arxiv: 2603.02070 · v2 · submitted 2026-03-02 · 💻 cs.AI · cs.CL· cs.HC· cs.MA

Recognition: no theorem link

Exploring Plan Space through Conversation: An Agentic Framework for LLM-Mediated Explanations in Planning

Guilhem Fouilh\'e , Rebecca Eifler , Antonin Poch\'e , Sylvie Thi\'ebaux , Nicholas Asher

Authors on Pith no claims yet

Pith reviewed 2026-05-15 17:50 UTC · model grok-4.3

classification 💻 cs.AI cs.CLcs.HCcs.MA

keywords AI planningexplanationslarge language modelsmulti-agent systemsinteractive interfacesgoal conflictsuser studiessequential decision making

0 comments

The pith

A multi-agent LLM architecture generates interactive explanations for AI plans that adapt to user questions and context.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces a framework where multiple LLM agents work together to produce explanations for plans generated by automated planners. The system remains independent of any specific explanation method and supports ongoing dialogue so users can ask follow-up questions shaped by their own goals. An implementation focused on goal-conflict explanations was tested against a standard template interface in a user study. If the approach holds, it would let humans steer planners more effectively through natural conversation instead of rigid output formats.

Core claim

The authors present a multi-agent Large Language Model architecture that is agnostic to the explanation framework and enables user- and context-dependent interactive explanations. They instantiate the framework for goal-conflict explanations and evaluate it through a user study that compares the LLM-mediated dialogue against a baseline template-based interface.

What carries the argument

multi-agent LLM architecture that mediates between the user, the planner, and explanation logic to produce context-aware responses

If this is right

The same agent structure can be reused for other explanation types such as plan repair or preference elicitation without rebuilding the core system.
Users receive explanations that evolve with their questions rather than fixed templates, supporting iterative refinement of plans.
The framework separates the explanation layer from the planner itself, allowing existing planners to gain conversational interfaces.
Human preferences and expertise can be incorporated more directly into the planning loop through dialogue.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach could be extended to domains like robotics or scheduling where plan revisions must respond to changing human constraints.
Adding lightweight verification steps between agents might address reliability concerns while preserving the conversational flow.
Longer-term use might reveal patterns in how users discover new preferences through interaction that static interfaces miss.

Load-bearing premise

LLM agents can reliably produce accurate, non-hallucinated explanations that correctly reflect the underlying planner's logic and user intent without additional verification steps.

What would settle it

A controlled test set of plans with known goal conflicts where LLM explanations are checked against ground-truth planner traces and user intent logs, and the explanations deviate in more than a small fraction of cases.

Figures

Figures reproduced from arXiv: 2603.02070 by Antonin Poch\'e, Guilhem Fouilh\'e, Nicholas Asher, Rebecca Eifler, Sylvie Thi\'ebaux.

**Figure 2.** Figure 2: Communication protocol between the user, the translators and the explanation framework. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: Comparison of maximum utility achieved over iteration steps between G LLM and G TPL. Scores are averaged by group at each iteration step. S-CAN S-HOW S-WHAT-IF S-WHY-NOT US-HOW US-WHY DIRECT FOLLOW-UP 0 20 40 60 Average Percentage per User (%) LLM-based Template-based [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

**Figure 5.** Figure 5: Feedback questionnaire results comparison for the two groups: G TPL and G LLM. Higher scores indicate stronger agreement. Violin plots show the distribution and mean score of the answers on a Likert scale. LLM-based significantly outperforms template-based in helping participants to improve their plans and reduce difficulty. 1 2 3 4 5 6 7 Likert scale score How easy was the task to you? (p-value: 0.26)… view at source ↗

**Figure 6.** Figure 6: Histogram of the number of turns in conversations from [PITH_FULL_IMAGE:figures/full_fig_p013_6.png] view at source ↗

**Figure 7.** Figure 7: Dialogue acts (left) and explanation moves (right) distribution in [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗

**Figure 8.** Figure 8: Architecture of the web platform The system has a modular architecture. Explanations and plans are provided by individual services that in the future can be extended with explanations and plans based on different approaches. The back-end and the explanation and planning services implement REST APIs. The communication between the back-end and a service is asynchronous. This means the service notifies the ba… view at source ↗

**Figure 9.** Figure 9: Screenshot of the goal creation using templates baseline (left) and LLM-based goal translator [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗

