{"id":"d735084f-efc8-4d45-ac4d-c944a76d31d9","arxiv_id":"2505.03941","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"GRAML uses a Siamese LSTM to learn a distance metric over observation traces, enabling one-shot adaptation to newly emerged goals in goal recognition.","lead":"This paper presents GRAML, a metric-learning approach to goal recognition that embeds observation traces so that traces leading to the same goal are close and traces to different goals are distant. The method is intended to adapt to new goals with a single example trace, and it shows competitive accuracy across six simulated environments.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The most load-bearing issue is that one-shot transfer to unseen goals is only tested for active goals whose distance to the base-goal set is unreported, and Section 6 concedes base-goal choice may be crucial, so the scope of the central generalization claim is not yet pinned down.","rationale":"The reader's weakest assumption and my concern coincide: the load-bearing premise is that a metric learned on base goals transfers to arbitrary new goals, and the paper itself acknowledges that base-goal coverage may be crucial. The held-out goal evaluations are relevant but incomplete, because they do not control or report the distance between base goals and active goals. The missing GRNet comparison and the free pass given to expert sequences in runtime comparisons are real secondary issues, but they mostly affect efficiency claims and comparison coverage rather than the core one-shot transfer claim. Given the diverse environments, the self-supervised training procedure, and the base-goal count experiments in Figure 6, there is partial empirical support, so rejection would be too strong. However, the central generalization claim is not yet bounded, so moving beyond conditional acceptance would also be premature. I therefore keep the reader's CONDITIONAL verdict.","tokens_in":12809,"tokens_out":4622,"duration_ms":48469,"concrete_test":"Re-run the Parking and PointMaze evaluations with a controlled split: train base goals only in one region (e.g., one half or one corner of the goal space), then sample active goals separately in the near region and in a far region, and report accuracy as a function of the minimum distance from each active goal to the base-goal set. Also report explicitly whether active goals are disjoint from base goals in the current Appendix results. If far-region accuracy approaches chance while near-region accuracy remains high, the one-shot generalization claim is bounded by base-goal coverage; if far-region accuracy stays high, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the learned embedding metric transfers to goals not seen during domain learning, so one reference trace per new goal suffices at inference, as formalized by Eq. (3) with s=1. For this to hold, the LSTM must place traces to an arbitrary new goal closer to same-goal reference traces than to other goals' reference traces. But the training objective in Eqs. (1)-(2) only constrains distances among the chosen base goals in \\bar{G}; nothing in the loss controls the embedding of traces whose goals lie outside \\bar{G}. The held-out evaluation in Section 4.3 is the right kind of evidence, but it does not report the spatial or behavioral relationship between \\bar{G} and the active goal set G0. If the active goals are near to, or overlap with, the base-goal distribution, then the measured one-shot accuracy is inflated relative to genuinely distant goals. Section 6 explicitly concedes that base-goal selection 'may be crucial for the model's success' and that uncovered regions of the state space may not be recognized correctly. Figure 6 only varies the number of base goals, not their coverage region or their distance to the tested goals. The paper therefore demonstrates transfer for the specific base-goal choices used, but it does not establish the boundary of the claimed generalization to newly introduced goals.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces GRAML, a goal-recognition method that casts recognition as metric learning. A Siamese LSTM is trained so that observation traces leading to the same goal are embedded closer to each other than traces leading to different goals; at inference, a new observation is classified by nearest-neighbor comparison against one or more reference traces per candidate goal. Two variants are proposed: BG-GRAML, which trains separate agents for a set of base goals, and GC-GRAML, which trains a single goal-conditioned policy. The method is evaluated on six discrete and continuous environments, under partial consecutive and non-consecutive observations, stochastic suboptimal traces, and held-out goals not used during training. The paper reports accuracy comparable to or better than the GR-as-RL baselines GRAQL and DRACO, with reduced goal-adaptation time, and analyzes the effect of the number of base and active goals.","tokens_in":13093,"tokens_out":3560,"duration_ms":37630,"significance":"If the central claim holds—that a metric learned on a set of base goals transfers to newly introduced goals from a single reference trace—GRAML would be a useful contribution to data-driven goal recognition, particularly for online dynamic goal recognition in continuous domains. The paper has notable strengths: it evaluates on six environments spanning discrete and continuous state/action spaces; it tests partial, non-consecutive, and stochastic suboptimal observations; and it uses held-out goals in the evaluation, which is the right kind of evidence against circularity. The proposed ODGR phase decomposition is also clearly articulated. However, the generalization claim is