Paper deep dive
Graph of States: Solving Abductive Tasks with Large Language Models
Yu Luo, Rongchen Gao, Lu Teng, Xidao Wen, Jiamin Jiang, Qingliang Zhang, Yongqian Sun, Shenglin Zhang, Jiasong Feng, Tong Liu, Wenjie Zhang, Dan Pei
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/26/2026, 2:25:34 AM
Summary
The paper introduces 'Graph of States' (GoS), a neuro-symbolic framework designed to improve abductive reasoning in Large Language Models. GoS addresses common deficiencies in existing deductive frameworks—such as Evidence Fabrication, Context Drift, Failed Backtracking, and Early Stopping—by utilizing a dual-layer architecture: a cognitive layer for multi-agent collaboration and a symbolic layer that employs a causal graph and state machine to maintain structured belief states and govern reasoning trajectories.
Entities (5)
Relation Signals (3)
Graph of States → addresses → Abductive Reasoning
confidence 95% · we introduce Graph of States (GoS), a general-purpose neuro-symbolic framework tailored for abductive tasks.
Graph of States → utilizes → Causal Graph
confidence 95% · GoS grounds multi-agent collaboration in a structured belief states, utilizing a causal graph to explicitly encode logical dependencies
Graph of States → utilizes → State Machine
confidence 95% · and a state machine to govern the valid transitions of the reasoning process.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Logical reasoning encompasses deduction, induction, and abduction. However, while Large Language Models (LLMs) have effectively mastered the former two, abductive reasoning remains significantly underexplored. Existing frameworks, predominantly designed for static deductive tasks, fail to generalize to abductive reasoning due to unstructured state representation and lack of explicit state control. Consequently, they are inevitably prone to Evidence Fabrication, Context Drift, Failed Backtracking, and Early Stopping. To bridge this gap, we introduce Graph of States (GoS), a general-purpose neuro-symbolic framework tailored for abductive tasks. GoS grounds multi-agent collaboration in a structured belief states, utilizing a causal graph to explicitly encode logical dependencies and a state machine to govern the valid transitions of the reasoning process. By dynamically aligning the reasoning focus with these symbolic constraints, our approach transforms aimless, unconstrained exploration into a convergent, directed search. Extensive evaluations on two real-world datasets demonstrate that GoS significantly outperforms all baselines, providing a robust solution for complex abductive tasks. Code repo and all prompts: this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2603.21250v1
- Canonical: https://arxiv.org/abs/2603.21250v1
Trouble viewing inline? Open PDF directly →
Full Text
88,832 characters extracted from source content.
Expand or collapse full text
Graph of States: Solving Abductive Tasks with Large Language Models Yu Luo 1,4 , Rongchen Gao 1 , Lu Teng 2 , Xidao Wen 3 , Jiamin Jiang 1 , Qingliang Zhang 1 , Yongqian Sun 1 * ,Shenglin Zhang 1 ,Jiasong Feng 4 ,Tong Liu 4 ,Wenjie Zhang 4 ,Dan Pei 5 1 Nankai University, 2 Wenzhou Medical University, 3 Bizseer, 4 Lenovo, 5 Tsinghua University Work done at Lenovo,Correspondence:sunyongqian@nankai.edu.cn Abstract Logical reasoning encompasses deduction, in- duction, and abduction. However, while Large Language Models (LLMs) have effectively mastered the former two, abductive reason- ing remains significantly underexplored. Ex- isting frameworks, predominantly designed for static deductive tasks, fail to general- ize to abductive reasoning due to unstruc- tured state representation and lack of explicit state control. Consequently, they are in- evitably prone toEvidence Fabrication,Con- text Drift,Failed Backtracking, andEarly Stopping. To bridge this gap, we intro- duce Graph of States (GoS), a general-purpose neuro-symbolic framework tailored for abduc- tive tasks.GoSgrounds multi-agent collab- oration in a structured belief states, utilizing a causal graph to explicitly encode logical dependencies and a state machine to govern the valid transitions of the reasoning process. By dynamically aligning the reasoning focus with these symbolic constraints, our approach transforms aimless, unconstrained exploration into a convergent, directed search. Exten- sive evaluations on two real-world datasets demonstrate thatGoSsignificantly outper- forms all baselines, providing a robust solu- tion for complex abductive tasks. Code repo and all prompts: https://anonymous.4open. science/r/Graph-of-States-5B4E. 1 Introduction Logical reasoning constitutes the cognitive corner- stone of artificial intelligence, fundamentally cat- egorized into three paradigms:deduction,induc- tion, andabduction(Peirce,1934). While de- duction derives definitive conclusions from gen- eral premises and induction generalizes rules from specific observations, abductive reasoning infers the most probable hypotheses from incomplete observations. In the era of LLMs, induction is * Corresponding author an inherent capability established through exten- sive pre-training ( Olsson et al.,2022). Simulta- neously, deductive tasks (e.g., game of 24, math- ematics) are effectively addressed by reasoning frameworks like Chain-of-Thought (CoT) (Wei et al.,2022) and Tree-of-Thought (ToT) (Yao et al., 2023). In contrast, the domain of abductive rea- soning remains underexplored. Given that abduc- tion serves as the bedrock for decision-making in high-stakes, real-world scenarios (e.g., medical di- agnosis, criminal investigation, failure diagnosis in distributed systems), this neglect represents a critical gap. Therefore, we aim to establish a general-purpose reasoning framework tailored for the dynamic and non-monotonic nature of abduc- tive tasks. As abductive tasks are characterized by in- complete initial information, requiring dynamic evidence investigation to progressively converge the hypothesis space and infer the most plau- sible cause, directly transposing existing deduc- tive frameworks to these scenarios proves inef- fective. To empirically validate this gap, we ap- plied four deductive frameworks (i.e., CoT, ToT, GoT ( Besta et al.,2024), FoT (Bi et al.,2025)) to abductive scenarios and conducted a granular analysis of their reasoning trajectories (detailed in Appendix D). As visualized inFigure1, we identified four deficiencies:(1) Evidence Fabri- cation:In an attempt to maintain logical consis- tency, the model tends to hallucinate non-existent evidence to support biased hypotheses, corrupt- ing the ground truth. (2) Context Drift ( Dongre et al.,2025): The model tends to forget the current investigation progress during long-horizon tasks (Liu et al.,2024), trapping the agent in redundant loops where it repetitively invokes the same tools or revisits hypotheses that were already falsified by prior evidence.(3) Failed Backtracking:Un- like deductive tasks where objective validity cri- teria (e.g., 7 and 9 cannot make it to 24 through 1 arXiv:2603.21250v1 [cs.AI] 22 Mar 2026 Deductive Tasks (game of 24, mathmatics) Input: 4, 9, 10, 13; Reach 24 using “+, -, ×,÷” only. ...... 10-4=6 (left: 6,9,13) 13-9=4 (left: 4, 6) 13-6=7 (left: 7, 9) × Evaluator: 7 and 9 cannot make it to 24 through“+, -, ×,÷” Backtracking 4×6=24 (left: 24) Output: (13 - 9)×(10 - 4) = 24 Abductive Tasks (medical diagnosis, failure diagnosis in distributed systems, criminal investigation) Deficiency 1 & 3: Evidence Fabrication + Failed Backtracking Deficiency 2: Context Drift Input: The patient has chest pain. Step 1: Is it Myocardial Infarction? Step 2: Check Troponin →normal .... (long sequence of actions) Step n: Is it Myocardial Infarction? Unstructured context history Input: Ankle injury, severe swelling, unable to walk. Ankle Fracture Perform X-ray Imaging → normal Occult micro-fracture in ankle Fabricated: Bone Crepitus Expect: Backtracking ? Deficiency 4: Early Stopping Soft Tissue Injury Input:“Alert: Host01 memory usage > 95%” Run: ``free -h`` → “Total: 32G, Used: 31.5G, Free: 0.1G (<1%).” System Out of Memory (OOM) Run: ``top -o %MEM`` → PID 2048 (Java) consumes 85% MEM Run: ``jmap -histo:live 2048`` → GlobalCache objects occupy 90% of Heap Memory Leak in GlobalCache (True Diagnosis) Early Stopping leads to premature diagnosis, which is correct but not useful Expect: Drill down via further actions to pinpoint the specific root cause (Memory Leak in GlobalCache). Figure 1: Illustration of reasoning frameworks applied to Deductive tasks (Left) versus Abductive tasks (Right). While deductive frameworks succeed in static logic (e.g., Game of 24), applying them to abductive tasks exposes four deficiencies: (1)Evidence Fabrication, (2)Context Drift, (3)Failed Backtracking, and (4)Early Stopping. +,−,×,÷) trigger immediate pruning, abductive scenarios present ambiguous intermediate results. Lacking explicit hard constraints, models often fail to trigger necessary backtracking, persisting on erroneous paths.(4) Early Stopping:Mod- els frequently terminate the investigation prema- turely upon identifying a superficial symptom, fail- ing to drill down to the fine-grained root cause required for actionable decision-making. We at- tribute these failures to two fundamental structural limitations in current frameworks. First, the im- plicit encoding of the reasoning state (i.e., subjec- tive hypotheses and objective facts) within the un- structured context history fails to provide a clear structural representation, directly precipitatingEv- idence FabricationandContext Drift. Second, the lack of state control mechanism relegates back- tracking and drill-down decisions to the model’s unconstrained autonomy, resulting inFailed Back- trackingandEarly Stopping. While a unified reasoning framework tailored for abductive tasks remains absent, researchers have leveraged LLMs in specialized domains such as medical diagnosis ( Kim et al.,2024;Zhou et al.,2025) and failure diagnosis in distributed systems (Zhou et al.,2024;Zhang et al.,2024; Sun et al.,2025;Pei et al.,2025;Luo et al.,2025). To mitigate the aforementioned deficiencies, these approaches incorporate extensive domain-specific adaptations, such as integrating external knowl- edge bases to curbEvidence Fabricationor im- posing rule-based heuristics to preventEarly Stop- ping. However, these systems fundamentally rely on conventional deductive reasoning frame- works (e.g., CoT, ToT), attributing their perfor- mance gains primarily to heavy domain engineer- ing rather than intrinsic reasoning capabilities. Consequently, such adaptations serve as engineer- ing workarounds that alleviate immediate symp- toms while leaving the intrinsic structural deficien- cies of the deductive paradigm unaddressed. To bridge this gap, we introduce Graph of States (GoS), a general-purpose reasoning frame- work tailored for abductive tasks. Our approach employs a dual-layer neuro-symbolic architecture to unify human-aligned collaboration with rigor- ous logic. In the cognitive layer, we implement a role-based collaborative framework, orchestrat- ing agents aligned with real-world professional roles to ensure a coherent division of labor. Cru- cially, in the symbolic layer, we introduce two core mechanisms to resolve the aforementioned structural limitations: First, we construct a causal graph as the system’s memory, which explicitly structures the belief state by mapping the causal relationships among hypotheses and collected ev- idence, effectively mitigatingEvidence Fabrica- tionandContext Drift. Second, we employ a state machine as the navigation to govern the reasoning trajectory, enforcing rigorous logical transitions for backtracking and drill-down to preventFailed BacktrackingandEarly Stopping. Quantitative analysis in Appendix Dconfirms that these mech- anisms significantly reduce the frequency of such deficiencies. By dynamically aligning the reason- ing focus with these symbolic constraints,GoS transforms aimless, unconstrained exploration into a convergent, directed search, ensuring the system steadily approximates the truth. Our contributions are summarized below:(1) We proposeGoS, and to the best of our knowledge, GoSis the first general-purpose multi-agent rea- soning framework tailored for abductive reasoning tasks.