**Figure 9.** Figure 9: The LLM infers the delivery location from the planning task. The multi-step interaction to [PITH_FULL_IMAGE:figures/full_fig_p019_9.png] view at source ↗

**Figure 10.** Figure 10: Screenshot of an open goal panel. Goals can be locked (indicated by a small lock as shown in the top right in [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗

**Figure 11.** Figure 11: Screenshot of the iteration steps overview left with the interface to create new steps on the [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗

**Figure 12.** Figure 12: Screenshot of the details view of a solvable iteration step (left) and an unsolvable step (right). [PITH_FULL_IMAGE:figures/full_fig_p021_12.png] view at source ↗

**Figure 13.** Figure 13: Screenshot of the template-based explanations for a solvable step (left) and an unsolvable [PITH_FULL_IMAGE:figures/full_fig_p021_13.png] view at source ↗

**Figure 14.** Figure 14: Screenshot of the LLM-based explanations. [PITH_FULL_IMAGE:figures/full_fig_p022_14.png] view at source ↗

**Figure 15.** Figure 15: Iteration steps list for a demo during the user study. [PITH_FULL_IMAGE:figures/full_fig_p023_15.png] view at source ↗

**Figure 16.** Figure 16: Comparison of max utility achieved over time between the two groups: [PITH_FULL_IMAGE:figures/full_fig_p025_16.png] view at source ↗

**Figure 18.** Figure 18: Dialogue acts in conversations from [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗

**Figure 19.** Figure 19: Explanation moves in conversations from G LLM group. C Examples LLM Interaction In this section, we show examples of the LLM conversations as seen by the user. These are not examples from the user study, but they use the same instance and domain [PITH_FULL_IMAGE:figures/full_fig_p027_19.png] view at source ↗

**Figure 20.** Figure 20: Example of a conversation in a solvable iteration step with the LLM-based interface [PITH_FULL_IMAGE:figures/full_fig_p028_20.png] view at source ↗

**Figure 21.** Figure 21: Here the LLM-based version gives information much more efficiently [PITH_FULL_IMAGE:figures/full_fig_p029_21.png] view at source ↗

read the original abstract

When automating plan generation for a real-world sequential decision problem, the goal is often not to replace the human planner, but to facilitate an iterative reasoning and elicitation process, where the human's role is to guide the AI planner according to their preferences and expertise. In this context, explanations that respond to users' questions are crucial to improve their understanding of potential solutions and increase their trust in the system. To enable natural interaction with such a system, we present a multi-agent Large Language Model (LLM) architecture that is agnostic to the explanation framework and enables user- and context-dependent interactive explanations. We also describe an instantiation of this framework for goal-conflict explanations, which we use to conduct a user study comparing the LLM-powered interaction with a baseline template-based explanation interface.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a clean multi-agent LLM wrapper for turning planning explanations into conversations, but the user study results and any checks against planner traces are missing from what is shown.

read the letter

The main thing to know is that the authors describe a multi-agent LLM setup that sits on top of existing explanation methods and lets users ask follow-up questions in natural language. They keep the core planner and explanation logic untouched, which is a sensible separation, and they instantiate the idea for goal-conflict explanations before running a user study against a template baseline. That modularity is the clearest practical contribution here and could be reused with other planners without much rework. The study itself is the right move for this kind of interface work, since it tries to measure whether the conversational version actually helps users understand plans better than static text. The architecture description is straightforward and avoids overclaiming new planning algorithms, which keeps the scope honest. The soft spots are straightforward too. The abstract and available description give no quantitative outcomes from the study, no methodology details, and no error analysis, so the claim that the system enables effective interactive explanations rests on evidence that is not yet visible. There is also no mention of any grounding step that would verify the LLM outputs against the actual planner traces or causal structure. If the agents produce plausible but inaccurate explanations, the interactivity benefit collapses, and that risk is left unaddressed. This paper is aimed at people already working on explainable AI for automated planning and human-in-the-loop decision support. A reader who needs new theoretical results or stronger planning algorithms will not find them, but someone building or evaluating explanation interfaces could adapt the multi-agent pattern without starting from scratch. The thinking is coherent on its own terms and engages the relevant literature on explanations in planning. It deserves a serious referee who can request the full study numbers, the exact comparison metrics, and some form of verification against the planner state. I would send it to peer review with those requests rather than desk reject.

