- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2(+1)
Read review
Paper summary
This paper proposes MOSAIC, a benchmark for evaluating whether multimodal models can act on social inferences rather than only report them. Most theory-of-mind benchmarks ask a model questions about someone else's mental state. Here, two embodied agents share a virtual room, and one of them is responsible for changing what the other believes. The informed agent knows which of two boxes hides a reward. It must help its partner in cooperative mode and work against it in competitive mode, where actively misleading the partner is expected only under the deeper-reasoning condition. The uninformed agent has to work out where the reward is from the informed agent's behavior and words. Signals travel through several channels at once. These are where the informed agent walks, where it looks, its facial and physiological expression, and three short spoken exchanges. The outcome is decided by which box the uninformed agent ends up closest to. The authors cross the cooperative/competitive framing with how deeply the informed agent is asked to reason, either acting on its own preferences (ToM-0), or deliberately modeling and steering its partner's beliefs (ToM-1). This gives four conditions, labeled Cooperative-ToM0 through Competitive-ToM1. They evaluate a range of small open vision-language models (1B–14B), a text-only model, and PCM-LLM that serves as a reference point.
The most notable aspect is how the evaluation separates outcome from mechanism. The primary score, TOCS, measures how often a trial's outcome matches what its condition predicts. It is read alongside two checks on whether apparent success is real. One is positional bias, a fixed habit of picking the same box. The other is an uncertainty rate, the share of trials where the agent ends equidistant from both boxes, which the authors read as a failure to move. The separate signal scores ask whether the informed agent produced a usable signal. They cover how often it shows facial expression and whether its gaze and trajectory point toward the reward. A final score asks whether the partner's choice followed that signal. This shifts the question from whether a model can win the game to where it breaks along the chain from inference to signal, to response. The central claim comes out of this framing. No vision-language model's outcome differs significantly from chance after multiple-comparison correction, and telling the informed agent to reason more deeply does not reliably change its behavior. Most models fail to produce coherent nonverbal signals in the first place, and when signals are present, the partner does not act on them. Only PCM-LLM gets through both steps in the original layout, and even its deception falls apart when the room layout changes. The authors also report that removing the rendered image does not hurt the two models it was tested on, and that imitation fine-tuning on PCM-LLM's behavior does not close the gap.
Critical strengths
Critical Strengths
- Social reasoning is tested through action, not question answering.
- The informed agent must change its partner's belief through movement, gaze, expression and speech, and the score depends on hat the partner finally does. Most ToM benchmarks only ask about mental states, and most multi-agent ones stay in language.
- The 2×2 design gives each condition a clear expected outcome. The task is solvable, since the reference architecture succeeds in most conditions, yet no VLM differs from chance.
- This is the paper's main contribution.
- Diagnostics that catch shortcuts and show where a model fails.
- Positional bias and uncertainty rate reveal when apparent success comes from a fixed habit or a failure to move.
- Separate scores for the sender's signals and the receiver's response show which side of the exchange breaks.
- Changing the room layout breaks the reference architecture's deception, and the authors trace this to a specific flaw in how it models visibility.
Critical weaknesses
-
Using one model for both agents prevents a clean separation of signal production and signal decoding.
- The same model checkpoint plays both the subject and the participant in every run (Sec. 3). FES, GAS and TAS measure subject-side signal production, but SSS and TOCS depend on both agents together. Participant-side decoding is therefore observable only when the same model also produces clear subject-side signals, which the paper reports for very few models. A model that produces clear signals but decodes poorly can also obtain the same TOCS as a model that fails at both.
- Evaluating each model against a fixed reference agent in the other role would separate subject-side and participant-side performance and make both comparable across models.
- The title claim, that models form correct beliefs but fail to act on them, has no corresponding metric. Sec. 4 describes its evidence as an exploratory analysis of one hand-picked case per failure type, with no frequency estimates, although the logged reasoning traces would allow a trial-level count. A benchmark's value rests on scores that can be attributed to the capability it claims to measure.
-
The task design confounds social reasoning with other factors.
