AIPOM: Agent-Aware Planning & Text Detection

• AIPOM is a dual-use acronym representing both an agent-aware interactive planning system for orchestrated multi-agent workflows and a SemEval system for detecting human-machine text boundaries.
• The planning system leverages a Directed Acyclic Graph and dual-panel interface to enable transparent, editable, and collaborative LLM-driven multi-agent planning.
• The SemEval approach combines a LoRA-tuned decoder with dual encoder sequence taggers to precisely localize the change-point between human and machine text, achieving a leaderboard MAE of 15.94.

Searching arXiv for relevant papers on “AIPOM” and closely related variants. tool call: arxiv_search {"^^^^^^^^8query8 OR AIpom arXiv8^^^^^^^^", "^^^^^^^^8max_results8 ^^^^8AIPOM OR AIpom arXiv8query8^^^^, "^^^^^^^^8sort_by8 "^^^^^^^^8relevance8 Searching again with exact titles to disambiguate the acronym and capitalization. {"^^^^^^^^8query8 OR ^^^^^^^^8ti:\8 "^^^^^^^^8max_results8 ^^^^8AIPOM OR AIpom arXiv8query8^^^^, "^^^^^^^^8sort_by8 "^^^^^^^^8relevance8 AIPOM denotes two distinct systems in recent arXiv literature. In one usage, “AIPOM” refers to Agent-aware Interactive Planning for Multi-Agent Systems, a human-in-the-loop planning environment for orchestrated multi-agent systems that combines conversational and graph-based control of LLM-generated plans (&&&^^^^8query8^^^^&&&). In another, “AIpom” refers to a SemEval-^^^^8max_results8query8max_results8relevance8^^^^ system for detecting boundaries between human-written and machine-generated text in mixed documents, implemented as a two-stage pipeline that combines an instruction-tuned decoder-only model with encoder-only sequence taggers (&&&^^^^8AIPOM OR AIpom arXiv8^^^^&&&). The shared acronym creates a nomenclatural collision across otherwise unrelated research areas: interactive multi-agent planning and AI-generated text detection.

^^^^8AIPOM OR AIpom arXiv8^^^^. Disambiguation and referential scope

The acronym is used for two separate technical artifacts, each defined by a different title, task formulation, and evaluation regime.

Term	Expanded form / title	Domain
AIPOM	“Agent-aware Interactive Planning for Multi-Agent Systems”	LLM-based planning in orchestrated multi-agent systems
AIpom	“AIpom at SemEval-^^^^8max_results8query8max_results8relevance8^^^^ Task 8: Detecting AI-produced Outputs in M^^^^8relevance8^^^^”	Human-machine mixed text detection

In the planning sense, AIPOM is a system intended to improve transparency, controllability, and human oversight in LLM-driven multi-agent workflows by enabling users to inspect, refine, and collaboratively guide generated plans (&&&^^^^8query8^^^^&&&). In the SemEval sense, AIpom is a leaderboard system for Human-Machine Mixed Text Detection on the M^^^^8relevance8^^^^ dataset, where the objective is to localize the boundary between human-authored and machine-generated spans (&&&^^^^8AIPOM OR AIpom arXiv8^^^^&&&).

A common misconception is to treat the acronym as univocal. The literature instead supports a disambiguated reading: the planning system is a UI- and workflow-centric platform for mixed-initiative orchestration, whereas the SemEval system is a hybrid decoder–encoder pipeline for boundary estimation.

^^^^8max_results8^^^^. AIPOM as agent-aware interactive planning

In the multi-agent planning literature, AIPOM is defined as a system for human-in-the-loop planning in orchestrated multi-agent systems. Its central premise is that existing LLM-based approaches often lack effective mechanisms for users to inspect, understand, and control generated plans, motivating a system that supports conversational guidance together with explicit workflow inspection and editing (&&&^^^^8query8^^^^&&&).

The core plan representation is a Directed Acyclic Graph (DAG). Nodes correspond to subtasks assigned to agents and include descriptions, inputs, outputs, and execution status. Edges encode data dependencies, so that output from one agent becomes input to another. This representation is not merely presentational; plan generation, refinement, and execution are all organized around the DAG and an agent registry. The paper’s prompt structuring example expresses plans as JSON with "nodes" and "edges", including node ids, agent names, task descriptions, input/output fields, and source–destination mappings between outputs and inputs.

