Probing Embodied LLMs: When Higher Observation Fidelity Hurts Problem Solving Oussama Zenkri1,2,3 1
arXiv:2605.20072v1 [cs.AI] 19 May 2026
2
and Oliver Brock1,2,3
Robotics and Biology Laboratory, Technische Universität Berlin, Germany Science of Intelligence, Research Cluster of Excellence, Berlin, Germany⋆ 3 Robotics Institute Germany (RIG)
Abstract. Large Language Models are increasingly proposed as cognitive components for robotic systems, yet their opaque decision processes make it difficult to explain success or failure in closed-loop embodied tasks. Following an empirical AI methodology, we study embodied LLM agents behaviorally by varying the information available to the agent and measuring the resulting changes in behavior. Using the Lockbox, a sequential mechanical puzzle with hidden interdependencies, we evaluate LLMs across RGB, RGB-D, and ground-truth symbolic observations in a physical robotic setup and use controlled simulation to probe the resulting behavior. Counterintuitively, agents perform best under raw RGB input and worst under perfect ground-truth observations. In simulation, we probe this effect by randomly flipping perceived action outcomes and find that moderate noise improves performance, peaking at a 40% flip probability with a 2.85-fold success rate increase over the noise-free baseline. Further analysis links this gain to a reduction in repetitive action loops. These findings suggest that success rates alone are insufficient for evaluating LLMs, as measured performance may reflect the interaction between perceptual errors and reasoning failures rather than robust problem solving. Keywords: Problem Solving · Embodied LLMs · Observational Noise
1
Introduction
Large Language Models (LLMs) are increasingly proposed as cognitive components for robots operating in unstructured environments, where successful task execution requires more than perception or language understanding alone [9]. Many robotic tasks demand sequential problem solving over extended interactions: agents must integrate observations over time, infer hidden structure, and ⋆
Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy - EXC 2002/1 “Science of Intelligence" project number 390523135. ChatGPT 5.5 and Claude Sonnet 4.6 were used to assist with grammatical corrections and text improvement. The lockbox and robotic arm sketches in Fig. 2 were generated using ChatGPT 5.5.
2
Zenkri O., and Brock O.
adapt their behavior based on action outcomes. These abilities are central to embodied intelligence, where problem solving unfolds as a closed-loop process between agent and environment. Analyzing embodied LLM agents requires moving beyond outcome-based evaluation toward methods that reveal the behavioral mechanisms underlying success and failure. However, LLMs are largely opaque, as their reasoning processes remain inaccessible, and the providers of the most capable models do not expose their weights. In embodied settings, performance further reflects the coupled effects of perception, reasoning, and environmental dynamics. A success or failure therefore cannot be attributed directly to reasoning ability alone, and aggregate success rates provide only a limited view of performance. They do not reveal whether an agent solved a task through coherent inference, accidental exploration, perceptual error, or an interaction between these effects. We therefore adopt a behavioral probing approach, aligned with empirical AI methodology [4] and inspired by the logic of system identification: rather than estimating an explicit dynamical model, we perturb an opaque agent and measure how its behavior changes. We use observation fidelity to denote how faithfully an agent’s observations represent the true state of its environment. Observation fidelity is a useful intervention variable because it affects the information on which the model bases its decisions while leaving the underlying task structure unchanged. We study this problem using the Lockbox, a sequential mechanical puzzle widely used to study problem solving in biological agents and increasingly useful as a diagnostic task for embodied AI systems. The task captures key challenges of real-world problems: actions have state-dependent outcomes, and optimal actions depend on the integration of past interactions. We present the same task through different observation channels, ranging from raw RGB and RGB-D inputs to symbolic ground-truth state descriptions, and evaluate LLM agents in a physical robotic setup and use controlled simulation to probe the resulting behavior. Our results reveal a counterintuitive pattern: increasing observation fidelity does not improve performance. Agents perform best with raw RGB input and worst when given symbolic ground-truth state. To probe this effect, we introduce controlled observational noise in simulation by randomly flipping perceived action outcomes. We find that moderate noise improves success relative to the noise-free baseline, whereas higher noise levels degrade performance again. Further analysis suggests that this improvement is associated with a reduction in repetitive action loops, a failure mode in which the agent repeats near-identical action subsequences despite making little or no progress [5]. Taken together, these findings show that observations are not merely passive inputs to an embodied reasoning process. They actively shape the dynamics of decision making. In closed-loop interaction, measured performance may therefore reflect an accidental interplay between perceptual errors, reasoning failures, and environmental dynamics rather than robust problem solving. This highlights the need to evaluate embodied LLM agents not only by final success rates, but by probing the behavioral mechanisms that give rise to success or failure.
