{"id":"78e6b92c-666f-404b-974c-70fe562b3ace","arxiv_id":"2507.21846","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An observer that actively moves and uses 'not seen' signals as evidence can infer a hidden goal faster than passive recognition, with MCTS planning matching a domain-specific greedy baseline on grid-world tasks.","lead":"This paper builds an active goal recognition system where an observer agent chooses where to look so it can identify another agent's hidden goal faster. It combines a Bayesian belief update with Monte Carlo Tree Search and tests it in grid-world simulations, reporting that the belief update beats passive goal recognition and that the planner matches a hand-crafted greedy baseline in easy and normal settings.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim relies on an unexamined actor-policy model: Eq. 2 uses an epsilon-greedy pi_hat although actors plan optimally under private costs, and no robustness analysis shows the belief/planning gains survive mismatch.","rationale":"The paper's central claim is that an observer using the proposed joint belief update and MCTS can recognize a hidden goal more effectively than passive goal recognition, and that the domain-independent MCTS is comparable to a domain-specific greedy baseline. Both halves of this claim depend on the same belief update, and the belief update depends on an assumed actor policy pi_hat via Eqs. 2 and 9. The paper explicitly states that the observer does not have access to the actor's policy, so pi_hat is an unverified modeling choice. In the experiments, the actor optimally follows a private cost map while the observer assumes an epsilon-greedy policy; this is exactly the kind of mismatch that can make a Bayes filter miscalibrated. The lack of any sensitivity analysis means we cannot tell whether the reported improvements are intrinsic to the framework or an artifact of the specific mismatch chosen. The formal derivation of the belief update is internally consistent conditional on pi_hat, and the paper is honest about MCTS underperforming in the large-hard configuration, which is creditworthy. The reader's conditional verdict is appropriate: the concern is substantial but addressable with additional robustness experiments, and it does not invalidate the overall framework. I therefore agree with the reader's weakest assumption and recommend no change to the conditional verdict.","tokens_in":13608,"tokens_out":10343,"duration_ms":127873,"concrete_test":"Re-run the six configurations under two diagnostic conditions: (A) generate actor behavior from the observer's own epsilon-greedy model, so there is no model mismatch; and (B) give the observer's belief update the true private-cost-optimal policy as pi_hat (oracle), while keeping everything else identical. Report CV, SR, and FP for AGR-MCTS and Belief-Greedy, with bootstrap 95% confidence intervals over the 50 instances per configuration. If condition B materially improves performance or condition A materially changes the numbers in Table 2, the reported results are sensitive to the actor-model assumption; if all differences are within the confidence intervals, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.1 introduces a private cost map known only to the actor, and has the actor execute an optimal path to its goal under that map. The observer is never given the map or the policy, yet the belief update in Eq. 9 and the planner's observation likelihood in Eqs. 19-20 are both built from P(st | st-1, g) = sum_a P(st | st-1, a) pi_hat(a | st-1, g) in Eq. 2, where pi_hat is set to an epsilon-greedy policy (Section 5.1). The true actor policy is optimal with respect to the private costs, so pi_hat differs from the real policy. Because the joint belief jt is the component credited with the gains over passive recognition (Figure 3), and because MCTS forward simulations sample observations from beliefs computed with this same pi_hat, a miscalibrated actor model can both distort the posterior over goals and steer the observer toward actions that are uninformative under the true actor behavior. The paper reports no sensitivity analysis over epsilon, no comparison with an oracle policy model, and no measurement of belief calibration. The central claim therefore rests on an unexamined assumption that the chosen pi_hat is a serviceable approximation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Probabilistic Active Goal Recognition (PAGR), a POMDP-based framework in which an observer maintains a joint belief over the actor's state and hidden goal, updates this belief by Bayes' rule, and selects informative observer actions via Monte Carlo Tree Search (MCTS). The squared marginal belief over goals is proposed as a reward signal, and the framework is evaluated in a grid-world domain with private actor cost maps. The central claims are that the joint belief update significantly outperforms passive goal recognition and that the domain-independent MCTS planner performs comparably to a domain-specific greedy baseline.","tokens_in":13918,"tokens_out":3940,"duration_ms":46334,"significance":"The formal contribution is sound in its core derivation: the joint belief update in Eq. (9) is a correct recursive Bayes update, and the squared-belief reward in Eq. (12) is derived from expected confidence rather than fitted. The paper also states a clear POMDP formulation that separates inference from planning, which is a useful step for active goal recognition. If the