- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces NormGym, a multi-agent interactive benchmark for normative competence evaluation. The main purpose is to evaluate competence under norms that agents are unlikely to have seen during pretraining to test adaptability to new environments with custom social norms. For that purpose, they design 50 “silly rules”, described as arbitrary norms.
The suite is composed of 10 layouts (e.g. a cafe). Each episode is assigned a layout and a hidden norm, and all agents have two consecutive goals. The agents can take actions that violate the norm. However, such actions trigger an environment-issued, globally visible sanction.
The evaluation consists of two main steps: interactive learning, and post-interaction evaluation. In the former, an agent interacts with a fixed layout for 200 steps, simultaneously inferring the norm and taking actions to achieve its goals. For the post-interaction evaluation, the agent is given a frozen belief, and its performance is evaluated both in the “training” layout, and an unseen one.
The authors evaluate 3 closed source models and 4 open-source models, concluding that even the strongest models at the task struggle to infer the social norms from interaction. The agents are said to mainly learn through their own sanctions, rather than from observing other agents’ behaviours. Lastly, they attempt training Qwen3-VL-4B-Instruct on normative competence subtasks and show little impact on NormGym performance.
Critical strengths
- The authors collect human data on their benchmark to verify that the tasks are solvable and to give an indicator of what good performance may look like. Additionally, to separate norm inference from execution consistent with the norm, they test the performance of GPT-5.6 Sol when given the true norm of the environment.
- The manuscript includes a process-level analysis of belief inference. Rather than simply reporting performance on the benchmark, the authors analyse the belief acquisition trajectory throughout the interaction, and investigate how different sources of evidence (own violation, observed violation) impact belief change.
Critical weaknesses
- Section 3.3 states that agents receive a textual state description including, among other things, all agents’ sanction signals. However, no such information seems to be included in the Appendix H prompts. My current understanding is that agents receive information about other agents’ sanctions purely from a red cross on top of the sanctioned agent in the rendered frames, and that these are not recorded in memory. By contrast, the agent’s own sanctions are included in the text, flagged in the 10 most recent steps in memory, and in a buffer containing up to 5 of the sanctioned earlier steps (Appendix C.2). My concern with this is that one of the main conclusions of RQ2 is that agents mainly learn norms via their own sanctions. However, if the setup is as I described above, there is a clear asymmetry in the way that both sources of evidence are presented to the agent, with its own sanctions being much more salient.
- Each norm has a single interactive learning phase, whose final belief is frozen for the evaluation phase. The 6 test episodes therefore measure the variability in belief execution rather than belief acquisition, even though studying this acquisition is the main purpose of the benchmark. Additionally, Table 1 reports no confidence intervals, making it difficult to interpret differences across models and types of norms.
Questions for authors
- Could the authors state clearly what information the agent receives via text (including per-step prompt and memory) vs through the rendered image? And, if the setup is as described in my first weaknesses bullet point, could the authors run an ablation in which other agents’ sanctions are presented in the same way as the agents’ own sanctions, and report the resulting change in belief correctness per source of evidence (same as Figure 4c)?
- Could the authors report confidence intervals for the main results in Table 1? And, if there is one single interaction phase per norm (i.e. one belief per evaluation), could the authors run multiple interaction phases to estimate the variability in belief acquisition?
6 comments
W2, Q2 — Acquisition variability
(1) Thank you for raising this point. We would like to clarify that the learning phase is not a single episode, but it is a continuous process over 200 steps where the agents can update their belief about the norm at each step. At every step, the agent gets new evidence from their action, and there are new background agents coming in with different identities and goals. This gives the agent repeated opportunities to observe different situations and refine its understanding. We set a fixed 200-step budget to keep the learning conditions consistent across models and norms. We have clarified the evaluation protocol in Appendix C.1 of the revised manuscript (lines 1058-1062).
