acceptodds
Under review as a conference paper at ICLR 2027

World Models as Version Spaces: Safe Tool Use under Hidden Rules

Abstract

World models enable tool-using agents to anticipate action outcomes and plan before execution. However, identical interfaces and similar observations can conceal different execution rules, making a plan safe in one deployment and unsafe in another. Overlooking this ambiguity can lead to irreversible errors that later replanning cannot undo. To address this challenge, we introduce VERSA, a world model that represents a version space of execution rules that remain plausible given the available evidence. VERSA learns a joint rule posterior through conditional flow matching and predicts plan outcomes with a learned rule-conditioned transition model. VERSA uses these predictions to select a plan under a risk constraint and decide whether further information is worth acquiring before execution. Theoretically, we derive upper bounds on execution risk and the value of further information, characterizing conditions under which near-optimal decisions can be made without fully identifying the deployment rules. Empirical evaluations show that VERSA achieves superior task performance and execution safety in hidden-rule environments.

Primary area:
foundation or frontier models, including LLMs
Keywords:
World Models, Version Spaces, Tool-Using Agents, Flow Matching

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_j5LtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VERSA, a world-model framework for safe tool use under hidden execution rules. It uses conditional flow matching to infer a joint posterior over plausible rule configurations and a rule-conditioned transition model to predict candidate-plan outcomes. Based on these predictions, VERSA performs risk-constrained planning and selectively acquires decision-relevant information. Experiments on retail and airline workflows demonstrate improvements in safe task completion over the selected baselines.

Critical strengths
  1. The paper identifies a practical challenge in tool-using agents: identical observations and tool interfaces may conceal different execution rules, potentially causing irreversible errors. Explicitly modeling these uncertainties is valuable for safe decision-making.

  2. VERSA combines rule posterior inference, rule-conditioned outcome prediction, and risk-aware decision-making into a coherent framework. The experiments and ablations demonstrate substantial improvements in safe success rates, particularly from considering multiple rule hypotheses.

Critical weaknesses
  1. The main baselines primarily consist of conventional LLM agents and earlier world-model planners, while several recent studies have explicitly addressed uncertainty in planning and decision-making. For example, BB-WMs[1] introduce explicit belief representations alongside world-model simulations to support decision-making under partial observability; PlanU[2] incorporates uncertainty-aware planning into MCTS as RAP's optimization; and DreamPhase[3] uses a learned latent world model to simulate multiple future trajectories with uncertainty-aware evaluation. The absence of comparisons with these methods limits the evidence for VERSA's advantages over existing uncertainty-aware planning approaches.
  2. Insufficiently differentiated methodological contributions. Although the paper addresses an interesting application scenario involving hidden execution rules, its methodological novelty relative to existing uncertainty-aware planning approaches remains unclear. The paper does not sufficiently explain why existing methods cannot be directly applied or adapted to this setting, or what fundamental limitations VERSA specifically overcomes.

[1] Towards a Belief-Based World Model for LLM Agents, arXiv:2609.00455.

[2] PlanU: Large Language Model Reasoning through Planning under Uncertainty, NeurIPS 2025.

[3] DreamPhase: Offline Imagination and Uncertainty-Guided Planning for Large-Language-Model Agents, ICLR 2026.

Questions for authors
  1. Could the authors provide additional comparisons with recent uncertainty-aware planning methods, such as BB-WMs, PlanU, and DreamPhase, or suitable adaptations of their core mechanisms, to better demonstrate VERSA's advantages?
  2. What are the essential methodological differences between VERSA and existing Bayesian or uncertainty-aware model-based planning approaches? Why are these approaches insufficient for the proposed hidden-rule setting, and what specific limitations does VERSA address?
5 comments

We thank the reviewer for the constructive suggestions. We adapted all three methods to our tasks, keeping the core mechanism of each and using Qwen2.5-7B-Instruct, the backbone of Table 1, and evaluated them on the complete retail, airline and new-user panels (SS: safe success; U: violation rate).

Q1.

VERSA achieves higher safe success, fewer violations and lower tool cost than all three methods on every panel.