Probing Embodied LLMs
2
3
Methodology
To study embodied LLM agents as opaque systems, we need a task that is both physically grounded and experimentally controllable. The task must expose the agent to key challenges of real-world problem solving, while also allowing the observation channel to be varied without changing the underlying reasoning problem. This section first introduces the Lockbox as such a diagnostic task, then describes the physical and simulated environments used to study agent behavior under different observation conditions.
Fig. 1: Our robotic system manipulating the Lockbox. Our Lockbox comprises two prismatic joints (sliding bars in the middle) and two revolute joints. The Lockbox is unlocked when the leftmost revolute joint, which we refer to as the target joint, is pulled. The robot employs a soft-hand end effector for manipulating the joints, an RGB-D camera for acquiring visual data, and a force-torque sensor for assessing the joint movability and guiding their manipulation. 2.1
The Lockbox
Lockboxes are mechanical puzzles where the objective is to get access to a reward by manipulating a set of moving parts (called joints) that are constrained in a structured but non-obvious way. Solving them therefore requires discovering the underlying constraint pattern and executing a sequence of actions accordingly. Lockboxes have been widely used as an experimental paradigm to study problem-solving behavior across a range of domains, including cognitive biology [2, 6, 11], cognitive psychology [14], and robotics [3, 8, 12]. The sequential, constraint-driven nature of lockboxes makes them a natural testbed for probing the behavior of LLM agents in embodied problem-solving settings. Moreover, the same underlying task can be presented across observation modalities (symbolic state descriptions, RGB, or RGB-D) without altering the latent reasoning challenge, making lockboxes a useful diagnostic task for probing how changes in the observation channel alter closed-loop problem-solving behavior.
4
Zenkri O., and Brock O.
Physical Lockbox Our physical setup, illustrated in Fig. 1, consists of four joints mounted on a wooden panel. All joints have binary states, meaning they can only occupy one of two possible positions: either endpoint of their motion range. The movability of each joint depends on the state of the others, as dictated by the underlying interlocking mechanism. Crucially, this dependency structure is hidden from the agent and must be inferred from the relationship between actions and outcomes. The physical setup therefore provides a setting in which observation, action execution, and subsequent model decisions are tightly coupled. After each selected action, the robot manipulates the corresponding joint and returns a new observation, allowing us to study how the agent updates its behavior from interaction history. The physical Lockbox also exposes the agent to sources of uncertainty that are difficult to reproduce faithfully in simulation, including sensor noise, contact variability, and imperfect manipulation. These factors can affect the agent’s ability to infer latent dependencies, making the task qualitatively harder. The physical setup therefore serves as the real-world setting in which behavioral phenomena can be observed before specific factors are isolated through controlled interventions in simulation. Lockbox Simulation Besides the physical Lockbox setup, we employ a simulated Lockbox environment that complements the real-world experiments. In our behavioral probing approach, the simulation serves as a controlled intervention environment: it isolates the underlying reasoning problem from the perceptual and actuation variability of the physical setup, while allowing the agent’s observation channel to be perturbed systematically. Agents can therefore be evaluated under perfect observation conditions, where the state transitions are deterministic, as well as under deliberately altered observations. The Lockbox simulation preserves the essential properties of the physical setup, including the number of joints and the underlying interlocking structure. The LLM acts purely at the symbolic level, receiving a textual description of the joints’ state and selecting actions accordingly. Since the simulated Lockbox provides no visual cues about the dependency structure, this structure can only be inferred from past interactions. This makes the simulation suitable for testing hypotheses generated from the physical experiments under conditions in which the true environment dynamics are fixed and only the agent’s perceived state is manipulated. 2.2
Experimental Setup
The experiments are conducted using an existing robotic manipulation system [7], consisting of a Franka Emika Panda robotic arm equipped with an RBO Hand 3 soft end-effector [10], an RGB-D camera for visual observation, and a force-torque sensor for manipulation feedback (see Fig. 1). A system integration layer mediates all interactions between the LLM and the physical environment, translating the model’s joint selections into physical
Probing Embodied LLMs
5
Fig. 2: Experimental setup with three input modalities for the LLM. Through our system integration, the LLM interacts with the Lockbox by selecting a joint to be manipulated, after which the corresponding new observation is forwarded to the model. Bottom: three instances of the considered input modalities. The ground-truth state (left) consists of a structured representation specifying the predefined labels of each Lockbox joint and its corresponding state. The RGB observation (middle) is captured before each interaction from a fixed viewpoint, enabling direct comparison of successive observations to infer action outcomes. The RGB-D observation (right) augments the RGB modality with an aligned depth channel, providing additional geometric information about the scene.
