{"id":"a4d41c5d-935f-4c24-9387-90d6bb5e2ea4","arxiv_id":"2509.05091","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Theory of Mind-inspired facilitator that infers agents' goals and sends utility-ranked feedback can raise task success and speed in two multi-agent environments, outperforming large language model facilitators.","lead":"This paper introduces ProToM, an AI observer that watches two agents with separate goals in a shared gridworld, guesses what each is trying to do, and sends short feedback messages that help one agent help the other. If it holds up, it is a step toward AI systems that actively nudge people into cooperative or helpful behavior, rather than only avoiding harm.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Simulation is a closed loop: the same planner both selects feedback and simulates the 'human' agents; only the 18-person human study tests transfer, so the central generalization is conditional.","rationale":"The paper's central contribution is a facilitator that selects feedback by expected utility under an inferred goal distribution. The strongest claim is that ProToM provides helpful feedback and is preferred by humans. I read the simulation as internal validation: it shows that when agents are perfectly compliant and plan exactly with the model ProToM uses to evaluate feedback, the method outperforms LLM baselines in choosing which messages to send. That is valuable but does not test the model's realism. The human study is the external anchor, and it is genuinely positive—participants could ignore messages and still rated ProToM higher and achieved better outcomes. However, with 18 participants, a single baseline, and one task, it cannot establish that the simulation's speedup numbers transfer, particularly because the explanation template may reveal private goal information. The threshold inconsistency (φ=2 vs 0–1 search) is a concrete reporting flaw that should be corrected, but it does not by itself overturn the human results. I therefore agree with the reader's weakest assumption: the compliance and planner-identity assumption is load-bearing and untested in the simulation. The proposed computational test—misspecifying the simulated agent's planner—directly targets that assumption and can be run with the released code. If ProToM remains robust under a different planner model, the concern is resolved; if not, the claim should be downgraded to 'works under perfect model match, with promising but preliminary human evidence.'","tokens_in":14443,"tokens_out":7231,"duration_ms":86897,"concrete_test":"Re-run the Overcooked simulation replacing the simulated human agents with a planner that differs from the π used in utility computation (e.g., a Boltzmann-rational policy with temperature τ=1, or a planner with altered movement/chopping cost weights), while keeping ProToM's feedback selection and thresholds unchanged. Compare success rate and speedup against the No-Facilitator condition and against the strongest LLM baseline (GPT-4o or o3). If ProToM's advantage disappears or drops below significance, the simulation results are an artifact of the same-planner loop; if the advantage holds, the model is robust to misspecification. Also report the φ/ϵ selection grid to resolve the φ=2 inconsistency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the transfer from ProToM's internal model to real agents. Section 3.2 states 'We assume that agent i will follow the feedback upon receiving it' and Eq. 12–14 define utility using planners π. In the simulation experiments (§4.4), the 'human' agents are exactly those same planners: A* for mDKG and the heuristic planner for Overcooked. Consequently, the reported simulation success rates and speedups measure how well feedback chosen under π is executed by π—a self-consistency check, not evidence that the model matches human planning or compliance. The only independent evidence is the human study (§4.5): 18 participants, one environment, and one LLM baseline (GPT-4o). That study also conflates the feedback-selection mechanism with explanatory templates (Eq. 17, §B.2) that explicitly reveal the other agent's inferred goal, which may be the dominant source of helpfulness. Reproducibility is further weakened by an internal inconsistency: φ=2 for Overcooked (§4.2) lies outside the stated grid search range 0–1.0 (§B.1). Thus the central claim that ProToM 'provides targeted and helpful feedback' in general rests on an untested assumption that human planning and compliance resemble π; the simulation numbers cannot establish that.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces ProToM, a facilitator that observes two agents pursuing independent goals, maintains a Bayesian belief over their goals via inverse planning, evaluates candidate feedback messages by expected utility (defined as step-count reduction under the agents' planners), and communicates only when utility and divergence thresholds are met. ProToM is evaluated in mDKG and Overcooked with simulated agents and compared with VLMs/RMs, and in an Overcooked human study with 18 participants (9 pairs) against GPT-4o and no-facilitator controls. The paper reports higher success rates, larger speedups, lower communication overhead, and higher subjective ratings for ProToM, and concludes that ProToM provides targeted, helpful feedback.","tokens_in":14736,"tokens_out":5649,"duration_ms":60996,"significance":"If