{"id":"b3c792a4-b0d7-4d13-998a-3e9b457078ab","arxiv_id":"1908.02948","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A group activity recognition framework uses two reinforcement-learning agents to distill informative frames and gate relation edges in a semantic graph, outperforming prior methods on Volleyball and Collective Activity datasets.","lead":"This paper presents a method for recognizing group activities in video that uses two reinforcement-learning agents to progressively focus on the most informative frames and person-to-person relations. The approach achieves competitive accuracy on two standard benchmarks, on par with or better than prior methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sign-based rewards in Eqs. 7-8 may yield a noisy or oscillating learning signal, so the reported 3.1% PRL gain over SRG may not replicate without seed-level validation.","rationale":"The reader identified the sign-based reward instability as the weakest assumption, and my reading agrees. The paper's own strongest claim is about the magnitude and super-additivity of the PRL improvement over SRG. That improvement is produced by a complex, sparsely documented RL training loop. The rewards in Eqs. 7 and 8 are sign functions of per-step changes; for the RG agent, each step changes one edge of a 15-edge graph, so the probability delta is likely tiny and noisy. Without learning curves, reward traces, or seed-level repeats, the ablation increments in Table 1 cannot be confidently attributed to a stable learned gating policy rather than to variance or reward hacking. The concern is not that the method is impossible, but that the reported numerical gains lack the evidence needed to establish replicability. The recommended fix is a direct sensitivity check: run with multiple seeds and with a smooth reward to see whether the sign function is essential. If the gains vanish, the central claim is weakened; if they persist, the concern is resolved. Therefore the appropriate verdict remains CONDITIONAL, i.e., unchanged from the reader's recommendation.","tokens_in":12754,"tokens_out":3207,"duration_ms":38782,"concrete_test":"Re-run the Volleyball ablations for Ours-SRG, Ours-SRG+RG, and Ours-PRL with five random seeds while logging per-step r_sparse, r_ascend, gate norms, and p_c. Then replace the sign functions in Eqs. 7 and 8 with the raw differences (or a scaled tanh) and repeat. If the smoothed-reward variant no longer reproduces the 1.5% and 3.1% gains over SRG, or if the sign-reward runs show that r_ascend is positive less than a majority of steps, the sign-based reward is the actual source of the claimed improvement and the contribution is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that progressive RL agents (FD and RG) materially improve group activity recognition over the SRG baseline, with PRL improving MCA by 3.1% on Volleyball and MPCA by 1.3% on CAD. The mechanism for that improvement is the alternate training loop in Sec. 3.5, whose learning signal is defined by Eqs. 7-10. The load-bearing assumption is that r_sparse = -sgn(L2,1(G_tau)-L2,1(G_tau-1)) and r_ascend = sgn(p_c_tau - p_c_tau-1) provide stable, informative rewards across the nine training stages. This is not secure. In the RG agent, one gate is emitted per edge at a time, so after a single edge update the change in groundtruth probability p_c is typically tiny and dominated by discretization and stochasticity. A sign function then converts this small delta into a ±1 reward that can be essentially random. Similarly, r_sparse rewards any decrease in the L2,1 norm, even if the pruned edges are group-relevant, and it can oscillate if the norm is near a local minimum. Because the SRG, FD, and RG components are alternately frozen and retrained, a reward that overfits to the currently frozen SRG may not transfer when that component is later trained. The paper reports no learning curves, reward traces, gate-value statistics, or multiple seeds, so the 3.1% and 1.3% gains could be the result of reward hacking or a single favorable run rather than a stable learned policy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Progressive Relation Learning (PRL) for group activity recognition. It constructs a Semantic Relation Graph (SRG) over person nodes and relation edges, then introduces two reinforcement-learning agents: a feature-distilling (FD) agent that selects informative frames from the low-level spatiotemporal features, and a relation-gating (RG) agent that multiplies each relation edge by a continuous gate. The graph and agents are trained alternately over nine stages. Experiments on the Volleyball and Collective Activity datasets report MCA/MPCA improvements over the authors' own SRG baseline and results on par with or better than prior methods. The main empirical claims are that PRL improves MCA by 3.1% over SRG on Volleyball and MPCA by 1.3% over SPA+KD on CAD.","tokens_in":13158,"tokens_out":4378,"duration_ms":42900,"significance":"If the empirical results are stable, the paper makes a useful contribution: it demonstrates that explicit relation-level gating learned by reinforcement learning can improve group activity recognition over a graph-only baseline, and the ablation decomposition (SRG vs. SRG+FD vs. SRG+RG vs. PRL) is informative. The paper also provides qualitative visualizations of the learned relation graphs. However, the central empirical claims rest on single-run results and on a sign-based reward whose stability is not demonstrated; these issues must be addressed before the contribution