empirical claims were fully supported, the framework would offer a practical, knowledge-light approach to AGR. However, the current evidence is incomplete: there are no error bars or significance tests, no previously published AGR baselines, and the 'comparable' parity claim in the abstract does not survive the large-hard configuration. These gaps concern the paper's load-bearing empirical claims rather than the derivations.","major_comments":[{"comment":"The abstract's claim that domain-independent MCTS 'performs comparably' to the domain-specific greedy baseline is contradicted by Table 2 in the large-hard configuration, where AGR-MCTS obtains CV 0.12 vs. Belief-Greedy's 0.22 and SR 0.30 vs. 0.60; AGR-MCTS is also worse on S-N and L-N. Since the parity claim is one of the paper's two headline contributions, it should be tempered, or the configuration-level analysis should explain why parity fails precisely where the environment is hardest.","section":"Abstract; §5.3, Table 2"},{"comment":"The belief update and the MCTS forward model assume an epsilon-greedy actor policy pi_hat in Eq. (2), while the actual actor computes an optimal path under a private cost map unknown to the observer. Because Eq. (9) and Eqs. (19)-(20) all condition on pi_hat, the reported gains of joint belief updating over passive inference may depend on this assumed policy model. The paper provides no robustness analysis over epsilon, no comparison with an oracle or learned policy model, and no calibration measurements for bt(g). Please add sensitivity experiments (e.g., epsilon in {0, 0.1, 0.5}) or otherwise demonstrate that the belief-update gains survive model mismatch.","section":"§4.1 Eq. (2); §5.1"},{"comment":"The claim that joint belief update 'significantly outperforms' passive goal recognition is not supported statistically. Figure 3 plots means without error bars or confidence intervals, and no significance tests are reported. With 10 layouts and 5 instances per configuration, the per-configuration sample is n=50, and some differences in Table 2 (e.g., S-N CV 0.22 vs. 0.14) could easily be within noise. Report variances and run paired tests (bootstrap or Wilcoxon) over layout-instance pairs.","section":"§5.3, Figure 3"},{"comment":"The evaluation does not include any previously published AGR method, such as the landmark-based approach of Shvo and McIlraith (2020) or the POMDP/SARSOP approach of Amato and Baisero (2019), both discussed in §2.2. The baselines are passive recognizers or the authors' own greedy method. Consequently, the experiments show improvement over passive recognition in this grid domain, but they do not establish that AGR-MCTS advances the state of the art in AGR itself.","section":"§5.1, Table 2"}],"minor_comments":[{"comment":"Table 2 reports the third metric as 'FP' in the table but 'FB' in the caption; make the notation consistent.","section":"§5.3, Table 2"},{"comment":"The PAGR tuple is typeset as 'P P AGR' with an awkward line break; fix the formatting.","section":"§3.1"},{"comment":"The text refers to 'Coverage (CV)' while the metric is defined as 'convergence (CV)' in §3; use a single term throughout.","section":"§5.3"},{"comment":"Hyperparameters such as the discount factor gamma, the UCB exploration constant, and the entropy-regularization weight are not specified; please state their values and, if possible, include a sensitivity check.","section":"§5.1"},{"comment":"The y-axis scale and exact values are hard to read from the grouped plot; annotating the means or printing the numbers would improve interpretability.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The core derivations are clean, but the empirical section needs substantial strengthening before this can be considered for publication. In particular, the 'comparable' claim in the abstract should be revised or removed, and the model-mismatch robustness question should be addressed with experiments. I would recommend requiring a statistical analysis and at least one prior AGR baseline before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know: the paper combines a joint belief state-goal update with MCTS for active goal recognition, and that specific combination is new relative to the cited work. The math is clean: Eq. 9 is a correct recursive Bayes update, and the squared-belief reward is derived, not fitted. The ablation comparing joint belief to passive recognition on the same observation traces is the strongest part; it demonstrates that treating unobserved steps as negative evidence helps. The authors are also unusually candid in Section 6.2 about MCTS search depth and about what the greedy baseline does better.\n\nThe soft spots are real but manageable. First, the abstract's 'comparable' claim does not survive the large-hard configuration: CV 0.12 vs 0.22 and SR 0.30 vs 0.60 against their own Belief-Greedy, and the discussion effectively concedes this. That should be fixed at the language level. Second, the actor-policy model. The observer updates beliefs using Eq. 2 with an epsilon-greedy policy model, while the actor actually plans optimally under a private cost map. No sensitivity analysis over epsilon, no oracle-model comparison, no calibration check. That is an unexamined assumption, and it is the weakest link because the reported gains come from the belief update. I do not think it is fatal—the same model is used throughout, and the comparison against passive recognition is still informative—but it needs robustness evidence. Third, there are no error bars or significance tests despite stochastic instance generation, and no comparison against the actual prior AGR approaches (Amato and Baisero; Shvo and McIlraith). Code is not released. All three are fixable.