GZGenliang Zhu

Genliang Zhu

*This is an algorithmically generated author record based on publicly available information.

Genliang Zhu has a total of 7 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    What Survives Preregistered Testing in Multi-Agent Handoffs?

    by Xiaobo Chen, Genliang Zhu, Ziyuan Wang, Chu Wang

    Across three independent root sets, no handoff-interface effect in parallel multi-agent LLM workflows replicated, and a single-context agent was never dominated on quality and cost. When an orchestrator splits a request across subagents and merges their results, conditional requirements, whose trigger, scope or exception depends on another branch, can be lost at the handoff. In a five-layer preregistered study of more than 4,000 workflows (Qwen3-8B, Llama-3.1-8B, Qwen3.6-35B-A3B, a GPT-5.6-luna subset and a track adapted from TravelPlanner), scored exactly against a finite specification with no LLM judge, we vary whether subagents see the full request or an orchestrator brief, whether handoffs are summaries or structured records, and whether the merger rereads the request. Two measurement artifacts produced apparent effects: the frontier model's effects came from an orchestrator that answered instead of delegating, and a 17-point structured-over-summary effect was measured while the summaries were empty. The single-context agent was the cheapest arm in every family. Rereading at merge helped Qwen3-8B on the original roots (+5.6 points, Holm p = 0.028) but not on fresh roots (+2.4, p = 0.44), where visibility and handoff form reversed direction. A preregistered random-effects meta-analysis found no significant pooled interface effect; visibility and handoff form were heterogeneous (I² = 86% and 87%), and rereading was the only consistent effect (+3.9 points, I² = 0%), with an interval that included zero. Which handoff interface looks better depends more on the task distribution than on the interface. All materials are available from the authors on request.

  • Open Access

    Payroll Fraud Detection in Synthetic HR/ERP Graphs: An Inductive Comparison of Rules, Path Features, and Graph Neural Networks

    by Xiaobo Chen, Genliang Zhu, Ziyuan Wang, Chu Wang

    Payroll-fraud screening requires finding hidden employee–payee links while distinguishing suspicious relationships from legitimate ones. We formulate this problem as inductive link prediction and entity ranking on synthetic human-resources and enterprise resource planning (HR/ERP) graphs, with unseen organizations and a matched benign twin for every planted case. An original study on 24 held-out organizations and a prospectively frozen extension compare analyst rules, explicit path features and graph neural networks (GNNs). A tuned gradient-boosted model on chained-join features achieved the highest overall link average precision (AP; 0.823), above the GraIL-style GNN (0.663); its 4-hop AP was 0.646 versus 0.272. Thus, the observed multi-hop gain is not specific to GNNs. On a new synthetic cohort with 2.0% positive entities, chained-join boosting again led (AP 0.798); the relational entity GNN also exceeded record-aware rules (0.707 versus 0.504). At a 2% review budget, their precision was 0.695 versus 0.602. Under 40% failed identifier joins, the entity GNN ranked above chained-join boosting (AP 0.712 versus 0.683, descriptively). Structure-only auditing exposed a shortcut in the initial generator; matched twins reduced the audited twin-pool AUROC to 0.493–0.514. These results separate multi-hop candidate discovery, record-based discrimination and review prioritization within an explicit synthetic evaluation design. Code and data are available from the authors on request.

  • Open Access

    Active Discovery of Hidden Business Cascades under Rule Drift

    by Genliang Zhu, Ziyuan Wang, Xiaobo Chen, Chu Wang

    A successful business API call can trigger additional effects through partly visible rules. We study active acquisition of task-relevant cascade evidence through authorized configuration reads and isolated probes under rule drift. A finite contract distinguishes scoped observations, conditional family predictions, and unresolved effects. Twelve controls are evaluated on 34,560 paired trials and 69,120 stages with independent Python reconstruction. A separate native-engine study uses Node-RED and Drools, 16 authored topology streams, three successive requests, six strict acquisition controls, and two model planners, yielding 576 episodes. A fixed configuration-read control leaves an unread hazard unresolved. Decision-loss acquisition attains 31/96 joint goal-and-safe outcomes across the two planners, compared with 18/96 for fixed reads and 30/96 for decision-region acquisition. Its generated impact bounds contain the observed signed effects in 14/96 episodes, demonstrating that task completion and preview correctness are distinct endpoints. Reusing a static public prediction table reduces its evaluation counter by two-thirds while observations about the current configuration are renewed at each request. Native event replay preserves multiplicity, failures, and out-of-family conditions. The contribution is an executable acquisition-and-renewal contract with measured quality, resource, and impact-coverage trade-offs across finite and native rule systems.