can be considered established.","major_comments":[{"comment":"The rewards r_sparse and r_ascend are sign functions of single-step differences in the L2,1 norm and in the groundtruth-label probability p_c. Because the RG agent emits one gate at a time (Sec. 3.3, 'Relation-gating Agent'), the change in p_c after a single edge update is typically small relative to classifier stochasticity, and the sign transform can turn this small delta into an essentially random ±1 reward. This is load-bearing because the reported 3.1% MCA gain of PRL over SRG (Table 1) is attributed to the progressive training loop of Sec. 3.5. The manuscript reports no learning curves, reward traces, gate-value statistics, or sensitivity analysis for the reward weights, so the stability and informativeness of this reward are not established. I request such diagnostics, and also a concrete test of the sign nonlinearity, e.g., ablating it to a saturating linear reward or reporting the variance of r_ascend across steps for the trained agents.","section":"§3.3, Eqs. (7)-(10)"},{"comment":"The headline improvements (PRL over SRG by 3.1% MCA on Volleyball; PRL over SPA+KD by 1.3% MPCA on CAD) are presented without error bars, multiple seeds, or a description of how the reported checkpoint was selected. Given the stochasticity of A3C training (16 asynchronous workers, sampled actions, Eq. 12), the differences between variants, which are often around 1%, may be within run-to-run variance. The central empirical claim therefore needs support from means and standard deviations over at least three seeds, or an explicit statement of the evaluation protocol and number of runs.","section":"§4.4, Tables 1-2"},{"comment":"In the nine-stage alternate training, the reward for each agent is generated by the currently frozen SRG and FD features, but the agent's policy is later used with a retrained SRG. The paper does not report the evolution of training rewards or validation accuracy across the nine stages, nor does it show that the alternating procedure converges rather than oscillating or overfitting to the frozen component at each stage. Please report per-stage validation accuracy (at least after stages 3, 6, and 9) and the reward curves for both agents, and state the stopping criterion used for the 2-hour agent training phases.","section":"§3.5, Alternate Training"}],"minor_comments":[{"comment":"The text says 'we normalize the values of gates before gating operation, i.e., gij = gji = (gij+gji)/2'; this is symmetrization, not normalization, and the terminology should be corrected.","section":"§3.3"},{"comment":"There are several typos, e.g., 'strucural RNNs' and 'the course individual level' should read 'structural RNNs' and 'the coarse individual level'.","section":"§1"},{"comment":"The implementation details state hyperparameter values (tau_max = 5, Omega = 15/20, Td = 5) but do not report any sensitivity analysis; a short discussion or reference to supplementary experiments would help the reader judge robustness.","section":"§4.2"},{"comment":"The sentence 'the PRL eventually improves 3.1% over the original SRG, which is even larger than the sum of increments from the two agents, 2.7%' is presented as evidence of mutual boosting; however, incremental gains in an ablation are not additive by construction, so this argument should be softened or supported by explicitly comparing the gain of PRL over SRG+FD and SRG+RG.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's approach is reasonable and the ablation study is helpful, but the empirical evidence is not yet at the level required for the claims made. The sign-based reward stability and the absence of multi-seed results are the main barriers. The novelty relative to prior progressive RL work (e.g., [30]) should also be clarified, and the authors should be explicit about what is new beyond combining A3C with graph gating."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper proposes PRL: a semantic relation graph (SRG) on top of per-person CNN-LSTM features, refined by two A3C agents—a feature-distilling (FD) agent that selects informative frames and a relation-gating (RG) agent that adjusts edge weights. The two agents are trained alternately with the SRG over nine stages. Reported results beat their own graph-only baseline by 3.1% MCA on Volleyball and 1.3% MPCA on CAD, and are competitive with state-of-the-art without optical flow.\n\nWhat is new and good: the specific combination of two RL agents in discrete and continuous action spaces, plus the alternate training loop, is a genuine departure from standard attention/graph methods in this subfield. The L2,1 sparsity reward for the gate matrix is a nice idea—it encourages structured row sparsity so the agent focuses on key participants rather than uniformly pruning edges. The ablation study is thorough: the RL agents beat temporal and relation attention variants by ~1% each. The graph visualizations are convincing and align with intuition. Writing is clear and the literature coverage is fair.\n\nSoft spots: the reward design is the main worry. Equations 7 and 8 take the sign of a difference in the L2,1 norm and in groundtruth probability. As the stress-test note says, per-edge updates make the probability delta tiny, so the sign reward can be noisy or oscillatory. The paper reports no learning curves, reward traces, gate statistics, or multiple seeds. In a subfield where 1–3% is the typical margin, the absence of error bars is a real weakness. The 3.1% gain over SRG, claimed to exceed the sum of individual agent gains, is exactly the kind of claim that needs seed-level variance to be credible. It is not fatal—the method is plausible and mutual boosting could occur—but the central quantitative claim rests on single runs. No code is released either.