Paper deep dive
EmbodiedCLBench: Evaluating Continual Learning for Self-Evolving Embodied Agents
Jie Huang, Yanan Chen, Ruixun Liu, Kaichen He, Xinyi He, Rui Huang, Zilong Zheng, Zhenliang Zhang, Yiwu Zhong
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%
Last extracted: 10/12/2026, 6:47:12 AM
Summary
The paper introduces EmbodiedCLBench, the first benchmark for evaluating the continual learning capabilities of self-evolving embodied agents in simulation. Using compositional generalization, the benchmark assesses whether agents can transfer skills from basic tasks to solve complex, unseen advanced tasks. Experiments across multiple agent harnesses and models reveal that while zero-shot performance on advanced tasks is poor, experience from basic tasks leads to consistent improvements, though this improvement does not scale with the volume of learning samples.
Entities (13)
Relation Signals (13)
Jie Huang → authored → EmbodiedCLBench
confidence 99% · EmbodiedCLBench: EVALUATING CONTINUAL LEARNING FOR SELF-EVOLVING EMBODIED AGENTS Jie Huang ⋆,1,2...
Yiwu Zhong → authored → EmbodiedCLBench
confidence 99% · Yiwu Zhong 1,2,† ... † Corresponding author
EmbodiedCLBench → evaluates → Continual Learning
confidence 98% · EmbodiedCLBench, the first benchmark in embodied simulation for evaluating the continual learning capability of self-evolving agents.
EmbodiedCLBench → basedon → Compositional Generalization
confidence 96% · With compositional generalization as the core evaluation principle, our benchmark evaluates whether agents can first learn from basic tasks and then recombine the acquired skills
EmbodiedCLBench → uses → TongSim
confidence 95% · We thus adopt TongSim as the simulation platform... It provides a controllable environment that supports rich physical interactions
Skill Learning → usedin → EmbodiedCLBench
confidence 93% · We conduct extensive evaluation across multiple agent harnesses and learning methods... Skill Learning...
In-Context Learning → usedin → EmbodiedCLBench
confidence 93% · We conduct extensive evaluation across multiple agent harnesses and learning methods... In-Context Learning (ICL)...
→ →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Continual learning requires models to accumulate past experience and transfer it to new tasks. Despite substantial research on continual learning, existing work remains largely confined to non-embodied settings, without extending to embodied environments. To address this gap, we introduce EmbodiedCLBench, the first benchmark in embodied simulation for evaluating the continual learning capability of self-evolving agents. With compositional generalization as the core evaluation principle, our benchmark evaluates whether agents can first learn from basic tasks and then recombine the acquired skills to solve complex, unseen advanced tasks. Based on EmbodiedCLBench, we conduct extensive experiments across multiple agent harnesses and learning methods, revealing consistent limitations and distinctive behaviors. Our results show that while advanced tasks can be hardly solved under zero-shot setting, the experience from basic tasks enables consistent improvement. Such benefit from the experience remains stable even if the task complexity increases. However, this improvement does not scale accordingly with the number of learning samples, and current agents cannot extract meaningful information from additional experience. Collectively, our findings highlight continual learning as a fundamental yet underexplored capability, and our benchmark offers valuable resources for designing effective self-evolving agents in embodied environments. Our code and dataset are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.13021v1
- Canonical: https://arxiv.org/abs/2610.13021v1
Trouble viewing inline? Open PDF directly →
Full Text
71,609 characters extracted from source content.
Expand or collapse full text
EmbodiedCLBench: EVALUATING CONTINUAL LEARN- ING FOR SELF-EVOLVING EMBODIED AGENTS Jie Huang ⋆,1,2 Yanan Chen ⋆,1 Ruixun Liu ⋆,1 Kaichen He 1 Xinyi He 1 Rui Huang 3 Zilong Zheng 2 Zhenliang Zhang 2 Yiwu Zhong 1,2,† ⋆ Equal contribution † Corresponding author 1 School of Intelligence Science and Technology, Peking University 2 State Key Laboratory of General Artificial Intelligence, BIGAI, Beijing, China 3 University of Hong Kong ABSTRACT Continual learning requires models to accumulate past experience and transfer it to new tasks. Despite substantial research on continual learning, existing work remains largely confined to non-embodied settings, without extending to embod- ied environments. To address this gap, we introduce EmbodiedCLBench, the first benchmark in embodied simulation for evaluating the continual learning capabil- ity of self-evolving agents. With compositional generalization as the core eval- uation principle, our benchmark evaluates whether agents can first learn from basic tasks and then recombine the acquired skills to solve complex, unseen ad- vanced tasks. Based on EmbodiedCLBench, we conduct extensive experiments across multiple agent harnesses and learning methods, revealing consistent lim- itations and distinctive behaviors. Our results show that while advanced tasks can be hardly solved under zero-shot setting, the experience from basic tasks en- ables consistent improvement. Such benefit from the experience remains stable even if the task complexity increases. However, this improvement does not scale accordingly with the number of learning samples, and current agents cannot ex- tract meaningful information from additional experience. Collectively, our find- ings highlight continual learning as a fundamental yet underexplored capability, and our benchmark offers valuable resources for designing effective self-evolving agents in embodied environments. Our code and dataset are available at https: //pku-value-lab.github.io/EmbodiedCLBench-homepage. 1INTRODUCTION Self-evolving agents (Gao et al., 2025; Fang et al., 2025) can accumulate experience from past interactions and progressively enhance their task performance without additional human supervi- sion. Such agents are capable of retaining and transferring skills across tasks over time, a capability broadly termed continual learning (Zheng et al., 2025b). This capability is particularly important for embodied intelligence (Liu et al., 2025a; Yifan et al., 2025; Liu et al., 2025b), where agents are expected to operate in complex, ever-changing physical environments that cannot be fully antici- pated at training time. It is therefore important to develop a principled approach for evaluating the capability to learn from prior interactions and transfers the acquired skills to new tasks. Existing benchmarks in continual learning and embodied intelligence partially address this evalua- tion goal. Research on continual learning has largely focused on non-embodied settings, including knowledge-intensive tasks (Gao et al., 2026), interactive digital environments (Jiang et al., 2026), and long-term simulated worlds (Cai et al., 2025). These platforms do not capture the spatial com- plexity and physical constraints inherent in embodied tasks. Meanwhile, embodied benchmarks evaluate capabilities such as scene perception (Wang et al., 2024), visual navigation (Anderson et al., 2018), robotic manipulation (Gu et al., 2023), and long-horizon task planning (Shridhar et al., 2020), but are not specifically designed to assess continual learning. Therefore, there remains a lack of a dedicated benchmark for evaluating whether self-evolving agents can accumulate experience and transfer it to solve novel tasks in embodied environments. 1 arXiv:2610.13021v1 [cs.AI] 9 Oct 2026 Basic TasksContinual Learning Advanced Tasks (Unseen, Compositional) 1 Pick up object 2 Walk to object 3 Open drawer 1 Clean up the room 2 Set the table StartWalk toPick upOpenGoal StartWalk toOpenPick upGoal Experience Accumulation Practice and refine basic tasks through interaction. Continual Learning Leverage the learned experience to generalize to novel, harder tasks. Advanced Task Solving Solve unseen advanced tasks by combining learned basic tasks. Figure 1: Compositional generalization is leveraged as a principled way to measure continual learning capability. An agent that has experience from basic tasks should recombine previously ac- quired skills to solve novel, unseen advanced tasks, rather than learning each new task from scratch. To address this gap, we introduce EmbodiedCLBench, the first benchmark in embodied simulation for evaluating the continual learning capability of self-evolving agents. As illustrated in Figure 1, our design stems from compositional generalization (Furuta et al., 2023; Guo et al., 2025). Instead of learning every new task from scratch, a self-evolving agent should recombine previously acquired skill elements to solve novel, unseen tasks. We thus consider compositional generalization as a principled way to measure agent continual learning capability. Specifically, we first assign a set of basic tasks for the agent to execute and learn the skills. We then test whether the agent can effectively leverage such experience to solve unseen advanced tasks that require recombining the