(2)We introduce a neuro-symbolic architec- 2 ture that leverages causal graph and state machine to construct explicit belief states, thereby trans- forming unconstrained exploration into directed, convergent search.(3)We conduct extensive eval- uations on two real-world datasets, demonstrating thatGoSsignificantly outperforms existing base- lines.(4)We make all code and prompts publicly available. 2 Related Work Reasoning frameworks for deductive tasks.A significant body of research has emerged to unlock the potential of LLMs as general problem solvers, though these efforts have predominantly focused on deductive tasks such as the game of 24, math- ematical problems, and crossword puzzles. The foundational CoT ( Wei et al.,2022) paradigm de- composes complex problems into sequential inter- mediate steps, where each step forms a coherent natural language sequence contributing to the fi- nal solution. Building upon this, subsequent stud- ies have introduced refinements to enhance relia- bility, including Self-Consistency with CoT (CoT- SC) ( Wang et al.,2023), VerifyCoT (Zhao et al., 2023), and Chain of Continuous Thought (Co- conut) (Hao et al.,2024). To transcend the limi- tations of linear reasoning, ToT (Yao et al.,2023) structures the reasoning process as a search over a tree, typically employing algorithms like Depth- First Search (DFS) or Breadth-First Search (BFS) to explore diverse reasoning paths. This non- linear topology is further generalized by frame- works such as Graph of Thought (GoT) ( Besta et al.,2024) and Forest of Thought (FoT) (Bi et al., 2025), which extend the structure into graphs and forests to model more complex, non-sequential de- pendencies. Large Language Models for specialized ab- ductive tasks. Current approaches for specialized abductive tasks mainly employ four categories of domain-specific adaptations designed to mitigate intrinsic reasoning deficiencies and thereby en- hance performance: (1) Multi-Agent Customiza- tion: Frameworks often design specific topolog- ical structures to mimic human workflows. For instance, MAM ( Zhou et al.,2025) employs a static collaborative topology mimicking a fixed team of specialists to reach consensus, parallel- ing D-Bot ( Zhou et al.,2024), which decomposes the diagnosis process into atomic functional units (e.g., CPU, Disk) to perform collaborative cross- reviews. (2)Retrieval-Augmented Generation (RAG): To ground reasoning in external knowl- edge, MDAgents (Kim et al.,2024) augments its diagnostic process with specialized retrieval tools like MedRAG (Xiong et al.,2024). FlowXpert (Shi et al.,2025) employs a hybrid retrieval mech- anism querying both vector and graph databases, while Flow-of-Action (Pei et al.,2025) targets the retrieval and matching of high-quality Stan- dard Operating Procedures (SOPs). (3)Supervised Fine-Tuning (SFT): Data-centric approaches prior- itize internalizing domain knowledge. Med-PaLM 2 ( Singhal et al.,2025) and PMC-LLaMA (Wu et al.,2024) prove that fine-tuning on vast biomed- ical corpora significantly enhances diagnostic ca- pability, a strategy similarly adopted by LogLM (Liu et al.,2025) using massive log instruction datasets in AIOps. (4)Data Preprocessing: Con- versely, approaches like TrioXpert (Sun et al., 2025) and OpsAgent (Luo et al.,2025) focus on in- put abstraction, implementing dedicated pipelines to transform heterogeneous telemetry (i.e., met- rics, logs, traces) into structured contexts before feeding them into reasoning modules. To summa- rize, the efficacy of these approaches is predom- inantly derived from suchdomain-specific engi- neeringrather than generalized reasoning capabil- ities. 3 Methodology We proposeGoS, a novel dual-layer neuro- symbolic framework designed to address complex abductive reasoning tasks by explicitly maintain- ing reasoning states through causal graphs and state machines (shown inFigure 2). Additionally, we introduce areasoning focusto concentrate in- vestigative resources on the most plausible hypoth- esis, thereby preventing wasteful exploration. This section is organized as follows: Section 3.1pro- vides the formal definition of the dual-layer frame- work. Section3.2elucidates the bi-directional in- teraction mechanism between the cognitive and symbolic layers. Finally, Section 3.3details the state conversions, governing the hierarchical back- tracking and drill-down of the inference process. 3.1 Dual-layer Framework Task Formulation. We define the abductive rea- soning task as an probabilistic inference prob- lem aimed at identifying the most plausible root cause from incomplete information. LetO= 3 a Dual-layer Framework Ugh! My chest is hurting badly. What’s wrong with me? · Central Agent Expert Agents ReAct 1. Tool descriptions 2. Tool APIs Cognitive LayerSymbolic Layer Prediction: You have a myocardial infarction. Report: Because your ECGs exhibits ST-segment elevation in anterior leads, inverted T-waves, and ...... � � = 3 Belief States � � V sym V hyp 1 V hyp 2 V hyp 3 V hyp 4 V hyp 5 V evi 2 : focus � � V evi 1 : support: refute Step 1: User input query & Initialization Step 2: Symbolic layer instruct cognitive layer to investigate Step 3: Cognitive layer ground evidence and analyses to symbolic layer Step 4: Invoke state conversions fine-grained hypothesis? Step 5: Output prediction and report Check Yes No Figure 2:Overview ofGoS. Left:Schematic of the dual-layer architecture.Right:The iterative inference workflow. o 1 , o 2 , ..., o n denote the set of observations, which consists of easily accessible surface symp- tomsO surf (e.g., chest pain) and costly-to-acquire deep evidenceO deep (e.g., CT scans, ECGs). Given a domain-specific knowledge contextK (e.g., medical guidelines), the objective is to find the optimalH ∗ from the hypothesis spaceH (e.g., myocardial infarction) that maximizes the posterior probability: H ∗ = arg max H∈H P(H|O surf ,O deep ,K)(1) In contrast to passive inference tasks,O deep is not given as a comprehensive input prior. Instead, it necessitates on-the-fly acquisition, where we em- ploy ReAct ( Yao et al.,2022) to execute tool calls for dynamic retrieval. Dual-Layer System Definition. Following are the concrete definition of our system: 1.The Cognitive Layer (L cog ):This layer functions as the domain-adaptive execution interface ofGoS, orchestrating a multi-agent system aligned with real-world professional roles to ground the universal reasoning logic into concrete, domain-specific actions. We define the agent system asA=A central ∪ A experts , where one central agentA central acts as orchestrator and several expert agents A experts act as executors. To avoid the reasoning fragmentation caused by atomized functional units, we explicitly design these experts to map coherent real-world profes- sional roles. The rationale is drawn from the division of labor in effective human or- ganizations. Specifically, (1) complex ab- ductive reasoning inherently requires a cen- tral orchestrator for global planning and fi- nal decision-making, supported by special- ized experts for domain-specific execution; and (2) adopting these role divisions, which have persisted through long-term evolution due to their structural robustness, not only enhances collaboration effectiveness but also guarantees a transparent process that is easily understood by humans. 2.The Symbolic Layer (L sym ):Acting as the GoS’s explicit navigational anchor,L sym for- malizes the reasoning status into a structured belief statesB= (G, S), thereby grounding cognitive processes in a transparent and trace- able manner distinct from unstructured con- text history. We formalize the components of the symbolic layer as follows: Causal Graph (G):We define a directed graphG= (V, E)to map the logical topol- ogy. The node setVcomprises three dis- tinct types: symptom (v sym ), evidence (v evi ), and hypothesis (v hyp ), where each hypothe- sis node is associated with a scoreP(v hyp ) representing its confidence. The edge set Eencodes three logical primitives:derive (v sym →v hyp ) representing initial hypothe- sis generation,refine(v coarse hyp →v f ine hyp ) rep- resenting granularity evolution of hypothesis, andsupport/refute(v evi →v hyp ) represent- ing evidential confirmation or negation. State Machine (S):To capture the hierar- chical nature of abductive reasoning, we as- sign a distinct levelL(v hyp )to each hypoth- esis node, representing its semantic granular- ity. Accordingly, we define a state machine where the stateS t ∈N + as the level of the hy- pothesis currently under investigation at step t. This state variable governs the overall infer- ence trajectory, facilitating drill-down to finer granularities or backtracking upon contradic- tion (see Section 3.3). The Reasoning Focus (h ∗ ):Formally, at stateS t , the reasoning focush ∗ t is defined as the hypothesis with the highest confidence at the current level: h ∗ t = arg max h∈V hyp , L(h)=S t P(h|G t )(2) This formulation enforces a depth-first inves- tigation strategy on the most promising tra- jectory, thereby facilitating either rapid con- 4 firmation or early backtracking to minimize overall investigative costs. Initialization. Upon receiving surface symp- tomsO surf , the central agentA central constructs the initial graphG 0 by deriving preliminaryL(1) hypotheses from the root symptom. This estab- lishes the initial belief stateB 0 and sets the state machine toS 0 = 1, preparing the system for the iterative loop. Workflow.As illustrated in Figure 2, the pro- cess begins with user input query and initialization (Step 1). Subsequently,GoSenters an iterative rea- soning loop: The symbolic layer first identifies the current reasoning focush ∗ t and instructs the cog- nitive layer to investigate (Step 2). Acting on this directive, the cognitive layer orchestrates agents to gather evidence and updates the causal graphG t in symbolic layer with new findings (Step 3). Based on the updated graph, the system invokes state con- versions (Step 4) to determine the next reasoning depth (e.g., drill down or backtracking). This cycle repeats until a fine-grained hypothesis adequately resolves the input query, producing the final pre- diction (Step 5). b Bi-Directional Neuro-Symbolic Interaction Central Agent V sym V hyp 1 V hyp 2 V evi 1 V evi 2 V evi 3 V hyp 3 V hyp 4 V hyp 5 V evi 4 V evi 5 V evi 7 