Method (SS ↑ % / U ↓ % / tool cost ↓)Retail, 122 episodesAirline, 574 episodesNew users (airline), 224 episodes
VERSA97.92 / 2.08 / 5.8188.09 / 3.52 / 3.2086.97 / 1.56 / 3.24
BB-WM61.20 / 26.04 / 10.3579.10 / 13.77 / 4.4275.56 / 19.95 / 4.61
PlanU65.10 / 30.99 / 10.9662.03 / 28.34 / 5.2564.36 / 23.38 / 5.25
DreamPhase47.53 / 16.15 / 10.4855.65 / 28.64 / 5.8155.72 / 26.03 / 5.79

VERSA is ahead by 9.0 to 50.4 points in safe success, and all differences in safe success and violation rate are statistically significant.

The advantage is largest on delayed-effect tasks. On airline, the four methods are within 84-92% when effects are immediate. When effects are delayed, VERSA reaches 92.5% safe success with no violation, against 73.9%, 39.2% and 34.3% for the three adapted methods.

Method (SS ↑ % / U ↓ %)Airline, immediate (252 episodes)Airline, delayed (273 episodes)Retail, delayed (47 episodes)
VERSA90.1 / 4.292.5 / 0.094.6 / 5.4
BB-WM87.9 / 5.973.9 / 18.930.1 / 59.1
PlanU91.6 / 1.139.2 / 57.721.7 / 75.0
DreamPhase84.3 / 6.234.3 / 51.626.3 / 30.8

Q2.

Existing uncertainty-aware planners model uncertainty in the state, in returns or in imagined trajectories. In the hidden-rule setting the uncertainty lies in the rules: a hidden configuration that is fixed for the whole episode and determines the consequence of every action. VERSA differs in three respects.

  1. It infers the rule configuration and plans over several rule hypotheses instead of one, which accounts for +41.4 points of safe success on retail and +29.4 on airline (Table 2).
  2. It selects plans under a bound on the violation probability, which also applies to stopping and is evaluated after delayed effects settle (Prop. 1). When effects are delayed, the three adapted methods lose 14 to 52 points on airline, whereas VERSA loses none (Q1).
  3. It queries only when the answer can change the next decision (Prop. 2), which yields +3.6 and +4.9 points over entropy-based information gain on the controlled benchmark.

The rule posterior alone is not sufficient: the adapted BB-WM and DreamPhase receive VERSA's posterior and still trail by 9.0 to 50.4 points.

We will add these comparisons to Sec. 5 and discuss the three papers in Related Work.

Thank you for the detailed response. The additional comparisons in Q1 have addressed my concern about the missing relevant baselines.

However, my concern regarding Q2 remains. While the authors clarify VERSA's design choices, I am still unclear about why existing Bayesian or uncertainty-aware planning methods cannot be directly adapted to the proposed hidden-rule setting, and what specific methodological limitations VERSA overcomes. I hope you can list a method feature to more clearly explain this point when creating the table. I believe this distinction is crucial to establishing the paper's core contribution.

If the authors can convincingly clarify this point, I would be willing to raise my rating to 3.

Thank you for the follow-up.

VERSA contributes a decision procedure for tool use when the rule in force has not been identified. Before acting, it predicts the outcome of every candidate plan, and of stopping, under every plausible rule and after pending effects settle, and it selects under a risk constraint on these predictions.

Why existing methods cannot be applied directly. Three of their premises do not hold under hidden rules.

  1. A single dynamics model. BB-WM specifies its belief and update for known environment rules, and PlanU and DreamPhase plan within one world model. Here deployments with identical interfaces follow different rules, and the rule in force is itself unknown.

  2. Responses that reflect outcomes. All three treat the response to an action as its outcome. Effects may be delayed and failures unreported, so the response does not determine the outcome.

  3. Safe exploration. PlanU explores, DreamPhase falls back to real interaction, and BB-WM gathers information by acting. A tool call may be irreversible, and pending effects still take place after the agent stops. VERSA therefore converts rule uncertainty into the risk of an irreversible action and bounds it before acting.

Quantities required by a risk-constrained decision. Such a decision requires, for each plan, the probabilities of a violation and of safe success. Neither is determined by an expected return, by a value penalized by policy uncertainty, or by the LLM's judgment, since a single number merges the two events. VERSA computes both: it infers a joint posterior over rule configurations, reconstructs the current state under each hypothesis by replaying the history, and rolls out every plan with the hypothesis held fixed until pending effects settle.