\n\nCitations and math are fine: the method is empirically evaluated on external benchmarks, hyperparameters are reported, and there is no circularity. Hyperparameter sensitivity is not studied, but that is a minor omission.\n\nWho benefits: anyone in group activity recognition or using RL for structured prediction. The paper deserves a serious referee; I would send it out. Before acceptance, ask for multiple seeds, significance tests, and ideally learning curves or reward traces.\n\nRecommendation: engage with the idea, but don't take the numbers at face value.","headline":"A plausible RL-based progressive relation learning method for group activity recognition with modest gains; deserves a serious referee but needs seed-level validation before the numbers are trusted.","tokens_in":13603,"tokens_out":2419,"would_cite":true,"duration_ms":25493,"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":"Two reinforcement-learning agents progressively select the defining frames and relations, lifting group activity recognition accuracy above its graph-only baseline.","keywords":["group activity recognition","semantic relation graph","deep reinforcement learning","feature distillation","relation gating","A3C","video understanding","graph neural networks"],"falsifier":"Train the same nine-stage procedure with the RG agent's gates replaced by uniform random gates in [0,1] (or with the sign rewards replaced by their raw magnitudes) and compare Volleyball MCA; if accuracy does not drop from 91.4%, the progressive relation-gating mechanism is not load-bearing.","tokens_in":12579,"feed_emoji":"🎯","tokens_out":6973,"duration_ms":66572,"temperature":0.7,"pith_summary":"Group activity recognition often fails because a scene contains many people doing unrelated things, and only a few interactions in a few frames define the activity. This paper tries to establish that a graph that explicitly stores person-to-person relations, refined by two reinforcement-learning agents, can progressively focus on the group-relevant relations and frames. On the Volleyball benchmark the full method raises multi-class accuracy by 3.1 percentage points over its own graph-only baseline, and it matches or exceeds prior methods that use no optical flow on both Volleyball and Collective Activity datasets. The authors argue that alternate training lets the graph, the frame-distilling agent, and the relation-gating agent boost each other.","feed_headline":"RL agents prune frames and relations to spot the group action","feed_subtitle":"On Volleyball, the full method beats its own graph baseline by 3.1 accuracy points, without optical flow.","key_machinery":"The central object is the semantic relation graph (SRG), a graph network in which nodes are people, edges are relation vectors, and a global attribute carries the activity score; messages pass along edges for several iterations. Around it sit two reinforcement-learning agents trained with the A3C algorithm. The feature-distilling (FD) agent works in a discrete action space, keeping or replacing candidate frames. The relation-gating (RG) agent works in a continuous action space, producing a gate in [0,1] per edge, and its rewards combine the sign of the change in the L2,1 norm of the gate matrix with the sign of the change in the predicted probability of the ground-truth label. Alternate nine-stage training, freezing two components while updating the third, is what lets the graph and the two agents improve each other.","core_discovery":"On its own terms, the paper claims that group activity recognition improves when the model explicitly represents each pairwise relation as a learnable vector in a semantic relation graph and then progressively refines both the input features and the graph. A feature-distilling agent selects which frames of the per-person spatiotemporal features to keep, while a relation-gating agent multiplies each edge by a continuous gate, with rewards that encourage structured row sparsity and an ascending probability for the true activity label. The three components are trained alternately in nine stages. The resulting PRL framework reaches 91.4% MCA on Volleyball, 1.3% above the best prior no-optical-flow method on Collective Activity, and its gain over the graph-only baseline (3.1% MCA) is larger than the sum of the gains from the two agents alone (2.7%), which the paper reads as evidence that the two agents reinforce each other.","pith_inferences":["A consequence the authors do not draw: the gate matrix itself is a per-relation importance map, so the framework can be read as an interpretability tool for why a group activity was predicted.","A testable extension: feed the distilled frames and gated relations into an online recognition loop, using the FD queue as a temporal buffer, to see if progressive refinement helps before the full clip is seen.","If the sign rewards are the fragile part, a natural variant replaces them with scaled probability-difference rewards; comparing training stability would separate the contribution of progressive refinement from the reward shaping."],"forward_implications":["Explicitly representing relations as edge vectors in a graph network outperforms passing messages implicitly: the SRG alone beats the message-passing baseline on Volleyball (88.3 vs 87.9 MCA).","The RL-based agents