not yet fully delimited: the relationship between the base-goal set and the held-out active goals is not measured, and the paper itself concedes that base-goal selection may be crucial. The evaluation also omits GRNet, the closest model-free dynamic-goal baseline, and the runtime comparison assigns zero cost to expert-provided reference sequences. These issues are addressable but currently leave the scope of the central claim and the speed advantage not fully established.","major_comments":[{"comment":"The central claim that the learned metric transfers to newly introduced goals is not yet delimited. The training loss in Eqs. (1)-(2) only constrains distances among traces to the chosen base goals, and Eq. (3) assumes that reference traces for previously unseen goals are embedded near same-goal traces and far from other-goal traces. The held-out evaluation in Section 4.3 never reports the relationship between the base-goal set and the active goal set G0, such as distance in state space, overlap with the base-goal distribution, or difficulty of the goals. Section 6 explicitly concedes that base-goal choice 'may be crucial for the model's success' and that uncovered areas of the state space may not be recognized correctly. The paper therefore demonstrates one-shot transfer for its specific base-goal choices, but it does not establish the boundary of the claimed generalization. Please report the coverage of the base goals relative to G0 (e.g., distance distributions, or probes into regions deliberately far from all base goals), or restrict the claim to goal sets sampled near the base-goal distribution.","section":"Section 3.1.1, Section 4.3, Section 6"},{"comment":"The closest dynamic-goal baseline, GRNet, is discussed in Related Work but is not included in the accuracy or runtime comparisons. GRNet is described as generalizing to any set of goals without additional learning, which is exactly the capability GRAML claims to improve. Without a comparison to GRNet, the phrases 'state-of-the-art GR' and the speed/runtime improvements are not demonstrated against the most relevant method. Please add GRNet to at least the discrete benchmarks, or provide a concrete incompatibility justification with evidence rather than a qualitative remark about fluent enumeration.","section":"Section 5, Section 4.3"},{"comment":"The runtime comparison credits BG-GRAML with G = 0 when expert sequences are provided, while the same paragraph reports 75 seconds per goal for MCTS planning and 18 minutes for a five-goal problem. The phase accounting omits the cost of producing the reference library and the domain-learning cost of training base-goal agents, which can dominate in BG-GRAML. Please report a complete wall-clock breakdown for all phases—domain learning, library generation, goal adaptation, and inference—on the same hardware, and state whether expert/oracle time is included or excluded. Without this, the speed advantage is partly an artifact of zero-cost assumptions.","section":"Section 4.3 (runtime paragraph)"}],"minor_comments":[{"comment":"The comparison between GC-GRAML and BG-GRAML is confounded because the two variants use different numbers of base goals (|G| = 20 for GC-GRAML and |G| = 5 for BG-GRAML). The conclusion that GC-GRAML benefits from the variety of sequences is therefore not isolated; please report both variants with matched |G| or explicitly analyze the effect.","section":"Section 4.2, Figure 5"},{"comment":"The accuracy table reports means and standard deviations but no significance tests or confidence intervals. Several claimed improvements appear to be within one standard deviation; please add statistical testing or clearly state which differences are reliable.","section":"Table 1"},{"comment":"The notation 'G = 0' in the runtime paragraph is confusing because G is also used for the goal set and for the goals-adaptation interval; please use a distinct symbol such as T_G or t_adapt.","section":"Section 4.3 and Figure 1"},{"comment":"There is a typo: 'a single GCRL agent is trained on ¯G, which can also be defined as a continuum or discrete set. The sample generation is also done similarly, but in this case, a single goal-conditioned policy generates the traces.' The phrase 'buts instead of multiple agents' appears garbled and should be corrected.","section":"Section 3.1.1"},{"comment":"The repository link is promised only post-acceptance, so the current manuscript does not provide the artifacts needed to reproduce the experiments. Please include an anonymized repository link or a detailed appendix with hyperparameters and environment configurations.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal and the core idea is promising, but the transfer claim needs to be empirically delimited before publication. The missing GRNet comparison is the most important gap from a related-literature perspective; the runtime accounting is the second. I do not see grounds for rejection, since the issues are addressable with additional experiments and analysis rather than a fundamental flaw. One small concern for the editor: the ODGR problem definition is self-cited from the authors' prior workshop paper; given the centrality of that definition, a more independent anchor or a fuller formal treatment would strengthen the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The GRAML paper is a genuine step forward for goal recognition under changing goal sets: it reformulates the problem as metric learning, shows one-shot adaptation on held-out goals, and tests on six environments covering discrete and continuous control. The core idea is simple and transferable: train a Siamese LSTM so traces to the same goal embed near each other, then at inference just embed one reference trace per new goal and do nearest-neighbor. That is actually new for this problem, and the held-out evaluation is the right kind of evidence, so the central claim is not circular.