Method features

BB-WMPlanUDreamPhaseVERSA
Environment rulesone set, knownone world modelone learned world modelseveral plausible sets; the one in force is hidden
Predicted before actingvalidity of the next actionreturn of action sequences; outcomes sampled per stepimagined branches of fixed length, scored by a value headsuccess and violation of every plan, and of stopping, under every plausible rule, after effects settle
Decision based onthe LLM's judgmentexpected return plus an exploration bonusvalue penalized by policy uncertainty; confidence gateviolation probability within a threshold, then highest safe success
Information acquired byactingexplorationreal interaction when uncertaina query, only when the answer can change the decision

Decisions without identifying the rule. On the main panels, the most probable rule differs from the true rule in 110 of 122 retail and 475 of 574 airline episodes, and VERSA queries in 14 and 6 episodes. The table compares it with a planner using only the most probable rule (Table 2, row 2), given the same posterior and transition model.

SS ↑ (%)Retail, estimate incorrectRetail, estimate correctAirline, estimate incorrectAirline, estimate correct
Planner using the most probable rule46.46100.0053.3290.00
VERSA97.92100.0087.7892.50

The single-rule planner is reliable only when its estimate is correct, whereas the safe success of VERSA is nearly unaffected. The differences for incorrect estimates are statistically significant.

Supplying VERSA's predictions to an existing method. The adapted BB-WM of Q1 receives VERSA's rule posterior and its predictions for every plan and for stopping. A new experiment also imposes VERSA's risk constraint in two steps, leaving the choice to the LLM.

Adapted BB-WMRetail SS ↑ / U ↓ (%)Airline SS ↑ / U ↓ (%)
with VERSA's posterior and predictions (Q1)61.20 / 26.0479.10 / 13.77
+ one plan only; told which plans meet the threshold60.68 / 14.3275.32 / 9.10
+ threshold enforced64.32 / 8.3374.94 / 7.76
VERSA97.92 / 2.0888.09 / 3.52

The constraint lowers its violation rate in both domains. VERSA remains ahead of every version, by 33.6 and 13.2 points of safe success over the last one and by 16.0 points on the new-user panel, all statistically significant.

Relation to Bayesian planning. VERSA shares with Bayesian planning the posterior and the practice of holding a hypothesis fixed within a simulation. It differs in three respects.

  1. The posterior over rules is learned. No likelihood model of the rules is available, so it is inferred from interaction histories.

  2. The decision is defined on the probabilities of violation and safe success, not on expected return as in the standard formulation. The former is constrained, for stopping as well as for execution, and Prop. 1 bounds the resulting risk.

  3. Information is obtained by queries rather than by trial and error, and only when the answer can change the decision. Prop. 2 and Cor. 1 bound the loss from deciding before the rule is identified.

Reviewer_xHyNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the author propose VERSA, a world model framework that proposes the applicable rule distribution based on the observation for safe tool using. Since point estimation is risky (relying on one rule may cause unsafe move) and full identification is costly, the authors use a flow-matching model to propose the real underlying rule of the current state, use a transition model to simulate outcomes across hypotheses, and apply risk-constrained planning using the simulated outcomes. The author evaluate the framework on retail and airline tasks and VERSA outperforms existing agentic baselines on both safety and success.

Critical strengths
  1. The method is well-motivated. Given that planning / simulating with only one rule might be risky, the authors proposed a method to discover a weighted rule distribution to explore more rules that can be useful and also safe.

  2. The authors did extensive experiments / ablations and showed that VERSA outperforms strong baselines (including ReAct, WebDreamer, and RAP) on retail and airline workflows, and shows that all designed components work.

Critical weaknesses
  1. The design of “flow-matching” is not well explained and the experiments also don’t support that it is a good method. Given that the model input the current observation and at last output a "distribution of rule” with rule and its weight, the flow-matching idea is a heavy model with longer inference time (256 samples * 8 steps) compared to simply using the model to predict a softmax distribution. The experiment in Appendix C also shows that the flow-matching method doesn't have much superior performance compared to enumeration and categorizing. So the design choice of such a heavy module should be explained, and other alternatives should be compared to?

  2. The current evaluation restricts rule configurations to a fixed catalogue of 36 rules. It remains unclear how VERSA scales as the size of the rule space grows combinatorially (e.g., to hundreds of possible rules) or how it handles out-of-catalogue rules that were not part of the training catalogue? The author should add experiment with more (or less) rules and show the experiment of VERSA?

Questions for authors
  1. The paper uses Conditional Flow Matching (256 samples * 8 steps) to generate continuous rule vectors before decoding them into a 36-rule catalogue. However, Appendix C results show that simpler alternatives (enumeration and categorizing) achieve statistically tied performance with much lower overhead. Why is this heavy module necessary, and in what specific setting does flow matching outperform standard categorical prediction with softmax? Seems that flow matching is better in fitting continuous things, so why use it on discrete rules with limited numbers?