the claim is established, the paper makes a useful contribution: it formalizes a third role for AI beyond human-AI coordination and team intervention, and it offers a modular theory-of-mind method that is more transparent than LLM-only feedback. Among the paper's strengths are the public code, the explicit two-level POMDP formulation, and a human study that goes beyond simulation. The core limitation is that the simulation evidence is close to circular; because the utility model and the simulated agents use the same planners, the simulated speedups are largely built into the objective. The human study is the only independent evidence, and it is small and partially confounded by the explanation templates. The conclusion as written overstates the support.","major_comments":[{"comment":"The load-bearing simulation evidence is self-consistency rather than independent validation. U(f) in Eqs. (12)–(13) is the expected reduction in completion cost C(f,g)=ℓ(π|f,g) (Eq. 14) under the same planner π that is used to simulate the agents in §4.4 (A* in mDKG and the heuristic planner in Overcooked). Section 3.2 further assumes the agent follows the feedback. Thus a positive utility almost mechanically implies that the simulated 'human' agent, which is π executing the selected feedback, completes the task faster. The simulation therefore cannot establish that ProToM's goal inference or feedback selection transfers to people. I recommend either evaluating with agents whose policies differ from the planner used in U(·) (e.g., noisy, model-mismatch, or independently trained agents), or explicitly presenting the simulation as a sanity check and placing the burden on the human study.","section":"§3.2, Eq. (12)–(14); §4.4"},{"comment":"Threshold tuning is reported inconsistently, which undermines reproducibility. Section 4.2 states that for Overcooked ProToM uses φ=2 (and ε=0.3), while §B.1 says φ and ε were chosen by a small search on a held-out set with values from 0 to 1.0 in increments of 0.1. A threshold of 2 is outside the stated search range. Please correct the reported value or the search range; otherwise the reader cannot reconstruct the experiment, and the selection of φ=2 appears post hoc.","section":"§4.2 vs. §B.1"},{"comment":"The only independent evidence for the central claim comes from the human study, but that study conflates the feedback-selection mechanism with the explanation templates. ProToM's communicated feedback is always accompanied by an explanation that explicitly states the inferred goal of the other agent (Eq. 17, templates in §B.2). Participants' high ratings on 'understood my goals' and 'helpfulness' could therefore be driven by the goal-revealing explanation rather than by the expected-utility selection. Since the LLM baseline is also asked for explanations, a cleaner comparison would require an ablation that delivers the same explanations with random or oracle-selected feedback, or that removes explanations from both arms.","section":"§4.5; Eq. (17), §B.2"},{"comment":"The human study is small: 18 participants, 9 pairs, one environment, one LLM baseline, and two trials per condition. With these numbers, the reported speedup differences (0.75±0.17 vs −0.28±0.11) may be real but are fragile; the study is also the only evidence that real humans comply with feedback well enough for the utility model to transfer. The conclusions in §6 that ProToM 'provides targeted and helpful feedback' and is 'consistently preferred by human users' go beyond what this sample supports. Please either expand the study, report effect sizes and power, or temper the scope of the claims.","section":"§4.5"}],"minor_comments":[{"comment":"The phrase 'higher communication overhead and task speedup' appears to be a typo; presumably 'lower task speedup' or 'reduced speedup' is intended.","section":"Abstract; §1"},{"comment":"Please clarify in the text or figures which exact statistical test and per-condition sample sizes correspond to the significance markers, especially in Figure 6 where the perceived-rating comparisons are reported.","section":"§4.3"},{"comment":"The initialisation of the belief particles is not described in the text. A brief sentence on the initial proposal distribution would help reproducibility.","section":"Algorithm 1"},{"comment":"Consider using 'simulated agents' rather than 'human agents' for the A* and heuristic planners, to avoid the impression that the simulation involved actual human participants.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and within the scope of the journal, but the simulation evidence is largely circular and the only independent evidence is a small human study. I do not think the paper should be rejected, but the authors should be asked to reframe the simulation as a sanity check, resolve the φ threshold inconsistency, and address the explanation confound in the human study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: this paper introduces a paradigm that is actually missing from the literature—an AI observer that nudges multiple agents with independent goals toward prosocial behavior—and it backs it with a live human study, not just simulated agents. That is a real contribution, and the authors deserve credit for framing it clearly and for shipping code.