manipulations, as shown in Fig. 2. Before each action, the robot moves to a homing position, from which a new observation of the Lockbox is captured and forwarded to the LLM, ensuring consistent viewpoints across all steps. We employ OpenAI GPT o1 as the decision-making LLM, selected for its advanced reasoning capabilities in sequential decision-making tasks. We evaluate the model under three input modalities with varying levels of observation fidelity: (i) RGB images, (ii) RGB-D images, which add depth to the visual signal, and (iii) the ground-truth symbolic state of all joints, which bypasses the vision pipeline entirely. We treat these modalities as interventions on the agent’s observation channel: the underlying task and action space remain fixed, while the information available for decision making changes. In the visual modalities, the model must infer action outcomes by comparing successive observations. In the ground-truth condition, the task reduces to symbolic reasoning over explicitly provided state descriptions.
6
Zenkri O., and Brock O.
Each trial begins with an instruction prompt describing the Lockbox problem and the task objective. The prompt is consistent across trials, with modalityspecific adjustments limited to the observation format. All trials start from the same initial Lockbox configuration. A trial proceeds as a sequence of observationaction steps, with a maximum budget of 20 steps. Failure to move the target joint within this limit is counted as an unsuccessful trial. Since GPT o1 fixes temperature and Top-P parameters at 1 and does not expose these parameters, the model’s outputs are inherently stochastic. We therefore conduct 10 independent trials per modality to obtain reliable performance estimates.
3
Results and Discussion
We now use the Lockbox to identify how changes in observation fidelity affect the behavior of embodied LLM agents. The goal is not only to compare final task success across input modalities, but to understand which behavioral mechanisms may explain the observed differences. We first analyze performance in the physical robotic setup, then test the role of observational noise in simulation, and finally examine repetitive action loops as a candidate behavioral mechanism underlying the resulting performance trends. 3.1
Observation Fidelity Alters Closed-Loop Behavior
We first use input modality as an intervention on the agent’s observation channel. The Lockbox task and action space are held fixed, while the agent receives observations with different levels of fidelity. This allows us to ask whether more faithful observations lead to more effective closed-loop problem solving, and whether deviations from this expectation reveal systematic behavioral weaknesses. For reference, we include a human-inspired strategy for solving the Lockbox [7], which receives the ground-truth symbolic state and infers the dependency structure through trial and error. The heuristic was derived from behavioral data collected in a prior study on how humans approach Lockbox-like problems [13]. As shown in Fig. 3, which plots cumulative success rate as a function of interaction steps, the human-inspired strategy achieves a 100% success rate and requires fewer steps than GPT o1 under any input modality. By contrast, GPT o1 reaches a maximum success rate of 80% across all three modalities. However, identical final success rates mask substantial differences in efficiency: under RGB input, GPT o1 reaches 80% success after 11 steps, whereas the ground-truth state input requires 15 steps. The human-inspired strategy reaches the same success rate within 9 steps. These differences show that final success rate alone provides an incomplete picture of agent performance because it does not capture the efficiency or structure of the decision-making process that produced the outcome. Turning to the effect of input modality, the results run counter to the common assumption that richer or more accurate observations should improve performance. GPT o1 performs best with RGB input and worst with ground-truth state, with RGB-D falling in between. In Fig. 3, the RGB condition reaches
Probing Embodied LLMs
7
most success thresholds faster than RGB-D, which in turn mostly outpaces the ground-truth condition. These results suggest that additional or more accurate input is not only underutilized but may actively hinder problem solving.