Referee Report

3 major / 2 minor

Summary. The paper proposes a multi-agent LLM architecture for generating interactive, user- and context-dependent explanations in automated planning. It claims the architecture is agnostic to any specific explanation framework, instantiates it for goal-conflict explanations, and evaluates it via a user study against a template-based baseline interface.

Significance. If the empirical support holds, the work could advance human-AI collaboration in planning by enabling natural-language, adaptive explanations that incorporate user preferences, potentially increasing trust and iterative refinement in real-world sequential decision systems. The multi-agent design offers flexibility that rigid templates lack.

major comments (3)

[User Study] User Study section: The manuscript references a user study comparing LLM-powered interaction to a template-based baseline but provides no quantitative results, methodology details (e.g., participant count, tasks, measures), statistical analysis, or error rates. This leaves the central claim of effective interactive explanations without load-bearing evidence.
[Framework Architecture] Framework Architecture section: The assertion that the multi-agent architecture is agnostic to the explanation framework is stated but demonstrated only through a single instantiation for goal-conflict explanations. No additional cases or general argument is given to establish agnosticism.
[Explanation Generation] Explanation Generation: No verification step, planner-trace alignment check, or hallucination filter is described to ensure LLM outputs faithfully encode the underlying planner's logic rather than plausible approximations. This directly risks undermining the reliability of user- and context-dependent explanations.

minor comments (2)

[Abstract] Abstract: Key quantitative or qualitative findings from the user study should be summarized to give readers a concrete sense of outcomes.
[Framework Architecture] Notation: Clarify how agent roles and message passing are defined to improve reproducibility of the multi-agent setup.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and will revise the manuscript to strengthen the presentation of the user study, clarify the framework's generality, and improve safeguards for explanation fidelity.

read point-by-point responses

Referee: [User Study] User Study section: The manuscript references a user study comparing LLM-powered interaction to a template-based baseline but provides no quantitative results, methodology details (e.g., participant count, tasks, measures), statistical analysis, or error rates. This leaves the central claim of effective interactive explanations without load-bearing evidence.

Authors: We agree that the User Study section requires substantially more detail to support the central claims. We will revise this section to include the full methodology (participant recruitment, count, demographics, tasks, experimental design, and measures), quantitative results with statistical analysis, and any observed error rates or qualitative findings. This will provide the necessary load-bearing evidence. revision: yes
Referee: [Framework Architecture] Framework Architecture section: The assertion that the multi-agent architecture is agnostic to the explanation framework is stated but demonstrated only through a single instantiation for goal-conflict explanations. No additional cases or general argument is given to establish agnosticism.

Authors: The architecture separates concerns across agents (planner interface, explanation generator, user modeler, and dialogue manager) so that only the explanation generator needs to be swapped for a different framework. While the paper focuses on goal-conflict explanations, we will add an explicit general argument in the Framework Architecture section describing how the modular interfaces support other explanation types (e.g., plan repair or preference elicitation) without altering the overall structure. revision: partial
Referee: [Explanation Generation] Explanation Generation: No verification step, planner-trace alignment check, or hallucination filter is described to ensure LLM outputs faithfully encode the underlying planner's logic rather than plausible approximations. This directly risks undermining the reliability of user- and context-dependent explanations.

Authors: We acknowledge this limitation. The current implementation grounds the explanation agent via prompts that include the raw planner trace, but no explicit verification or hallucination filter is described. We will revise the Explanation Generation section to document the prompt-based grounding techniques used and will add a lightweight alignment check (e.g., keyword or entity matching against the trace) to mitigate the risk of plausible but unfaithful outputs. revision: yes

Circularity Check

0 steps flagged

No circularity: independent architectural proposal

full rationale

The paper presents a descriptive multi-agent LLM framework for interactive planning explanations, instantiated via a user study on goal-conflict cases. No equations, fitted parameters, predictions, or derivations appear in the provided text. The agnosticism claim and interactivity features are introduced as design choices rather than results derived from prior steps or self-citations. The architecture does not reduce to its own inputs by construction, and the user study serves as external evaluation rather than a self-referential loop. This is a standard non-circular framework paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The framework rests on the unproven assumption that LLMs can serve as reliable mediators for planning explanations; no free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption LLMs can generate accurate and context-appropriate explanations for planning problems without systematic errors or hallucinations
Required for the interactive explanations to be trustworthy; invoked implicitly when claiming the architecture enables useful user- and context-dependent interaction.