\n\nOverall: a legitimate extension with correct formal core and a fair ablation, but with an overreaching abstract and one unexamined load-bearing assumption. For a reader working on goal recognition or active perception, this is worth reading. It deserves a serious referee, not a desk reject, provided the revision asks for the missing analysis and softened claims.","headline":"A clean joint-belief+MCTS formulation for active goal recognition that deserves review, but the abstract overstates MCTS performance and the actor-model mismatch needs testing.","tokens_in":14442,"tokens_out":2627,"would_cite":true,"duration_ms":31478,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Planning where to look reveals hidden goals faster than passive observation.","keywords":["Active Goal Recognition","POMDP","Joint Belief Update","Monte Carlo Tree Search","Goal Recognition as Planning","Partial Observability","Multi-agent Systems","Active Information Gathering"],"falsifier":"Run the same grid-world configurations with the actor using exactly the observer's epsilon-greedy policy and compare convergence with the private-cost-map condition: the paper's mechanism predicts larger gains over passive recognition when the belief model matches the actor, while vanishing or negative gains under severe mismatch would show the gains depend on that assumption.","tokens_in":13408,"feed_emoji":"🎯","tokens_out":8130,"duration_ms":93684,"temperature":0.7,"pith_summary":"This paper asks how an observer should move, not just watch, in order to identify another agent's hidden goal. The authors propose a probabilistic active goal recognition framework in which the observer maintains a joint belief over the actor's state and candidate goal and uses a Monte Carlo Tree Search to choose actions that sharpen that belief. In grid-world experiments, the joint belief update outperforms passive goal recognition on convergence, success rate, and final goal probability, and it does so in part by treating 'actor not seen' as evidence against goals that would have been visible. The paper also reports that the domain-independent MCTS planner performs comparably to a strong domain-specific greedy baseline. If correct, this provides a knowledge-light method for active goal inference in partially observable multi-agent settings.","feed_headline":"Choosing where to look beats passive watching at goal recognition","feed_subtitle":"Joint belief updates turn 'actor not seen' into evidence, lifting convergence and success rates in grid-world tests.","key_machinery":"The load-bearing object is the joint belief matrix $j_t$ over actor states and candidate goals, together with its recursive update. In the prediction step the previous belief is pushed through a model of the actor's goal-conditioned behavior, $P(s_t\\mid s_{t-1},g)$; in the update step the observation likelihood $P(o_t\\mid s_t,u_t)$, which depends on the observer's own state, is applied by Bayes' rule. Because the observation likelihood depends on where the observer is looking, the same 'no detection' observation carries different evidentiary weight for different goals, converting absence of evidence into negative evidence. This belief is used twice: it is the input to the MCTS search that selects observer actions, and the squared marginal goal belief $\\sum_g b_t(g)^2$ is the reward that guides the search. The MCTS algorithm, adapted from PFT-DPW without double progressive widening, alternates decision and chance nodes, samples future observations from the current belief, and backs up averaged values to approximate the Bellman optimality equations online.","core_discovery":"The paper's central claim is that active goal recognition can be solved by coupling a Bayesian belief update with sample-based planning, and that this coupling beats passive recognition. Formally, the observer tracks $j_t(s_t,g)=P(s_t,g\\mid o_{0:t},u_{0:t})$, a joint distribution over the actor's state $s_t$ and goal $g$ conditioned on the observation history and the observer's own trajectory, updated recursively by $j_t=h(j_{t-1},u_t,o_t)$ using Bayes' rule. Marginalizing gives $b_t(g)=\\sum_{s_t}j_t(s_t,g)$, and the planner maximizes expected squared marginal belief $\\sum_g b_t(g)^2$, which rewards actions that concentrate probability on one goal. The empirical claim is that this joint update outperforms a passive cost-based recognizer in every tested configuration, and that the MCTS variant matches a domain-specific greedy baseline without using any domain-specific heuristics. The decisive mechanism is that the update exploits negative observations: when the observer looks in a direction and does not see the actor, goals that would have been visible in that direction lose probability.","pith_inferences":["If the same update rule were applied to active sensor placement beyond goal recognition, any task where hidden state is inferred from a movable observer would inherit the ability to use 'not detected' as a signal, suggesting a general principle for cooperative monitoring.","Because the squared-belief reward is exactly the expected confidence under a random goal draw, one testable extension is to reward convergence directly, which may overcome the sparse-reward problem at shallow search