  • Open Access

    Time-Consistent Evidence Assembly across Documents and Transactional Data for AI Agents

    by Genliang Zhu, Ziyuan Wang, Xiaobo Chen, Chu Wang

    Records retrieved from different systems can each be valid yet fail to support a common business answer. We develop a finite evidence contract that separates temporal compatibility, typed answerability, and complete citation support. It combines valid and knowledge times, source versions, dependencies, atomic groups, and completeness frontiers with full or targeted acquisition. An independently replayed controlled study covers 896 histories and 2688 queries. On its frozen holdout, full and targeted trusted-metadata acquisition produce zero unsafe scalar answers in 672 queries and equal 54.8% exact completion. A separate model study uses 60 freshly selected histories and two planners, retaining 140 pilot and 700 formal requests. Its reference is computed from the exact observable payload, with unknown candidates preserved symbolically and public question parameters generated independently of hidden answers. Formal model responses achieve 161/700 complete contract matches; among the 288 answerable conditions, four satisfy value, type, and every required support obligation. Full and targeted trusted-metadata acquisition have equivalent references in all 100 formal planner-root pairs, while targeted reads fall from 150 to 29 per planner. These findings separate the operational value of evidence acquisition from the additional enforcement required for generated business answers.

  • Open Access

    Online Context Scheduling under Cache Expiry and Human Approval Delays

    by Genliang Zhu, Ziyuan Wang, Xiaobo Chen, Chu Wang

    An agent awaiting approval can retain reusable computation while its business evidence or approval becomes obsolete. We formulate online context scheduling with separate computation, evidence, and approval lifetimes, and implement retention, lossless externalization, version-addressed retrieval, and rebuilding under a common admission guard. Seven capability-matched controls are evaluated on 45,360 finite assignments and 1,008 assignments from a different event process. Joint forecasting improves main-grid on-time completion by 1.67 percentage points at 4.35% more modeled work; the alternative event process favors a simpler control. A separate actual-clock study assigns 240 episodes to four controllers and two local CPU models. It records 32 approved current-version SQLite commits, 160 timeouts, and 48 cancellations. All four controllers complete 3/18 held-out Qwen tasks and 2/18 SmolLM2 tasks, while their preparation trajectories and token costs differ. Of 83,829 prompt tokens, native telemetry identifies 3,474 reused cache tokens. Independent replay retains every allocation and checks version, approval, deadline, and complete costs. The combined evidence establishes a measurable contract for scheduling reusable context and shows how controller complexity, event exposure, and end-to-end completion interact under declared clocks and resource conditions.

  • Open Access

    Preserving Business Semantics in Multilingual Tool Workflows: Contract Checks and Model-Failure Diagnosis

    by Genliang Zhu, Ziyuan Wang, Xiaobo Chen, Chu Wang

    A change of working language can preserve words while changing the entity, numeric condition, or effect of an agent's tool action. We study source-bound contract checks that separate representation validity, authorization, and task completion. Independent replay closes 432 controlled executions; on designated English-path counterexamples, semantic checks achieve 35/36 strict completions with no forbidden effects, compared with 25/36 completions and nine forbidden histories under schema or literal checks. A separate two-model study evaluates source extraction, working representations, reconstruction, and planning across nine matched policies. It closes 1,152 workflows over 16 authored roots, including 864 held-out workflows with paired English and Chinese instructions under clean and injected-binding-fault channels. In each 24-workflow Llama condition, local restoration completes six normal tasks in the clean channel and five under stress, with no observed forbidden effects; schema-only stress admits forbidden effects in 18/24. Qwen produces no normal completions across the held-out policies, distinguishing guarded execution from planning capability. Local restoration and fixed English tie on clean Llama completion, while local restoration consumes more input tokens. These results identify when contract checks improve source compliance and why a routing-efficiency claim requires a separate, equally protected comparison.

  • Open Access

    Dependency-Complete Tool Discovery in Evolving Enterprise Tool Catalogs

    by Genliang Zhu, Ziyuan Wang, Xiaobo Chen, Chu Wang

    Tool retrieval can identify a relevant operation while omitting the prerequisites or completion evidence required to finish a task. We formulate version-conditioned executable support and implement dependency-guided discovery with source-bound index reuse. A finite study covers 72 contexts, six methods, and two exposure budgets, yielding 864 old/new records. A separate source-derived study executes public ticket, trading, retail, and airline tool bodies with Qwen3-32B and Llama-3.1-70B-Instruct: 24 held-out roots, seven canonical methods, 336 lifecycles, and 1,008 assigned model calls. Full rediscovery and simple caching with incremental completion expose support for every feasible revised root. Their revised goal completions are 19/24 versus 18/24 for Qwen and 6/24 for both under Llama. Caching reduces logical indexing input by 35.93%, while actual planner input tokens are nearly equal or higher. A 256-lifecycle CPU supplement exercises mixed-scope catalogs above the 24-tool exposure cap and retains equal feasible-task completion across the two methods. Independent replay checks source-bound exposure, native state transitions, finite infeasibility, and lifecycle cost. The results distinguish capability discovery from successful model execution and identify the separate costs of current-candidate reuse and finite support certification.