learned skills. From this perspective, basic tasks serve as test-time learning (Sun et al., 2020; Liu et al., 2021) data for the agent to build experience before attempting advanced tasks. To this end, EmbodiedCLBench organizes tasks into three compositional levels: (1) atomic actions from TongSim (Sun et al., 2025), representing basic object–action interactions; (2) basic tasks, with each composing a small number of atomic actions into practical household operations across multiple categories; and (3) advanced tasks, with each combining elements from multiple basic-task categories. Based on EmbodiedCLBench, we conduct extensive evaluation across multiple agent harnesses (e.g., OPENCLAW and CODEX) and learning methods (In-Context Learning and Skill Learning). Our re- sults reveal several notable and, in some cases, unexpected findings. First, while advanced tasks can hardly be solved in the zero-shot setting, agents with experience from related basic tasks achieve sub- stantially higher success rates. This validates that experience from simpler tasks can be effectively transferred to unseen, harder tasks. Second, the performance gain over zero-shot agents remains at a stable level as task complexity increases. This suggests that continual learning capability is not sensitive to task difficulty. Third, the improvement from learning does not scale accordingly with the number of learning samples. A single basic task significantly improves performance, and additional samples yield no consistent further improvement. Fourth, past experience facilitates the decision behavior of agents by selecting more meaningful actions with fewer unexecutable ones. Our contributions are summarized as follows: • We introduce EmbodiedCLBench, the first benchmark in embodied simulation for evalu- ating the continual learning capability of self-evolving agents. • Our key innovation lies in leveraging compositional generalization as a principled way to construct tasks, design evaluation protocols, and evaluate continual learning capability. • We conduct a comprehensive evaluation across multiple agent harnesses and learning methods, revealing limitations of current agents and providing insights for future design. 2 2RELATED WORK Continual Learning Benchmarks. Evaluating continual learning capabilities has become an in- creasingly active topic for LLMs and agents (Gao et al., 2025; Fang et al., 2025; Zheng et al., 2025b). Existing benchmarks span several distinct domains. In knowledge-intensive tasks, prior work evaluates whether agents can effectively transfer procedural experience across coding, reason- ing, and question-answering tasks (Gao et al., 2026; Wu et al., 2024; Shu et al., 2026). In interactive digital environments, benchmarks test continual learning capabilities in web automation, operating systems, and knowledge graphs (Jiang et al., 2026; Zheng et al., 2025a). In long-term simulated environments, benchmarks model persistent virtual worlds where past behaviors continuously shape future states and decisions (Cai et al., 2025). Another major direction focuses on skill-centric evolu- tion, studying whether agents can abstract reusable skills from experience and transfer them across tasks (Zhong et al., 2026; Zhang et al., 2026; Guan et al., 2026; Lei et al., 2026). Despite this progress, these benchmarks primarily operate in digital or symbolic environments, leaving contin- ual learning in embodied scenarios largely unexplored. Embodied AI Benchmarks. Embodied AI benchmarks evaluate the capabilities ranging from per- ception to long-horizon task execution (Liu et al., 2025a; Yifan et al., 2025; Liu et al., 2025b). At the perception level, benchmarks test 3D scene understanding (Wang et al., 2024), spatial reason- ing (Du et al., 2024), physical world understanding (Chow et al., 2025), and scene dynamics (Wang et al., 2025). For navigation, the field has progressed from language-conditioned navigation (An- derson et al., 2018; Ku et al., 2020) to sound-guided navigation Liu et al. (2026), object ground- ing (Qi et al., 2020), multi-modal lifelong navigation (Khanna et al., 2024), and active exploration for question answering (Majumdar et al., 2024). In manipulation, benchmarks evaluate multi-task generalization (Yu et al., 2020; James et al., 2020), object-level skill transfer (Gu et al., 2023), life- long knowledge transfer (Liu et al., 2023), and everyday task scalability (Nasiriany et al., 2024). For long-horizon planning, benchmarks require agents to follow compositional instructions (Shridhar et al., 2020), interact through dialogue (Padmakumar et al., 2021), compose language-conditioned skills (Mees et al., 2022), handle diverse household activities (Li et al., 2023), and adapt plans based on visual feedback (Tupini et al., 2026). These benchmarks comprehensively cover perception, nav- igation, manipulation, and planning. Recent works further explore general-purpose embodied agent evaluation and realistic simulation environments (Peng et al., 2024; Ren et al., 2025). However, none is designed for evaluating the continual learning capability of skills in long-horizon tasks. Compositional Generalization. Compositional generalization, the ability to recombine learned elements for solving novel problems, has been studied across multiple domains. In digital envi- ronments, benchmarks test whether mobile and web agents can compose atomic operations into multi-step tasks (Guo et al., 2025; Furuta et al., 2023). In video understanding, benchmarks study whether models can recognize unseen action–object combinations from separately learned compo- nents (Li et al., 2024; Ahn et al., 2026). For reasoning tasks, benchmarks show that models proficient in individual abilities still fail when these abilities must be combined (Alazraki et al., 2026; Li et al., 2025). In embodied manipulation, benchmarks evaluate whether robot policies can generalize to un- seen skill compositions after training on atomic skills (Wu et al., 2026; Chen et al., 2025a;b; Haresh et al., 2024). In contrast, EmbodiedCLBench evaluates the continual learning capability in embodied simulation without any benchmark-specific training. 3BENCHMARK CONSTRUCTION As illustrated in Figure 2, the construction of EmbodiedCLBench consists of four stages: (1) envi- ronment setup, (2) task generation, (3) graph construction, and (4) evaluation protocol. We introduce each stage in the following subsections. 3.1ENVIRONMENT SETUP: EMBODIED SIMULATION Evaluating compositional generalization requires a large combinatorial space of object–action pairs. We thus adopt TongSim as the simulation platform (Sun et al., 2025). It provides a controllable environment that supports rich physical interactions and scales to diverse task configurations. It was also used in recent embodied agent evaluation studies (He et al., 2026; Zhao et al., 2026). 3 A. A. Open cabinet Goal B. Look at lamp S₂ C. Switch on TV Distractor State: Near cabinet S₁ ... Task: Open the cabinet. Init State: Far from cabinet S₀ ... Far from cabinet Near cabinet Looking at lamp Cabinet open S₀ S₁ S₂ Goal walk to cabinet look at lamp open cabinet Embodied Simulation A controllable environment that supports rich physical interactions. Interactive Objects ... chairplateTV Executable Actions ... walk topick upopen Set the table Open drawer Task Generation Basic Task Walk to object Advanced Task Start Walk to Pick up Goal ... Evaluation Graph Construction Figure 2: Overview of the construction process of EmbodiedCLBench. During task generation, basic tasks connect a few atomic actions into simple household operations, and advanced tasks com- bine elements from multiple basic tasks into compositional tasks. For each task, a state-transition graph is built, with each graph node as an intermediate state and paired with action options and a first-person-view image. During evaluation, the agent observes the current state and selects from candidate actions. After each selection, the agent transitions to the next node on the graph. Specifically, TongSim is a simulation with household scenarios and supports a wide range of inter- active objects (such as cups, plates, cabinets, lamps, and appliances) and executable actions (such as walk to, look at, pick up, put down, open, close, and switch on, etc.). This simulation offers three key advantages: (1) the platform supports a large number of object types and action primitives, providing sufficient combinatorial diversity; (2) all scenes are indoor domestic environments, so that the tasks are grounded in meaningful, real-world household activities; and (3) TongSim has not been widely used as training data for existing foundation models, making it suitable for evaluating the generalization capability of agents with reduced risk of data contamination. 