  2. Evaluations are restricted to a fixed set of 36 rules. How does VERSA scale as the rule space grows combinatorially into hundreds of possibilities (or a smaller list of rules)?

See more details in the weakness section.

3 comments

We thank the reviewer for the encouraging assessment and the two concrete questions. We answer both below with new results (SS: safe success; U: violation rate).

Q1.

The overhead of the Flow posterior is small. With the paper's setting (256 samples × 8 steps), posterior inference takes 6.4 ms and adds 5-6% to the time of a complete decision.

Airline, time per decision (ms)Flow, 256 × 8Joint categorical (softmax)Enumeration
Posterior inference6.43.93.4
Complete decision51.749.348.7

The number of steps can also be reduced: on retail, 2 Euler steps give the same success and violation outcome as 8 on every episode.

Flow outperforms softmax when effects are delayed and on rule configurations never used in training. In controlled comparisons where only the posterior differs, we compare Flow with a softmax posterior over the same rule codes.

Metric (%)FlowSoftmaxΔ
SS ↑, all rules76.5073.85+2.65
SS ↑, delayed-effect rules69.9064.64+5.25
U ↓, delayed-effect rules1.985.22-3.24
SS ↑, 4,096 rules, configurations never used in training43.7542.35+1.40

All four differences are statistically significant, and the gain in the first row is +2.6 to +2.8 points at each of five risk thresholds.

Why flow matching for rules. A rule configuration is a combination of factors (preconditions, cascades, effect timing and failure reporting; 3 × 2 × 3 × 2 = 36 in the paper), and such factors can also be continuous, as the processing times in the settlement environment (Appendix B.1). VERSA therefore infers a continuous rule code and decodes it to the catalogue. The same posterior then applies unchanged to larger combinatorial rule spaces and to continuous rule parameters, and its inference time stays under 5 ms from 16 to 4,096 rules (Q2).

Q2.

We evaluated VERSA with 16, 256 and 4,096 rule configurations on a controlled two-stage workflow with four hidden rule parameters, using the same amount of training data at every size. Safe success is unchanged from 256 to 4,096 rules, and posterior inference stays under 5 ms.

Rule configurations162564,096
SS ↑ (%)50.0744.6644.66
U ↓ (%)3.394.176.64
Posterior inference per episode (ms)4.14.34.6
Total time per episode (ms), VERSA27.345.467.4
Total time per episode (ms), softmax posterior26.548.871.4

Rules not used in training. VERSA maintains its performance on rule configurations that never appear in training.

Configurations never used in trainingSS ↑ (%)U ↓ (%)Reference SS (%)
Airline workflows: 15 of 36, 5% prior mixture80.6-85.22.2-4.188.2, seen configurations
Controlled benchmark, new deployments: 819 of 4,09643.754.1045.70, trained configurations; 42.35, softmax posterior

VERSA reasons over the rules in its catalogue. A new rule type is supported by adding it to the catalogue and to the training of the transition model; the rest of the framework is unchanged.

Reviewer_xHyNOpenReview ↗

reply to author response

Thanks to the authors for the additional experiments and the detailed rebuttal.

My concern 1 is resolved. But for concern 2, it seems that the safe success rate (ss%) drops as the rule number goes up. Could the authors explain this performance degradation? These results do not necessarily demonstrate that the framework scales effectively to larger rule spaces.

Thank you for the follow-up. The decrease from 16 to 256 rules comes from the inference problem itself and not from the framework. To separate the two, we add two oracle references and a planner without rule inference, on the same test deployments as in our previous table. The oracles use the true rule or the exact environment model, which are not available at deployment; they show what is achievable.

SS ↑ (%) by number of rule configurations162564,096
Oracle: rule known59.1860.1661.72
Oracle: exact Bayesian inference53.3248.8348.44
VERSA50.0744.6644.66
VERSA relative to exact inference93.9%91.5%92.2%
No rule inference (prior only)41.4137.5026.56
  1. The tasks do not become harder. With the rule known, safe success is about 60% at every size; the remaining tasks cannot be completed within the deadline and budget.

  2. The decrease comes from identifying the rule. The interaction history has the same length at every size, so a finer rule space leaves more configurations compatible with the evidence. Exact Bayesian inference shows the same pattern as VERSA: a decrease from 16 to 256 rules and almost none from 256 to 4,096.