We also understand this is a valid concern, so we ran an additional experiment to further check the consistency of norm acquisition. We re-ran the 200-step interactive learning phase of GPT 5.6 Sol on 20 norms, randomly selecting 5 from each norm family. We used the same environment and seed as in our new ablation experiment above with the other-agent violation information. We then used the LLM judge to score the final beliefs from both runs on the same 1-7 scale. The average score was 3.65 in both runs, with no significant difference between them (paired Wilcoxon signed-rank test, p = 0.89). This shows that norm acquisition is reasonably consistent across repeated learning runs. We have also added this repeated-learning experiment and its results in Appendix D.5 of the revised manuscript.
(2) We would also like to clarify that the purpose of having six test episodes here is not just to evaluate whether the agent can execute its final belief, but to test whether what it learned during interaction can generalize to new situations. In other words, we want to see if it really learns the norm, not just memorizing patterns and behaviors from the learning phase. We have 3 test episodes that are in the same layout seen by the agent in the learning phase, and 3 test episodes that are in a different layout. For each test episode, all agents in the environment have different initial positions and goals. By having independent test episodes with different initial setups, we can know whether the agent has learned the underlying norm. We design it this way to eliminate the possibility that agents just memorize specific behaviors. We have clarified the evaluation protocol in Appendix C.1 of the revised manuscript (lines 1083-1088).
In our experiments, we also did the Know-Norm baseline, where the model is directly given the true norm, to help separate failures in learning the norm from failures in executing it, which indeed shows that the model can achieve near-perfect joint success when given norms, but not when not given them.
W2, Q2 — confidence intervals
We report 95% percentile-bootstrap CIs over norms in Appendix D.4. Human results cover 25 randomly selected norms and report the best observed performance per norm, so no CI is reported. These intervals help show the uncertainty in the performance estimates, so we should be cautious about drawing strong conclusions from small differences between models or norm families. At the same time, our main finding remains clear that learning and following unfamiliar community norms remains challenging.
Table 2. Performance on NormGym (%). Each cell shows performance in seen / unseen layout, as percentages with 95% CIs in brackets. The columns Access, Procedure, Coordination, and Allocation report joint success within each norm family.
| Baseline | Access | Procedure | Coord. | Alloc. | Joint Success, all | Task Success, all |
|---|---|---|---|---|---|---|
| Claude Opus 5 | 69.4 [50.0, 88.9] / 69.4 [44.4, 91.7] | 48.5 [21.2, 75.8] / 54.5 [27.3, 81.8] | 66.7 [47.9, 83.3] / 54.2 [37.5, 70.8] | 30.3 [21.2, 39.4] / 24.2 [9.1, 39.4] | 55.3 [44.7, 66.0] / 51.3 [39.3, 62.7] | 95.7 [91.7, 98.7] / 96.7 [92.7, 99.3] |
| GPT-5.6 Sol | 69.4 [52.8, 86.1] / 44.4 [22.2, 69.4] | 45.5 [21.2, 69.8] / 57.6 [30.3, 84.8] | 20.8 [8.3, 33.3] / 29.2 [18.8, 41.7] | 48.5 [27.3, 72.7] / 36.4 [12.1, 60.6] | 44.0 [33.3, 54.7] / 40.7 [30.0, 52.0] | 86.3 [78.3, 93.3] / 83.0 [74.0, 91.3] |
| Gemini 3.1 Pro | 36.1 [11.1, 63.9] / 36.1 [13.9, 61.1] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 37.5 [16.7, 58.3] / 14.6 [6.2, 25.0] | 15.2 [0.0, 36.4] / 15.2 [0.0, 33.3] | 24.0 [14.0, 35.3] / 16.7 [8.7, 25.3] | 90.3 [82.7, 96.7] / 87.7 [79.3, 94.7] |