pith-pipeline@v0.9.0 · 5455 in / 1082 out tokens · 49009 ms · 2026-05-15T17:50:20.483941+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

53 extracted references · 53 canonical work pages · 2 internal anchors

[1]

Oxford University Press (1980)

Achinstein, P.: The Nature of Explanation. Oxford University Press (1980)

work page 1980
[2]

In: Calzolari, N., Kan, M.Y., Hoste, V., Lenci, A., Sakti, S., Xue, N

Alshomary, M., Lange, F., Booshehri, M., Sengupta, M., Cimiano, P., Wachsmuth, H.: Modeling the quality of dialogical explanations. In: Calzolari, N., Kan, M.Y., Hoste, V., Lenci, A., Sakti, S., Xue, N. (eds.) Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). pp. 1152...

work page 2024
[3]

In: Butler, R

Bromberger, S.: An approach to explanation. In: Butler, R. (ed.) Analytical Philsophy. Oxford University Press (1962)

work page 1962
[4]

In: Proc

Brunello, A., Montanari, A., Reynolds, M.: Synthesis of ltl formulas from natural language texts: State of the art and research directions. In: Proc. TIME. Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik (2019)

work page 2019
[5]

In: IJCAI (2017)

Chakraborti,T.,Sreedharan,S.,Zhang,Y.,Kambhampati,S.:Planexplanationsasmodelreconciliation: Moving beyond explanation as soliloquy. In: IJCAI (2017)

work page 2017
[6]

org/abs/2501.08406

Chen, H., Constante-Flores, G.E., Mantri, K.S.I., Kompalli, S.M., Ahluwalia, A.S., Li, C.: Optichat: Bridging optimization models and practitioners with large language models (2025),https://arxiv. org/abs/2501.08406

work page arXiv 2025
[7]

Chen, W., Su, Y., Zuo, J., Yang, C., Yuan, C., Chan, C.M., Yu, H., Lu, Y., Hung, Y.H., Qian, C., Qin, Y., Cong, X., Xie, R., Liu, Z., Sun, M., Zhou, J.: AgentVerse: Facilitating Multi-Agent Col- laboration and Exploring Emergent Behaviors. International Conference on Representation Learn- ing2024, 20094–20136 (May 2024),https://proceedings.iclr.cc/paper_f...

work page 2024
[8]

https://doi.org/10.48550/arXiv.2511.09378

Corrêa, A.B., Pereira, A.G., Seipp, J.: The 2025 Planning Performance of Frontier Large Language Models (Nov 2025). https://doi.org/10.48550/arXiv.2511.09378

work page doi:10.48550/arxiv.2511.09378 2025
[9]

In: Proc

Cosler, M., Hahn, C., Mendoza, D., Schmitt, F., Trippel, C.: nl2spec: interactively translating unstruc- tured natural language to temporal logics with large language models. In: Proc. CAV (2023)

work page 2023
[10]

AIJ (2021)

Dazeley, R., Vamplew, P., Foale, C., Young, C., Aryal, S., Cruz, F.: Levels of explainable artificial intelligence for human-aligned conversational explanations. AIJ (2021)

work page 2021
[11]

In: AAAI (2014)

De Giacomo, G., De Masellis, R., Montali, M.: Reasoning on LTL on finite traces: Insensitivity to infiniteness. In: AAAI (2014)

work page 2014
[12]

JAIR (2015)

Domshlak, C., Mirkis, V.: Deterministic oversubscription planning as heuristic search: Abstractions and reformulations. JAIR (2015)

work page 2015
[13]

In: ICAPS (2022) Exploring Plan Space through Conversation 15

Eifler, R., Brandao, M., Coles, A., Frank, J., Hoffmann, J.: Evaluating plan-property dependencies: A web-based platform and user study. In: ICAPS (2022) Exploring Plan Space through Conversation 15

work page 2022
[14]

In: AAAI (2020)

Eifler, R., Cashmore, M., Hoffmann, J., Magazzeni, D., Steinmetz, M.: A new approach to plan-space explanation: Analyzing plan-property dependencies in oversubscription planning. In: AAAI (2020)

work page 2020
[15]

In: IJCAI (2020)