The system’s planning is explicitly agent-aware. The planning module decomposes user goals into plans based on agent capabilities and input/output specifications from a registry. This shifts plan synthesis away from unconstrained natural-language decomposition toward structured assignment of subtasks, explicit interface matching, and traceable dataflow. A plausible implication is that the DAG formalism functions as a verification substrate as much as a planning substrate, because it renders agent choice and dependency structure inspectable before and after execution.

^^^^8sort_by8^^^^. Architecture and interaction model of the planning system

AIPOM’s architecture is organized into four modules: a Planning Module, a Conversation Module, an Execution Coordinator, and a Controller (&&&^^^^8query8^^^^&&&). The planning module uses LLMs to construct agent-aware plans. The conversation module parses user utterances for plan creation or refinement. The execution coordinator executes subtasks across agents and monitors outputs at every step. The controller synchronizes communication among these components.

Its interface is explicitly dual-panel. The conversational interface supports natural-language commands, explanations, and high-level feedback, with queries, system explanations, and plan updates shown as chat bubbles with execution trace. The graph-based interface renders the current plan as an editable node-link DAG and supports direct manipulation: adding or removing nodes and edges, editing agent assignments and task properties, adjusting input/output specifications, rearranging graph layout, and executing whole plans or specific segments.

Several mechanisms operationalize transparency and control. Plan graph visualization makes data flows and agent assignments explicit. Node cards expose subtask details, assignments, and intermediate or final outputs. Execution inspection shows node outputs after execution. Edit history and feedback loops surface refinement actions in both graph and chat views. The system also includes LLM assistance on demand, described as a help button that can “fix” or complete a partially edited or underspecified plan. This mixed-initiative design allows control to alternate between natural-language steering and direct graph manipulation without collapsing one into the other.

The paper’s job-search example illustrates this interaction pattern. A user begins with a broad request and receives a three-step plan—web search, job extraction, and summarization—then inspects outputs in node cards, adjusts search settings, edits extraction fields, and adds conversational feedback such as filtering by location. The same refinement is reflected in both chat and plan views, preserving cross-modal consistency between language-level intent and graph-level execution structure.

^^^^8relevance8^^^^. Empirical findings, usability claims, and limitations of the planning system

The empirical evaluation combines quantitative plan-refinement experiments with a pilot user study (&&&^^^^8query8^^^^&&&). The quantitative experiment uses GSM8K and Multi-Step Arithmetic under three feedback conditions: detailed natural language, vague natural language, and partial direct manipulation plus LLM fix (DM+Fix). The reported finding is that DM+Fix outperformed vague NL feedback and matched detailed NL feedback for complex plan modifications while reducing user effort. For simple local changes, direct manipulation alone was sufficient. The unreliability of vague feedback is presented as evidence of the limits of conversational interfaces when structural visibility is absent or weak.

The pilot study compares graph representation vs. text and direct manipulation vs. natural-language feedback. Users inspected, understood, and edited plans more correctly and quickly with graph visualization than with textual plans. The study also reports that users more often missed errors in textual plans, whereas graph-based presentation yielded higher accuracy and less false acceptance. In refinement, direct manipulation was strongly preferred and judged more intuitive, less mentally demanding, and “more suitable for detailed editing.”

These results are used to support claims about plan correctness, debuggability, and user trust. The paper’s interpretation is that explicit visualization, iterative correction, and hybrid feedback mechanisms improve oversight relative to black-box or chat-only planning tools. This suggests a shift from prompt-centric interaction toward workflow-centric interaction, where transparency is implemented through inspectable intermediate structure rather than post hoc explanation alone.

The limitations discussed by the authors are structurally important. Complexity scaling remains difficult for intricate multi-step workflows. The system still depends on plan structure correctness, especially with respect to input-output mappings and agent assignments. The paper also notes data privacy and responsible use concerns, including the possibility that human involvement may introduce privacy risks or biased planning. Future directions include application to healthcare and finance, more sophisticated graph edits such as freezing, merging, and constraint enforcement, stronger LLM verification of plans, and expanded user studies in real-world settings.

^^^^8query8^^^^. AIpom as a system for human-machine mixed text detection

In SemEval-^^^^8max_results8query8max_results8relevance8^^^^ Task 8, Subtask C, AIpom is defined as a system for detecting the boundary between human-written and machine-generated text in the multi-generator, multi-domain, multi-lingual M^^^^8relevance8^^^^ dataset (&&&^^^^8AIPOM OR AIpom arXiv8^^^^&&&). The task is formulated as boundary localization rather than document-level classification, so the key prediction target is a change-point separating human and machine segments.