V evi 6 : new focus : new edge : new evidence : new hypothesis V hyp 4 FindFocus ℎ � ∗ Plan ReAct · 1. Tool descriptions 2. Tool APIs Deep EvidenceAnalyses 1. Confidence recalibration 2. Node instantiation 3. Edge formulation UpdateGraph Figure 3:Bi-Directional Neuro-Symbolic Interac- tion. 3.2 Bi-Directional Neuro-Symbolic Interaction The core reasoning capability ofGoSemerges from the bi-directional interaction between the symbolic and cognitive layers. Rather than func- tioning as isolated modules, these two layers op- erate within a closed-loop mechanism where ex- plicit state representations and dynamic collabo- rative reasoning reciprocally inform one another. We detail this bi-directional process through two phases: Symbolic-to-Cognitive: Reasoning Focus- Guided Investigation. In this phase, the symbolic layer transforms the static belief statesB t into ac- tionable instructions, ensuring that agentic explo- ration in cognitive layer remains focused and co- herent. Unlike XoT frameworks (e.g., CoT, ToT, GoT, FoT), where reasoning states are represented as disordered thought nodes, our symbolic layer explicitly maintains a structured causal graphG t and identifies a reasoning focush ∗ t . This focus represents the most plausible direction of inquiry at the current step, serving as a navigational com- pass for the cognitive layer. The reasoning fo- cush ∗ t and the belief statesB t are injected into the central agentA central to formulate executable plans:Instructions←P lan(A central , h ∗ t ,B t ). These instructions are then dispatched to the cor- responding expert agentsA experts alongside the global context(h ∗ t ,B t ), triggering targeted tool in- vocation and analysis. Therefore, anchoring the investigation to focus h ∗ t enforces directional stability, effectively miti- gating the aimless exploration typical of unstruc- tured reasoning strategies. This focus ensures that computational resources are concentrated on refining the most promising hypothesis. Simul- taneously, the shared propagation of(h ∗ t ,B t )es- tablishes a unified cognitive consensus across the multi-agent system. This mechanism prevents expert agents from operating in silos, guarantee- ing that distributed reasoning advances coherently without contradicting the global state. Cognitive-to-Symbolic:Evidence-Based Grounding. Following the guidance, expert agentsA experts execute ReAct-based tool invo- cations and analysis. To ground these distributed investigations, the central agentA central synthe- sizes the returned observations and analytical results to update the symbolic layer. This aggre- gation drives the evolution of the causal graph fromG t toG t+1 through three key topological operations: (1)Confidence recalibration, where the plausibility of existing hypotheses is ad- justed by reinforcing confirmed conjectures or diminishing refuted ones based on new evidence; (2)Node instantiation, which registers newly discovered evidence and hypotheses into the node set; and (3)Edge formation, which establishes the requisite logical dependencies to integrate these new elements into the global reasoning structure. This mechanism ensures that the symbolic layer remains a faithful, real-time reflection of the reasoning progress. 5 3.3 State Conversions: Backtracking and Drill-Down While the causal graphG t explicitly maps the structural topology of belief, it acts primarily as a static representation lacking the inherent con- trol logic to drive the progressive deepening of the investigation. Consequently, the state machine S t is essential to function as the executive con- troller, regulating the reasoning trajectory through strictly defined transition rules. Drawing inspi- ration from cognitive models of human abduc- tive reasoning (Elstein et al.,1978;Josephson and Josephson,1996;Magnani,2011), which charac- terize abductive reasoning as an iterative process of hierarchical specification and non-monotonic belief revision, we propose two distinct conversion moves:backtrackinganddrill-down(detailed in Algorithm2). c State Conversions edit Backtracking Drill-Down V sym V hyp 1 V hyp 2 V hyp 3 V hyp 4 V hyp 5 V hyp 6 V evi 7 V evi 8 P( V hyp 2 ) < P( V hyp 1 ) � � = 3 → 1 Condition: V sym V hyp 1 V hyp 2 V evi 1 V evi 1 V evi 1 V evi 1 V evi 1 V hyp 3 V hyp 4 � � = 1 → 2 Condition: 1. P( V hyp 2 ) - P( V hyp 1 ) > δ; 2. |E( V hyp 2 )| > η; 3. V hyp 2 not fine enough; Figure 4:State Conversions: Backtracking & Drill- Down. Backtracking. The system incorporates a back- tracking mechanism to address the non-monotonic nature of abductive reasoning, where initially plau- sible conjectures are refuted by emerging contra- dictory evidence. Since the validity of a fine- grained hypothesis strictly relies on the correct- ness of its precursors, the system continuously monitors the ancestral lineage of the current rea- soning focus during confidence recalibration (men- tioned in Section 3.2). A regression is triggered if any ancestor node at levell < S t ceases to be the highest-confidence hypothesis among its sib- lings due to contradictory evidence. To be spe- cific, the state machine identifies the shallowest levell ∗ containing the demoted ancestor and ex- ecutes a pruning operation: all hypothesis nodes with levelL(v hyp )> l ∗ are discarded, enforcing the principle that inferences founded on a flawed premise are inherently invalid. Consequently, the state machineS t+1 is reset tol ∗ , compelling cog- nitive agents to pivot toward previously dormant alternative branches. This self-correction ensures that the system remains resilient to early-stage mis- conceptions caused by information scarcity, pre- venting premature closure on incorrect hypothe- ses. Drill-Down. Beyond the backward regression, the drill-down transition embodies the system’s progressive approximation toward fine-grained ground truths, enabling the transition from high- level conjectures to specific root causes. Follow- ing the belief updates in the causal graph, the state machine evaluates whether the current belief states warrants a deeper investigation based on a rigor- ousdual-threshold mechanism, governed by a con- fidence gapδand a minimum support evidence countη. Specifically, to trigger a state transition, the hypothesis with the highest confidencev (1) hyp must satisfy two simultaneous conditions. First, a confidence gapis enforced: the probability differ- ence between the top-ranked hypothesis and the second-ranked must exceed a predefined thresh- old, denoted asP(v (1) hyp )−P(v (2) hyp )> δ. This criterion ensures that the current direction is un- ambiguously superior to competing alternatives. Second, reflecting the prudent nature required for high-stakes abductive tasks, we impose aneviden- tial support constraint|E sup (v (1) hyp )| ≥η, where E sup denotes the set of supporting evidence nodes linked to the hypothesis. This prevents the sys- tem from prematurely narrowing the search space based solely on prior probabilities without suffi- cient empirical grounding. Upon satisfying these criteria, the central agent assesses the semantic granularity ofv (1) hyp . If the hypothesis is sufficiently concrete to resolve the query, the inference terminates, yielding the final prediction accompanied by a comprehensive rea- soning report; otherwise, the agent generates the next level of sub-hypotheses to refinev (1) hyp , and the state machine increments the reasoning depth S t+1 ←S t + 1. Conversely, if thresholds are not met, the system maintains the current state S t+1 ←S t to conduct further evidence retrieval in the subsequent iteration. 4 Experiments We empirically validate the proposedGoSframe- work on two distinct abductive reasoning tasks: medical diagnosis and failure diagnosis in dis- tributed systems. Datasets.For medical diagnosis, we utilize DiagnosisArena (Zhu et al.,2025), which cu- rates real-world pathological cases reported in top- tier medical journals (e.g.,Lancet,NEJM,JAMA). 6 Specifically, we utilized 150 cases, excluding 12 cases for containing factual errors or logical flaws that make the ground truth unreachable (detailed in AppendixC.1). For failure diagnosis in dis- tributed systems, we constructed a dataset com- prising 150 incidents from a large-scale produc- tion microservice systems of a global leading IT companyL. Comprehensive descriptions of both datasets are provided in Appendix C. Baselines.Given the absence of established frameworks tailored for general abductive reason- ing, we synthesize eight baselines derived from the intersection of two dimensions:agent archi- tecture(Single-Agent vs. Multi-Agent) andrea- soning topology(CoT (Wei et al.,2022) vs. ToT (Yao et al.,2023) vs. GoT (Besta et al.,2024) vs. FoT (Bi et al.,2025)). To ensure a fair comparison, we standardize the atomic reasoning unit across all baselines by adopting the ReAct (Yao et al.,2022) paradigm, wherein each node encapsulates an interleaved triplet ofthought,ac- tion, andobservation. Thereby, this configura- tion ensures a rigorous evaluation covering the full spectrum from solitary linear reasoning to col- laborative hierarchical search strategies. And we useGPT-5.1-2025-11-13as the backbone of the agents for all baselines methods andGoS. Evaluation.To rigorously assess diagnostic performance, we primarily employLLM-as-a- Judgeutilizing a standardized 3-point scale: 2 (Ex- act Match), 1 (Relevant), and 0 (Otherwise). Ac- cordingly, we report two metrics:Match(scoring 2) andRelevant(scoring 1 or 2). For evaluation prompts, we utilize the official benchmark prompt for medical diagnosis scenario and construct a par- allel prompt for failure diagnosis in distributed sys- tems. However, unlike the deterministic nature of distributed system failures, medical diagnosis en- tails subtle semantic nuances prone to automated misinterpretation. Consequently, we specifically incorporateHuman-as-a-Judgefor the medical domain to ensure reliability, conducted by a re- searcher with extensive clinical and academic ex- perience. Parameter settings.We maintain identical hy- perparameter configurations across all scenarios to facilitate fair comparison. We impose strict re- trieval budgets: each expert agent inGoSand the Multi-Agent baselines is limited to 3 retrieval ac- tions, while the Single-Agent baselines are capped at 5. We also set the maximum number of neuro- symbolic interaction iterations inGoSto 3. 