100%
100
80%
Success Probability [%]
80
60
40 Human-Inspired Strategy GPT o1 RGB GPT o1 RGB-D GPT o1 Ground Truth State
20
6
8
10
12
14
16
18
20
Number of Interactions N
Fig. 3: The human-inspired strategy outperforms GPT o1 across all input modalities. The figure shows the success probability as a function of the number of interactions, comparing the human-inspired strategy to GPT o1 under different input modalities. Counterintuitively, higher input fidelity is associated with lower performance. The model performs worst under ground truth state observations, suggesting limitations in its ability to use structured state information for effective decision making. To better understand this effect, we analyze the model’s ability to correctly interpret action outcomes. We define a misinterpretation as an incorrect assessment of a state change following a manipulation. For visual inputs, misinterpretation rates vary substantially across trials and modalities, with a wide spread suggesting inconsistent perceptual grounding rather than a stable failure mode (Table 1). Why depth information worsens performance relative to RGB remains unclear. The added channel possibly introduces noise into the model’s visual processing or shifts its attention toward spatial features that are not task-relevant. Surprisingly, misinterpretations also occur in the ground-truth condition, yielding an average error rate of approximately 5% at the step level. Since ground-truth input requires no perceptual inference, these errors point to failures in the model’s reasoning over structured representations, most plausibly hallucination of state changes, which stands in contradiction to the inputs.
8
Zenkri O., and Brock O.
Table 1: State Misinterpretation Rates Across Input Modalities Input Modality
Average State Misinterpretation Rate per Trial
Total State Misinterpretations
RGB RGB-D Ground Truth
21.7% ± 12.2% 26.8% ± 12.5% 4.9% ± 6.9%
22∗∗ 35∗∗ 6∗
∗ : occurred in 4 out of 10 trials
∗∗ : occurred in all 10 trials
Taken together, these findings reveal a counterintuitive pattern: higher input fidelity corresponds to worse task performance. If lower-fidelity visual inputs improve performance by introducing stochasticity into the perceived state, then a similar effect should appear when state uncertainty is introduced directly in a controlled symbolic simulation. The following section tests this prediction. 3.2
Perturbing Observations Reveals a Non-Monotonic Response
To test whether lower-fidelity inputs improve performance by introducing stochasticity into the perceived Lockbox state, rather than by providing additional visual cues, we design a controlled experiment that removes visual observations entirely. Instead, we use a simulated Lockbox with a structured state representation as model input and introduce noise directly into the state observations through random state-flips applied at controlled probabilities. Specifically, with a probability p, the agent receives the inverted outcome of an action it took. This perturbation affects only what the agent perceives, not the actual state of the Lockbox, thereby mimicking the effect of misinterpreting a visual observation. We evaluate state-flip probabilities ranging from 0% to 60% in increments of 10%, examining how varying levels of observational noise affect the model’s ability to solve the Lockbox. For each flip probability, the experiment is repeated three times, with each repetition consisting of 10 independent trials (210 trials in total). We use GPT-4o for this study because the experiment operates purely over symbolic states rather than multimodal observations, and because the larger number of controlled intervention trials would otherwise be prohibitively costly. Accordingly, the simulation results should be interpreted as evidence for a plausible mechanism, not as a direct explanation of the physical results. Figure 4 shows the observed success rates across all repetitions for each state flip probability, together with a second-order polynomial fitting to better capture the underlying trend (the order of the fit was determined through AIC and cross-validation RMSE minimization). The fit reveals a non-monotonic relationship between observational noise and task performance: starting from an average success rate of 23.3% under ground-truth observations, performance rises to a peak at a flip probability of 40% before declining at higher noise levels. The improvement relative to the noise-free baseline indicates that moderate observation noise can improve measured task performance in this controlled setting. These results support the hypothesis that part of the performance difference between observation modalities is linked to stochasticity in the perceived
Probing Embodied LLMs
9
100
Observed Success Rates
Line Fit
95% CI
Success Rate [%]
80
60
40
20
0
10
20
30
40
50
60
State-Flip Probability [%]
Fig. 4: Performance exhibits a non-monotonic dependence on perceptual noise. Success rates increase with state uncertainty, peaking at a 40% state-flip probability (≈ 2.85× the ground-truth baseline, on average), before declining at higher noise levels, suggesting that moderate noise can improve observed task performance without necessarily reflecting improved reasoning. state, rather than to visual information alone. However, they offer no mechanistic explanation for why perceptual noise should improve performance in the first place. In the following section, we investigate one possible mechanism that may underlie this effect. 3.3