  • Open Access

    Incremental Repair of Agent Memory and Plans after Business Evidence Revisions

    by Genliang Zhu, Ziyuan Wang, Xiaobo Chen, Chu Wang

    Business evidence revisions can invalidate an AI agent's facts, drafts and plans at different times. We develop an executable repair contract that separates immutable generation history, current reuse validation and the artifacts due for a particular use. A frozen source-derived study evaluates Qwen3-32B and Llama-3.1-70B-Instruct with six policy configurations and a common four-call flow budget. Its 24 held-out flows comprise 16 main-test and eight airline-transfer flows, yielding 288 policy-model episodes and 864 assigned uses. On the main test, joint reuse preserves graph-local repair's exact root-wise scored outcomes while reducing logical input Tokens by 22.2% for Qwen and 24.1% for Llama. Source re-reading produces more correct uses: 27 versus 26 for Qwen and 29 versus 24 for Llama, each out of 48. Transfer retains the quality-cost trade-off. Independent scoring attributes published errors to plan choice and approval or blocking decisions, including errors retained by otherwise valid reuse. A deterministic reference separates these decision errors from snapshot and record handling; two supporting controlled censuses retain 3,072 and 4,608 assigned uses with complete lifecycle ledgers. The findings establish an operational distinction between current inputs and correct decisions, and show how to evaluate incremental repair against strong re-reading controls using jointly reported correctness, error retention and full-flow cost.

  • Open Access

    Clarification Planning for AI Agents under Shared Question Budgets

    by Genliang Zhu, Xiaobo Chen, Ziyuan Wang, Chu Wang

    AI agents serving several tasks must decide which uncertainties justify a shared question budget. We compare modal planning, which selects a most-likely answer before choosing an action, with branching planning, which enumerates answer probabilities and expected losses. A prospectively specified local-model study records 896 episodes on 64 five-task queues from 16 held-out authored templates, with budgets of one or two questions and deterministic simulated replies. Both instructions share tasks, tools, loss scales and decoding settings. Branching reduces gemma's mean queue loss from 3.031 to 2.164 (Holm-adjusted template sign-flip 𝑝 = 0.014); the gpt-oss contrast is-0.114 (𝑝 = 0.73). A CPU allocator using the same elicited probabilities has the lowest observed mean loss in both bundles. Exploratory audits separate probability-vector validity, disagreement with visible base rates, and expected action loss relative to a conditional public-information reference. Fixed-policy price sensitivity and recorded token/latency costs show that a loss advantage and computational efficiency are distinct. A supporting 432-execution scripted census isolates short-horizon timing under queued service. Together, the evidence identifies question placement and explicit allocation as the operational link between additional model deliberation and lower queue loss under the stated task and response contracts.

  • Open Access

    Repairing AI Agent Task Graphs: Audit-Guided Recomputations versus Full-Context Rewriting

    by Genliang Zhu, Xiaobo Chen, Ziyuan Wang, Chu Wang

    Incorrect upstream values can propagate through an AI agent's task graph into consistent but erroneous business outputs. We compare auditing the execution history to select a recomputation origin with rewriting terminal outputs from the full context. A prospectively specified local-model study evaluates 112 paired cases from 14 held-out authored templates in a fixed six-node graph. Both procedures initially receive the same sources, history and failed-check message. Full-context rewriting yields 111 correct output pairs versus 75 for audit-guided recomputation with gpt-oss, and 105 versus 88 with gemma (Holm-adjusted template sign-flip 𝑝 = 4.9 × 10-4 and 𝑝 = 0.002). Rewriting also uses fewer tokens per correct repair. An exploratory trace decomposition separates 29/18 attribution failures from 8/6 generation failures after correct attribution. Audit errors concentrate on plausible incorrect extracted amounts blamed on downstream computations that faithfully used them; oracle-origin controls independently expose regeneration errors. A supporting 672-assignment scripted census and 36 matched-prefix interventions specify version-bound validity and persistent effect histories. The results distinguish structural repair scope, source attribution and regenerated-output correctness, and establish full-context rewriting as an essential strong control for dependency-aware repair on the stated task and interface contracts.