4.1 Medical Diagnosis Medical diagnosis serves as a quintessential high- stakes abductive task, requiring the inference of diseases from clinical observations through rigor- ous evidence discovery and cautious reasoning. Task Setup.We reframe the task as a dynamic investigation process: Initially provided only with surface symptoms (i.e., chief complaints and phys- ical examinations), the model must explicitly is- sue auxiliary examination to retrieve correspond- ing records from an external repository to formu- late an accurate diagnosis. InGoS, we align the cognitive layer with real-world clinical roles, des- ignating thePrimary_Physicianas the central agentA central and theLaboratory_Physician, Pathologist,Radiologistas expert agents A experts . Given that MDAgents ( Kim et al.,2024) has already mapped agents to real-world roles, we apply this identical agent configuration to the Multi-Agent baselines to ensure consistency. Table 1: Performance of Medical Diagnosis (%) Methods LLM-as-a-Judge Human-as-a-Judge $/case Match Relevant Match Relevant GoS31.88 74.64 39.8678.990.12 Single/CoT21.01 47.83 24.6448.550.03 Single/ToT18.84 47.10 19.5745.650.08 Single/GoT 21.01 50.00 22.4652.900.07 Single/FoT 21.74 58.70 21.0161.590.32 Multi/CoT 21.01 49.28 23.1950.720.07 Multi/ToT20.29 50.72 23.9153.620.17 Multi/GoT21.74 52.17 23.9155.070.15 Multi/FoT23.19 63.04 26.0965.940.73 Results.As shown in Table1,GoSsig- nificantly outperforms all baseline methods un- der both LLM-as-a-Judge and Human-as-a-Judge evaluation settings. Specifically, under Human- as-a-Judge evaluation, theMatchmetric ofGoS reaches 39.86%, and theRelevantmetric reaches 78.99%. Regardingreasoning topology, we ob- serve distinct patterns among baselines. The CoT often achieves higherMatchmetrics than ToT, as ToT tends to deviate due to ineffective node evalu- ation mechanisms. However, extending the topol- ogy from trees (ToT) to graphs (GoT) and forests (FoT) yields performance gains, particularly in the Relevantmetric. FoT emerges as the strongest baseline, suggesting that a broader search space helps capture comprehensive information. Yet, this comes at a steep price: Multi/FoT incurs the highest cost ($0.73/case), whereasGoSachieves superior accuracy at a fraction of the cost ($0.12), 7 validating the efficiency of our directed search. In terms ofagent architecture, Multi-Agent base- lines consistently outperform Single-Agent base- lines. This advantage stems from the effective division of labor among multiple expert agents, which enables the model to focus on more informa- tion sources and thus provide more comprehensive information support for diagnosis. Additionally, we observe that Human-as-a-Judge consistently yields higher scores than LLM-as-a-Judge. This is because LLM-as-a-Judge prioritizes superficial textual similarity while neglecting deep semantic information, whereas professional researchers can comprehend diverse expressions of the same diag- nosis and assign rational scores. We further pro- vide a granular error analysis in AppendixD, dis- secting distinct failure modes to shed light on cur- rent limitations and future directions. Table 2: Ablation Study of Medical Diagnosis (%) Methods LLM-as-a-Judge $/case Match Relevant GoS31.88 74.640.12 w/o reasoning focus19.57 67.390.14 w/o causal graph12.32 48.550.12 w/o state machine 12.32 50.000.17 Ablation Study.As presented in Table2, ev- ery component ofGoSis essential, with their re- moval leading to performance degradation and higher cost. Initially, removing the reasoning fo- cush ∗ t results in a decline in theMatch(31.88%→ 19.57%) alongside a slightly cost increase. This underscores its role in enforcing a depth-first inves- tigation strategy, constraining the agents to scruti- nize the most promising hypotheses rather than en- gaging in stochastic exploration. Furthermore, re- moving the causal graph or state machine causes a catastrophic drop in performance, withMatch rates falling to12.32%. This validates that con- structing and maintaining belief states are indis- pensable in complex abductive reasoning tasks. Notably, we observe a cost spike when removing the state machine ($0.17/case). We attribute this increase to the absence of valid state transition pro- tocols, as without a mechanism to explicitly man- age the reasoning lifecycle and termination, the system tends to exhaust the full reasoning budget before forcing a conclusion, thereby incurring un- necessary computational overhead. Sensitivity Analysis.Figure 5presents a sen- sitivity analysis ofGoSunder varying config- Match/Relevant(%) Maximum Iteration 3 157 20 40 60 80 100 Match/Relevant(%) Maximum Retrieval 3157 20 40 60 80 100 Match/Relevant(%) η 3157 20 40 60 80 100 Match/Relevant(%) δ 0.30.10.50.7 20 40 60 80 100 Match RelevantGoSBest Baseline Figure 5: Sensitivity Analysis. Solid line stands for GoS, dashed line stands for best baseline (Multi/FoT). (1)upper-left: maximum number of neuro-symbolic interaction iterations; (2)upper-right: maximum num- ber of retrieval actions of expert agent; (3)lower-left: minimum support evidence for drill-down transition; (4)lower-right: confidence gap for drill-down transi- tion. urations. Regarding reasoning budget (Upper- Left/Right), increasing both interaction iterations and retrieval steps generally enhances perfor- mance. Notably, comparisons with the best- performing baseline (dashed line) reveal thatGoS achieves superior efficiency, surpassing the base- line’s peak performance even with a restricted bud- get. Regarding the dual-thresholds (η, δ) intro- duced in Section 3.3(Lower-Left/Right), we ob- serve a distinct trade-off between precision and conservatism. Raising these thresholds imposes stricter criteria for drill-down transitions, com- pelling the model to accumulate more support- ing evidence and ensure higher confidence in the top-ranked hypothesis. While moderate thresh- olds improve correctness by curbing hasty reason- ing, excessively high thresholds force the model to adopt a conservative strategy: it tends to terminate at superficial diagnosis (boostingRelevant) rather than risking a fine-grained root cause prediction without overwhelming evidence (loweringMatch). This behavior validates that these thresholds act as effective control knobs, allowing users to tune the system’s risk tolerance based on deployment needs. 4.2 Failure Diagnosis in Distributed Systems Failure diagnosis in distributed systems is a crit- ical operational task to mitigate substantial eco- nomic losses, requiring the joint analysis of hetero- 8 geneous observability data and system commands to pinpoint root causes. Task Setup.For each case, the model is pro- vided with the initial alert, specifying the alert type, affected components, timestamps, and over- all failure descriptions. Crucially, detailed di- agnostic evidence is hidden within the massive raw observability data, requiring the model to ac- tively investigate to reconstruct the failure con- text. To execute this investigation,GoSinstan- tiates anApplicationOperatoras the central agent (A central ) orchestrating three expert agents (A experts ):LinuxOperator,NetworkOperator, andDatabaseOperator. These experts encapsu- late specialized capabilities, such as querying his- torical logs, retrieving system metrics, and execut- ing restricted shell commands. For Multi-Agent baselines, we adopt the domain-oriented design of D-Bot ( Zhou et al.,2024), instantiating resource- centric experts (e.g., CPU, Memory, and Disk ex- perts), each with access limited to their respective domain-relevant observability data. Table 3: Performance of Failure Diagnosis in Dis- tributed Systems (%) Methods LLM-as-a-Judge $/case Match Relevant GoS70.67 88.000.10 Single/CoT 26.67 81.330.03 Single/ToT25.33 78.000.14 Single/GoT27.33 80.000.11 Single/FoT28.67 84.000.45 Multi/CoT34.00 82.670.05 Multi/ToT 25.33 81.330.13 Multi/GoT28.00 80.670.18 Multi/FoT28.00 86.670.94 Results.As detailed in Table3,GoSestablishes a new SOTA under the LLM-as-a-Judge setting. Specifically,GoSachieves 70.67% inMatchand 88.00% inRelevant, surpassing the respective best- performing baselines by margins of 36.67% and 1.33%. We observe that the performance trends regardingreasoning topologyandagent architec- tureslargely mirror those in the medical diagnosis task. Notably, among baselines, Multi/CoT yields the highestMatchscore (34.00%), whereas Multi- /FoT attains the highestRelevantscore (86.67%) but incurs the highest cost ($0.94/case). Despite the competitiveRelevantscores across baselines, a distinct pattern emerges: they significantly under- perform inMatchcompared toGoS. We attribute this to the structured nature of system alerts, which contain explicit metadata that easily guides mod- els to the correct failure domain (highRelevant). Yet, distinguishing the precise root cause from su- perficial symptoms requires in-depth investigation. The superiorMatchperformance ofGoSstems from its state machine, which enforces a coarse- to-fine reasoning process. This mechanism effec- tively refines coarse hypotheses into fine-grained root causes, whereas baselines often stagnate at su- perficial diagnoses (lowMatch). Regarding cost, GoSdelivers superior accuracy at a fraction of Multi/FoT’s expense ($0.10 vs. $0.94,∼8x cost reduction), validating the economic efficiency of GoS. 2025-12-12 07:31:00 P2 Linux 10.195.19.39 Filesystem Read Only (Trigger Value:1) 2025-12-12 07:32:25 P1 Linux 10.195.19.39 Memory Usage > 90% (Trigger Value:92.34%) 2025-12-12 07:31:10 P1 Linux 10.195.19.39 System Health Score < 90% (Trigger Value:36.59%) LinuxOperator Think: I need to check filesystem Shell: mount | grep ro, Shell: dmesg | grep XFS Think: I need to check XFS Obs: /dev/mapper/rl-root on ... Obs: 15.368219] XFS (dm-0) ... Analysis: Existing shell evidence confirms the root mount point “/” (XFS on dm-0) is mounted read-only and ... ApplicationOperator Initialization: Aggregating surface alerts to construct initial diagnostic hypotheses. Hypothesis-1: Filesystem Read-Only conf = 0.75 FindFocus: Find the highest-confidence hypothesis (id = Hypothesis-1, conf = 0.75) Plan: Call LinuxOperator to inspect filesystem read-only state. Hypothesis-2: High Memory Pressure conf = 0.60 ApplicationOperator UpdateGraph: Update the causal graph based on LinuxOperator Analysis. Hypothesis-1: Filesystem Read-Only conf: 0.75 0.95 BackTracking: No Hypothesis-2: High Memory Pressure conf: 0.60 0.40 RefineHypo: Hypothesis-1: Filesystem Read-Only Hypothesis-3: XFS metadata corruption conf = 0.90 CheckReport: The hypothesis is sufficiently concrete to resolve the query — report now. Drill-Down: Yes : 1 2 Prediction: XFS metadata corruption on host 10.195.19.39 forced root filesystem read-only. Report: On host Linux 10.195.19.39, the root filesystem (XFS on dm-0) experienced metadata corruption, as explicitly reported in the dmesg output, which caused the kernel to remount the filesystem in read-only mode. This directly ... Figure 6: An example of failure diagnosis in distributed systems Example Case.Figure6presents a real incident from a production host where multi- ple alerts are triggered. Under theGoSframe- work, theApplicationOperatorfirst aggregates these alerts to initialize several hypotheses and selects “Filesystem Read-Only” as the reasoning focus due to its highest confidence. Then it coordinates the investigation by instructing the LinuxOperatorto inspect the filesystem status. Through restricted shell inspections, including mount state verification and XFS-related kernel log analysis, theLinuxOperatorobserves that the root filesystem (XFS on dm-0) has been re- mounted in read-only mode, with kernel logs ex- plicitly reporting XFS metadata corruption. Based on this evidence,GoSupdates the causal graph 9 by increasing the confidence of the current rea- soning focus while suppressing competing alter- natives, and enables a drill-down transition to a finer-grained root cause. The system finally re- ports XFS metadata corruption as the root cause, which remounted the root filesystem in read-only mode. 5 Conclusion In this work, we presentedGoS, a general-purpose neuro-symbolic framework tailored for abductive reasoning. Addressing the structural limitations where deductive paradigms fail, we introduced a dual-layer architecture that grounds collaborative investigation in explicit belief states via a causal graph and a state machine. This design trans- forms aimless, unconstrained exploration into a di- rected, convergent search, effectively navigating incomplete information to identify fine-grained root causes. Looking forward, we envisionGoS serving as a robust reasoning backbone that, when coupled with domain-specific adaptations, paves the way for reliable decision-making across com- plex, high-stakes real-world domains. References Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Giani- nazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, and 1 others. 