The system uses a two-stage pipeline. The first stage is an instruction-tuned decoder-only LLM, specifically Mistral-^^^^8 OR ti:\8^^^^B-OpenOrca, fine-tuned with LoRA (Low-Rank Adaptation). The decoder is trained to output only the machine-generated part of the input text according to a prompt that requires the output to begin with "Answer: ", separates tokens by spaces, and outputs "None" if the whole text is human-written. Post-processing identifies the boundary by finding where the decoder begins outputting the machine-generated portion in the original text, and a special <BREAK> token is inserted at that point.

The second stage consists of two encoder-only sequence taggers, both based on DeBERTaV^^^^8sort_by8^^^^-Large in a classic sequence labeling setup. Tokens in the human-written segment receive label ^^^^8query8^^^^ and tokens in the machine-generated segment receive label ^^^^8AIPOM OR AIpom arXiv8^^^^. The first encoder is trained on texts marked up by the decoder with <BREAK>. The second encoder is trained on the union of original training data and texts labeled by the decoder. At inference time, the position of the first token predicted as ^^^^8AIPOM OR AIpom arXiv8^^^^ is taken as the boundary estimate.

Final prediction is obtained by averaging the two encoder boundary indices:

PRESERVED_PLACEHOLDER_^^^^8query8^^^^

The system also uses two-fold cross-labeling: the training set is split in two, and each half is labeled with a decoder fine-tuned on the other half. This design is presented as a robustness measure intended to limit overfitting while exploiting decoder-generated supervision.

^^^^8ti:\8^^^^. Evaluation, ablations, and technical significance of AIpom

AIpom was ranked second on the leaderboard in Subtask C, placing ^^^^8max_results8^^^^nd out of ^^^^8sort_by8sort_by8^^^^ teams, with a Mean Absolute Error (MAE) of ^^^^8AIPOM OR AIpom arXiv8query8^^^^.9^^^^8relevance8^^^^ on the test set (&&&^^^^8AIPOM OR AIpom arXiv8^^^^&&&). The evaluation metric is

PRESERVED_PLACEHOLDER_^^^^8AIPOM OR AIpom arXiv8^^^^

where lower MAE indicates more accurate boundary localization.

The reported results differentiate sharply between model configurations. The LoRA Mistral decoder attains Dev MAE ^^^^8max_results8^^^^.^^^^8relevance8AIPOM OR AIpom arXiv8^^^^ and Test MAE ^^^^8AIPOM OR AIpom arXiv8 OR ti:\8^^^^.^^^^8query8query8^^^^. Zero-shot Mistral performs much worse, with Dev MAE ^^^^8query8ti:\8^^^^.^^^^8query8AIPOM OR AIpom arXiv8^^^^ and Test MAE 8^^^^8query8^^^^.8^^^^8AIPOM OR AIpom arXiv8^^^^. A DeBERTa on training set baseline yields Dev MAE ^^^^8max_results8^^^^.^^^^8AIPOM OR AIpom arXiv8query8^^^^ and Test MAE ^^^^8AIPOM OR AIpom arXiv8^^^^9.9^^^^8 OR ti:\8^^^^. The submitted AIpom ensemble reaches Dev MAE ^^^^8AIPOM OR AIpom arXiv8^^^^.^^^^8ti:\88^^^^ and Test MAE ^^^^8AIPOM OR AIpom arXiv8query8^^^^.9^^^^8relevance8^^^^, the best reported score among the submitted variants in the summary. The paper also notes that additional fine-tuning led to a lower MAE of ^^^^8AIPOM OR AIpom arXiv8query8^^^^.^^^^8max_results8AIPOM OR AIpom arXiv8^^^^ than the submitted system.

The ablation study attributes the gains to the pipelined encoder–decoder design. The decoder alone improves substantially after LoRA tuning relative to zero-shot inference, but the encoders trained on decoder predictions perform better than the decoder alone. Encoders trained only on original human labels perform worse than encoders trained on decoder-labeled data. The best performance arises from integrating the decoder’s change-point predictions into encoder inputs and averaging the outputs of two encoders. The paper interprets this as evidence that the task has a generative aspect—identifying the onset of machine text—but also benefits from the token-level precision of discriminative sequence labeling.

The broader significance of AIpom lies in demonstrating model synergy between instruction-tuned decoder-only and encoder-only architectures for boundary detection in mixed-authorship text. The reported performance drop from development to test set is described as evidence of domain shift and motivates future work on out-of-domain robustness. Within the scope of the SemEval task, AIpom is presented as a high-performing hybrid approach whose gains derive from staged prediction, decoder-assisted relabeling, and simple averaging rather than from a single monolithic model.