  3. VERSA retains 92-94% of the safe success of exact inference at every size, although the training examples per rule fall from 1,250 to 130 to 9. Its approximation error does not grow with the rule space, and posterior inference stays between 4.1 and 4.6 ms.

  4. Rule inference matters more as the space grows. Without it, safe success falls to 26.56% and the violation rate rises to 25.78% at 4,096 rules, against 44.66% and 6.64% for VERSA. The margin of VERSA grows from 8.7 to 18.1 points.

Reviewer_KMRhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes VERSA for tool environments in which shared interfaces conceal deployment-specific execution rules. It maintains a joint posterior over rules using conditional flow matching, predicts action consequences with a rule-conditioned transition model, and chooses among candidate plans, information queries, and stopping. The controller considers both safety violations and delayed effects. The paper derives conditional bounds on execution risk and the remaining value of rule information. Evaluation uses modified retail and airline workflows built from public τ-bench resources, together with constructed benchmarks for posterior inference and information acquisition.

Critical strengths
  1. The decision-oriented treatment of hidden rules is useful. The paper distinguishes identifying every hidden rule from identifying enough to choose a safe plan. It accounts for irreversible violations, pending effects, and the fact that stopping need not be safe.

  2. The appendix and controls strengthen the documentation and empirical support.

Critical weaknesses
  1. The posterior-specific contribution and generalization scope need qualification. a) Enumeration exactly matches Flow's retail safe success and violations at lower tool cost (Tables 7-8). b) On airline, Flow improves safe success by only 0.12 percentage points over enumeration and 0.67 points over joint classification; both confidence intervals include zero. Its violation rate is 3.52%, versus 0.52% for joint classification, with a paired increase of 2.99 points [0.65, 5.99]. These are trade-offs rather than evidence of an unequivocally better posterior. c) Flow-based posterior estimation and decision-directed information gathering have established precedents, which the paper cites. The contribution is most convincing as their integration for hidden-rule tool execution. d) The workflow split is by source user, and composite airline tasks reuse the test-user pool. This supports user-level and task-composition transfer, but does not establish transfer to unseen rule configurations or open-ended enterprise deployments.

  2. Independent reproduction and the practical safety claim remain incomplete. a) Section B.4 describes retail adaptation from previously trained checkpoints without identifying their training data or original training procedure. Implementation detail is substantially more limited for airline and the controlled environment. b) Although the paper states that supplementary source code is supplied, the submission provides neither a code attachment nor an accessible repository. The implementation is unavailable for independent reproduction; the statement should be corrected or the missing code supplied.

Questions for authors

See Critical Weaknesses.

4 comments

We thank the reviewer for the careful reading. We answer each point below (SS: safe success; U: violation rate).

1a.

With 2 Euler steps instead of 8, Flow reaches the same success and violation outcome on every retail episode, at the lowest tool cost of the three posteriors.

Retail, 122 episodesSS ↑ (%)U ↓ (%)Tool cost ↓
Flow, 2 Euler steps97.922.084.82
Enumeration97.922.085.32
Joint categorical95.573.135.69

In a controlled comparison where only the posterior differs, Flow improves over a softmax posterior on the same rule codes, most strongly on delayed-effect rules.

Metric (%)FlowSoftmaxΔ
SS ↑, all rules (2,400 episodes)76.5073.85+2.65
SS ↑, delayed-effect rules (1,020 episodes)69.9064.64+5.25
U ↓, delayed-effect rules1.985.22-3.24

All three differences are statistically significant, and the gain in safe success is +2.6 to +2.8 points at each of five risk thresholds.

1b.

Posterior smoothing is already part of VERSA's inference (Appendix B.4): the retail configuration mixes the posterior with a 5% uniform component. With a 5% mixture of the public rule prior on airline as well, Flow has no violation on the full airline panel, compared with 0.26% for joint classification, while safe success stays at 88%.

Airline, 574 episodesSS ↑ (%)U ↓ (%)Tool cost ↓
Flow, paper configuration88.093.523.20
Flow, 5% mixture88.000.003.77
Joint categorical, paper configuration87.420.523.50
Joint categorical, 5% mixture87.980.263.77

The reduction in Flow's violation rate, -3.52 points, is statistically significant. Across five independently trained airline models, safe success is 87.4-89.2%, and the mean violation rate with the mixture is 0.1%.

1c.

VERSA makes four contributions.