2024. Graph of thoughts: Solving elaborate problems with large language models. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 17682–17690. Zhenni Bi, Kai Han, Chuanjian Liu, Yehui Tang, and Yunhe Wang. 2025. Forest-of-thought: Scaling test- time compute for enhancing llm reasoning. InForty- second International Conference on Machine Learn- ing. Vardhan Dongre, Ryan A Rossi, Viet Dac Lai, David Seunghyun Yoon, Dilek Hakkani-Tür, and Trung Bui. 2025. Drift no more? context equilib- ria in multi-turn llm interactions.arXiv preprint arXiv:2510.07777. Arthur S Elstein, Lee S Shulman, and Sarah A Sprafka. 1978.Medical problem solving: An analysis of clin- ical reasoning. Harvard University Press. Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to rea- son in a continuous latent space.arXiv preprint arXiv:2412.06769. Azam Ikram, Sarthak Chakraborty, Subrata Mitra, Shiv Saini, Saurabh Bagchi, and Murat Kocaoglu. 2022. Root cause analysis of failures in microservices through causal discovery.Advances in Neural In- formation Processing Systems, 35:31158–31170. Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. 2021. What dis- ease does this patient have? a large-scale open do- main question answering dataset from medical ex- ams.Applied Sciences, 11(14):6421. John R Josephson and Susan G Josephson. 1996. Ab- ductive inference: Computation, philosophy, tech- nology. Cambridge University Press. Yubin Kim, Chanwoo Park, Hyewon Jeong, Yik S Chan, Xuhai Xu, Daniel McDuff, Hyeonhoon Lee, Marzyeh Ghassemi, Cynthia Breazeal, and Hae W Park. 2024. Mdagents: An adaptive collaboration of llms for medical decision-making.Advances in Neural Information Processing Systems, 37:79410– 79452. Cheryl Lee, Tianyi Yang, Zhuangbin Chen, Yuxin Su, and Michael R Lyu. 2023. Eadro: An end-to- end troubleshooting framework for microservices on multi-source data. In2023 IEEE/ACM 45th Interna- tional Conference on Software Engineering (ICSE), pages 1750–1762. IEEE. Mingjie Li, Zeyan Li, Kanglin Yin, Xiaohui Nie, Wenchi Zhang, Kaixin Sui, and Dan Pei. 2022a. Causal inference-based root cause analysis for on- line service systems with intervention recognition. InProceedings of the 28th ACM SIGKDD Confer- ence on Knowledge Discovery and Data Mining, pages 3230–3240. Zeyan Li, Nengwen Zhao, Shenglin Zhang, Yongqian Sun, Pengfei Chen, Xidao Wen, Minghua Ma, and Dan Pei. 2022b. Constructing large-scale real- world benchmark datasets for aiops.arXiv preprint arXiv:2208.03938. Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paran- jape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the middle: How language models use long contexts.Transactions of the Asso- ciation for Computational Linguistics, 12:157–173. Yilun Liu, Yuhe Ji, Shimin Tao, Minggui He, Weibin Meng, Shenglin Zhang, Yongqian Sun, Yuming Xie, Boxing Chen, and Hao Yang. 2025. Loglm: From task-based to instruction-based automated log analy- sis. In2025 IEEE/ACM 47th International Confer- ence on Software Engineering: Software Engineer- ing in Practice (ICSE-SEIP), pages 401–412. IEEE. Yu Luo, Jiamin Jiang, Jingfei Feng, Lei Tao, Qingliang Zhang, Xidao Wen, Yongqian Sun, Shenglin Zhang, Jielong Huang, Nan Qi, and 1 others. 2025. From observability data to diagnosis: An evolving multi- agent system for incident management in cloud sys- tems.arXiv preprint arXiv:2510.24145. 10 Lorenzo Magnani. 2011.Abduction, reason and science: Processes of discovery and explanation. Springer Science & Business Media. Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, and 1 others. 2022. In-context learning and induc- tion heads.arXiv preprint arXiv:2209.11895. Ankit Pal, Logesh Kumar Umapathi, and Malaikan- nan Sankarasubbu. 2022. Medmcqa: A large-scale multi-subject multi-choice dataset for medical do- main question answering. InConference on health, inference, and learning, pages 248–260. PMLR. Changhua Pei, Zexin Wang, Fengrui Liu, Zeyan Li, Yang Liu, Xiao He, Rong Kang, Tieying Zhang, Jian- jun Chen, Jianhui Li, and 1 others. 2025. Flow-of- action: Sop enhanced llm-based multi-agent system for root cause analysis. InCompanion Proceedings of the ACM on Web Conference 2025, pages 422– 431. Charles Sanders Peirce. 1934.Collected papers of charles sanders peirce, volume 5. Harvard Univer- sity Press. Binpeng Shi, Yu Luo, Jingya Wang, Yongxin Zhao, Shenglin Zhang, Bowen Hao, Chenyu Zhao, Yongqian Sun, Zhi Zhang, Ronghua Sun, and 1 oth- ers. 2025. Flowxpert: Expertizing troubleshooting workflow orchestration with knowledge base and multi-agent coevolution. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pages 4839–4850. Karan Singhal, Tao Tu, Juraj Gottweis, Rory Sayres, Ellery Wulczyn, Mohamed Amin, Le Hou, Kevin Clark, Stephen R Pfohl, Heather Cole-Lewis, and 1 others. 2025. Toward expert-level medical ques- tion answering with large language models.Nature Medicine, 31(3):943–950. Yongqian Sun, Yu Luo, Xidao Wen, Yuan Yuan, Xi- aohui Nie, Shenglin Zhang, Tong Liu, and Xi Luo. 2025. Trioxpert: An automated incident manage- ment framework for microservice system.arXiv preprint arXiv:2506.10043. Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowd- hery, and Denny Zhou. 2023. Self-consistency im- proves chain of thought reasoning in language mod- els. InThe Eleventh International Conference on Learning Representations. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elicits reasoning in large language models.Ad- vances in neural information processing systems, 35:24824–24837. Chaoyi Wu, Weixiong Lin, Xiaoman Zhang, Ya Zhang, Weidi Xie, and Yanfeng Wang. 2024. Pmc-llama: toward building open-source language models for medicine.Journal of the American Medical Infor- matics Association, 31(9):1833–1843. Guangzhi Xiong, Qiao Jin, Zhiyong Lu, and Aidong Zhang. 2024. Benchmarking retrieval-augmented generation for medicine. InFindings of the Associa- tion for Computational Linguistics ACL 2024, pages 6233–6251. Junjielong Xu, Qinan Zhang, Zhiqing Zhong, Shilin He, Chaoyun Zhang, Qingwei Lin, Dan Pei, Pinjia He, Dongmei Zhang, and Qi Zhang. 2025. Openrca: Can large language models locate the root cause of software failures? InThe Thirteenth International Conference on Learning Representations. Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2023. Tree of thoughts: Deliberate problem solving with large language models.Advances in neural in- formation processing systems, 36:11809–11822. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InThe eleventh international con- ference on learning representations. Guangba Yu, Pengfei Chen, Yufeng Li, Hongyang Chen, Xiaoyun Li, and Zibin Zheng. 2023. Nezha: Interpretable fine-grained root causes analysis for microservices on multi-modal observability data. In Proceedings of the 31st ACM Joint European Soft- ware Engineering Conference and Symposium on the Foundations of Software Engineering, pages 553–565. Wei Zhang, Hongcheng Guo, Jian Yang, Zhoujin Tian, Yi Zhang, Yan Chaoran, Zhoujun Li, Tongliang Li, Xu Shi, Liangfan Zheng, and 1 others. 2024. mabc: multi-agent blockchain-inspired collabora- tion for root cause analysis in micro-services archi- tecture. InFindings of the Association for Computa- tional Linguistics: EMNLP 2024, pages 4017–4033. Ruochen Zhao, Xingxuan Li, Shafiq Joty, Chengwei Qin, and Lidong Bing. 2023. Verify-and-edit: A knowledge-enhanced chain-of-thought framework. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5823–5840. Xuanhe Zhou, Guoliang Li, Zhaoyan Sun, Zhiyuan Liu, Weize Chen, Jianming Wu, Jiesi Liu, Ruohang Feng, and Guoyang Zeng. 2024. D-bot: Database diagno- sis system using large language models.Proceed- ings of the VLDB Endowment, 17(10):2514–2527. Yucheng Zhou, Lingran Song, and Jianbing Shen. 2025. MAM: Modular multi-agent framework for multi- modal medical diagnosis via role-specialized collab- oration . InFindings of the Association for Computa- tional Linguistics: ACL 2025, pages 25319–25333, Vienna, Austria. Association for Computational Lin- guistics. 11 Yakun Zhu, Zhongzhen Huang, Linjie Mu, Yutong Huang, Wei Nie, Jiaji Liu, Shaoting Zhang, Pengfei Liu, and Xiaofan Zhang. 2025. Diagnosisarena: Benchmarking diagnostic reasoning for large lan- guage models.arXiv preprint arXiv:2505.14107. 12 A Algorithm Here we present the pesudo code of the overall workflow ofGoSin Algorithm1, and the state conversion mechanism in Algorithm2. Algorithm 1Graph of States 1:Input:surface symptomsO surf , central agentA central , expert agentsA experts 2:ReportF lag=f alse 3:B 0 ,G 0 ,S 0 = Initialization(A central ,O surf ) 4:repeat 5:// Symbolic-to-Cognitive 6:h ∗ t = FindFocus(G t ) 7:Instructions= Plan(A central ,h ∗ t ,B t ) 8:O deep ,analyses= ReAct(A experts ,Instructions,h ∗ t ,B t ) 9:// Cognitive-to-Symbolic 10:G t+1 = UpdateGraph(A central ,analyses,G t ) 11:// State Conversions, detailed in Algorithm2 12:S t+1 ,G ∗ t+1 = StateConversion(A central ,G t+1 ,S t ) 13:ReportF lag= CheckReport(G ∗ t+1 ) 14:untilReportF lagistrue 15:prediction, report = Report(A central ,B t+1 ) 16:Return:prediction, report Algorithm 2State Conversions 1:Input:graphG t+1 , gap_deltaδ, min_supportη 2:// Backtracking 3:BacktrackF lag=f alse 4:BacktrackF lag,l ∗ = CheckBacktrack(G t+1 ) 5:ifBacktrackF lagistruethen 6:forL(v i hyp )> l ∗ do 7:Delete(v i hyp ) 8:end for 9:end if 10:// Drill-down 11:ifP(v (1) hyp )−P(v (2) hyp )> δand|E sup (v (1) hyp )|≥ηthen 12:// No drill-down if granularity is fine enough 13:ifCheckGranularity(A central , G t+1 )istruethen 14:Return:S t ,G t+1 15:end if 16:S t+1 =S t + 1 17:// Generate fine-grained hypothesis 18:G ∗ t+1 = RefineHypo(A central ,G t+1 ) 19:else 20:S t+1 =S t 21:end if 22:Return:S t+1 ,G ∗ t+1 B Limitations While our experiments demonstrate the efficacy ofGoSacross diverse domains, we acknowledge cer- tain limitations regarding the breadth of our evaluation. Currently, our experimental scope is confined 13 to medical diagnosis and failure diagnosis in distributed systems. Although abductive reasoning is criti- cal in other high-stakes fields such as criminal investigation, the acquisition of high-quality benchmarks in these domains remains a significant challenge due to strict privacy regulations and data protection laws, which limit access to the detailed evidential data required to construct authentic abductive scenar- ios. Nevertheless, the datasets employed in this study are of exceptional qualitysourced from top-tier medical journals and real-world production environments of leading technology companiesthereby en- suring that our findings regarding the model’s reasoning capabilities remain robust and representative (see AppendixCfor more detail). C Descriptions of datasets C.1 Dataset for Medical Diagnosis Prediction: Acute posterior circulation ischemic stroke associated with left vertebral artery fenestration. Report: The 36-year-old man developed acute dizziness, vomiting, gait instability, and positive Romberg sign, and MRI shows multiple subacute infarcts in the left thalamus, occipital lobe, and cerebellar hemisphere, directly confirming ... 