3.2TASK GENERATION: COMPOSITIONAL GENERALIZATION Our goal is to evaluate the capability of continual learning. However, it is an abstract concept that is challenging to measure directly. We operationalize it through compositional generalization: if an agent has learned from basic tasks, it should recombine previously mastered skill elements to solve novel tasks rather than learning each new task from scratch. This leads to a natural evaluation protocol—let the agent first complete simpler tasks to accumulate experience, then test whether it can leverage that experience to solve more complex, unseen tasks. We design two levels of tasks based on atomic actions available in TongSim. Specifically, an atomic action is a single interaction with an object (e.g., “pick up the cup”). It consists of a sequence of steps (e.g., walking to the cup, looking at it, and picking it up), each of which is one executable action. Building on it, a basic task composes a small number of atomic actions into a practical household operation (e.g., “put the cup on the table”, which consists of “pick up the cup” and “put down the cup on the table”). Further, an advanced task composes more atomic actions than a basic task (e.g., “tidy up the room”, which requires picking up the book, putting the book on the shelf, closing the cabinet, and switching off the lamp). We provide more details in the benchmark overview (Section 4.1). All task descriptions are generated by Qwen3.7-Plus and manually verified for correctness (see Appendix B). 3.3GRAPH CONSTRUCTION: ACTION TRAJECTORY ANNOTATION For TongSim simulation, we chose its closed-source version, which provides more diverse actions than the open-source version. As a result, online simulation is not practical for the benchmark evaluation. Therefore, we construct an offline, reproducible evaluation process in the form of a state- 4 transition graph, as shown in Figure 2. A graph structure naturally represents the task procedure: at each state, an agent selects an action and transitions to the next state. This graph structure offers three advantages: (1) offline evaluation can be run in parallel without running the simulator; (2) flexible control over the task complexity through graph branching and depth; and (3) fine-grained analysis of agent behavior by checking each decision against the pre-computed graph. Specifically, each graph node represents a unique environment state (e.g., “near the desk and look- ing at the lamp”). It is paired with an image captured from TongSim and a set of action candidates presented to the agent during evaluation. These candidates are generated by Qwen3.7-Plus and manually verified for correctness. Each executable candidate corresponds to an edge that determin- istically transitions to the next state. Given a task, we first define its initial and goal states and then use breadth-first search to find the shortest path between them. Since each state supports only a bounded set of objects and actions, the reachable state space is compact enough for reliable graph search. We construct the distractor paths by expanding alternative executable actions from nodes along the shortest path. These paths are executable but deviate from the shortest path. The detailed construction pipeline and task examples are provided in Appendices B and C. 3.4EVALUATION PROTOCOL With the task graphs built, we define how agents are evaluated on them. The evaluation consists of a learning phase and a testing phase. In the learning phase, the agent completes basic tasks to accumulate experience. In the testing phase, the agent attempts unseen advanced tasks with this experience. In this setup, basic tasks serve as the memory (Xu et al., 2026; Huang et al., 2026) and test-time learning data (Sun et al., 2020; Liu et al., 2021) for the self-evolving agents. For each task, the evaluation follows an interactive, multi-turn format: (1) the agent receives a task description and an image of the initial state; (2) at each step, the agent selects one action from a set of action candidates; (3) based on the action decision, the agent transitions to the next state and receives a new image with updated objects and action candidates; (4) this repeats until the agent reaches the goal (success) or exhausts the step budget (failure). The action candidates at each step are designed into three types: optimal actions that follow the shortest path, suboptimal actions that are executable but deviate from the shortest path, and invalid actions that cannot be executed in the current state. This interactive format preserves the sequential decision-making structure of embodied tasks while enabling fully offline, deterministic, and reproducible evaluation. 4MAIN EXPERIMENTS In this section, we first describe the experimental setup, including the benchmark overview, evalua- tion settings, agent configurations, and metrics. We then report and discuss the main results. 4.1EXPERIMENTAL SETUP Benchmark overview. Figure 3 summarizes the overall statistics of EmbodiedCLBench. The en- vironment covers diverse scenes, object categories, and atomic actions, providing a broad combi- (a) Environment(b) Task Composition(c) Task Complexity 6 Scenes 2 CameraHeights 51 Object Categories 7 AtomicActions Atomic Actions: pick upplace onplace inopen closeswitch onswitch off Basic Tasks 168 basic tasks 28 28 2828 28 28 Object Manipulation Spatial Placement Accessibility Control Appliance Operation Activity-Driven Scene Configuration Household State Restoration Advanced Tasks 100 advanced tasks Composed by combining 2–4 basic categories. 0246 Basic Tasks Advanced Tasks Number of Atomic Actions 051015 Basic Tasks Advanced Tasks Number of Steps Basic TasksAdvanced Tasks 2 4 6 12 0246 051015 Shortest-Path Steps Atomic Actions per Task Figure 3: Benchmark statistics of EmbodiedCLBench: environment configuration (a), task compo- sition across six categories (b), and task complexity at two levels (c). 5 natorial space for task construction. The benchmark contains 268 tasks: 168 basic tasks and 100 advanced tasks. Basic tasks span six categories, with each grounded in a distinct type of daily house- hold activity. Advanced tasks require substantially more atomic actions and shortest-path steps. The atomic actions of each advanced task come from two to four categories of basic tasks. Each atomic action consists of 3 steps. Therefore, the shortest path has 6 steps for a basic task (2 atomic actions) and 12 steps for an advanced task (4 atomic actions). With this scale, real-world grounding, and structured difficulty, EmbodiedCLBench can reveal meaningful differences in the continual learning capability of agents. Examples of basic and advanced tasks are provided in Appendix C. Evaluation settings. We evaluate agents under three settings: (1) Zero-shot: the agent performs basic and advanced tasks directly, without a learning phase. (2) In-Context Learning (ICL): in the learning phase, the agent completes two related basic tasks for each advanced task. In the testing phase, the full interaction trajectories of these basic tasks are retained in context. (3) Skill Learning: in the learning phase, the agent completes the same basic tasks as in ICL. It then writes a concise skill summary of its experience in Markdown format. In the testing phase, only this skill summary is provided in context, instead of the interaction trajectories. The step budget of each task is the length of its shortest path plus extra steps. In the testing phase, we allow 3 extra steps as a small tolerance for suboptimal decisions. In the learning phase, we allow 6 extra steps for more exploration. Agent configurations. We evaluate all tasks under each of four agent harnesses: OPEN- CLAW (OpenClaw, 2026), CLAUDE CODE (Anthropic, 2026), CODEX (OpenAI, 2026b), and HER- MES (Nous Research, 2026). We pair each harness with two backbone models through their official APIs: Qwen3.8-Flash (Qwen, 2026) and GPT5.6-Terra (OpenAI, 2026a). Table 1: Main results on EmbodiedCLBench. For In-Context Learning and Skill Learning, agents first learn from two basic tasks and then attempt one advanced task. Success Rate (%)BASIC TASKADVANCED TASK HarnessModelZero-shotZero-shotIn-Context LearningSkill Learning HERMES Qwen3.8-Flash58.98.044.033.0 GPT5.6-Terra53.011.016.034.0 OPENCLAW Qwen3.8-Flash50.65.035.038.0 GPT5.6-Terra47.612.039.022.0 CODEX Qwen3.8-Flash52.44.034.033.0 GPT5.6-Terra54.811.052.026.0 CLAUDE CODE Qwen3.8-Flash42.93.027.031.0 GPT5.6-Terra51.89.031.025.0 Metrics. We report Success Rate (SR), which measures whether an agent can complete a task within the allotted step budget: SR = 1 N N X i=1 c i , c i ∈ 0, 1 indicates whether task i is completed.(1) Success rate directly reflects continual learning capability—whether accumulated experience en- ables the agent to solve novel, compositionally harder tasks. 4.2RESULTS Table 1 presents the main results. Agents achieve a moderate zero-shot success rate on basic tasks (around 50%), indicating that current agents can execute simple household tasks. Zero-shot success rate on advanced tasks drops to around 10%, validating that advanced tasks are genuinely difficult. Both ICL and Skill Learning substantially improve advanced task success rate, raising it from around 6 10% to above 30%. This improvement is consistent across all harnesses, suggesting that experience from basic tasks can effectively transfer to unseen advanced tasks. Finding 1: While advanced tasks are hardly solvable under zero-shot setting, experience from basic tasks enables consistent improvement across different harnesses and models. Discussion. Between ICL and Skill Learning, there is no consistent advantage of one over the other. We conjecture that both methods can extract meaningful information from the agent experi- ence. Besides, the performance across harnesses is close yet far from saturation, suggesting great potential for improvement. Designing harnesses for continual learning can be a future direction. 