\n\nWhat is new and good: the facilitator framing (observer, no shared goal, feedback to individuals) is distinct from the cited human-AI coordination and team-intervention work. The method is coherent: Bayesian inverse planning for goal inference, expected-utility feedback selection, a divergence gate to avoid over-communicating, and templated explanations grounded in inferred goals. The human study (18 participants, 9 pairs) is genuinely informative: participants could ignore the feedback, and they still preferred ProToM over GPT-4o on all four subjective ratings, with better task success and speedup. That is independent evidence that the approach transfers beyond the simulation.\n\nNow the soft spots. The simulation results are largely a closed loop. The utility U(f) is the expected step reduction under planner π, and the simulated \"human\" agents are exactly those planners, executing the selected feedback. So the reported speedups mostly confirm that feedback chosen by π is followed by π—a self-consistency check, not evidence about real human planning. The human study is the only non-circular test, and it is small, uses a single LLM baseline, and the explanation templates (which explicitly state the other agent's goal) might be the main driver of perceived helpfulness, not the feedback selection itself. There is also a concrete reporting inconsistency: for Overcooked the main text sets φ=2, while the supplementary says the grid search was 0 to 1.0. That should be cleaned up. The paper's own limitations section is honest about not testing real-world settings and about ProToM-Oracle sending more messages than ProToM.\n\nVerdict: the central claim—that a ToM-informed facilitator can promote prosocial behavior among independent agents—holds up in human data, but the simulation evidence overstates the strength. The paper deserves a serious referee and likely a revision that (a) breaks the simulation circularity by using held-out planner types or real human trajectories, (b) reports threshold selection consistently, and (c) considers a comparison that isolates the explanation templates. I would cite this for the paradigm and the human study, and bring it up in a reading group with a warning about the simulation protocol.","headline":"A genuinely new facilitator paradigm with a real human study, but the simulation results are largely internal to the method's own assumptions.","tokens_in":15215,"tokens_out":922,"would_cite":true,"duration_ms":11653,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"An AI that observes, rather than joins, a group can nudge people toward prosocial actions by sending rare, goal-grounded feedback.","keywords":["prosocial behaviour","Theory of Mind","Bayesian inverse planning","multi-agent systems","facilitator agent","feedback generation","human-AI interaction","expected utility"],"falsifier":"Run the same Overcooked comparison with human participants told truthfully that the facilitator's suggestions are sometimes wrong, and track how often they follow them; if the perfect success rate and speedup over no-feedback disappear once users start ignoring messages, the compliance assumption is the load-bearing part. A cheaper check: replace the simulated agent's planner with another planner of similar competence and see whether ProToM's utility scores still rank feedback correctly.","tokens_in":14344,"feed_emoji":"🤝","tokens_out":9004,"duration_ms":93552,"temperature":0.7,"pith_summary":"This paper asks whether an AI can stand outside a group of humans who have different goals and, simply by sending a few words of feedback, get them to help one another. It proposes ProToM, which reads each agent's goals from observed behaviour with Bayesian inverse planning, then scores each possible message by how many steps it would save if obeyed, and sends it only if it would change the agent's plan and is expected to help. In simulation on two multi-agent environments, ProToM reached perfect success and clearly faster completion with much less chatter than GPT-4o and other large vision-language and reasoning models. An 18-person Overcooked study found the same pattern: humans succeeded more, finished faster, and rated ProToM's feedback as more helpful and better explained. The claim, if right, is that prosocial behaviour can be promoted by a purely observational AI, and that Bayesian goal inference plus expected-utility message selection does this better than end-to-end language reasoning.","feed_headline":"Goal-reading AI nudges humans to help each other, beating GPT-4o","feed_subtitle":"A facilitator that infers goals beats GPT-4o by saying less, not more.","key_machinery":"The load-bearing machinery is Bayesian inverse planning over a particle-filtered belief of each agent, feeding an expected-utility message selector. Inverse planning means working backwards from observed actions to a distribution over goals; the particle filter lets the facilitator track what each agent believes while only seeing their observations. The selector evaluates each candidate message by simulating the agent's plan with and without it under sampled goals and scoring the reduction in steps-to-goal; a divergence gate suppresses messages that would not alter the agent's behaviour, which is why ProToM communicates rarely. The formal wrapper is a two-level POMDP: the inner level is the","core_discovery":"ProToM claims that an AI facilitator needs no shared goal with the people it