depth.","A sharper falsification target than averaged success is calibration: one could check whether $b_t(g^*)$ at each time step matches empirical frequency under the actor's true private-cost policy, because systematic overconfidence would indicate the epsilon-greedy model is the source of the gains.","Extending the observation model to noisy or partially occluded sensors would test whether the negative-evidence mechanism survives degraded 'not seen' signals, since real sensors produce false negatives."],"forward_implications":["Under the same observation sequence, the joint belief update assigns probability mass in favor of goals consistent with periods of non-detection, so an observer that moves to create informative non-detections can converge sooner.","Because the belief update and planner are domain-independent, the same code can be applied to new grid-like environments without redesigning the reward function.","The MCTS planner's shallow lookahead, about four steps in large grids, is the main bottleneck in hard scenarios, so deeper search methods are a direct route to closing the gap with the greedy baseline.","The reported gains quantify the value of active sensing: the observer's trajectory is part of the inference, not just the observations collected along the way.","Relaxing the keyhole assumption, so that the actor reacts to being observed, would extend the formulation to collaborative and adversarial settings."],"supporting_citations":[{"why":"introduces the POMDP formulation of Active Goal Recognition that this paper adapts and generalizes to a joint observation function.","marker":"Amato and Baisero (2019)"},{"why":"supplies the cost-based goal recognition method and single-observation inference approach used as the passive baseline.","marker":"Masters and Sardina (2019)"},{"why":"provides the probabilistic plan recognition model with cost differences used in the passive baseline's likelihood formula.","marker":"Ramírez and Geffner (2010)"},{"why":"formalizes active goal recognition with an observer that acts, the problem definition this work reworks probabilistically.","marker":"Shvo and McIlraith (2020)"},{"why":"contributes the PFT-DPW online POMDP solver structure on which the paper's MCTS adaptation is based.","marker":"Sunberg and Kochenderfer (2018)"},{"why":"introduces POMCP, the sample-based MCTS planning paradigm that grounds the online planning component.","marker":"Silver and Veness (2010)"},{"why":"defines the convergence measure used to evaluate timeliness and confidence of goal inference.","marker":"Vered et al. (2018)"},{"why":"identifies the keyhole observation assumption that the paper adopts and later identifies as a limitation.","marker":"Masters and Vered (2021)"}],"fun_headline_variants":["Bayesian active sensing beats passive goal recognition","Plan where to look: active goal inference wins","MCTS and belief updates outperform passive goal reading","Active search: better goal recognition than passive watch","Bayesian beliefs + MCTS plan where to look"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The observer models the actor with a fixed epsilon-greedy policy, but the actor actually plans optimally under a private cost map; if that modeled policy is a poor match to real behavior, the joint belief can become miscalibrated and the measured advantage over passive recognition may not hold.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian active sensing beats passive goal recognition","Plan where to look: active goal inference wins","MCTS and belief updates outperform passive goal reading","Active search: better goal recognition than passive watch","Bayesian beliefs + MCTS plan where to look"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000644,"raw_usage":{"total_tokens":2942,"prompt_tokens":911,"completion_tokens":2031,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":1960}},"tokens_in":527,"tokens_out":2031,"duration_ms":17152,"temperature":1.0,"reasoning_tokens":1960,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:17:52.351532+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same grid-world configurations with the actor using exactly the observer's epsilon-greedy policy and compare convergence with the private-cost-map condition: the paper's mechanism predicts larger gains over passive recognition when the belief model matches the actor, while vanishing or negative gains under severe mismatch would show the gains depend on that assumption.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the cost-based goal recognition method and single-observation inference approach used as the passive baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"formalizes active goal recognition with an observer that acts, the problem definition this work reworks probabilistically."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"contributes the PFT-DPW online POMDP solver structure on which the paper's MCTS adaptation is based."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"introduces POMCP, the sample-based MCTS planning paradigm that grounds the online planning component."},{"cited_title":"F.; Kaminka, G.; and Meneguzzi, F","cited_arxiv_id":null,"evidence_quote":"defines the convergence measure used to evaluate timeliness and confidence of goal inference."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"identifies the keyhole observation assumption that the paper adopts and later identifies as a limitation."}],"review_version":1}