be restored after at least one knowledge processing action is completed; The actions analyzed compared to the alarms were: expanding the search space, changing the query keywords, or switching the data source.

8. The non-productive cycle determination method based on information discovery rate adaptive gating according to claim 1, characterized in that, The estimated progress of the research report ranges from 0 to 1, and serves as the basis for the completeness of the output report content when the recovery action of triggering the termination branch and outputting the report is executed.

9. A non-productive cyclic decision system based on adaptive gating of information discovery rate, implementing the method of any one of claims 1 to 8, characterized in that, include: The research event stream parsing module receives the original tool call event stream, performs structured parsing and temporal denoising, and outputs a standardized research event sequence; The entity set maintenance module maintains the cumulative knowledge entity set and the sliding window near window entity set, and outputs the two sets and the total number of entities they contain; The rate calculation module counts the number of new knowledge entities within the near window and calculates the information discovery rate index. The multi-mode evaluation module performs four non-productive mode detections in parallel and outputs alarm signals corresponding to each mode. The adaptive gating module performs gating suppression processing on alarm signals based on the information detection rate and outputs a comprehensive judgment signal; The workspace verification module compares the differences in workspace paths between adjacent events and outputs a signal indicating the actual change. The loop determination and recovery module comprehensively determines the running status and matches the corresponding recovery action; each module is connected sequentially according to the data flow direction, and the output of the previous module serves as the input of the subsequent module; The operation log recording module records the information discovery rate, comprehensive judgment signal, actual change signal of the working area, running status, and recovery action five-tuple data for each time step, which is used for analysis and threshold parameter tuning.

10. A computer storage medium storing a computer program executable on a processor, characterized in that, When the processor executes the computer program, it implements the unproductive cyclic decision method based on information discovery rate adaptive gating as described in any one of claims 1-8.

Images