- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces ROBOT VALUES, a benchmark containing approximately 8,700 synthetic household scenarios and 59,000 candidate robot actions representing different human values. The authors evaluate 10 VLMs, finding consistent preferences for safety and accommodation over privacy, and limited success in following value-specific instructions that conflict with default choices. Fine-tuning on the benchmark improves value-conditioned action selection, with some transfer to external datasets.
The paper's strongest contribution is a large-scale benchmark for evaluating how robot planners make decisions when human values conflict. The authors demonstrate that VLMs exhibit systematic preferences for certain values and struggle to follow instructions that conflict with these preferences. While the findings support the utility of the benchmark, the broader claims about human values and robot decision-making depend on how faithfully the synthetic scenarios and annotations represent real-world value conflicts.
Critical strengths
-
The paper addresses how autonomous systems should navigate situations where multiple human values are relevant and potentially conflicting. This is particularly important for robots operating in domestic environments.
-
The benchmark construction is systematic and reasonably well documented, combining diverse scenario generation, stakeholder-perspective modeling, and multiple stages of quality control. I appreciate the effort to validate automated filtering against human annotations.
-
The experimental evaluation is comprehensive, spanning multiple VLM architectures, default and value-conditioned decisions, and fine-tuning experiments. The finding that models struggle to prioritize explicitly requested values is particularly interesting and has meaningful implications for robot planning.
Critical weaknesses
-
Clarify the operationalization of stakeholder-grounded human values. The manuscript emphasizes grounding robot decisions in stakeholder perspectives, but these perspectives are themselves generated by LLMs rather than collected from actual stakeholders. While the human annotation audit supports the consistency of the generated data, it does not necessarily establish that the identified conflicts reflect human judgments or preferences. I encourage the authors to clarify this distinction and discuss how biases in simulated stakeholder reasoning might influence the benchmark.
-
Clarify the interpretation of default preferences and value steering. The authors interpret action selections as evidence of model value preferences, but these may also reflect differences in perceived action appropriateness or contextual relevance. Similarly, the distinction between matched and conflicting conditions is defined relative to the model's default choice, making some difference in accuracy expected by construction. The authors should more carefully distinguish failure to recognize a target value, preference for another action, and failure to follow explicit instructions. This would strengthen the interpretation of the reported steering failures.
-
Strengthen the evidence for visual grounding. The benchmark is presented as image-grounded, yet the modality ablations in Appendix E.3 suggest that the general patterns of value preferences remain relatively similar even when images or contextual descriptions are removed. To what extent does the visual information actually influence action selection? A clearer analysis of decisions that depend specifically on visual evidence would help establish what the multimodal benchmark contributes beyond text-based value reasoning.
Questions for authors
See areas for improvement ('critical weaknesses') above.
4 comments
Part 2 (2/3)
W2. Interpretation of default preferences and value steering
Thank you for the constructive feedback on the cause of value steering failures. We agree that the accuracy gap between matched and conflicting conditions is partly expected by definition. Our intention was to measure whether models can follow an explicit user-specified value priority even when it conflicts with the model’s default choice.
To better understand why models fail to steer, we conducted additional experiments.
1. Instruction-Following Capabilities
| Model | IFEval ↑ | MIA-Bench ↑ |
|---|---|---|
| Models evaluated in our paper | ||
| Molmo2-8B | 81.0 | 85.5 |
| Cosmos-Reason2-8B | 79.9 | 86.5 |
| Molmo2-ER | 78.2 | 85.4 |
| InternVL3.5-8B | 78.2 | 87.2 |
| RLDX-1-VLM | 76.7 | 76.0 |
| Qwen3-VL-2B-Instruct | 68.4 | 81.6 |
| Cosmos-Reason2-2B | 56.9 | 78.6 |
| InternVL3-8B | 51.8 | 87.1 |
| InternVL3-2B | 42.9 | 76.1 |
| RoboBrain2.0-7B | 42.3 | 80.6 |
| Reference models | ||
| GPT-4o | – | 88.6 |
| GPT-4o-mini | 80.4 | – |
| LLaVA-NeXT-110B | – | 79.8 |
| Llama-3.1-8B-Instruct | 75.9 | – |
| Qwen2.5-7B-Instruct | 71.2 | – |
We examined whether low steering accuracy could reflect broadly weak instruction-following performance. We find that several models have decent instruction following capability but they still fail to steer their actions. On IFEval[1] and the multimodal MIA-Bench[2], several models evaluated in our paper show instruction-following ability comparable to the reference models (table above). However they follow conflicting value requests in only 13.4--24.6% of cases. These results suggest that low steering accuracy cannot be attributed simply to weaknesses in general instruction following. We have updated the results in Section 5, with details in Appendix F.1.
2. Action-Value Recognition
We examined whether steering failures primarily reflect difficulty recognizing which actions prioritize the requested value. As reported in Section 5 (Figure 5), conflicting steering accuracy is higher when models correctly recognize the action–value mapping, but remains low even in those cases (15.9--29.3%). Thus, difficulty recognizing action–value mappings alone does not adequately explain the low steering accuracy.
3. Inappropriate Actions
To also check whether models are not following the requested value priority because the requested action is inappropriate, we checked how many candidate actions are considered inappropriate using an LLM judge. We used the Gemini 4 Carbon model to rate every candidate action either as appropriate, appropriate with caution, or inappropriate. Overall, 4.6% of actions were annotated as inappropriate. Conflicting accuracy is 19.3% across all items and 19.5% after excluding items where all actions associated with the target value are rated as inappropriate, showing a subtle difference. Default value rankings are also similar between scenarios with and without an inappropriate action (mean Spearman ).