\n\nThe paper also does a few things well. It introduces two variants, BG and GC, the goal-conditioned one is a natural fit, and the benchmark suite (Minigrid, PointMaze, Parking, Panda) gives the field a common testbed that was missing for ODGR. The ablation on number of base goals is useful, and the explicit discussion of observability and suboptimal traces is honest. Section 6 is unusually candid about the base-goal coverage limitation.\n\nThe soft spots are real but not fatal. The abstract's 'runtime improvements' line is too broad: BG-GRAML's zero adaptation time comes from expert-provided sequences, and the paper later admits that planning for the library takes about four times less than retraining. That is still an improvement, but the framing in the abstract overstates it. A bigger issue is the missing comparison to GRNet, the closest dynamic-goal baseline. The paper mentions GRNet's fluent enumeration weakness, but never runs it, so we don't know if the added complexity buys accuracy. The stress-test concern about base-goal coverage is legitimate: the loss only constrains distances among base goals, and Section 6 concedes that uncovered state-space regions may not be recognized. The held-out evaluation should report how far the active goals lie from the base-goal distribution. That is the main missing number.\n\nOverall, the paper shows clear thinking and honest engagement with its own limitations. I would send it to peer review, but the authors should be asked to (1) report base-goal distances/coverage, (2) add a GRNet comparison where feasible, and (3) tone down the abstract's runtime claim. For a reader working on plan or goal recognition, this is worth reading now; for others, it is a solid application of metric learning with a clean ODGR formulation.","headline":"A real, if narrowly bounded, contribution to online dynamic goal recognition: the metric-learning idea works and is evaluated on held-out goals, but the abstract overstates runtime and the base-goal coverage boundary needs to be pinned down.","tokens_in":13595,"tokens_out":1845,"would_cite":true,"duration_ms":18447,"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":"The paper claims that goal recognition can be cast as metric learning, so a newly introduced goal is recognized from a single example trace in a learned embedding space.","keywords":["goal recognition","metric learning","online dynamic goal recognition","one-shot adaptation","Siamese network","LSTM embeddings","goal-conditioned reinforcement learning"],"falsifier":"Train GRAML on base goals restricted to one corner of a gridworld, then introduce a goal in the opposite corner and measure one-shot accuracy on partial traces; if accuracy stays at chance while base-goal traces are still recognized, the metric-transfer claim is falsified.","tokens_in":12625,"feed_emoji":"🎯","tokens_out":7803,"duration_ms":70875,"temperature":0.7,"pith_summary":"The paper aims to turn goal recognition—inferring what an agent is trying to achieve from the actions it has taken—into a metric-learning problem. Its claim is that a recurrent network trained on traces that lead to a fixed set of base goals learns a distance in embedding space in which traces to the same goal are close and traces to different goals are far apart. Once that metric exists, a newly introduced goal is handled by embedding a single example trace for that goal and comparing new observations against it, so no retraining per goal is needed. The authors evaluate this on discrete and continuous environments with partial, noisy, and suboptimal observations, and report one-shot adaptation speeds that avoid the policy-retraining cost of previous approaches. If the claim holds, dynamic environments whose goal sets change over time gain a recognition method whose adaptation phase is a cheap library update rather than a full learning problem.","feed_headline":"A single example trace is enough to recognize a new goal","feed_subtitle":"GRAML learns a distance between observation traces, so unseen goals are recognized by nearest neighbor in embedding space.","key_machinery":"The load-bearing object is the embedding space $\\mathbb{R}^k$ produced by the Siamese LSTM. The similarity function is $e^{-\\sum_{i} |v_1[i]-v_2[i]|}$ (Eq. 1), trained by binary cross-entropy on same-goal and different-goal trace pairs (Eq. 2). Inference is nearest-neighbor in that space with averaging over $s$ example embeddings per goal (Eq. 3). The mechanism's work is to convert a goal recognition problem into a nearest-neighbor lookup, so that adapting to new goals only requires embedding their example traces through the fixed LSTM.","core_discovery":"GRAML's central discovery is that the recognition step can be decoupled from goal identity: instead of learning what each goal looks like, learn a metric over the space of observation traces. Concretely, a Siamese network with two shared-weight LSTM branches is trained with binary cross-entropy to produce embeddings $v$ such that the exponential of negative Manhattan distance (Eq. 1) is near 1 for same-goal pairs and near 0 for different-goal pairs. At inference, each active goal is represented by a small library of embeddings of example traces, and a query trace is assigned to the goal whose average similarity is maximal (Eq. 3). The paper claims this metric transfers to goals unseen in training, which is what makes one-shot adaptation possible: one example trace per new goal is enough to locate that goal in the learned space.","pith_inferences":["Editorial