- Movement actions are specified in egocentric directions, while positions are given in world-frame coordinates (App. B). The paper's own cases show the reasoning trace naming the correct box, followed by movement toward the wrong one because world-frame coordinates are mapped onto egocentric directions incorrectly (Sec. 4, App. E). TOCS, therefore, mixes social reasoning failures with spatial grounding failures, and no ablation with object-targeted movement separates the two.
- The ToM conditions are set only through prompt text, with no check that models follow them. The subject is told its strategy "should match" its ToM order and that it "may" mislead, and the prompt defines ToM order 0 only by how much an agent trusts others, not by how the subject should act. The authors note that this does not establish that models follow the specified reasoning (Sec. 3.2). The null effect of ToM order therefore cannot distinguish an inability to reason at ToM-1 from not following the instruction.
- The benchmark is framed as multimodal, but positions, orientations and preferences are also provided in the structured belief-and-observation state. The text-only Llama-8B reaches comparable TOCS and higher TAS than most VLMs (Table 4, Fig. 3), and removing the rendered image does not degrade either model tested. The necessity of the visual channel is not established.
- The subject-side metrics do not distinguish deceptive from honest signaling. Under Competitive-ToM1, PCM-LLM obtains TAS 1.00 and GAS 0.74, the same sign as a subject the paper describes as following an honest strategy (App. I.2), whereas Table 1 defines misdirection as a negative score. SSS is also affected by fixed spatial biases of both agents in the canonical layout (App. D.1), contrary to the statement in Sec. 4.3.1 that this layout largely avoids the problem. - TOCS places every VLM at chance, so differences between models appear only in the signal-level metrics (App. D.2.5).
-
Several reported results are inconsistent with the paper's own tables and figures.
- PCM-LLM's Competitive-ToM1 result in that the canonical layout is not significant in Table 4 (p_adj = 0.136) but is significant in Table 7 (p_adj = 0.046), likely because the two tables apply Holm correction over different sets of tests. The chi-square tests diverge in the same way (Tables 5 and 8). The paper does not state which correction applies.
-
Several values reported in the text differ from the corresponding figures and tables.
Questions for authors
- In the participant's belief states shown in App. E.2, fields such as "Marie | felt emotion valence" and "Marie | preference towards Dark brown box" appear. Are these the participants' own estimates or the subject's true values?
- After each verbal exchange, the subject's preference-updating module can change its preference toward the participant and toward the boxes. Both the interaction mode and the subject's belief about where the reward is are encoded as these preferences. How often does either one change during a trial, and how are such trials handled in the analysis? Are there any statistics present on this? If not, can the authors provide?
- App. C.3 defines GAS using a gaze certainty for each box and a threshold θ_g. How is this certainty computed from the subject's behavior, and what value of θ_g is used?
- App. B.1 says reward locations were assigned without a fixed seed, while App. F.2 says identical seeds were used across models. Which is correct? Since each cell appears to have been run once at temperature 0.7, could you report confidence intervals for the signal-level metrics as well as for TOCS?
- App. F.3.2 reports InternVL-14B SSS values from −0.227 to −0.731 in the aligned layout, and App. F.3.3 describes its TAS and GAS as showing strongly negative correlations in the orthogonal layout. Neither matches the per-condition values in Fig. 6. Are these values computed separately by reward location?
- The Discussion mentions a study with 35 human participants under a related nine-round protocol. Could you report its methods, ethics approval, and results, and explain how its protocol differs from MOSAIC?
- For the two fine-tuned InternVL models (App. H), how many demonstrations were used, which agent role was trained, and what training settings were used?
- Using the logged reasoning traces, please report per model and condition how often the box named in the reasoning differs from the box the agent then moves toward.
- Please report SSS only for the trials in which the subject produced a clear directional signal, so that participant-side decoding is measured only where a signal exists.
- Why are the evaluations not presented against a fixed reference agent in the subject role?
- Please see the Weaknesses section. I have mentioned some issues there.
10 comments
Response to Q1-Q4
Response to Q1:
Thanks for this question. We clarify that they are the participant's own estimates, not the subject's true values, so there is no information leakage that would undermine attributing the participant's success to social reasoning. Each agent receives only its own belief state. Felt emotion and preference are the participant's estimates of the subject, updated by its own belief module from observations (or dialogue), never copied from the subject's state. "Marie | felt emotion valence" makes this clearest: felt emotion is internal, so others can only infer it from facial and physiological cues.