5ANALYSIS 5.1EFFECTS OF COMPOSITIONAL COMPLEXITY Motivation and Setting. Compositional generalization is the core approach for evaluating con- tinual learning in EmbodiedCLBench. A natural question follows: does more complex composition make continual learning harder? In the main experiment, advanced tasks consist of 4 atomic ac- tions. Here, we extend our evaluation data to advanced tasks with 3 to 6 atomic actions to control compositional complexity. We measure both the raw success rate at each level and the gain over zero shot, defined as the success rate of In-Context Learning or Skill Learning minus the zero-shot success rate. We use HERMES with Qwen3.8-Flash for this analysis. Figure 4: Effect of compositional complexity on continual learning. The x-axis varies the number of atomic actions per advanced task from 3 to 6. (a) reports the raw success rate achieved by each setting. (b) reports the gain over zero shot, defined as the success rate of ICL or Skill Learning minus that of zero shot, isolating the improvement attributed to the learning phase. Results and Analysis. As shown in Figure 4(a), success rate decreases across all three settings as atomic actions increase from 3 to 6. The largest drop occurs between 3 and 4 atomic actions (e.g., ICL falls from 72.0% to 42.0%), after which the success rate levels off. In the zero-shot set- ting, success rate decreases sharply (e.g., from 28.0% to 4.0%). In contrast, ICL and Skill Learning remain more stable in relative terms. Specifically, Figure 4(b) shows that the gain over zero shot drops initially but then stabilizes at around 30%. This indicates that experience from basic tasks re- mains helpful as compositional complexity increases. At the highest complexity (6 atomic actions), zero-shot success rate drops to 4.0%, while ICL and Skill Learning still achieve 32.0% and 26.0%, respectively. This shows that experience from basic tasks still brings a substantial gain over zero shot, even when zero-shot agents can barely solve the tasks. Finding 2: In zero-shot setting, success rate decreases sharply as complexity increases. In contrast, the gain over zero-shot setting remains stable, indicating that the benefit of continual learning is less sensitive to task difficulty. 7 5.2SCALING WITH THE NUMBER OF LEARNING SAMPLES Motivation and Setting. Continual learning assumes that accumulated experience improves an agent’s performance on harder tasks. If this holds, more experience should lead to a higher success rate on advanced tasks. In the main experiment, agents learn from 2 basic tasks before attempting advanced tasks. Here, we increase the number of learning samples (i.e., basic tasks in the learning phase) from k=0 (zero shot) to k=6. We measure the success rate on the same set of advanced tasks under both In-Context Learning and Skill Learning settings. We evaluate two agent harnesses, HERMES and OPENCLAW, both using Qwen3.8-Flash as the backbone model. Figure 5: Success rate versus the number of learning samples. Each learning sample is a basic task that the agent completes before being tested on advanced tasks. Results and Analysis. As shown in Figure 5, the most substantial improvement occurs between k=0 and k=1. A single learning sample raises the success rate from 8.0%–10.0% to 28.0%– 40.0% across both harnesses and learning meth- ods. This initial jump accounts for most of the total improvement across the k range.Even minimal experience (completing just one basic task) provides sufficient new information for the agent to substantially improve on unseen ad- vanced tasks. Beyond k=1, however, success rate does not continue to rise but instead fluctuates without a clear upward trend. This pattern re- mains across ICL and Skill Learning, and across both harnesses. None of the four curves shows consistent improvement with additional learning samples. This suggests that current agents can- not extract sufficiently distinct information from additional learning samples. Therefore, scalable continual learning, where success rate keeps im- proving as learning samples increase, remains an open challenge for embodied agents. Finding 3: Success rate does not scale accordingly with the number of learning samples. A single learning sample significantly improves performance, while additional samples yield no consistent further gain. 5.3ERROR ANALYSIS Motivation and Setting. The main results show that ICL and Skill Learning improve success rate, but do not reveal what agents actually learn from experience. As defined in Section 3.4, the action candidates at each step fall into three types: optimal actions that follow the shortest path, suboptimal actions that are executable but deviate from the shortest path, and invalid actions that cannot be executed in the current state. We analyze how agents choose among these three action types. Instead of running new experiments, we use the selected actions recorded at each decision step in the main experiment (all four harnesses with Qwen3.8-Flash). Specifically, we compare the selection rate of each action type under ICL and Skill Learning with that in the zero-shot setting. Figure 6: Error analysis of decision behavior. Each group of bars shows the change in selection rate relative to the zero-shot setting for three action types: optimal actions follow the shortest path, suboptimal actions are executable but deviate from the shortest path, and invalid actions cannot be executed in the current state. 8 Results and Analysis. As shown in Figure 6, experience sharply reduces invalid actions across all harnesses (by 13%–20%). Fewer invalid actions allow agents to stay on executable trajectories and reach the goal within the step budget. This suggests that the reduction in invalid actions contributes to the higher success rates in Table 1. Meanwhile, optimal actions increase substantially. Suboptimal actions also increase slightly (by 3%–8%). Overall, experience shifts the selections of agents from invalid actions to optimal and suboptimal actions. ICL and Skill Learning show the same direction of change, but Skill Learning consistently produces smaller shifts across all harnesses. In contrast, Table 1 shows no consistent advantage between ICL and Skill Learning in success rate. This may reflect a difference in how experience is represented. Interaction trajectories may induce larger changes in decision behavior, while skill summaries may induce smaller but more focused changes. Finding 4: Continual learning alters the decision behavior of agents, resulting in an increase in optimal actions and a decrease in infeasible actions. 5.4STEP BUDGET SENSITIVITY Motivation and Setting. Each task has a shortest path to the goal, but agents often make subopti- mal decisions along the way. As described in Section 4.1, agents are allowed 3 extra steps by default as a small tolerance for suboptimal decisions. Here, we vary the number of extra steps from 0 to 6. We examine how the step budget affects success rate under different evaluation settings. We use HERMES with Qwen3.8-Flash for this analysis. Figure 7: Step budget sensitivity. The x-axis shows the number of extra steps allowed beyond the shortest path. By default, agents are allowed 3 extra steps in the testing phase. Results and Analysis. As shown in Figure 7, all three settings show a monotonically increas- ing success rate as the budget grows from 0 to 6 extra steps. This is expected, since a larger bud- get gives agents more room for suboptimal deci- sions. Across all budget values, the ranking of the three settings remains consistent. ICL achieves the highest success rate, followed by Skill Learn- ing and then zero shot. This indicates that the gain over zero shot holds across different budgets. At the strictest budget (0 extra steps), zero-shot suc- cess rate is 0.0%, while ICL and Skill Learning still achieve 15.0% and 7.0%, respectively. Even under this strictest budget, learning from experi- ence remains effective. Notably, compared to the zero-shot setting, ICL and Skill Learning exhibit a steeper increase in success rate as the budget grows. This indicates that the agents with prior experience can better benefit from additional tol- erance. Meanwhile, the gap between experienced agents and zero-shot agents remains substantial even at the most generous budget (6 extra steps). This suggests that additional tolerance cannot replace the role of prior experience. Finding 5: As the number of extra steps increases, success rate improves across all settings, and continual learning methods yield a steeper increase compared to the zero-shot setting. 