inference: if the transfer claim is right, the same embedding metric could be reused across many different goal sets in the same domain, making GRAML a pretrained 'goal-space prior' that is updated by appending embeddings rather than by re-optimizing.","Editorial inference: the method implicitly assumes that distinct goals are geometrically separable in trajectory space; goals that share all near-optimal paths with another goal would remain indistinguishable unless the library contains disambiguating traces, a limitation the paper does not analyze.","Editorial inference: the embedding space itself could be used to measure similarity between goals, which would give an automatic way to choose base goals that cover the goal space—directly addressing the base-goal selection issue the paper leaves open.","Editorial inference: a straightforward test of the mechanism is to probe whether the learned metric is approximately monotone with true plan distance; if it is, the same model could also serve as a planner heuristic, not just a recognizer."],"forward_implications":["Adding a goal to the active set costs one example trace per new goal: no retraining of the recognition network and no policy learning for that goal.","Continuous domains need no discretization, because the LSTM maps continuous state-action pairs into a fixed embedding space rather than predicting in state space.","Recognition accuracy is expected to rise with the number of base goals and fall as the number of simultaneous active goals grows, since more goals create more confusable traces.","The same learned metric can serve repeated goal-set changes in an online setting: each new set of goals only triggers the adaptation phase, not a new domain learning phase.","Larger example libraries per goal improve accuracy at linearly increasing adaptation and runtime cost, giving a tunable accuracy-speed trade-off."],"supporting_citations":[{"why":"Supplies the Siamese recurrent architecture that embeds two sequences so that distance in embedding space expresses similarity.","marker":"[Mueller and Thyagarajan, 2016]"},{"why":"Introduces goal recognition as reinforcement learning, the per-goal policy-training baseline GRAML is designed to avoid.","marker":"[Amado et al., 2022]"},{"why":"Provides the actor-critic policy-learning baseline for continuous domains that GRAML compares against.","marker":"[Nageris et al., 2024]"},{"why":"Formulates Online Dynamic Goal Recognition, the changing-goal problem setting GRAML addresses.","marker":"[Shamir et al., 2024]"},{"why":"Supplies goal-conditioned reinforcement learning, which GC-GRAML uses to generate diverse traces from a single policy.","marker":"[Liu et al., 2022]"},{"why":"Describes the closest model-free deep approach that generalizes to new goals, framing GRAML's continuous-domain contribution.","marker":"[Chiari et al., 2023]"},{"why":"Provides the plan-recognition-as-planning formulation from which the paper's goal recognition definition descends.","marker":"[Ramirez and Geffner, 2009]"}],"fun_headline_variants":["One trace per goal: metric learning makes it work","Goal recognition as metric learning: one-shot adaptation","Recognize new goals from a single observation trace","GRAML: distance metric enables one-shot goal recognition","Learn a metric, then spot unseen goals with one trace"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned metric transfers from the base goals used in training to any new goal introduced later, so if the base goals do not cover the goal space, an unseen goal's traces can land anywhere in embedding space and recognition fails.","fun_headline_variants_meta":{"raw":{"variants":["One trace per goal: metric learning makes it work","Goal recognition as metric learning: one-shot adaptation","Recognize new goals from a single observation trace","GRAML: distance metric enables one-shot goal recognition","Learn a metric, then spot unseen goals with one trace"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1506,"prompt_tokens":874,"completion_tokens":632,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":557}},"tokens_in":490,"tokens_out":632,"duration_ms":6307,"temperature":1.0,"reasoning_tokens":557,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:41:13.919832+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train GRAML on base goals restricted to one corner of a gridworld, then introduce a goal in the opposite corner and measure one-shot accuracy on partial traces; if accuracy stays at chance while base-goal traces are still recognized, the metric-transfer claim is falsified.","supporting_citations":[{"cited_title":"Siamese recurrent architectures for learning sentence similarity","cited_arxiv_id":null,"evidence_quote":"Supplies the Siamese recurrent architecture that embeds two sequences so that distance in embedding space expresses similarity."},{"cited_title":"Goal recognition as reinforcement learning","cited_arxiv_id":null,"evidence_quote":"Introduces goal recognition as reinforcement learning, the per-goal policy-training baseline GRAML is designed to avoid."},{"cited_title":"Odgr: Online dynamic goal recognition","cited_arxiv_id":null,"evidence_quote":"Formulates Online Dynamic Goal Recognition, the changing-goal problem setting GRAML addresses."},{"cited_title":"Goal recognition as a deep learning task: The grnet approach","cited_arxiv_id":null,"evidence_quote":"Describes the closest model-free deep approach that generalizes to new goals, framing GRAML's continuous-domain contribution."},{"cited_title":"Ramirez and H","cited_arxiv_id":null,"evidence_quote":"Provides the plan-recognition-as-planning formulation from which the paper's goal recognition definition descends."}],"review_version":1}