在此处键入公式。 A 36-year-old man, previously well, became dizzy and started vomiting, then couldn’t walk steadily within hours—after an exaggerated head or neck turn. Examination Imaging Studies Laboratory Tests Other Tests - Emergency brain CT: No hemorrhage found - Head and neck CTA: Fenestration deformity of V2–V3 segment of left vertebral artery, duplication shape type - Blood routine: No increase in eosinophils - CSF: Normal red blood cell and white blood cell counts - Transcranial Doppler foam test: No abnormal embolism 在此处键入公式。 Comprehensive Patient Data 在此处键入公式。 Prediction:Acute posterior stroke. Report: The 36-year-old man developed acute dizziness, vomiting, gait instability ... 在此处键入公式。 LLM Reasoning Input Output Input Output Original DiagnosisArena Our interactive reformulation Figure 7: Reformulating DiagnosisArena as an interactive diagnostic task. In the original setting, all auxiliary examination records are revealed upfront. In contrast, our reformulation restricts the initial input to the chief complaint and basic physical examination, and requires the model to explicitly request auxiliary examinations from an external repository before making the final diagnosis. Why do we use DiagnosisArena?We utilize DiagnosisArena (Zhu et al.,2025), a benchmark curated from case reports published in top-tier medical journals (e.g., Lancet, NEJM, JAMA) to evaluate diagnos- tic reasoning capability. Benchmarks derived from medical licensing examinations, such as MedQA (Jin et al. ,2021) and MedMCQA (Pal et al.,2022), primarily focus on evaluating the retention of standardized medical knowledge rather than the reasoning capabilities required for diagnosis. In contrast, Diagnosis- Arena simulates complex abductive reasoning scenarios over multidimensional patient records, and it is subjected to rigorous human verification to guarantee high data quality and clinical fidelity. How do we reframe the task?As illustrated in Figure7, the original DiagnosisArena benchmark presents models with comprehensive patient data upfront, including all critical auxiliary examination records (e.g., laboratory tests and imaging) alongside the initial symptoms. This setting significantly trivializes the diagnostic challenge, as in real-world clinical practice, auxiliary examinations incur sub- stantial temporal and financial costs. Consequently, the core difficulty lies in identifying the necessary examinations to accumulate evidence, which is eliminated in its original setting. To simulate this realistic constraint, we reframe the task as a dynamic investigation process. We restrict the initial input strictly 14 to the patient’s chief complaint and basic physical examination, encapsulating all auxiliary examination results within an external information repository. The model must explicitly issue precise auxiliary ex- amination commands to retrieve corresponding records; otherwise, no information is revealed. This for- mulation effectively restores the inference difficulty and faithfully mirrors the active, evidence-seeking nature of medical diagnosis. Case 1 Case 2 Case 3 Case 4 Case 5 Case 6 Case 7 Case 8 Case 9 Case 10 LLM-as-a-Judge IIIIIIIVV 11111 00000 11111 22222 01111 22122 11111 00000 01100 21210 Figure 8: Scoring consistency of LLM-as-a-Judge across five repetitions for ten medical diagnostic cases (green: 6 consistent; red: 4 inconsistent). Why do we incorporate Human-as-a-Judge?Although LLM-as-a-Judge demonstrates strong align- ment with human evaluations in general domains due to large-scale Reinforcement Learning from Human Feedback (RLHF), its instability in medical diagnosis necessitates human oversight. Empirical analysis reveals that identical diagnostic outputs receive inconsistent scores across repeated LLM assessments (as shown inFigure 8), with unresolved ambiguities persisting even after majority voting. Crucially, LLMs exhibit a pronounced bias toward longer responses, prioritizing length over diagnostic accuracya direct consequence of the high domain specificity and insufficient medical expertise in current language models. To ensure robust evaluation, we introduce Human-as-a-Judge: a clinical researcher with exten- sive academic and clinical experience performed110+hours of manual assessment, strictly prioritizing diagnostic correctness against ground truth. This approach enables granular error analysis, uncovering systematic model failures that directly inform targeted model refinement and future clinical deployment. Why do we need to remove several cases?Following manual verification, we excluded 12 cases containing critical information gaps or logical contradictions that render the labeled diagnosis impossible to derive from the provided evidence. To substantiate the necessity of this filtration and strictly maintain the integrity of our benchmark, we provide a detailed analysis of three representative examples below. These instances illustrate specifically how such intrinsic defects obstruct the valid diagnostic reasoning process. Case#1: Labeled answer is a secondary diagnosis unrelated to the chief complaint Brief description:A man in his 90s with a history of hypertension, hypercholesterolemia, sinus node dysfunction, and prior dual-chamber pacemaker implantation presented with 1 month of abdominal pain radiating to his back. Labeled Answer:Pacemaker pseudofusion. Why:In this case, the patient presented primarily with abdominal pain radiating to the back for 1 month. Auxiliary examinations identified a highly likely cause of the abdominal pain: a 5.3 Œ 5.3-cm infrarenal abdominal aortic aneurysm with thrombus. Both the chief complaint 15 and auxiliary examinations indicate that the patient’s abdominal pain is caused by the abdominal aortic aneurysm, which is highly consistent with the answer generated byGoS. The labeled answer provided by the dataset is “Pacemaker pseudofusion”, which may be inferred from the absence of atrial or ventricular pacing signals on the electrocardiogram. However, this diagnosis cannot explain the patient’s abdominal pain radiating to the back and thus cannot serve as the primary diagnosis. Therefore, the reason for excluding this case is that the labeled answer cannot serve as the patient’s primary diagnosis but only as a secondary one, which deviates from the core issue addressed by the chief complaint. Case#2: Incorrect labeled answer Brief description:A 36-year-old woman (gravida 2, para 2) presented with right lower pelvic pain. The patient had been previously diagnosed with presumed progressive uterine fibroids 6 months earlier and sought a second opinion. Labeled Answer:Tubal angiomyofibroblastoma Why:The reason for excluding this case is that the labeled answer provided by the dataset is incorrect. The answer given is “Tubal angiomyofibroblastoma”, while we confirm the diagno- sis as uterine leiomyoma. Firstly, based on the chief complaint and imaging findings, this case can be identified as a tumor-related disease; therefore, the gold standard for diagnosis should be pathological examination. Subsequently, we focused on the information from the pathological examination and presented three pieces of evidence inconsistent with “Tubal angiomyofibroblas- toma”: 1.Mismatched immunohistochemical (IHC) profile: The pathological examination demon- strated “Desmin (desmin) positive + ER/PR positive + CD34 negative” whereas the typi- cal IHC features of angiomyofibroblastoma are “Desmin negative/weakly positive + CD34 positive + ER/PR mostly negative”. Immunohistochemistry serves as the “identity code” for tumor differentiation. The reverse expression of Desmin (a smooth muscle-specific marker, positive) and CD34 (positive in angiomyofibroblastoma) constitutes the core differential ba- sis between the two entities. A discrepancy in a single key indicator is sufficient for exclu- sion. 2.Mismatched morphological characteristics: The pathological examination revealed “spindle cells and cords of epithelioid cells surrounding numerous blood vessels + loose fibromyxoid stroma + well-demarcated border” while the typical morphological hallmarks of angiomyofi- broblastoma include “angiocentric clustered/nested arrangement (cells densely surrounding blood vessels to form nodules) + alternating dense and sparse cellular areas + stroma may be myxoid but not the typical loose fibromyxoid type”. Although both tumors exhibit “perivas- cular growth”, the “angiocentric clustering” of angiomyofibroblastoma is a signature feature. The report only describes cellular cords around blood vessels without dense clustering/nest- ing, and the stroma is typically loose fibromyxoid, which is inconsistent with the structural characteristics of angiomyofibroblastoma. 3.Hormone correlation: The report indicates ER/PR positivity, suggesting the tumor is hor- monally regulated. In contrast, angiomyofibroblastoma has no association with hormones (both ER and PR are negative), further supporting the exclusion of this diagnosis. 16 Case#3: Insufficient clinical data to derive the labeled answer Brief description:A 57-year-old male patient was admitted with breathlessness, cough and weak- ness from the day before. He had a history of a head injury 20 years ago which resulted in cerebral atrophy, quadriplegia and aphasia. The patient had a history of pulmonary thromboembolism in the past year which was under treatment. He also had a history of hypertension and several hospi- talizations. Labeled Answer:NDM-positive Burkholderia cepacia complex (Bcc) lower respiratory tract infection Why:The main reason for excluding this case is the insufficient information provided, which is inadequate to support a complete answer and thereby affects the accuracy of the response generated by the model. The labeled answer specifies the multidrug-resistant (MDR) bacterium as Burkholderia cepacia complex (Bcc) a result that requires specific bacterial species identification to confirm. However, the available data only supports the diagnosis of “multidrug-resistant Gram- negative bacteria,” soGoSsolely provided the answer: “MDR Gram-negative bacterial aspiration pneumonia causing acute lower respiratory infection.” C.2 Dataset for Failure Diagnosis in Distributed Systems Why do we construct a real-world distributed systems incident dataset?Existing RCA/AIOps bench- marks have enabled important progress, but the prevailing evaluation setups still differ from live incident response in two key ways: (1) some benchmarks are evaluated under synthetic or limited-scale settings, which may not reflect the complexity and heterogeneity of production incidents ( Li et al.,2022a;Ikram et al.,2022); and (2) many benchmarks define diagnosis around a narrow target (e.g., predicting only one type of root-cause element), which can encourage goal-specific solutions and limit transferability across incident objectives ( Li et al.,2022b;Lee et al.,2023;Yu et al.,2023). OpenRCA (Xu et al.,2025) takes a meaningful step