6CONCLUSION We introduce EmbodiedCLBench, the first embodied simulation benchmark for evaluating the con- tinual learning capability of self-evolving agents. Our key idea is to formulate continual learning through compositional generalization: rather than solving each new task from scratch, an agent should accumulate experience from basic tasks and recombine previously acquired skills to solve un- seen, more complex tasks. Following this principle, EmbodiedCLBench organizes tasks into atomic actions, basic tasks, and advanced tasks, providing a structured protocol for measuring whether embodied agents can acquire, retain, and transfer experience over time. 9 Extensive evaluations across multiple agent harnesses and learning methods reveal both promising progress and clear limitations: (1) Experience from basic tasks substantially improves performance on unseen advanced tasks over zero-shot execution, showing that current agents can transfer previ- ously acquired knowledge to new compositions. (2) However, the performance of existing agents is far from saturation. Designing effective continual learning methods for embodied agents is a promising direction to explore. (3) Further, the benefit of learning does not consistently increase with the amount of experience. How to enable the agents to accumulate and exploit additional in- teractions is still an open problem. Overall, EmbodiedCLBench provides a principled testbed for studying self-evolution in embodied agents and highlights the need for stronger mechanisms that can continuously consolidate, generalize, and reuse experience across tasks. REFERENCES Geo Ahn, Inwoong Lee, Taeoh Kim, Minho Shim, Dongyoon Wee, and Jinwoo Choi. Why Can’t I Open My Drawer? Mitigating Object-Driven Shortcuts in Zero-Shot Compositional Action Recognition, July 2026. arXiv:2601.16211 [cs.CV]. Lisa Alazraki, Lihu Chen, Ana Brassard, Joe Stacey, Hossein A Rahmani, and Marek Rei. Agent- coma: A compositional benchmark mixing commonsense and mathematical reasoning in real- world scenarios. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p. 8383–8410, 2026. Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko S ̈ underhauf, Ian Reid, Stephen Gould, and Anton van den Hengel. Vision-and-Language Navigation: Interpreting Visually-Grounded Navigation Instructions in Real Environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. Anthropic. Claude Code, 2026. URL https://claude.com/product/claude-code. Yuxuan Cai, Yipeng Hao, Jie Zhou, Hang Yan, Zhikai Lei, Rui Zhen, Zhenhua Han, Yutao Yang, Junsong Li, Qianjun Pan, Tianyu Huai, Qin Chen, Xin Li, Kai Chen, Bo Zhang, Xipeng Qiu, and Liang He. Building Self-Evolving Agents via Experience-Driven Lifelong Learning: A Frame- work and Benchmark. arXiv preprint arXiv:2508.19005, 2025. Yangtao Chen, Zixuan Chen, Nga Teng Chan, Junting Chen, Junhui Yin, Jieqi Shi, Yang Gao, Yong-Lu Li, and Jing Huo. RoboHiMan: A Hierarchical Evaluation Paradigm for Compositional Generalization in Long-Horizon Manipulation. arXiv preprint arXiv:2510.13149, 2025a. Zixuan Chen, Junhui Yin, Yangtao Chen, Jing Huo, Pinzhuo Tian, Jieqi Shi, Yiwen Hou, Yinchuan Li, and Yang Gao. DeCo: Task Decomposition and Skill Composition for Zero-Shot Generaliza- tion in Long-Horizon 3D Manipulation. arXiv preprint arXiv:2505.00527, 2025b. Wei Chow, Jiageng Mao, Boyi Li, Daniel Seita, Vitor Guizilini, and Yue Wang. PhysBench: Benchmarking and Enhancing Vision-Language Models for Physical World Understanding. arXiv preprint arXiv:2501.16411, 2025. Mengfei Du, Binhao Wu, Zejun Li, Xuanjing Huang, and Zhongyu Wei. EmbSpatial-Bench: Benchmarking Spatial Understanding for Embodied Tasks with Large Vision-Language Mod- els. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), p. 346–355. Association for Computational Linguistics, 2024. doi: 10.18653/v1/2024.acl-short.33. Jinyuan Fang, Yanwen Peng, Xi Zhang, Yingxu Wang, Xinhao Yi, Guibin Zhang, Yi Xu, Bin Wu, Siwei Liu, Zihao Li, Zhaochun Ren, Nikos Aletras, Xi Wang, Han Zhou, and Zaiqiao Meng. A Comprehensive Survey of Self-Evolving AI Agents: A New Paradigm Bridging Foundation Models and Lifelong Agentic Systems. arXiv preprint arXiv:2508.07407, 2025. Hiroki Furuta, Yutaka Matsuo, Aleksandra Faust, and Izzeddin Gur.Exposing Limitations of Language Model Agents in Sequential-Task Compositions on the Web.arXiv preprint arXiv:2311.18751, 2023. 10 Huan-ang Gao, Jiayi Geng, Wenyue Hua, Mengkang Hu, Xinzhe Juan, Hongzhang Liu, Shilong Liu, Jiahao Qiu, Xuan Qi, Yiran Wu, Hongru Wang, Han Xiao, Yuhang Zhou, Shaokun Zhang, Jiayi Zhang, Jinyu Xiang, Yixiong Fang, Qiwen Zhao, Dongrui Liu, Qihan Ren, Cheng Qian, Zhenhailong Wang, Minda Hu, Huazheng Wang, Qingyun Wu, Heng Ji, and Mengdi Wang. A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence. arXiv preprint arXiv:2507.21046, 2025. Xingze Gao, Chuanrui Hu, Hongda Chen, Pengfei Yao, Zhao Wang, Yi Bai, Zhengwei Wu, Yunyun Han, Xiaofeng Cong, Jie Gui, Yafeng Deng, and Teng Li. EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer. arXiv preprint arXiv:2607.05202, 2026. Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. ManiSkill2: A Unified Benchmark for Generalizable Manipulation Skills. In International Conference on Learning Representations, 2023. Tianyi Guan, Yiding Wang, Haotong Yang, Siyuan Cao, Shirui Liu, Yi Hu, Jiaqi Li, and Muhan Zhang. ContinualSkillBench: Can LLM Agents Truly Evolve Their Capabilities? arXiv preprint arXiv:2608.03874, 2026. Yuan Guo, Tingjia Miao, Zheng Wu, Pengzhou Cheng, Ming Zhou, and Zhuosheng Zhang. Atomic- to-Compositional Generalization for Mobile Agents with A New Benchmark and Scheduling Sys- tem. arXiv preprint arXiv:2506.08972, 2025. Sanjay Haresh, Daniel Dijkman, Apratim Bhattacharyya, and Roland Memisevic. ClevrSkills: Com- positional Language And Visual Reasoning in Robotics. In Advances in Neural Information Pro- cessing Systems, volume 37, p. 38235–38266, 2024. doi: 10.52202/079017-1209. Xinyi He, Ying Yang, Chuanjian Fu, Sihan Guo, Songchun Zhu, Lifeng Fan, Zhenliang Zhang, and Yujia Peng. Automatic Cognitive Task Generation for In-Situ Evaluation of Embodied Agents. arXiv preprint arXiv:2602.05249, 2026. Jie Huang, Ruixun Liu, Sirui Sun, Xinyi Yang, Yin Li, Yixin Zhu, and Yiwu Zhong. M 3 eval: Multi-modal memory evaluation through cognitively-grounded video tasks.arXiv preprint arXiv:2606.05008, 2026. Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. RLBench: The Robot Learning Benchmark & Learning Environment. IEEE Robotics and Automation Letters, 5(2): 3019–3026, April 2020. doi: 10.1109/lra.2020.2974707. Sihang Jiang, Lipeng Ma, Zhonghua Hong, Keyi Wang, Zhiyu Lu, Tengfei Wang, Shisong Chen, Jinghao Zhang, Tianjun Pan, Weijia Li, Jiaqing Liang, and Yanghua Xiao.SEA-Eval: A Benchmark for Evaluating Self-Evolving Agents Beyond Episodic Assessment. arXiv preprint arXiv:2604.08988, 2026. Mukul Khanna, Ram Ramrakhya, Gunjan Chhablani, Sriram Yenamandra, Theophile Gervet, Matthew Chang, Zsolt Kira, Devendra Singh Chaplot, Dhruv Batra, and Roozbeh Mottaghi. GOAT-Bench: A Benchmark for Multi-Modal Lifelong Navigation.In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 16373–16383, 2024. Alexander Ku, Peter Anderson, Roma Patel, Eugene Ie, and Jason Baldridge. Room-Across- Room: Multilingual Vision-and-Language Navigation with Dense Spatiotemporal Grounding. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), p. 4392–4412. Association for Computational Linguistics, 2020. doi: 10.18653/v1/ 2020.emnlp-main.356. Yingtie Lei, Zhongwei Wan, Jiankun Zhang, Samiul Alam, Zixuan Zhong, Peizhou Huang, Xin Wang, Jingxuan Zhang, Donghao Zhou, Yunta Hsieh, et al. Skillevolbench: Benchmarking the evolution from episodic experience to procedural skills. arXiv preprint arXiv:2605.24117, 2026. 11 Chengshu Li, Ruohan Zhang, Josiah Wong, Cem Gokmen, Sanjana Srivastava, Roberto Mart ́ ın- Mart ́ ın, Chen Wang, Gabrael Levine, Michael Lingelbach, Jiankai Sun, Mona Anvari, Minjune Hwang, Manasi Sharma, Arman Aydin, Dhruva Bansal, Samuel Hunter, Kyu-Young Kim, Alan Lou, Caleb R Matthews, Ivan Villa-Renteria, Jerry Huayang Tang, Claire Tang, Fei Xia, Silvio Savarese, Hyowon Gweon, Karen Liu, Jiajun Wu, and Li Fei-Fei. BEHAVIOR-1K: A Benchmark for Embodied AI with 1,000 Everyday Activities and Realistic Simulation. In Proceedings of The 6th Conference on Robot Learning, volume 205 of Proceedings of Machine Learning Research, p. 80–93. PMLR, 2023. Month: 14–18 Dec. Rongchang Li, Zhenhua Feng, Tianyang Xu, Linze Li, Xiao-Jun Wu, Muhammad Awais, Sara Atito, and Josef Kittler. C2C: Component-to-Composition Learning for Zero-Shot Compositional Ac- tion Recognition. arXiv preprint arXiv:2407.06113, 2024. Tianle Li, Jihai Zhang, Yongming Rao, and Yu Cheng. Unveiling the Compositional Ability Gap in Vision-Language Reasoning Model. In Advances in Neural Information Processing Systems, volume 38, Main Conference, p. 126574–126592. Curran Associates, Inc., 2025. doi: 10.52202/ 085713-4222. Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: Benchmarking Knowledge Transfer for Lifelong Robot Learning.arXiv preprint arXiv:2306.03310, 2023. Huaping Liu, Di Guo, and Angelo Cangelosi. Embodied intelligence: A synergy of morphology, action, perception and learning. ACM Computing Surveys, 57(7):1–36, March 2025a. ISSN 1557-7341. doi: 10.1145/3717059. Ruixun Liu, Yuxuan Wang, Jiacheng Xie, Yuhuan You, Donghua Cai, Junming Lin, Xiong-Hui Chen, Zhifang Guo, Yunfei Chu, Qize Yang, Xize Cheng, Jin Xu, and Yiwu Zhong. Om- niecho: Audio-visual spatial understanding for omni-modal embodied agents. arXiv preprint arXiv:2609.23407, 2026. Yang Liu, Weixing Chen, Yongjie Bai, Xiaodan Liang, Guanbin Li, Wen Gao, and Liang Lin. Align- ing Cyber Space With Physical World: A Comprehensive Survey on Embodied AI. IEEE/ASME Transactions on Mechatronics, December 2025b. doi: 10.1109/tmech.2025.3574943. Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexan- dre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34:21808–21820, 2021. Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Alexander Sax, and Aravind Ra- jeswaran. OpenEQA: Embodied Question Answering in the Era of Foundation Models. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 16488–16498, 2024. Oier Mees, Lukas Hermann, Erick Rosete-Beas, and Wolfram Burgard. CALVIN: A Benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks. IEEE Robotics and Automation Letters, 7(3):7327–7334, July 2022. doi: 10.1109/lra.2022.3180108. Soroush Nasiriany, Abhiram Maddukuri, Lance Zhang, Adeet Parikh, Aaron Lo, Abhishek Joshi, Ajay Mandlekar, and Yuke Zhu. RoboCasa: Large-Scale Simulation of Everyday Tasks for Gen- eralist Robots. arXiv preprint arXiv:2406.02523, 2024. Nous Research.Hermes Agent, 2026.URL https://github.com/NousResearch/ hermes-agent. OpenAI. GPT-5.6 Terra Model | OpenAI API, 2026a. URL https://developers.openai. com/api/docs/models/gpt-5.6-terra. OpenAI. OpenAI Codex CLI - Getting Started, 2026b. URL https://help.openai.com/ en/articles/11096431. 12 OpenClaw. OpenClaw — open-source AI assistant, 2026. URL https://openclaw.ai/. Aishwarya Padmakumar, Jesse Thomason, Ayush Shrivastava, Patrick Lange, Anjali Narayan-Chen, Spandana Gella, Robinson Piramuthu, Gokhan Tur, and Dilek Hakkani-Tur. TEACh: Task-driven Embodied Agents that Chat. arXiv preprint arXiv:2110.00534, 2021. Yujia Peng, Jiaheng Han, Zhenliang Zhang, Lifeng Fan, Tengyu Liu, Siyuan Qi, Xue Feng, Yuxi Ma, Yizhou Wang, and Song-Chun Zhu. The tong test: Evaluating artificial general intelligence through dynamic embodied physical and social interactions. Engineering, 34:12–22, 2024. Yuankai Qi, Qi Wu, Peter Anderson, Xin Wang, William Yang Wang, Chunhua Shen, and An- ton van den Hengel. REVERIE: Remote Embodied Visual Referring Expression in Real Indoor Environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. Qwen.Qwen3.8-Flash, 2026.URL https://w.qwencloud.com/models/qwen3. 8-flash. Jiawei Ren, Yan Zhuang, Xiaokang Ye, Lingjun Mao, Xuhong He, Jianzhi Shen, Mrinaal Dogra, Yiming Liang, Ruixuan Zhang, Tianai Yue, et al. Simworld: An open-ended realistic simulator for autonomous agents in physical and social worlds. arXiv preprint arXiv:2512.01078, 2025. Mohit Shridhar, Jesse Thomason, Daniel Gordon, Yonatan Bisk, Winson Han, Roozbeh Mottaghi, Luke Zettlemoyer, and Dieter Fox. ALFRED: A Benchmark for Interpreting Grounded Instruc- tions for Everyday Tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. Yiheng Shu, Bernal Jim ́ enez Guti ́ errez, Saisri Padmaja Jonnalagedda, Yuguang Yao, Huan Sun, and Yu Su. AgentCL: Toward Rigorous Evaluation of Continual Learning in Language Agents. arXiv preprint arXiv:2606.02461, 2026. Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time train- ing with self-supervision for generalization under distribution shifts. In International conference on machine learning, p. 9229–9248. PMLR, 2020. Zhe Sun, Kunlun Wu, Chuanjian Fu, Zeming Song, Langyong Shi, Zihe Xue, Bohan Jing, Ying Yang, Xiaomeng Gao, Aijia Li, Tianyu Guo, Huiying Li, Xueyuan Yang, Rongkai Liu, Xinyi He, Yuxi Wang, Yue Li, Mingyuan Liu, Yujie Lu, Hongzhao Xie, Shiyun Zhao, Bo Dai, Wei Wang, Tao Yuan, Song-Chun Zhu, Yujia Peng, and Zhenliang Zhang. TongSIM: A General Platform for Simulating Intelligent Machines. arXiv preprint arXiv:2512.20206, 2025. Andrea Tupini, Lars Liden, Reuben Tan, Yu Wang, and Jianfeng Gao. AsgardBench – Eval- uating Visually Grounded Interactive Planning Under Minimal Feedback.arXiv preprint arXiv:2603.15888, 2026. Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodied- Scan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 19757–19767, 2024. Weizhen Wang, Chenda Duan, Zhenghao Peng, Yuxin Liu, and Bolei Zhou. Embodied Scene Un- derstanding for Vision Language Models via MetaVQA. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR), p. 22453–22464, 2025. Cheng-Kuang Wu, Zhi Rui Tam, Chieh-Yen Lin, Yun-Nung Chen, and Hung-yi Lee. Stream- Bench: Towards Benchmarking Continuous Improvement of Language Agents. arXiv preprint arXiv:2406.08747, 2024. Zenan Wu, Bingqing Wei, Lu Liu, Zheqi He, Xi Wang, Jiakang Liu, Zehui Li, Guocai Yao, Jing-Shu Zheng, Xi Yang, and Yongtao Wang. ATOM-Bench: A Real-World Benchmark for Atomic Skills and Compositional Generalization in Manipulation Policies. arXiv preprint arXiv:2606.16826, 2026. 13 Binyan Xu, Xilin Dai, and Kehuan Zhang. Contextual agentic memory is a memo, not true memory. arXiv preprint arXiv:2604.27707, 2026. Chen Yifan, Mingjie Wei, Xuesong Wang, Yuanxing Liu, Jizhe Wang, Hao Song, Longxuan Ma, Donglin Di, Churui Sun, Kaifeng Liu, Le Qi, Jianing Yu, Xiang Tian, Shiqi Liang, Chuanxi Duan, Zijian Hong, Weinan Zhang, and Ting Liu. Embodied AI: A Survey on the Evolution from Perceptive to Behavioral Intelligence. SmartBot, September 2025. doi: 10.1002/smb2.70003. Tianhe Yu, Deirdre Quillen, Zhanpeng He, Ryan Julian, Karol Hausman, Chelsea Finn, and Sergey Levine. Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning. In Proceedings of the Conference on Robot Learning, volume 100 of Proceedings of Machine Learning Research, p. 1094–1100. PMLR, 2020. Month: 30 Oct–01 Nov. Ziao Zhang, Kou Shi, Shiting Huang, Avery Nie, Yu Zeng, Yiming Zhao, Zhen Fang, Qishen Su, Haibo Qiu, Wei Yang, Qingnan Ren, Shun Zou, Wenxuan Huang, Lin Chen, Zehui Chen, and Feng Zhao. SkillFlow:Benchmarking Lifelong Skill Discovery and Evolution for Autonomous Agents. arXiv preprint arXiv:2604.17308, 2026. Shiyun Zhao, Xinwei Song, Tianyu Guo, Xiaomeng Gao, Mingyuan Liu, Xu Han, Yuanyuan Zhang, Zhenliang Zhang, Xue Feng, and Bo Dai. NormAct: Benchmarking Embodied Agents’ Proactive Compliance with Unspoken Social Norms. arXiv preprint arXiv:2606.27826, 2026. Junhao Zheng, Xidi Cai, Qiuke Li, Duzhen Zhang, ZhongZhi Li, Yingying Zhang, Le Song, and Qianli Ma. LifelongAgentBench: Evaluating LLM Agents as Lifelong Learners. arXiv preprint arXiv:2505.11942, 2025a. Junhao Zheng, Chengming Shi, Xidi Cai, Qiuke Li, Duzhen Zhang, Chenxing Li, Dong Yu, and Qianli Ma. Lifelong Learning of Large Language Model based Agents: A Roadmap. arXiv preprint arXiv:2501.07278, 2025b. Shanshan Zhong, Yi Lu, Jingjie Ning, Yibing Wan, Lihan Feng, Yuyi Ao, Leonardo F. R. Ribeiro, Markus Dreyer, Sean Ammirati, and Chenyan Xiong. SkillLearnBench: Benchmarking Con- tinual Learning Methods for Agent Skill Generation on Real-World Tasks.arXiv preprint arXiv:2604.20087, 2026. 14 ABASIC TASK CATEGORIES As described in Section 3.2, basic tasks in EmbodiedCLBench span six capability categories. Table 2 lists each category and its core evaluation question. Table 2: Six capability categories of basic tasks in EmbodiedCLBench. CategoryCore Question Object ManipulationCan the agent locate and manipulate everyday objects into a usable or manageable state? Spatial PlacementCan the agent understand spatial relationships between objects and surfaces or containers? Accessibility ControlCan the agent change the state of doors, drawers, and cabinet doors to enable or restrict access? Appliance OperationCan the agent toggle household appliances and environmental de- vices? Activity-Driven Scene ConfigurationCan the agent set up multiple objects and devices to support activ- ities such as reading, dining, or leisure? Household State RestorationCan the agent restore a room to a tidy, energy-efficient, and reusable state? BGRAPH CONSTRUCTION PIPELINE As described in Sections 3.2 and 3.4, EmbodiedCLBench builds tasks from atomic actions available in TongSim. Given the atomic actions for each task, we first construct the state-transition graph. We then generate two types of textual content: task descriptions and per-state action options. LLM Generation. Both task descriptions and action options are produced by a three-stage pipeline powered by Qwen3.7-Plus. The pipeline serves two purposes: converting low-level atomic actions into natural household scenarios, and ensuring that action options do not give away the cor- rect answer through shortcuts. The key prompt specifications for all three stages are provided in Section B.1. Human Verification. Three annotators independently verify each generated sample. The verifica- tion covers three aspects: captured images, task descriptions, and per-state action options. For each aspect, annotators check for factual errors, logical inconsistencies, and whether the content reflects realistic household activities. A sample is retained only if all three annotators approve it. B.1PROMPT SPECIFICATIONS Stage 1: Task Description Rewrite # Stage 1: Task Description Rewrite ## Objective Rewrite a baseline task description into a natural, scene-based household request. The rewritten description is what the tested model actually sees. You do not change the underlying task structure -- only the surface expression. ## Task Hierarchy - L1 (atomic goal): the smallest unit of household goal, e.g., "turn off the TV", "open the cabinet", "place the cup on the table". - L2 (basic task): composed of 2 L1 goals forming a coherent household activity. - L4 (advanced task): composed of 4 L1 goals combining multiple capability categories. ## Input Fields - level: L2 or L4. 