helps. It treats the observed agents as the hidden state of a POMDP, infers each agent's goal as a belief distribution via Bayesian inverse planning, and scores candidate feedback messages by the expected reduction in the number of steps to the agents' goals, assuming the recipient follows the message. It communicates only when the best message clears a utility threshold and the agent's predicted trajectory with feedback diverges from the one without. In the paper's evaluations, this produces perfect success rates in both mDKG and Overcooked, substantially faster completion than no feedback, and fewer messages than","pith_inferences":["The same expected-utility gate could decide when to interrupt in non-text assistive systems, such as highlighting a relevant object in an interface, because the feedback set is discrete and grounded in the environment state.","A natural next experiment is to make the facilitator adaptive: learn each user's compliance rate and the planner they actually use, then re-rank messages under that empirical model; ProToM's current utility computation assumes obedience and the paper's own planner.","Because the human study used 18 participants and one LLM baseline, the sharpest testable claim is whether the advantage survives with users who have reason to distrust the facilitator; that would isolate the compliance assumption from message quality."],"forward_implications":["An AI that never takes actions in the environment can still improve multi-agent success and speed by sending a small number of goal-grounded messages.","Feedback should be withheld unless it changes predicted behaviour; over-communication, the failure mode of the tested large vision-language and reasoning models, reduces or reverses the benefit.","Explicit goal inference plus expected-utility selection is sufficient to beat those models on these two grid-world benchmarks, suggesting that modular Theory of Mind is a viable alternative to end-to-end language reasoning for real-time assistance.","Pairing each message with an explanation of the other agent's inferred goal matters: human participants rated ProToM higher on clarity and goal understanding and ignored fewer of its messages."],"supporting_citations":[{"why":"Supplies the mDKG environment, the A* planner used to simulate agents, and the Bayesian inverse planning machinery ProToM builds on.","marker":"Zhi-Xuan et al. 2024"},{"why":"Supplies the Overcooked environment and the human-AI coordination setting that ProToM adapts to independent goals.","marker":"Carroll et al. 2019"},{"why":"Supplies the Bayesian social goal inference formulation that underpins ProToM's goal inference and its divergence check.","marker":"Ullman et al. 2009"},{"why":"Supplies the partial-observability Overcooked adaptation with doors that ProToM uses for its partially observable experiments.","marker":"Wu et al. 2021"},{"why":"Supplies GPT-4o, the vision-language baseline against which ProToM is compared in simulation and in the human study.","marker":"OpenAI 2024"}],"fun_headline_variants":["AI with Theory of Mind beats GPT-4o at helping teams","Mind-reading AI facilitator beats GPT-4o with less talk","Goal-inferring AI gives better feedback than GPT-4o","ProToM: Mind-reading AI helps humans cooperate better than GPT-4o"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a person who receives feedback follows it and then plans exactly like the simulator ProToM uses to score the feedback.","fun_headline_variants_meta":{"raw":{"variants":["AI with Theory of Mind beats GPT-4o at helping teams","Mind-reading AI facilitator beats GPT-4o with less talk","Goal-inferring AI gives better feedback than GPT-4o","ProToM: Mind-reading AI helps humans cooperate better than GPT-4o"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001276,"raw_usage":{"total_tokens":5045,"prompt_tokens":724,"completion_tokens":4321,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":4244}},"tokens_in":468,"tokens_out":4321,"duration_ms":31720,"temperature":1.0,"reasoning_tokens":4244,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:36:54.720577+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same Overcooked comparison with human participants told truthfully that the facilitator's suggestions are sometimes wrong, and track how often they follow them; if the perfect success rate and speedup over no-feedback disappear once users start ignoring messages, the compliance assumption is the load-bearing part. A cheaper check: replace the simulated agent's planner with another planner of similar competence and see whether ProToM's utility scores still rank feedback correctly.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the mDKG environment, the A* planner used to simulate agents, and the Bayesian inverse planning machinery ProToM builds on."},{"cited_title":"K.; Griffiths, T.; Seshia, S.; Abbeel, P.; and Dragan, A","cited_arxiv_id":null,"evidence_quote":"Supplies the Overcooked environment and the human-AI coordination setting that ProToM adapts to independent goals."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian social goal inference formulation that underpins ProToM's goal inference and its divergence check."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies GPT-4o, the vision-language baseline against which ProToM is compared in simulation and in the human study."}],"review_version":1}