{"id":"392888b1-d971-4a78-a24a-b9aa459b28a0","arxiv_id":"2412.02447","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A trajectory prediction model that decomposes forecasts into a linear base, a self-sourced vibration, and a social resonance vibration, achieving strong benchmark results with an interpretable decomposition.","lead":"This paper proposes a pedestrian trajectory prediction model that treats future paths as the sum of independent vibration components, separating an agent's own randomness from social influences. The model reports state-of-the-art results on several benchmarks and offers a way to visualize which prediction uncertainty comes from intention versus interaction.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed decoupling of self vs social randomness is not enforced by the architecture: the re-bias branch also uses ego-only features and z_r, so the 'social-sourced' label is unsupported.","rationale":"The reader's weakest assumption was Eq. (10), the element-wise product of spectral embeddings as the encoding of social influence. I see that as a consequence of a broader gap: the model's explanatory decomposition is not identifiable. Even if Eq. (10) were replaced by a more principled spectral-similarity kernel, the re-bias branch would still consume ego-only Δf^i_e and z_r, so nothing structurally prevents z_r from encoding intention randomness. This makes the 'decoupled randomness' claim the load-bearing one: if it fails, the paper is a competitive empirical trajectory predictor with an additive bias architecture, but not a method that separates causes of randomness. The paper deserves credit for extensive ablations, clear specification, and competitive numbers (Tables 2–4); the concern is not about benchmark results but about the interpretability conclusion. A concrete computational check on the released code can settle the question. If the check shows that ΔY_r responds to neighbors and z_r's effect disappears when neighbors are removed, the concern is largely mitigated; if not, the central explainability claim should be withdrawn or substantially weakened. This supports a CONDITIONAL verdict with explicit conditions, matching but sharpening the reader's assessment.","tokens_in":27919,"tokens_out":5457,"duration_ms":61032,"concrete_test":"Run a disentanglement check on the provided code (github.com/cocoon2wong/Re). For a fixed set of test scenes, (i) vary z_s with z_r fixed to its zero/mean value and record the distribution of ΔY^i_s; (ii) vary z_r with z_s fixed and record the distribution of ΔY^i_r. Then repeat (ii) with all neighbor trajectories removed from the input (so F^i_R is zero) while keeping the ego trajectory identical. If the variance of ΔY^i_r under z_r sampling remains substantial when neighbors are removed, the 'social-sourced' branch is not actually social. Additionally, compute the principal angle between the two sampled-bias subspaces; if it is not consistently near 90° on ETH-UCY and SDD, the claimed vertical independence does not hold. Report these numbers as a table.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is the superposition in Eq. (1): Y^i = Y^i_l + ΔY^i_s + ΔY^i_r, with ΔY^i_s carrying self/intention randomness and ΔY^i_r carrying social randomness. For this decomposition to be explanatory, the social-sourced branch must be driven by social context, and z_r should modulate only social effects. The architecture does not enforce this. In Eq. (13), f^i_r(z_r) is produced by Transformer Tr from Concat(Δf^i_e, F^i_R, z_r), where Δf^i_e is the ego's own differential spectrum (the same signal used in Eq. (7) for the self-bias). Thus the re-bias branch has direct access to ego-only information and to z_r even when F^i_R carries no neighbor information; with no neighbors, F^i_R is zero-padded but z_r still enters the computation. No term in the training loss (Eq. (15)) penalizes overlap between the two bias terms, enforces independence of z_s and z_r, or requires ΔY^i_r to vanish when neighbors are absent. The min-over-K objective can be satisfied by arbitrary allocation of randomness between the two branches. The appendix's perpendicular-vibration analysis is explicitly labeled 'not strict proofs' and fails on NBA and nuScenes, so the only evidence for decoupling is post-hoc visualization. Consequently, the paper's central interpretability claim—that intention randomness and social randomness are separately simulated—is not established by the model design or the reported experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Resonance (Re), a trajectory prediction model that represents an agent's future trajectory as the superposition of a linear base, a self-sourced bias, and a social/resonance-sourced bias (Eq. 1). The architecture uses spectral (Haar) transforms, two mirrored Transformer branches, and an angle-based resonance gathering mechanism (Eqs. 10–13) to compute the two biases. The authors claim that this decomposition separately simulates