  1. We formulate the problem of safe tool use when shared interfaces hide deployment-specific rules and effects are delayed or irreversible.
  2. We represent the world model as a version space of plausible rules and let one decision objective drive rule inference, consequence prediction and information acquisition.
  3. We prove bounds that tie posterior and prediction error to execution risk and to the remaining value of rule information (Props. 1-2, Cor. 1).
  4. We build a system in which each component is supported by evidence:
ComponentEvidenceSS ↑
Planning over multiple rule hypothesesTable 2+41.4 retail, +29.4 airline
Anticipating delayed effectsTable 11, delayed-execution tasks94.6% retail, 92.5% airline
Decision-value queriesTable 2+3.9 retail, +1.1 airline
vs. entropy-based information gain (controlled benchmark)+3.6 and +4.9
Flow posterior1a+2.65 overall, +5.25 delayed-effect rules

1d.

VERSA keeps its performance on new users and on new task compositions, stays above 80% on rule combinations never used in training, and is within 2 points of the trained configurations in a 4,096-rule space.

SettingEvaluationSS ↑ (%)U ↓ (%)Reference SS (%)
New users24 users, 224 episodes86.971.5688.09, main airline panel
New task compositions618 episodes (Table 3)87.750.1688.09, main airline panel
Unseen rule combinations (15 of 36), 5% mixture333 episodes81.72.488.2, seen combinations
804 episodes85.22.2
147 episodes, new users80.64.1
4,096-rule controlled benchmark (819 unseen)unseen configurations43.754.1045.70, trained configurations; 42.35, softmax posterior

2a.

Each row covers both the rule posterior and the transition model.

ModelInitializationDataStepsCheckpoint
Retail, pretrainingfrom scratchsynthetic deployments from the four domains of Appendix C.1: 1,347 (202,050 episodes), then 13,635 (374,082 episodes)10,000 (posterior 12,000), then 8,000final EMA, then validation loss
Retail, adaptationpretrained checkpoints3,072 retail deployments from 132 training users (43,008 episodes), plus replay of the first-stage synthetic data8,000validation loss, EMA 0.999
Airlinefrom scratch224 deployments (33,600 episodes), 21 rule configurations10,000final EMA
Controlled environmentfrom scratchdeployments simulated from the training prior40,000final weights

All models are trained with AdamW, warm-up and cosine learning-rate decay. No test user appears in any training stage.

2b.

The source code was submitted together with the paper. As stated in the Reproducibility Statement, it is in the supplementary material on the submission page. The ZIP contains the VERSA implementation for both domains (versa/retail, versa/airline), the rule engine and models (versa/src), training and data tools (versa/tools), run scripts (versa/scripts) and evaluation code (versa/eval).

Thank you again for your feedback. We would appreciate your thoughts on our response and the additional results, and would be glad to clarify anything further.

Thank you for your reply. Having read the reviews and replies, I decided to raise my score.

We sincerely thank you for carefully reading our rebuttal and providing timely feedback. We express our gratitude once again!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for AC

We sincerely thank the AC for their time and effort in reviewing our work. We summarize below the revisions made in response to the reviewers' comments:

  1. We clarified VERSA's methodological distinctions in the Introduction and Section 2.2: candidate plans and stopping are evaluated across plausible execution rules after delayed effects settle, and information acquisition uses the same risk-constrained decision criterion. (For Reviewer j5Lt\textcolor{blue}{\text{j5Lt}}.)

  2. We added comparisons with our tool-domain adaptations of BB-WM, PlanU, and DreamPhase on the retail, airline, and new-user airline panels. VERSA achieves higher safe success and lower violation rates across all three panels. These results are summarized in Section 5.2 and reported in Appendix C.6, Table 12. (For Reviewer j5Lt\textcolor{blue}{\text{j5Lt}}.)

  3. We added a controlled comparison between Flow and softmax posteriors, alongside separate airline measurements of posterior inference and complete-decision runtime, addressing posterior quality and computational overhead (Appendix C.7, Table 13). (For Reviewer xHyN\textcolor{blue}{\text{xHyN}}.)

  4. We added rule-space scaling and generalization results: performance and runtime for 16, 256, and 4,096 rule configurations, plus results on 819 combinations excluded from training within the modeled catalogue (Appendix C.8, Table 14). (For Reviewers xHyN\textcolor{blue}{\text{xHyN}} and KMRh\textcolor{blue}{\text{KMRh}}.)

The additions are highlighted in blue in the revised manuscript, including the newly added references.

Best regards, Authors of Submission 55147

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.