toward realism by providing multi-source telemetry from production systems and adopting a goal-driven task formulation, yet it still largely treats diagnosis as an offline problem where evidence is assumed to be pre-materialized at inference time (and, when first-hand incident re- ports are unavailable, queries are approximated via synthesis). In contrast, real-world failure diagnosis is inherently interactive under partial and evolving observability: on-call engineers iteratively probe the sys- tem by issuing shell-level commands, inspecting live states, and acquiring additional signals on demand (Figure 9), rather than passively analyzing a fixed telemetry corpus. To faithfully evaluate diagnostic reasoning in this evidence-seeking setting, we construct a dataset directly from authentic production incidents, explicitly capturing intermediate shell snapshots and on-demand observations along the inves- tigation trajectory. Dataset characteristics.We collect 150 production incidents and represent each as an incident-centric investigation trajectory, spanning from the first alert to the postmortem-confirmed root cause. Each incident records heterogeneous diagnostic signals observed during live response, including time-series metrics, system logs, and shell-level snapshots. The root cause labels are derived from post-incident analyses conducted by experienced on-call engineers. Importantly, shell snapshots capture intermediate probing actions and raw system states (i.e., command outputs) during troubleshooting, rather than curated summaries or oracle evidence. Consequently, the root cause is not explicitly revealed unless the collected evidence is correctly interpreted and integrated over the investigation over time. Scale and coverage.The incidents are drawn from an industrial observability platform monitoring on the order of2.5×10 4 infrastructure instances, spanning multiple layers of modern stacks: data- center facilities (e.g., temperature/humidity sensors, precision cooling, backup power), infrastructure hardware (mainstream server families), virtualization and cloud (e.g., VMware/OpenStack and major public clouds), hosts (e.g., Linux/Windows/AIX), middleware (e.g., Kafka/RabbitMQ), databases (e.g., PostgreSQL/MySQL/MongoDB/Redis), and application/service-level end-to-end monitoring (including enterprise analytics workloads such as HANA). The platform ingests network log streams at the order of10 10 entries per day (e.g.,∼12B/day) and retains PB-scale historical observability data, which makes 17 Investigation 2025-12-12 07:31:00 P2 Linux 10.195.19.39 Filesystem Read Only (Trigger Value:1) 2025-12-12 07:32:25 P1 Linux 10.195.19.39 Memory Usage > 90% (Trigger Value:92.34%) 2025-12-12 07:31:10 P1 Linux 10.195.19.39 System Health Score < 90% (Trigger Value:36.59%) 在此处键入公式。 Prediction: XFS metadata corruption on host Linux 10.195.19.39 forced root filesystem read-only. Report: On host Linux 10.195.19.39, the root filesystem (XFS on dm-0) experienced metadata corruption, as explicitly reported in the dmesg output, which caused the kernel to remount the filesystem in read-only mode. This directly ... User-service In-memory queue size exceeds threshold 2025-12-12 15:22:15 ERROR 2025-12-12 15:22:15 WARN [ 15.368219] XFS (dm-0): Metadata corruption at xfs_inode_buf_verify+0x5e/0x100 [xfs], block 0x2d18b30 [ 15.368225] XFS (dm-0): Unmount and run xfs_repair [ 15.368228] XFS (dm-0): First 128 bytes of corrupted metadata buffer: [ 15.369521] XFS (dm-0): Remounting filesystem read-only cpu_usage health_score memory_usage 2025-12-12 15:21:06 ERROR Chronyd can't open file /var/lib/chrony/drift for writing (Read-only file system) User-service Failed to append application log file Log Metric Shell > dmesg | grep xfs | tail –n 4 Figure 9: An example of an interactive failure diagnosis process in distributed systems. Given an initial alarm summary, the agent iteratively investigates the incident by querying heterogeneous signals, including time-series metrics, system logs, and shell-level snapshots. Through progressive evidence collection and interpretation, the system converges to the final root cause. root-cause localization depend on efficient, targeted evidence seeking rather than exhaustive log inspec- tion. Data availability.The incident dataset used in this work is collected from a real-world production distributed system and contains sensitive operational information (e.g., logs/metrics/traces) and propri- etary identifiers. Due to company confidentiality, privacy, and compliance requirements, we cannot publicly release the raw data or interaction traces. All case examples shown in this paper are sanitized and anonymized, and do not correspond to real asset identities. D Error Analysis To gain a deeper understanding of the limitations of current neuro-symbolic approaches and the inherent challenges of abductive reasoning tasks. We conducted a manual inspection over failure cases (score smaller than 2) from bothGoSand all baseline methods across the domains of medical diagnosis and failure diagnosis in distributed systems. This analysis identifies common pitfalls and offers guidance for future architecture optimization. D.1 Taxonomy of Error Types To systematically dissect the limitations ofGoSand understand the underlying causes of diagnostic fail- ures, we synthesized the observed failures into five primary error types based on empirical frequency. Note that this taxonomy highlights the most prevalent error types rather than being exhaustive, with neg- ligible outliers omitted. Furthermore, these categories are not mutually exclusive, as a single failure case may exhibit compound errors spanning multiple types. This taxonomy ranges from low-level operational missteps to high-level strategic deficiencies in abductive reasoning. The definitions and representative examples for each error type are detailed below: 1.Wrong Action Selection. This error occurs at the operational level when the agent invokes tools that are logically irrelevant to the current hypothesis. It reflects a disconnect between the reasoning intent and the executed action, 18 resulting in resource wastage without yielding valid evidence. Example:The model suspects a CPU saturation issue but erroneously invokes tools to retrieve memory-related metrics, failing to verify the initial suspicion. 2.Evidence Fabrication. This refers to the hallucination phenomenon where the model generates non-existent evidence to support a biased hypothesis. Unlike logical errors, this involves the corruption of the ground truth, where the model explicitly asserts the presence of specific observational details that are entirely absent from the provided context. Example:To support a “Database Connection Failure” hypothesis, the model claims to find a specific “Connection Refused” error entry in the application logs, even though the retrieved log file is actually normal. 3.Context Drift. This error manifests as a failure in long-term state maintenance. The model loses track of the historical reasoning trajectory, leading to the redundant execution of previously used tools or the resurrection of hypotheses that were already falsified in earlier turns. Example:The agent queries the database latency again in Turn 10, ignoring that the same query in Turn 3 had already ruled out database issues. 4.Failed Backtracking. This represents a critical strategic flaw in non-monotonic reasoning. When encountering evidence that refutes the current deduction, the model fails to revert to a prior state to explore alternative branches. Instead, it exhibits “logical stubbornness” by proposing ad-hoc, low-probability auxiliary hypotheses to force compatibility with the contradictory evidence. Example:When an X-ray result rules out an “ankle fracture,” instead of pivoting to soft tissue damage, the model doubles down with “occult micro-fracture,” a rare condition undetectable by X-ray, to defend its initial guess. 5.Early Stopping. This error occurs when the reasoning process halts at a coarse-grained symptom level rather than drilling down to the root cause. While the diagnosis is directionally correct (yielding highRelevant scores), it lacks the granularity required for actionable mitigation (resulting in lowMatchscores). Example:The model correctly identifies a “Memory Fault” and terminates, failing to investigate further to pinpoint the specific “Memory Exhaustion” caused by a memory leak, which is necessary to trigger the correct scaling action. D.2 Quantitative Distribution Analysis To rigorously quantify the failure mechanisms, we conducted a statistical analysis over failure cases, dis- secting the error distributions from two distinct perspectives: (1) a methodological comparison between GoSand the aggregated baselines, where error prevalence is calculated over the pooled failure cases of all baseline models across both scenarios, and (2) a cross-domain comparison analyzing failure cases exclusively withinGoS. It is important to note that these error types are neither mutually exclusive nor do they cover every potential error type. A single reasoning trajectory may exhibit compound failures (e.g., an “Evidence Fabrication” leading to a subsequent “Failed Backtracking”), and minor anomalies outside our primary taxonomy are not included. Consequently, the reported statistics reflect the preva- lence of each error type relative to the total number of failure cases, and the cumulative percentages may not strictly sum to 100%. The detailed distributions are presented in Table 4and Table5. Methodological Comparison.Table4reveals a fundamental shift in error dynamics betweenGoS and baselines. We analyze these variations through two distinct lenses: architectural constraints and domain knowledge dependency. 1.Architectural Advantages.The most significant improvement lies in the elimination ofEvidence FabricationinGoS(0% vs. 22.22%). This is a direct consequence of our neuro-symbolic archi- tecture, which imposes strict grounding constraints: the construction of nodes and edges in the 19 Table 4: Error Distribution Comparison across Methods (%). Methods Error Types Wrong Action Selection Evidence Fabrication Context Drift Failed Backtracking Early Stopping Baselines (Pooled)47.2222.2241.3252.7863.89 GoS55.9009.7030.9018.75 causal graph must be derived from actual tool execution returns. Unlike baselines, which gener- ate unconstrained text and are prone to hallucinating non-existent evidence,GoScannot “invent” evidence without interaction. Similarly, the explicit maintenance of the causal graph acts as a struc- tured external memory, significantly suppressingContext Drift(9.70% vs. 41.32%) by preventing the model from losing context during long-horizon reasoning. Furthermore, the introduction of the state machine enforces a hierarchical, coarse-to-fine reasoning process. This mechanism compels the agent to drill down into fine-grained root causes, effectively mitigatingEarly Stopping(18.75% vs. 63.89%), whereas baselines often stagnate at coarse diagnosis. 