Eifler, R., Steinmetz, M., Torralba, A., Hoffmann, J.: Plan-space explanation via plan-property depen- dencies: Faster algorithms & more powerful properties. In: IJCAI (2020)

work page 2020
[16]

In: Proc

Feldhus, N., Wang, Q., Anikina, T., Chopra, S., Oguz, C., Möller, S.: Interrolang: Exploring nlp models and datasets through dialogue-based explanations. In: Proc. EMNLP (2023)

work page 2023
[17]

In: Béchet, F., Lefèvre, F., Asher, N., Kim, S., Merlin, T

Fichtel, L., Spliethöver, M., Hüllermeier, E., Jimenez, P., Klowait, N., Kopp, S., Ngonga Ngomo, A.C., Robrecht, A., Scharlau, I., Terfloth, L., Vollmer, A.L., Wachsmuth, H.: Investigating co-constructive behavior of large language models in explanation dialogues. In: Béchet, F., Lefèvre, F., Asher, N., Kim, S., Merlin, T. (eds.) Proceedings of the 26th A...

work page 2025
[18]

In: ICAPS (2023)

Fuggitti, F., Chakraborti, T.: NL2LTL – a python package for converting natural language (NL) instruc- tions to linear temporal logic (LTL) formulas. In: ICAPS (2023)

work page 2023
[19]

JAIR (2023)

Gamba, E., Bogaerts, B., Guns, T.: Efficiently explaining csps with unsatisfiable subset optimization. JAIR (2023)

work page 2023
[20]

Large Language Model based Multi-Agents: A Survey of Progress and Challenges

Guo, T., Chen, X., Wang, Y., Chang, R., Pei, S., Chawla, N.V., Wiest, O., Zhang, X.: Large language model based multi-agents: A survey of progress and challenges. arXiv:2402.01680 (2024)

work page internal anchor Pith review Pith/arXiv arXiv 2024
[21]

JAIR26, 191–246 (2006)

Helmert, M.: The Fast Downward planning system. JAIR26, 191–246 (2006)

work page 2006
[22]

In: Proc

Junker, U.: Preferred explanations and relaxations for over-constrained problems. In: Proc. AAAI (2004)

work page 2004
[23]

In: ICML (2024)

Kambhampati, S., Valmeekam, K., Guan, L., Verma, M., Stechly, K., Bhambri, S., Saldyt, L., Murthy, A.: Position: Llms can’t plan, but can help planning in llm-modulo frameworks. In: ICML (2024)

work page 2024
[24]

In: ICAPS (2024)

Krarup, B., Coles, A.J., Long, D., Smith, D.E.: Explaining plan quality differences. In: ICAPS (2024)

work page 2024
[25]

JAIR (2021)

Krarup, B., Krivic, S., Magazzeni, D., Long, D., Cashmore, M., Smith, D.E.: Contrastive explanations of plans through model restrictions. JAIR (2021)

work page 2021
[26]

In: NeurIPS Workshop on Human Centered AI (2022)

Lakkaraju, H., Slack, D., Chen, Y., Tan, C., Singh, S.: Rethinking explainability as a dialogue: A practitioner’s perspective. In: NeurIPS Workshop on Human Centered AI (2022)

work page 2022
[27]

In: Proc

Liao, Q.V., Gruen, D., Miller, S.: Questioning the ai: informing design practices for explainable ai user experiences. In: Proc. CHI (2020)

work page 2020
[28]

Archives of Psychology (1932)

Likert, R.: A technique for the measurement of attitudes. Archives of Psychology (1932)

work page 1932
[29]

In: Workshop on Language and Robotics at CoRL (2022)

Liu, J.X., Yang, Z., Schornstein, B., Liang, S., Idrees, I., Tellex, S., Shah, A.: Lang2ltl: Translating natural language commands to temporal specification with large language models. In: Workshop on Language and Robotics at CoRL (2022)

work page 2022
[30]

AI (2019)

Miller, T.: Explanation in artificial intelligence: Insights from the social sciences. AI (2019)

work page 2019
[31]

Menu- based interfaces: An empirical study

Nguyen, Q.N., Sidorova, A., Torres, R.: User interactions with chatbot interfaces vs. Menu- based interfaces: An empirical study. Computers in Human Behavior128, 107093 (Mar 2022). https://doi.org/10.1016/j.chb.2021.107093

work page doi:10.1016/j.chb.2021.107093 2022
[32]