15 - ability_categories: capability categories involved (e.g., appliance operation, spatial placement). - l1_goals: list of constituent L1 atomic goals with descriptions, action types, and target objects. - baseline_task_description: a direct concatenation of all L1 goals (e.g., "Turn off the TV, open the lamp, close the cabinet door"). ## Rewrite Principles 1. Merge multiple L1 goals into a single coherent living scenario (e.g., "A guest is arriving soon -- make the living room presentable and quiet"). The model should need to reason about what to do from the scene image, not read the steps directly from the text. 2. Do NOT produce a checklist, enumerate target objects by name, expose execution order, or translate L1 goals one-by-one with synonyms. 3. Replace specific object names with functional or contextual references (e.g., TV -> "noise source"; table lamp -> "reading light"; cabinet -> "storage area"). 4. Add a plausible living scenario (guest arriving, bedtime routine, post-cooking cleanup, etc.) that makes all L1 goals co-exist naturally. ## Output A single JSON object with fields: description_highlevel (the rewritten task description), rationale (why the rewrite is natural and leak-free), and anti_leak_check (boolean flags for exposes_order, is_checklist, mentions_specific_objects). Stage 2: Per-State Option Rewrite # Stage 2: Per-State Option Rewrite ## Objective Rewrite the four hard-coded action options at a single graph state into natural language. After rewriting, all four options should look like plausible next steps, forcing the model to rely on the scene image and current state rather than linguistic cues. ## Underlying Execution Pattern (must be hidden) Each household goal typically requires three phases: 1. Observe: look at the target object (look_at). 2. Approach: move to the target object (walk_to). 3. Operate: execute the action (open / close / switch_on / etc.). If option wording reveals which phase it belongs to (e.g., observe options always say "check the status of X"), the model can guess the answer without seeing the image. The core challenge is to make all options sound like purposeful household actions regardless of their underlying phase. ## Option Roles - success_path: the optimal next action. Must not stand out among the four options. - valid_detour: executable and task-relevant, but not optimal. Must read like active task progress, not like a detour. - invalid: cannot be executed in the current state. Must read like a plausible next step; its appeal comes from semantic relevance to the task, not from low-quality wording. ## Rewrite Rules 1. Each option: 1 sentence, 10-25 English words, naming the involved object or area and connecting to the high-level task scenario. 2. Two-layer fit: every option (including distractors) must fit both the current sub-goal and the high-level task description. 3. Hide phases: do not let observe options read as "status checks", approach options as "moving closer to operate", or operate options as the only ones with a sense of completion. 4. Uniform style: all four options should be similar in length, detail, and directness. The correct option must not be the only one containing task keywords. 5. Invalid options must not contain words like "wrong", "invalid", or "not needed", and must not explain why the action cannot be executed. ## Output A single JSON object containing: state_id, options (each with id, text, role, design_note), and self_check (correct_stands_out, current_goal_fit, highlevel_task_fit, phase_exposed, has_implementation_terms). 16 Stage 3: Cross-State Option Hardening # Stage 3: Cross-State Option Hardening ## Objective Given a batch of states from the same task graph, detect and eliminate cross-state linguistic patterns that the tested model could exploit. The model sees its full interaction history, so any recurring pattern across consecutive steps can be discovered and used to bypass visual reasoning. ## Context Each household goal takes three phases (observe -> approach -> operate). A basic task has 2 such goals, producing a repeating 2-beat rhythm on the shortest path: Goal 1: observe -> approach -> operate Goal 2: observe -> approach -> operate Stage 2 made each state's options internally deceptive but could not address cross-state patterns. This stage fixes that. ## Input A batch of states (either the full shortest path or a branch cluster), each containing: state_id, state_context, belongs_to_l1 (which sub-goal this state belongs to), and options (with baseline_text, rewritten_text, role, action_type, object). ## Patterns to Detect and Eliminate 1. Three-beat rhythm: correct options (success_path) across consecutive states show a cyclic "attention/focus -> head/move -> execute/switch" pattern that repeats per sub-goal. 2. Style bias: correct options are systematically shorter, more direct, or contain more task keywords than distractors. 3. Expression repetition: the same phrasing or sentence structure appears across multiple states. ## Hardening Strategies A. Blur phase boundaries: make observe-phase correct options sound like preparation or positioning rather than looking. Make approach-phase correct options sound like executing part of an action rather than walking. B. Mix phase styles in distractors: in an observe state, give some distractors an operate-style wording; in an operate state, give some distractors an observe-style wording. C. Break repetition across sub-goals: if sub-goal 1's correct options follow pattern X, sub-goal 2's correct options must use a different pattern. D. Vary detail level: some correct options should be short and direct, others should include contextual detail, matching the variation in distractors. E. Vary verbs and sentence structures: the same action type appearing in different states must use different verbs and phrasings. ## Constraints - Modify rewritten_text only (not state_id, id, role, action_type, or baseline_text). - Edits are adjustments to Stage 2 output, not full rewrites from baseline_text. - Options must remain specific, natural, and connected to the high-level task scenario. - Do not use implementation terms (graph, state, precondition, phase). ## Output A single JSON object containing: states (each with state_id, modified flag, options, modification_note) and pattern_report (documenting which patterns were found and how they were fixed). CTASK EXAMPLES We provide task examples at both compositional levels to illustrate the evaluation format. The agent first receives a natural-language task description. At each step, it observes the current-state image and selects one action from multiple candidates. Specifically, we include two basic-task examples (Figures 8 and 9; Tables 3 and 4) and one advanced-task example (Figure 10; Table 5). 17 State 0State 1State 2State 3 State 4State 5State 6 Figure 8: Visual states along the shortest path for a basic-task example. Table 3: Basic-task example: action options and hidden graph roles at each state. Task: I’m about to watch a movie in the living room. Turn off the overhead light so the screen isn’t washed out, and open up the media cabinet so everything’s within reach. StateOptionText shown to the agentHidden role State 0AHead over to the ceiling light area to get ready to turn it off.Invalid in current state State 0BSwitch on the overhead light to brighten the living room before the movie.Invalid in current state State 0CShut off the ceiling light right away to set the mood for the movie.Invalid in current state State 0DThe overhead light is still on — that’s the first thing to deal with before the movie. Correct shortest-path action State 1ACheck the ceiling light to see if it’s already been switched off.Invalid in current state State 1BSwitch off the ceiling light to darken the room for movie viewing.Invalid in current state State 1CTurn on the ceiling light to brighten the space before settling in for the movie.Invalid in current state State 1DWalk over to the ceiling light so you can reach the switch.Correct shortest-path action State 2ASwitch off the overhead light — the room needs to be dark for the movie.Correct shortest-path action State 2BStep closer to the ceiling fixture to get a better grip on the switch.Executable detour State 2CGlance up at the ceiling light to confirm how bright it still is.Invalid in current state State 2DCheck the television area to see if the media setup is ready.Executable detour State 3AHead over to the television area to get the media setup going.Invalid in current state State 3BWalk toward the ceiling light to start working on the room’s ambiance.Executable detour State 3CSwitch on the overhead light to brighten things up before the movie.Executable detour State 3DWith the light off, the media cabinet by the TV is the next thing to handle.Correct shortest-path action