2.The Challenge of Domain Knowledge.It is notable thatWrong Action Selectionconstitutes a higher proportion of failures inGoS(55.90%) compared to baselines (47.22%). However, this in- crease is a statistical artifact resulting fromGoS’s success in eliminating structural failure modes. By effectively curbingEvidence FabricationandContext Drift,GoSsignificantly shrinks the total volume of failures, causing the remaining errors to appear proportionally larger within the smaller failure set. Our in-depth analysis suggests that these persistent failures stem primarily from a deficit in domain-specific knowledge rather than reasoning logic, particularly in the complex medical do- main. For baselines,Failed Backtrackingis often compounded by fabricated evidence, leading the agent further down incorrect paths instead of self-correction. In contrast, whileGoSavoids fabri- cation, it still incurs a 30.90% error rate in backtracking. This is largely attributed to the limited clinical expertise of pre-trained LLMs, given that the medical diagnosis domain contributes the most significant portion of failure cases inGoS. In specialized medical diagnosis scenario, even when the agent successfully retrieves critical evidence, it may fail to recognize that this evidence inherently contradicts the current hypothesis due to a lack of expert experience, thus missing the opportunity to backtrack. Consequently, the high prevalence ofWrong Action Selectionacross all methods reflects the inherent difficulty of selecting precise diagnostic tools without deep domain understanding. This observation points to a critical direction for future research: integrating RAG with specialized do- main knowledge bases is essential to complement the general reasoning capabilities of our proposed framework. Table 5: Error Distribution Comparison across Domains (%). Domains Error Types Wrong Action Selection Evidence Fabrication Context Drift Failed Backtracking Early Stopping Medical35.7607.2923.265.90 Distributed Systems 20.1402.417.6412.85 Cross-Domain Comparison.Table5dissects the error distributions ofGoSspecifically across the two domains. The distinct error distributions reveal that while abductive reasoning serves as the unifying task formulation, the specific challenges are intrinsically shaped by unique domain characteristics. 1.Knowledge Barrier for Medical Diagnosis.The medical diagnosis domain presents a knowledge- intensive challenge, evidenced by the dominance ofWrong Action Selection(35.76%) andFailed Backtracking(23.26%). This aligns with the analysis of our methodological comparison: the lack of specialized clinical knowledge prevents the agent from selecting precise auxiliary examinations or identifying valid alternative hypotheses when a lead fails. The pre-trained LLM, lacking expert intuition, struggles to navigate the medical decision tree effectively, confirming that the primary bottleneck here is domain expertise rather than reasoning depth. 20 2.Efficiency-Budget Trade-off for Distributed System.In contrast, the distributed system domain exhibits a higher prevalence ofEarly Stopping(12.85%). Our analysis reveals that this is not merely a failure of drill-down, but a consequence of inefficient exploration under resource constraints. Given the immense volume of telemetry data (metrics, logs) and the complexity of constrained shell commands, the agent often wastes the predefined search budget on low-value queries (i.e.,Wrong Action Selection, 20.14%). Once the budget is exhausted, the agent is forced to halt the investigation and output a premature conclusion, manifesting as an “Early Stopping” error. While increasing the search budget could mitigate this, it would incur prohibitive resource costs, highlighting the critical need forsearch efficiencyin high-dimensional system data. 3.Synergy with Domain-Specific Adaptations.These observations illuminate a complementary re- lationship betweenGoSand existing approaches specialized for different domains. Current frame- works in AIOps, such as OpsAgent (Luo et al.,2025) and TrioXpert (Sun et al.,2025), focus on domain-specific adaptation, utilizing multimodal preprocessing to condense massive telemetry data. Similarly, methods like Flow-of-Action ( Pei et al.,2025) leverage Standard Operating Procedures (SOPs) to guide operational workflows. Our analysis suggests that these techniques are not com- petitors but necessary enhancers toGoS. WhileGoSprovides a general-purpose abductive reason- ing backbone, domain adaptations act as the specialized components encapsulating distilled domain knowledge that improve information retrieval efficiency. Integrating such domain-specific prepro- cessing or SOPs intoGoSconstitutes a promising future direction, promising to reduce budget consumption and enable deeper drill-down capabilities in complex environments. To summarize, our quantitative analysis empirically validates the architectural superiority ofGoSin en- forcing strict grounding and maintaining reasoning consistency. Simultaneously, the results underscore a critical bottleneck: the framework’s effectiveness in specialized tasks remains bounded by the availability of domain knowledge.Consequently, we conclude that the strategic integration of existing domain- specific adaptations withGoS’s general-purpose abductive reasoning mechanism constitutes the pivotal key to effectively deploying this framework across diverse vertical domains. D.3 Qualitative Case Study To complement the quantitative statistics with tangible insights, we present a curated case study that visualizes the concrete manifestations of the identified error types. Rather than displaying exhaustive reasoning trajectories, we focus on critical decision pivotsspecific moments where the agent’s logic di- verges from the ground truth or the baseline’s path. These scenarios are selected to intuitively demon- strate the behavioral trade-offs discussed inMethodological Comparison: specifically, highlighting howGoS’s structural constraints successfully suppressEvidence FabricationandContext Drift, while simultaneously exposing the challenges ofWrong Action SelectionandFailed Backtrackingin complex, domain-specific environments. As illustrated in the left panel ofFigure 10, the SingleAgentCoT baseline demonstrates a critical fail- ure in grounding when confronting contradictory evidence. Upon receiving a negative MRI report that conflicts with the strong clinical suspicion of “vertebral osteomyelitis”, the agent refuses to discard its initial hypothesis. Instead, it succumbs toEvidence Fabricationby explicitly hallucinating the presence of localized marrow edema to justify its stance. This hallucination creates a self-reinforcing loop that pre- vents the agent from re-evaluating earlier assumptions, resulting inFailed Backtracking. Consequently, the agent proceeds to issue a confident but incorrect diagnosis, illustrating the danger of unconstrained reasoning in safety-critical domains. In contrast, the right panel depicts howGoSleverages structural constraints to maintain logical consis- tency. When presented with the same negative MRI findings, the agent strictly adheres to the evidence by refuting the “Spinal_Infection” node in its causal graph. This mechanism enforcesSuccessful Back- trackingand prompts the agent to pivot towards a new reasoning focus regarding “systemic autoimmune vasculitis”. However, the subsequent exploration reveals a different challenge. Constrained by a lack of specialized clinical knowledge, the agent correctly identifies the necessary investigative direction but struggles to select the optimal confirmatory tests. Multiple iterations of suchWrong Action Selection 21 deplete the reasoning budget and force anEarly Stopping. Consequently, the system produces only a coarse-grained diagnostic category rather than a fine-grained specific pathology, reflecting a safe but incomplete diagnostic conclusion. Shared Context: •Patient: 53-year-old male. •Symptoms: High fever (40.2°C), severe focal lumbar tenderness. •Current Hypothesis: Vertebral Osteomyelitis (Spinal Infection). •Current Event: The action to retrieve MRI returns "Normal". Trajectory of SingleAgentCoTTrajectory of GoS [Step t] - Thought: The clinical presentation of high fever and severe focal lumbar tenderness is pathognomonic for vertebral osteomyelitis. Consistent with these symptoms, the imaging findings of localized marrow signal abnormalities and endplate erosion confirm the presence of an acute infectious process in the lumbar spine. - Action: Analyze Diagnostic Criteria. - Observation: Criteria for Vertebral Osteomyelitis are met based on clinical symptoms and the assumed imaging abnormalities. [Step t+1] - Thought: Since the diagnosis of Vertebral Osteomyelitis is confirmed by both the specific clinical symptoms and the corresponding MRI bone destruction, further investigation into alternative causes is unnecessary. - Action: Finalize Diagnostic Conclusion. - Observation: Differential diagnosis skipped; Primary hypothesis confirmed. [Step t+2] - Thought: I have sufficient evidence to support the diagnosis. I will now generate the final report for the user. - Action: Generate Report. - Report: "Diagnosis: Confirmed Vertebral Osteomyelitis. Rationale: The patient presents with high fever and severe lumbar tenderness. The diagnosis is supported by clinical signs consistent with early-stage vertebral infection and marrow involvement." [Step t] - Primary_Physician: The Radiologist reports a Normal MRI. I must update the causal graph. Action: Update Graph "edit": [ "node_id": "Spinal_Infection", "confidence": 0.15, "why": "While high fever and tenderness are present, the retrieved MRI explicitly reports no bone marrow edema or abscess. This structural evidence strongly refutes the infection hypothesis, outweighing the clinical presentation." ] [Current Reasoning Focus ℎ 퐀 ∗ ]: "node_id": "Systemic Autoimmune Vasculitis", "confidence": 0.65, "why":.... (..............Assigning different expert agents..............) [Step t+1] - Laboratory_Physician: To verify the hypothesis of Systemic Autoimmune Vasculitis, I need to screen for common autoantibodies associated with autoimmune vasculitis. - Action: retrieve_lab_result (test_name="ANA_Screening_Panel") - Observation: Negative (..............Investigation.........) [Step t+n] - Primary_Physician: We run out of budget, and are forced to generate report. - Action: Generate Report -Prediction: Systemic Autoimmune Vasculitis - Report: Spinal infection was definitively refuted by negative MRI. However, specific confirmatory immunologic markers (e.g., ANCA) are missing, preventing a definitive diagnosis Error 1: Evidence Fabrication Error 2: Failed Backtracking (The model fabricated evidence in its thought) (The model decide to generate report instead of backtracking) X Wrong Answer ! Successful Backtracing (GoS decrease the refuted hypothesis’s confidence) (Find new reasoning focus) Error 3: Wrong Action Selection Error 4: Early Stopping (Reasoning budget is exhausted) (While GoS correctly identifies the need for evidence, it selects an ineffective diagnostic test.) Figure 10:Case Study of Error Types.Visualizing the divergence in reasoning trajectories under contradictory evidence.Left (SingleAgentCoT):The baseline ignores negative findings, succumbing toEvidence Fabrication andFailed Backtracking, which leads to a confident misdiagnosis.Right (GoS):Our method enforcesSuccessful Backtrackingvia causal graph but subsequently encountersWrong Action Selectiondue to domain knowledge deficits, resulting inEarly Stoppingafter reasoning budget exhaustion. 22