outperform their attention-based counterparts: SRG+FD beats SRG+temporal attention by 0.9 MCA and SRG+RG beats SRG+relation attention by 1.1 MCA on Volleyball.","Progressive refinement of frames and relations is complementary: the full PRL improves over the SRG by 3.1 MCA, more than the sum of the two individual gains, implying mutual reinforcement through alternate training.","Without any optical flow input, the method is on par with or ahead of several prior systems that use optical flow, and the learned gate values concentrate on participants whose actions define the activity, as shown in the visualized graphs."],"supporting_citations":[{"why":"Supplies the message-passing baseline and the shared VGG16 feature pipeline that the SRG is compared against.","marker":"[22]"},{"why":"Provides the graph-network formulation of node, edge, and global-attribute updates that the semantic relation graph extends.","marker":"[24]"},{"why":"Provides the Volleyball dataset and the MCA/MPCA evaluation protocol used for the main benchmark.","marker":"[14]"},{"why":"Provides the Collective Activity dataset and the action-class merging convention used for the second benchmark.","marker":"[8]"},{"why":"Supplies the Asynchronous Advantage Actor-Critic algorithm used to optimize both the FD and RG agents.","marker":"[20]"},{"why":"Supplies the Actor-Critic architecture on which both agents' policy and value networks are built.","marker":"[16]"},{"why":"Supplies the person CNN+LSTM feature extraction and the key-frame/key-participant idea that the FD agent refines.","marker":"[36]"},{"why":"Supplies SPA+KD, the strongest no-optical-flow prior method that the paper aims to surpass on Collective Activity.","marker":"[31]"},{"why":"Supplies the shifted-class reward idea used in the RG and FD reward functions.","marker":"[30]"}],"fun_headline_variants":["Two RL agents refine graph and frames for group activity","Progressive pruning with RL boosts group activity recognition","Feature-distilling and relation-gating agents sharpen group actions","RL-tuned semantic graph lifts group activity accuracy","Frame and relation pruning via RL improves group recognition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that rewarding small improvements in the model's confidence and in the sparsity of the relation gates is a stable learning signal across all nine training stages, so the two agents genuinely improve rather than chase random fluctuations.","fun_headline_variants_meta":{"raw":{"variants":["Two RL agents refine graph and frames for group activity","Progressive pruning with RL boosts group activity recognition","Feature-distilling and relation-gating agents sharpen group actions","RL-tuned semantic graph lifts group activity accuracy","Frame and relation pruning via RL improves group recognition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1285,"prompt_tokens":914,"completion_tokens":371,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":297}},"tokens_in":530,"tokens_out":371,"duration_ms":4914,"temperature":1.0,"reasoning_tokens":297,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:28:04.909321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same nine-stage procedure with the RG agent's gates replaced by uniform random gates in [0,1] (or with the sign rewards replaced by their raw magnitudes) and compare Volleyball MCA; if accuracy does not drop from 91.4%, the progressive relation-gating mechanism is not load-bearing.","supporting_citations":[{"cited_title":"stagnet: An attentive semantic RNN for group activity recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the message-passing baseline and the shared VGG16 feature pipeline that the SRG is compared against."},{"cited_title":"Riedmiller, Raia Hadsell, and Peter Battaglia","cited_arxiv_id":null,"evidence_quote":"Provides the graph-network formulation of node, edge, and global-attribute updates that the semantic relation graph extends."},{"cited_title":"Ibrahim, Srikanth Muralidharan, Zhiwei Deng, Arash Vahdat, and Greg Mori","cited_arxiv_id":null,"evidence_quote":"Provides the Volleyball dataset and the MCA/MPCA evaluation protocol used for the main benchmark."},{"cited_title":"What are they doing?: Collective activity classification using spatio-temporal relationship among people","cited_arxiv_id":null,"evidence_quote":"Provides the Collective Activity dataset and the action-class merging convention used for the second benchmark."},{"cited_title":"Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu","cited_arxiv_id":null,"evidence_quote":"Supplies the Asynchronous Advantage Actor-Critic algorithm used to optimize both the FD and RG agents."},{"cited_title":"Konda and John N","cited_arxiv_id":null,"evidence_quote":"Supplies the Actor-Critic architecture on which both agents' policy and value networks are built."},{"cited_title":"Participation-contributed temporal dynamic model for group activity recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the person CNN+LSTM feature extraction and the key-frame/key-participant idea that the FD agent refines."},{"cited_title":"Mining semantics-preserving attention for group activity recognition","cited_arxiv_id":null,"evidence_quote":"Supplies SPA+KD, the strongest no-optical-flow prior method that the paper aims to surpass on Collective Activity."},{"cited_title":"Deep progressive reinforcement learning for skeleton-based action recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the shifted-class reward idea used in the RG and FD reward functions."}],"review_version":1}