PCM-LLM's successful deception also requires these fields to be estimates. In Comp-ToM1, the subject shows a negative expression while gazing at the reward box, the participant's estimate of "the subject prefers the reward box" drops to , and 70% of trials end at the empty box. A participant reading true values could not be misled.
Response to Q2:
We thank the reviewer for this question and provide the statistics below.
Frequency. Each trial has three verbal exchanges, each followed by one output of the preference-updating module. Per trial, the module requests a change in the subject's preference toward the participant 0.58 times and toward the boxes 1.65 times on average. These requests reverse the interaction mode 0.17 times and the reward belief 0.21 times per trial.
Handling. The original analysis kept all trials and did not treat trials with reversal requests separately, because verbal updates are a small share of the subject's preference updates. At each of the 10 timesteps, the subject re-outputs its preferences in action prediction from the nonverbal interaction, whereas verbal updates occur only 3 times (rounds 3, 6, and 9) and reverse the mode or reward belief less than once per trial. Verbally induced changes are therefore likely to be overwritten or diluted by the many nonverbal updates that follow. They may matter in individual trials, but we suppose that their effect on the overall statistics is limited.
Response to Q3:
We thank the reviewer for pointing out this missing piece of the GAS definition. With this addition, GAS is fully reproducible. Gaze certainty reuses the sensory-certainty function of the Projective Consciousness Model (Rudrauf et al., 2022, J. Theor. Biol., Section 4.2.2). It is visual acuity, which decays exponentially as the box moves away from the center of gaze, divided by a distance term. With the angle between the gaze direction and the direction to the box, and the box distance projected onto the gaze direction:
Values lie in (0, 1]. There is no gaze command. The subject controls its gaze by moving and turning, so GAS measures the gaze signal it produces with its body.
is fixed at 0.3 in all experiments, chosen from how certainty varies with pose.
- Initial pose. Facing the participant, subject has initial certainty about 0.28 towards both boxes.
- Turning toward a box in place raises that box's certainty to about 0.44.
0.3 lies in between. The default pose counts as gazing at no box, and turning toward a box counts as gazing at it even without approaching it. Appendix C.3 now gives the formula, source, parameters, and this rationale.
Response to Q4:
Random seeds. Appendix B.1 is correct. The "identical seeds" statement in F.2 was wrong and has been removed. No experiment used a fixed seed. The reward location is drawn independently in each trial, so reward sequences differ across models. We checked balance after the experiments. In the main experiment, the reward is at in 465, 447, 449, and 460 of the 900 trials per condition (p 0.33, Appendix D.2). In the layout experiment, it is at in 20–32 of the 50 trials per cell, and no cell deviates from 0.5 (uncorrected , Appendix F.3). In F.1, "counterbalanced" now reads "drawn at random in each trial and checked for balance after the experiment".
Confidence intervals. Each cell was not run only once. It contains 50 independent trials, each sampled at temperature 0.7. We added 95% intervals for TOCS (exact Clopper-Pearson) and for TAS, GAS, FES, and SSS (10,000 bootstrap resamples) in every cell. The median half-widths are 0.14 for TOCS, 0.22 for TAS, 0.24 for GAS, 0.17 for SSS, and 0.04 for FES. The full table for all 18 systems and four conditions is added to Appendix D.2, and the intervals for the layout experiment to Appendix F.3.
Response to Q5-Q7
Response to Q5:
- F.3.2. (aligned layout, Internvl-14B). These values were not computed by reward location. They are a typo: the text was not updated to match the figure. The per-condition SSS shown in Fig. 6 is to , and the revision now reports these values.
- F.3.3 (orthogonal layout, Internvl-14B). "Both TAS and GAS show strongly negative correlations" was ambiguous. It meant that the subject's trajectory and gaze signals correlate strongly negatively with the participant's choice, that is, a strongly negative SSS, not that TAS and GAS are negative. The participant picks the dark box in 44–49 of the 50 trials per condition, while the subject's signals lean toward the light box. The two fixed preferences point to opposite boxes, so SSS is strongly negative. The sentence now reports TAS, GAS, and SSS directly.