Second, we added a decline option to the candidate actions ("Decline to follow the requested priority because it would be inappropriate in this situation, and ask the household members how to proceed"). In conflicting cases, models chose it in only 0.3% of cases, and in only 2.6% even when we add a prompt that explicitly mentions declining is allowed ("If following this priority would be unsafe or otherwise inappropriate in this situation, you may choose the decline option instead").
Third, we asked each model whether each candidate action is appropriate in the scene. Among conflicting failures, the model itself judged every action corresponding to the requested value as inappropriate in 9.8% of cases. Also, even when models judge an action corresponding to the requested target value as appropriate, they follow the request only 17.5% of the time.
Together, these experiments show that inappropriate target actions do not explain most steering failures. Even when a model judges an action to be appropriate, it often fails to steer its decision toward that action. We have updated these analyses in Section 5 (Justifiable refusals paragraph) and Appendix F.2.
[1] Zhou, Jeffrey, et al. "Instruction-following evaluation for large language models." arXiv preprint arXiv:2311.07911 (2023).
[2] Qian, Yusu, et al. "Mia-bench: Towards better instruction following evaluation of multimodal llms." International Conference on Learning Representations. Vol. 2025. 2025.
Part 3 (3/3)
W3. Strengthen the evidence for visual grounding.
Thank you for the feedback on the visual grounding of our dataset. We first clarify why we include images in our evaluation. Household robot planners typically receive visual observations alongside task and contextual information, so we include realistic scene images to reflect this practical input setting.
Our evaluation provides candidate actions that describe concrete responses to the household situation. These descriptions may themselves convey relevant situational information, reducing the need to recover that information from the image. For example, the action "R1 moves the dripping boots to a mat to keep the floor safe for everyone" (Figure 1) already conveys both the presence of dripping boots and the associated safety concern, which the model need not infer solely from the image. Thus, the availability of textual cues does not by itself make the task independent of the household situation; relevant situational information can be conveyed not only by images but also by context descriptions and candidate actions.
That said, we agree that the similar aggregate value rankings in Figure 10 leave open how much images influence individual action choices. To examine this, we first analyzed how much decisions change depending on whether an image is present at the item level. We analyzed the most voted action across five runs for each item and model. An average of 35.2% of robot decisions changed when there is no image input. However, the default value preferences remained similar. We found that about a quarter (24.4%) of these changes were between actions that have the same value. Despite these item-level changes, aggregate value rankings remained similar.
We also tested how changes in visual input affect action choices in the steering task. To examine this, we conducted additional experiments measuring how often models change their choices when images are (1) removed or replaced with (2) a random image or (3) a grey image. The results show that models change their action choices by (1) 20.7--33.3%, (2) 13.2--28.9%, and (3) 15.6--29.6% in each setting. This shows the models' sensitivity to changes in visual input.
Furthermore, we tested whether the visual scene matters beyond the text of the candidate actions in the steering setting, by removing the household decision context, which also gives partial information that could be obtained from the image. This leaves the image as the only information beyond the candidate actions. In this setting, models change their action choices by (1) 31.9--54.0% with no image, (2) 29.4--56.5% with a random image, and (3) 27.9--44.3% with a grey image.
These experiments indicate that visual inputs influence model decisions even when the textual context is removed, suggesting that action choices are not solely decided by candidate action wording.
For the steering experiments above, we used 160 randomly sampled images, resulting in 114,560 steering queries across eight input conditions across 10 models. We additionally ran the context removal experiment on a separate set of 500 images and observed a similar trend. We incorporated these results in Appendix F.5 and a corresponding reference in Section 5.
Gentle follow-up
As the discussion phase is at the late stage, we would like to gently follow up and ask whether our rebuttal has addressed your concerns.
To briefly summarize, we have conducted human validation with 87 annotators. Also, we have conducted additional experiments to further clarify the interpretation of default preferences and value steering, as well as the visual grounding of our dataset.
Part 1 (1/3)
We thank the reviewer for stating that the problem is important, the benchmark construction systematic, and the evaluation comprehensive. Also, we appreciate the constructive feedback. In response, we ran additional experiments and human surveys to address the concerns. Unless otherwise noted, we use the 10 VLMs from the main paper.
W1. Validation on stakeholder perspectives.
We agree that our audit shows consistency with our own annotations. For a more thorough validation, we ran two human surveys on the Prolific survey platform with each item rated by 3 to 4 annotators balanced over age groups (18--29 / 30--49 / 50--64 / 65+) and gender. The two surveys cost £348 (about \460),andannotatorswerepaid£9perhour(about12).
For 60 random scenarios and a total of 302 stakeholder reactions, 47 annotators judged 94.1% of the situations and 89.1% of the stakeholder reactions realistic. We provided the annotator an image, household decision context and asked the annotator "Is this situation realistic for a real household?". Then we additionally provided robot action, each stakeholder information (e.g., P1: Juan (58-year-old husband), primary resident observing the robot; wants the birthday cake…), and the stakeholder's reaction (e.g., I understand the robot wants to avoid risk to the cake, but…). Then we asked the annotator "Is this a realistic reaction from P1?".
For 60 other scenarios, 40 annotators agreed with the benchmark value label for 89.2% of the actions. These results support the plausibility of our generated stakeholder reactions and scenarios. Also we would like to mention that our design choice of using five LLMs from three model families partially mitigate the potential bias caused by the generation model.
The above stakeholder reaction validation and value label validation experiments are updated in Section 4 (Benchmark Validation) with details provided in Appendix C (Human Validation).