acceptodds
Under review as a conference paper at ICLR 2027

CCE-Mobility: Counterfactual World–Agent Co-Evolution for Mobility Agents

Abstract

Large Language Model (LLM) agents are increasingly used for mobility planning and tool-based assistance, yet the world may change after a plan or commitment has already been formed. Roads may be blocked, services may become unavailable, or acquired information may become stale, requiring the agent to revise its decisions while preserving the original user need. While self-evolving agents learn from interaction experience and language world models simulate environment dynamics, an important question remains underexplored: what plausible environmental change would invalidate an already-made decision while leaving the original task achievable? We introduce CCE-Mobility, a counterfactual world–agent co-evolution framework for recovery under dynamic world changes. Given an observed trajectory, the world model reasons about decision-supporting dependencies and constructs a localized, executable intervention that invalidates one while preserving the original goal. The resulting counterfactual world provides recovery experience for agent adaptation, while persistent failures guide subsequent world-side evolution. Thus, the agent learns to recover from challenging world changes, while the world model learns to construct counterfactuals targeting unresolved weaknesses. Experiments on TravelPlanner and MobilityBench show consistent recovery gains across all evaluated backbones. On TravelPlanner, CCE-Mobility improves Recovery by 0.1092–1.0055 and reduces Low by 4.99–26.28 percentage points. Paired validation further shows an additional 4.05-point Low reduction from world-side evolution over agent-only adaptation. Code will be made publicly available at https://anonymous.4open.science/r/CCE-Mobility-4254.

Primary area:
applications to robotics, autonomy, planning
Keywords:
LLM agents, Self-evolving Agents, World Models, Counterfactual Interventions, World-agent Co-evolution, Mobility Planning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_GoHpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CCE-Mobility, a counterfactual world-agent co-evolution framework that enhances LLM mobility agents’ recovery against dynamic environmental changes. It constructs solvable counterfactual scenarios from real trajectories, learns recovery skills from failures, and leverages persistent unresolved failures to guide the world model for better challenge generation. Evaluated on TravelPlanner and MobilityBench, CCE-Mobility greatly boosts counterfactual recovery performance.

Critical strengths

The paper addresses a meaningful problem of adapting mobility agents to dynamic environmental changes, emphasizing recovery from invalidated decisions while preserving the original user intent.

Critical weaknesses
  1. The experimental evidence for world–agent co-evolution remains limited. The incremental benefit of world-side evolution is primarily demonstrated on a TravelPlanner validation panel, while the corresponding MobilityBench result is inconclusive.

  2. The paper lacks sufficient analysis of how different counterfactual intervention types contribute to recovery improvements. Without intervention-level comparisons, it is difficult to determine which environmental changes provide the most effective learning signals.

  3. The gains in recovery scores do not consistently translate into improvements in final task success. This raises concerns about whether the proposed framework improves practical task completion rather than primarily optimizing recovery-oriented metrics.

  4. The experiments lack cross-intervention generalization analysis. It remains unclear whether the learned recovery skills transfer to unseen types of environmental disruptions, rather than only improving performance on intervention patterns encountered during evolution.

Questions for authors
  1. Could the authors provide intervention-specific experiments comparing recovery performance across different types of environmental disruptions?

  2. Could the authors conduct additional experiments using unseen intervention types, either through a leave-one-type-out setting or newly designed disruptions?

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

This paper introduces the idea of agent-world co-evolution and learning from failures has been explored in previous work. It generates counterfactual world changes that make the agent's original plan no longer work while keeping the task achievable. The agent learns from these challenges, and its failures are used to generate new challenges.

Critical strengths
  1. This paper addresses a practical problem of how agents adapt when the environment changes.
  2. This paper evaluates the proposed CCE-Mobility framework on various LLM backbones for the mobility agent.
Critical weaknesses
  1. One of the main contributions from this paper is the proposed world-side LLM challenge generator that generates disruptions to update the task/mobility agent with recovery memories. The ideas of residual-guided evolution and counterfactual recovery are straightforward in the given setup, where challenges are generated dynamically and failures are used to generate more challenges. Moreover, the ideas of learning skills from failures, counterfactual data generation, agent-world co-evolution have already been widely explored in existing agent improvements literature. Overall, the technical novelty of the proposed co-evolution framework is limited and hard to distinguished from existing approaches.

  2. Although the method improves recovery scores, these improvements do not always lead to better final task success, specifically the results show substantial improvements in recovery quality, but they struggle to translate into final task acceptance improvement. Why might this happen? It would be useful to analyze why some tasks still fail after recovery improves. Does this reveal a drawback of self-evolution? Moreover, the incremental improvement from world-side evolution over agent-only adaptation is only demonstrated on the validation set of TravelPlanner, while the corresponding MobilityBench results remain inconclusive. Stronger evidence would be needed to support the effectiveness and generalization capability of CCE-Mobility's world-side evolution mechanism.

  3. The paper stores recovery skills in external agent memory. During inference, the agent retrieves skills from interaction history. I am concerned about how this approach scales as more skills are learned. A larger memory may increase retrieval latency and make it harder to find the right skills.

  4. This work provides limited analysis of what skills the agent actually learns, how these skills evolve across training rounds.

Questions for authors
  1. Can you make figure 1 caption more interpretable, with more context and explanations, particular on how the two approaches differ in their recovery behaviors?
  2. The implementation details of the world-side challenge generator is fairly limited, could you please provide some more technical details on this?
Reviewer_oJ1tOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how mobility agents recover when the world changes after a plan has been made. CCE-Mobility uses an LLM world model to read an observed trajectory and identify a dependency behind the agent's decision. The agent continues from the same history, and its failed recoveries are distilled into recovery skills stored in an external memory. Failures that persist after this update are sent back to the world side to shape the next round of challenges. The method is evaluated on counterfactually augmented versions of TravelPlanner and MobilityBench with several backbones.

Critical strengths
  1. Recovering from world changes that happen after planning is practically important and understudied. Challenges are frozen and shared across methods, and recovery contracts are hidden from the agent. The mechanism analysis uses paired comparisons with bootstrap intervals.
  2. The paper is clear about which results are mixed or inconclusive, and it evaluates several backbones.
Critical weaknesses
  1. Both training and test challenges appear to come from the authors' own intervention pipeline, so the evaluation measures recovery on the same kind of change the agent learned from. No external dynamic benchmark is used, such as TripTide or CostBench, both of which the paper cites.
  2. The headline Recovery metric is assigned by an evaluator that scores against a contract derived by the same pipeline. The main paper does not say which model serves as this evaluator or how well it agrees with human judgment.
  3. Some frontier backbones score near zero on the official TravelPlanner pass metrics even without evolution, which suggests a setup issue and makes those comparisons hard to interpret.
  4. SkillOpt is close to or better than CCE-Mobility in several settings. Much of the gain may therefore come from agent-side skill learning on counterfactual data rather than from co-evolution, and it is unclear whether the baselines learn from the same challenges.
Questions for authors
  1. Do SkillOpt, Reflexion, and Voyager learn from the same counterfactual challenges as CCE-Mobility? How does CCE-Mobility perform on an external dynamic benchmark that its own pipeline did not build?
  2. Which model is the recovery evaluator, and how well does it agree with human judgment? Why do the Recovery gains not translate into higher counterfactual pass rates?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.