Response to Q6:
This is a human-participant experiment we ran under a related protocol, and its full results are not in this paper. The main difference from MOSAIC is the number of rounds per trial.
Methods. Participants wore a VR headset (Meta Quest Pro) and replaced the participant agent in the competitive task against the subject agent, Marie. As in MOSAIC, Marie knows where the reward is, the participant infers it from Marie's verbal and nonverbal behavior, and the choice is the box the participant ends closer to. MOSAIC trials have 10 rounds, with verbal rounds at R3, R6, and R9. Human trials have 9 rounds, 7 nonverbal and 2 verbal (R3, R6), because each round gives both sides a fixed 10-second thinking period and 10 rounds would keep participants in VR too long. The design was within-subject, crossing Marie's architecture (PCM-LLM vs. Llama3.1-8B-Instruct) with ToM order (ToM-0 vs. ToM-1). Position, gaze target, and facial action units were recorded at every step, and participants answered a questionnaire in VR after each trial. As a reference, we ran 280 agent-agent simulations (70 per condition) under the same 9-round protocol.
Ethics and participants. The study was approved by the institutional ethics review board (institution and approval number withheld for anonymous review), and all participants gave written informed consent. Of 38 recruited participants, 3 were excluded because VR discomfort prevented them from completing all trials, leaving 35 (19 female, 16 male, age years).
Results. Task success was compared with a binomial GEE clustered on participant, and seven-round trajectories and gaze sequences with clustered bootstrap tests, with Bonferroni correction within each family. Task success did not differ between human and simulated participants in any condition (). Human/simulated TOCS: PCM-LLM ToM-0 0.67/0.64, PCM-LLM ToM-1 0.47/0.63, LLM ToM-0 0.49/0.44, LLM ToM-1 0.54/0.61. Behavior nonetheless differed: 9 of 12 continuous trajectory comparisons and all 4 gaze-sequence comparisons were significant.
Role in the paper. Because the round schedule differs and only the competitive condition was tested, the study cannot serve as a human baseline for MOSAIC, and the Discussion uses it only as human-level support. It shows that equal task success can coexist with different receiver behavior, in line with MOSAIC's point that outcome metrics are not self-validating and must be reported with signal production and receiver behavior.
Response to Q7:
The full setup is in the table in our response to Reviewer 2, W2, and is added to Appendix H.
Response to Weakness 1
Response to W1.1 and 1.2:
We thank the reviewer for the careful reading of our signal-level metrics. We agree that even role-specific metrics cannot cleanly separate the two roles, so we ran the suggested cross-play. The fixed reference agent is PCM-LLM, the only system that both produces and reads signals in self-play. It was paired with Internvl-4B, -8B, -14B, Minicpm-8B, and Qwen3.5-9B in both directions, with 50 trials per cell, otherwise as in the main experiment. We chose these five because their self-play subjects represent different channel profiles under cooperation: Internvl-14B gives clear trajectory and gaze signals (TAS 0.76–0.86, GAS 0.88–0.90), Qwen3.5-9B mainly gaze (GAS 0.52–0.58), Minicpm-8B gaze only in Coop-ToM1 (GAS 0.50, TAS ≤ 0.08), and Internvl-8B and Internvl-4B weak signals on both channels (TAS ≤ 0.26, GAS ≤ 0.42).
TOCS with the chance level in parentheses. Bold: above chance. ↓: below chance (Holm within each model's four conditions, p_adj < 0.05).
Therefore, the sender side and the receiver side can therefore be clearly distinguished.
Response to W1.3:
We accept this point and used all logs to add frequency statistics at two levels, now reported in Appendices E.4 and E.5.
Belief-to-action conversion is therefore a measurable bottleneck, concentrated on the participant side. It is, however, only part of the failure. For the other VLMs, the chain breaks earlier: the subject sends no directional signal, or the participant forms no correct belief. We therefore changed the title to a neutral one that covers every break point: "MOSAIC: Tracing Social Reasoning from Belief to Behavior in Embodied Multimodal Agents".