In: XAI (2023)

Nguyen, V.B., Schlötterer, J., Seifert, C.: From black boxes to conversations: Incorporating xai in a conversational agent. In: XAI (2023)

work page 2023
[33]

COLM (2024)

Peng, B., Narayanan, S., Papadimitriou, C.: On limitations of the transformer architecture. COLM (2024)

work page 2024
[34]

In: Workshop on Progress Towards the Holy Grail at CP (2024)

Povéda, G., Strahl, A., Hall, M., Boumazouza, R., Quintana-Amate, S., Alvarez, N., Bleukx, I., Tsouros, D., Verhaeghe, H., Guns, T.: Trustworthy and explainable decision-making for workforce allocation. In: Workshop on Progress Towards the Holy Grail at CP (2024)

work page 2024
[35]

Emergent Coordination in Multi-Agent Language Models

Riedl, C.: Emergent Coordination in Multi-Agent Language Models (Oct 2025). https://doi.org/10.48550/arXiv.2510.05174

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.05174 2025
[36]

Cambridge University Press (1979)

Searle, J.R.: Expression and meaning: Studies in the theory of speech acts. Cambridge University Press (1979)

work page 1979
[37]

CoRR (2023)

Shen, H., Huang, C., Wu, T., Huang, T.K.: Convxai: Delivering heterogeneous AI explanations via conversations to support human-ai scientific writing. CoRR (2023)

work page 2023
[38]

Nature Machine Intelligence (2023)

Slack, D., Krishna, S., Lakkaraju, H., Singh, S.: Explaining machine learning models with interactive natural language conversations using talktomodel. Nature Machine Intelligence (2023)

work page 2023
[39]

In: AAAI (2012)

Smith, D.: Planning as an iterative process. In: AAAI (2012)

work page 2012
[40]

In: ICAPS (2004)

Smith, D.E.: Choosing objectives in over-subscription planning. In: ICAPS (2004)

work page 2004
[41]

AIJ (2021)

Sreedharan, S., Chakraborti, T., Kambhampati, S.: Foundations of explanations as model reconciliation. AIJ (2021)

work page 2021
[42]

In: IJCAI (2019) 16 G

Sreedharan, S., Srivastava, S., Smith, D.E., Kambhampati, S.: Why can’t you do that hal? explaining unsolvability of planning tasks. In: IJCAI (2019) 16 G. Fouilhé et al

work page 2019
[43]

In: XAI (2023)

State, L., Ruggieri, S., Turini, F.: Reason to explain: Interactive contrastive explanations (reasonx). In: XAI (2023)

work page 2023
[44]

NeurIPS (2023)

Valmeekam, K., Marquez, M., Sreedharan, S., Kambhampati, S.: On the planning abilities of large language models-a critical investigation. NeurIPS (2023)

work page 2023
[45]

In: Proc

Vasileiou, S.L., Yeoh, W.: PLEASE: generating personalized explanations in human-aware planning. In: Proc. ECAI (2023)

work page 2023
[46]

Load package$Pi before package$P j into truck$T

Zhang, T., Yang, X.J., Li, B.: May i ask a follow-up question? understanding the benefits of conversations in neural network explainability. International Journal of Human–Computer Interaction (2024) Exploring Plan Space through Conversation 17 A Platform for Iterative Planning with Explanations The source code and pre-built docker images will be publishe...

work page 2024
[47]

**Context-Aware**: Base suggestions on the current state (enforced, satisfied, unsatisfied goals)

work page
[48]

**Conversation-Aware**: Consider what has already been discussed to avoid repetition

work page
[49]

**Actionable**: Suggest questions that help the user make progress toward finding an optimal combination of goals

work page
[50]

**Natural Language**: Use natural, conversational phrasing that sounds like a real user

work page
[51]

**Diverse**: Provide 1-3 questions covering different aspects or question types when pos- sible

work page
[52]

**Goal-Specific**: Reference actual goals from the current context by name

work page
[53]

question1

**Progressive**: If the user just learned something, suggest logical follow-up questions ## Suggestion Strategy - You can take inspiration from questions the user previously asked ### For Unsolvable Tasks: - Prioritize understanding why it’s unsolvable - Suggest questions about how to make it solvable - Suggest questions about which goals are conflicting ...

work page