Repetitive Action Loops as a Candidate Failure Mode
A recurring behavioral pattern observed across trials is the tendency of the LLM to produce repetitive action subsequences within a single trial. For example, an agent might select slider A, then slider B, then slider C, before looping back to the same subsequence at later timepoints of the trial. Although the agent eventually breaks out of such loops, the intervening redundant actions substantially degrade the problem-solving efficiency and frequently cause the agent to exhaust its interaction budget before solving the Lockbox. This observation raises the question of whether observational noise is associated with fewer repetitive action loops, and thereby increasing their probability of success. To evaluate this hypothesis, we examine whether the occurrence of action loops within a trial’s action sequence is correlated with success rate, and whether this relationship is modulated by the state-flip probability. We define an action loop as any subsequence of actions of length three or higher that appears at least twice within the same trial. For a given trial, the set
10
Zenkri O., and Brock O. Observations
Line Fit
95% CI
(a)
(b) 100
Action Loops Probability [%]
100
Success Rate [%]
80
60
40
20
40
50
60
70
80
Action Loops Probability [%]
90
80
60
40
20
0
10
20
30
40
50
60
State-Flip Probability [%]
Fig. 5: Perceptual noise is associated with fewer action loops, which are associated with reduced performance. (a) Action loop probability and success rate are negatively correlated (r = −0.69, Pearson correlation), with higher engagement in action loops corresponding to lower success rates. (b) The probability of engaging in repetitive action sequences decreases with increasing perceptual noise, reaching its minimum at a 40% state-flip probability (on average, ≈ 26% lower than the ground-truth baseline), before rising again at higher noise levels. Together, these trends suggest that the observed increase in success rates under moderate noise is consistent with the reduction in repetitive action sequences. of action loops covering that trial is determined as the non-overlapping collection of individual action loops that maximizes the total coverage of the action sequence in that trial. This combinatorial optimization problem is solved via binary Integer Linear Programming (ILP), applied over the set of action loops obtained by greedily enumerating all repeated subsequences in the trial. Figure 5(a) shows that action loop occurrence is negatively correlated with success rate: as the proportion of trials containing action loops increases, success probability decreases consistently. Figure 5(b) further shows that this occurrence is non-monotonically dependent on the state-flip probability. It decreases from a maximum average value of approximately 80% at zero noise, reaches a minimum near a state-flip probability of 40%, reflecting a decrease of more than 26%, and then rises again at higher noise levels. This pattern is consistent with the hypothesis that moderate perceptual noise reduces the likelihood of action loop formation and may contribute to the observed improvement in task success. Becoming trapped in recursive loops that repeatedly generate near-identical action sequences is a well-documented failure mode of LLM-based agents [5].
Probing Embodied LLMs
11
In interactive decision-making settings like the Lockbox, the tendency is compounded when repeated failed attempts to move a joint leave the environment state unchanged, causing the agent to observe the same state across consecutive steps and providing no novel contextual signal to drive exploration toward alternative actions. This interpretation aligns with prior findings that noise injected in the prompt can increase the diversity of LLM responses [1].
4
Limitations
First, our evaluation was restricted to models from a single provider, OpenAI. Given the field’s rapid development, these models no longer represent the state of the art. Whether the observed behavior generalizes to newer models, to models from other providers, or to open-source models remains an open question. Crossprovider comparison was deliberately outside the scope of this study. Future work should therefore assess whether our findings reflect a broader limitation of LLM-based agents or are specific to the models evaluated here. Second, our evaluation was conducted on a single visual instance of the Lockbox, meaning that the observed behavior could, in principle, be specific to this particular layout. However, additional simulated experiments with actively varied visual layouts [15] provide preliminary evidence that the findings are not layout-specific. A systematic study of visual variation in real-world settings is nonetheless needed to establish how broadly these findings generalize. Finally, model performance was benchmarked against a human-derived heuristic rather than empirical data from human participants solving the same task. Although this heuristic provides a principled approximation of human reasoning, it may not fully capture the variability, strategies, and error patterns of actual human performance. A direct comparison with human participants would therefore provide a stronger and more ecologically valid benchmark.