| GLM-4.6V-Flash | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 11.7 [8.3, 15.7] / 15.7 [11.0, 20.7] |
| Qwen3-VL-8B | 5.6 [0.0, 16.7] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 3.0 [0.0, 9.1] | 1.3 [0.0, 4.0] / 0.6 [0.0, 2.0] | 38.3 [31.7, 45.3] / 34.0 [26.7, 41.3] |
| Llama-3.2-11B | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 3.0 [1.0, 5.0] / 3.0 [1.3, 5.0] |
| Ministral-3-8B | 8.3 [0.0, 16.7] / 8.3 [0.0, 16.7] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 2.1 [0.0, 6.2] | 6.1 [0.0, 15.2] / 6.1 [0.0, 15.2] | 3.3 [0.7, 6.0] / 4.0 [1.3, 7.3] | 49.7 [42.0, 57.0] / 50.0 [41.7, 58.3] |
| RL (DQN-MLP) | 33.3 [8.3, 58.3] / 8.3 [0.0, 22.2] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 8.0 [2.0, 16.0] / 2.0 [0.0, 5.3] | 10.7 [3.7, 19.0] / 3.0 [0.0, 7.3] |
| RL (DQN-CNN) | 44.4 [19.4, 72.2] / 8.3 [0.0, 22.2] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 9.1 [0.0, 27.3] / 0.0 [0.0, 0.0] | 12.7 [4.7, 22.0] / 2.0 [0.0, 5.3] | 16.0 [7.3, 25.3] / 11.3 [6.0, 17.3] |
| Known-Norm LLM | 100.0 [100.0, 100.0] / 100.0 [100.0, 100.0] | 84.8 [63.6, 100.0] / 84.8 [63.6, 100.0] | 91.7 [77.1, 100.0] / 83.4 [72.9, 93.8] | 84.8 [69.7, 97.0] / 87.8 [78.8, 97.0] | 90.7 [83.3, 96.7] / 88.7 [82.0, 94.0] | 99.7 [99.0, 100.0] / 97.0 [94.3, 99.3] |
| Norm Oracle | 100.0 [100.0, 100.0] / 100.0 [100.0, 100.0] | 100.0 [100.0, 100.0] / 100.0 [100.0, 100.0] | 100.0 [100.0, 100.0] / 100.0 [100.0, 100.0] | 100.0 [100.0, 100.0] / 100.0 [100.0, 100.0] | 100.0 [100.0, 100.0] / 100.0 [100.0, 100.0] | 100.0 [100.0, 100.0] / 100.0 [100.0, 100.0] |
| Norm-Blind Planner | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 0.0 [0.0, 0.0] / 0.0 [0.0, 0.0] | 100.0 [100.0, 100.0] / 100.0 [100.0, 100.0] |
W1, Q1 — Missing information about other agents' violations in the prompt
Thank you for pointing this out. To address this, in the prompt we added an explicit note about other agents' violations everywhere the agent is told about its own (though the other agents’ violations have already appeared in the rendered image):
Nothing else changed, and we re-ran GPT-5.6 Sol on all 50 norms.
The same pattern remains after these changes. The table shows the average change in belief correctness immediately after each type of event (same measure as in the paper Figure 4 panel c), reported as mean ± SE. Explicitly including other agents' violations in the prompt and memory increased belief improvement after observed violations from 0.023 to 0.057. However, the model still updated its beliefs more strongly after its own violations (0.162), while belief changes following no violation remained minimal (0.004). This shows that the difficulty of learning from others is not solely due to how violation information was presented in the original prompt. We have added this ablation experiment and its results in Section 4.4.2 (lines 421 to 426) and Appendix D.3, with the modified prompt and memory format provided in Appendix H.1.
Table 1. Effect of explicitly recording other agents’ violations in the memory on belief revision
We also checked whether these prompt changes affected the model's overall performance. Joint success remained similar, changing from 44.0% to 48.7% on seen layouts and from 40.7% to 44.0% on unseen layouts. Paired Wilcoxon signed-rank tests across all 50 norms showed no statistically significant difference in joint success between the original run and the new run, on either seen layouts (p = 0.34) or unseen layouts (p = 0.43). Even with the additional information about other agents' violations in the prompt, the model still performs poorly. This gives us additional confidence that our overall findings are not driven solely by how other agents' violations were presented in the original prompt.