State 4APower down the television to eliminate any screen glare during the movie.Invalid in current state State 4BWalk up to the TV unit — the cabinet is right ahead.Correct shortest-path action State 4CFocus on the television to assess whether the screen needs attention.Invalid in current state State 4DSwitch on the television so you can test the viewing angle for the movie.Invalid in current state State 5AShift your focus to the ceiling light to confirm it’s off for movie time.Executable detour State 5BOpen the media cabinet so everything is within reach for the movie.Correct shortest-path action State 5CTake a closer look at the television area to plan the next move for the media setup. Invalid in current state State 5DStep over to the television unit to get a better position for the entertainment area.Executable detour 18 State 0State 1State 2State 3 State 4State 5State 6 Figure 9: Visual states along the shortest path for a basic-task example. Table 4: Basic-task example: action options and hidden graph roles at each state. Task: Get the living room ready for a relaxed evening — turn on some background entertainment and open up the storage cabinet so every- thing’s within reach. StateOptionText shown to the agentHidden role State 0AStart by checking the TV area — that’s where the background entertainment will come from. Correct shortest-path action State 0BShut off the TV screen since it’s not needed for the living room ambiance.Invalid in current state State 0CHead over to the TV wall to get closer to the entertainment controls.Invalid in current state State 0DPull open the storage cabinet to make everything accessible for the evening.Invalid in current state State 1AShut the storage cabinet to keep the living room looking tidy.Invalid in current state State 1BShift your focus to the cabinet area near the TV to assess what’s needed.Invalid in current state State 1CGet to the entertainment setup — the background noise needs to start soon.Correct shortest-path action State 1DPull open the storage compartment so things can be reached easily for the evening. Invalid in current state State 2ASwitch on the screen to bring some background entertainment into the living room. Correct shortest-path action State 2BHead over to the screen area to get a better position for the setup.Executable detour State 2CFocus on the storage cabinet to make sure it’s ready for easy access.Executable detour State 2DCheck the screen to see if it’s already running before proceeding.Invalid in current state State 3AThe storage cabinet still needs opening — that’s the next thing to handle.Correct shortest-path action State 3BHead over to the storage unit to get it ready for the evening.Invalid in current state State 3CSwitch off the ambient background element near the TV to set a cleaner mood.Executable detour State 3DWalk toward the entertainment area to adjust the atmosphere.Executable detour State 4AShut the storage cabinet to keep the living room looking tidy.Invalid in current state State 4BCheck the cabinet area to see what’s currently inside.Invalid in current state State 4CPull open the storage cabinet to make everything accessible for the evening.Invalid in current state State 4DMove into position near the storage cabinet — it needs to be opened.Correct shortest-path action State 5AStep closer to the storage cabinet to get a better grip on the handle.Executable detour State 5BScan the cabinet area to confirm everything is in order.Invalid in current state State 5CGlance toward the TV wall to take in the overall setup before finishing up.Executable detour State 5DOpen up the storage cabinet so everything is within reach for the evening.Correct shortest-path action 19 State 0State 1State 2State 3 State 4State 5State 6State 7 State 8State 9State 10State 11 State 12 Figure 10: Visual states along the shortest path for an advanced-task example, including the final state. Table 5: Advanced-task example: action options and hidden graph roles. Task: I’m done for the night and heading to bed — dim the bedroom lamp, open up the study desk so it’s ready for tomorrow, shut the bedside drawer, and toss the old remote into the bin. StateOptionText shown to the agentHidden role State 0ACheck the trash can area to see what needs to be tossed before bed.Executable detour State 0BSwitch on the table lamp to brighten the bedside area for cleanup.Invalid in current state State 0CThe lamp’s still glowing — that’s the first thing to sort out before the room can rest. Correct shortest-path action State 0DHead over to the table lamp to adjust the lighting for bedtime.Invalid in current state State 1ASwitch off the table lamp to dim the room for bedtime.Invalid in current state State 1BGet closer to the lamp — you’l need to be right there to kill the light.Correct shortest-path action State 1CBrighten the table lamp so you can see what’s left to tidy.Invalid in current state State 1DFocus on the table lamp area to assess what needs adjusting for the night.Invalid in current state State 2ABrighten the lamp a bit so you can see what’s left on the nightstand before closing it up. Invalid in current state State 2BKill the lamp light to start dimming the room for the night.Correct shortest-path action State 2CShift over to the lamp side to get the lighting sorted before settling in.Invalid in current state State 2DFocus on the lamp area to assess what still needs adjusting before bed.Invalid in current state State 3AGlance over at the trash can to see what needs tossing before bed.Executable detour State 3BThe study desk still has stuff on it — that’s the next area to sort out.Correct shortest-path action State 3CHead over to the study desk to get in position for clearing it off.Invalid in current state State 3DFlip open the study desk to start organizing items for tomorrow’s setup.Invalid in current state Continued on the next page. 20 Table 5 (continued). StateOptionText shown to the agentHidden role State 4AGlance over at the desk to see what’s still out on the surface.Invalid in current state State 4BFlip open the desk to start organizing what’s inside for tomorrow.Invalid in current state State 4CShut the desk drawers to keep the room tidy for the night.Invalid in current state State 4DStep up to the desk so you can reach across and clear it off properly.Correct shortest-path action State 5AShift your focus to the desk to assess what still needs clearing.Invalid in current state State 5BOpen the desk cabinet so you can stow tonight’s items inside.Correct shortest-path action State 5CGlance over at the table lamp to check whether the room is dim enough.Executable detour State 5DTake a moment to review the bedside cabinet before closing it up for the night.Executable detour State 6AShift your focus to the table lamp to see if it still needs attention for the night.Executable detour State 6BGlance over at the trash can to check if anything else needs tossing before bed.Executable detour State 6CTake a look at the remote control in your hand to decide where it should go next. Invalid in current state State 6DThe bedroom nightstand is still open — head back and deal with that next.Correct shortest-path action State 7AShut the bedside cabinet to keep the room tidy for the night.Invalid in current state State 7BPull the bedside cabinet open to check if anything needs to be put away.Invalid in current state State 7CGlance over at the bedside cabinet to see what’s still out.Invalid in current state State 7DMove in close to the nightstand — the cabinet needs shutting and you need to be right there. Correct shortest-path action State 8AShift focus to the desk area to see what needs clearing for tomorrow.Executable detour State 8BShut the bedside cabinet to keep the nightstand tidy before sleep.Correct shortest-path action State 8CCheck the table lamp status to see if it needs dimming for bed.Executable detour State 8DGlance at the bedside cabinet to confirm what’s inside before closing.Invalid in current state State 9AShift focus to the table lamp area to continue dimming the room for rest.Executable detour State 9BThe remote control still needs to go — spot the trash can so you know where it’s headed. Correct shortest-path action State 9CGlance over at the desk surface to assess what else still needs to be cleared away. Executable detour State 9DCheck the remote control in your hand to decide whether it should be put away.Invalid in current state State 10ASeal up the trash can so nothing spills out overnight.Invalid in current state State 10BGlance at the trash can to make sure it’s ready to receive items.Invalid in current state State 10CPop open the trash can lid so you can drop things in.Invalid in current state State 10DGet to the trash can — the remote control in your hand needs to go there.Correct shortest-path action State 11ASet the remote control down near the trash can to keep the nightstand clear.Correct shortest-path action State 11BCheck the bedside cabinet to make sure it’s properly closed for the night.Executable detour State 11CTake a closer look at the remote control to decide where it should go.Invalid in current state State 11DTurn your attention to the desk area to see what still needs clearing.Executable detour 21