5
Conclusion
This work set out to probe the behavior of LLMs as cognitive components in embodied systems. Specifically, we examined how their performance changes across different levels of observation fidelity, since real-world perception is inherently noisy. Rather than improving monotonically with increasing fidelity, performance peaks under moderate perceptual noise and degrades when agents are given perfect observations. This counterintuitive finding is consistent with a candidate mechanism: accurate feedback may sustain repetitive action loops, whereas erroneous observations can disrupt them. In other words, what looks like better performance under lower fidelity inputs should not be interpreted as evidence for stronger reasoning, but may instead reflect a side effect of compounding perceptual errors that happen to be beneficial for this task. This distinction matters beyond the Lockbox. In embodied settings where perception and reasoning are simultaneously imperfect and difficult to disentangle, success rates may reflect the accidental interplay of these imperfections as much as genuine
12
Zenkri O., and Brock O.
problem-solving ability. We conclude that LLMs are still subject to weaknesses at both the perceptual and reasoning level, and that evaluating them rigorously requires real-world closed-loop benchmarks combined with controlled behavioral interventions that reveal the mechanisms behind success and failure.
References 1. Agrawal, P., Goyal, P.: Addressing llm diversity by infusing random concepts (2026), https://arxiv.org/abs/2601.18053 2. Auersperg, A.M.I., Kacelnik, A., von Bayern, A.M.P.: Explorative learning and functional inferences on a five-step means-means-end problem in goffin’s cockatoos (cacatua goffini). PLOS ONE 8(7), e68979 (07 2013) 3. Baum, M., Bernstein, M., Martin-Martin, R., Höfer, S., Kulick, J., Toussaint, M., Kacelnik, A., Brock, O.: Opening a lockbox through physical exploration. In: IEEERAS International Conference on Humanoid Robotics (Humanoids). pp. 461–467 (2017) 4. Cohen, P.R.: Empirical methods for artificial intelligence, vol. 139. MIT press Cambridge, MA (1995) 5. Lai, H.: SpecRA: Monitor degenerative repetition in LLM agents using randomized FFT (2026), https://openreview.net/forum?id=xVO4BqmzVD 6. Lang, B., Kahnau, P., Hohlbaum, K., Mieske, P., Andresen, N.P., Boon, M.N., Thöne-Reineke, C., Lewejohann, L., Diederich, K.: Challenges and advanced concepts for the assessment of learning and memory function in mice. Frontiers in Behavioral Neuroscience 17, 1230082 (2023) 7. Li, X., Zenkri, O., Pfisterer, A., Brock, O.: A biologically inspired design principle for building robust robotic systems. arXiv preprint arXiv:2408.10192 (2024) 8. Liu, Z., Xu, Z., Song, S.: BusyBot: Learning to interact, reason, and plan in a BusyBoard environment. In: Conference on Robot Learning (CoRL). pp. 505–515. PMLR (2023) 9. Mon-Williams, R., Li, G., Long, R., Du, W., Lucas, C.G.: Embodied large language models enable robots to complete complex tasks in unpredictable environments. Nature Machine Intelligence pp. 1–10 (2025) 10. Puhlmann, S., Harris, J., Brock, O.: RBO Hand 3: A platform for soft dexterous manipulation. IEEE Transactions on Robotics 38(6), 3434–3449 (2022) 11. Stanton, L.A., Cooley-Ackermann, C., Davis, E.C., Fanelli, R.E., Benson-Amram, S.: Wild raccoons demonstrate flexibility and individuality in innovative problemsolving. Proceedings of the Royal Society B: Biological Sciences 291(2027), 20240911 (2024) 12. Verghese, M., Atkeson, C.: Using memory-based learning to solve tasks with stateaction constraints. In: IEEE International Conference on Robotics and Automation (ICRA). pp. 9558–9565 (2023) 13. Zenkri, O., Bolenz, F., Pachur, T., Brock, O.: Extracting principles of exploration strategies with a complex ecological task. In: International Conference on Simulation of Adaptive Behavior. pp. 289–300. Springer (2024) 14. Zenkri, O., Bolenz, F., Pachur, T., Brock, O.: Human exploration in complex problem-solving tasks: More effortful interaction leads to higher efficiency. Adaptive Behavior 33(5-6), 321–332 (2025). https://doi.org/10.1177/ 10597123251364738 15. Zenkri, O., Brock, O.: How far are llms from agi? evidence from a large world problem. In: Proceedings of the German Robotics Conference (GRC) (2026)