self/intention randomness and social randomness in an explainable way, and they support this with quantitative benchmarks on ETH-UCY, SDD, NBA, and nuScenes, plus ablations and visualizations. The model achieves strong or state-of-the-art results on several pedestrian benchmarks and is released with code.","tokens_in":28231,"tokens_out":3514,"duration_ms":39145,"significance":"If the claimed decomposition were rigorously established, the paper would make a valuable interpretability contribution to trajectory prediction: separating intention-driven from socially-driven randomness is a conceptually appealing and practically useful goal. The paper's strengths are its extensive quantitative comparisons, careful ablations (including interaction representations, waypoint counts, transforms, and angle partitions), efficiency analyses, and public code release. However, the central interpretability claims — that the two biases are causally decoupled, that self and social randomness are separately simulated, and that the two vibrations are nearly perpendicular — are supported mainly by post hoc visualizations. The appendix explicitly states that the perpendicular-vibration explanation is 'not strict proofs' and that the phenomenon fails on NBA and behaves differently on nuScenes. Because these claims are load-bearing for the paper's framing and abstract, they need either architectural enforcement, quantitative statistical validation, or a substantially weakened claim.","major_comments":[{"comment":"The claim that ΔY_r is 'social-sourced' is not enforced by the architecture. In Eq. (13), the re-bias branch receives Concat(Δf_e, F_i^R, z_r), where Δf_e is the ego's own differential spectrum also used in the self-bias branch (Eq. 7), and z_r is a free noise vector that remains active even when F_i^R is zero-padded because no neighbors exist. Nothing in the training loss (Eq. 15) prevents the model from allocating arbitrary randomness or ego-only information to the 'social' branch. The authors should provide a direct validation of the decomposition, for example by (a) measuring the change in ΔY_r when neighbors are removed while holding z_r fixed, (b) statistically testing dependence between the z_s and z_r branches, or (c) adding a regularization or architectural constraint that forces ΔY_r to vanish without neighbors and forces independence of the two noise channels.","section":"§3, Eq. (13); §4, Eq. (15)"},{"comment":"The paper's central interpretability observation — that self and social vibrations are 'almost vertical' — is explicitly stated in Appendix C.2 to be 'not strict proofs,' and the quantitative generality is weak: NBA samples show roughly 60-degree direction differences, and nuScenes shows the roles of the two biases switching (self-bias encodes velocity changes, re-bias encodes turning). Since the perpendicularity is presented in the main text as evidence of decoupling and independence, the authors should either provide statistical measurements (e.g., distributions of the angle θ_s+θ_r across test samples with confidence intervals) or downgrade this claim to an exploratory observation rather than a validated property of the model.","section":"§4.2, 'Discussions on Vibrations'; Appendix C.2"},{"comment":"The ablation results show that adding the linear base consistently hurts performance on NBA (rows a7–a9 drop by up to 7.4%/17.7% ADE/FDE) and on nuScenes (up to 7.6%/12.9%). This contradicts the narrative that the linear least-squares fit is a neutral 'reference' or 'equilibrium' for the vibration decomposition in all scenarios. While the quantitative SOTA results are not undermined, the paper should either explain why the linear base is inappropriate on these datasets and what replaces it as the equilibrium point, or revise the claim that the decomposition is a universal physical analogy.","section":"Table 4 and Table 11 (ablation rows a1–a0)"},{"comment":"The specific form of the resonance feature — an element-wise product of the ego and neighbor spectral embeddings — is a strong inductive assumption about how social interaction is computed, but the ablations in Table 8 only compare different gathering mechanisms (Social Pooling, GCN, SocialCircle, Resonance Gathering), not alternative pair-wise combination functions (e.g., concatenation, difference, or cross-attention). Since the resonance-like interaction model is one of the two main contributions, the paper should include an ablation that varies the combination function in Eq. (10) to support the claim that spectral similarity via multiplication is the right mechanism.","section":"§3, Eq. (10)"}],"minor_comments":[{"comment":"The formatting of row a2 shows '✓ ✓×' with a missing space before the × symbol; this should be corrected for consistency with the other rows.","section":"Table 11"},{"comment":"The GCN row reports '0.240/0.272' for the first entry, which appears to be a typo or a misaligned value (the FDE is implausibly low compared to neighboring rows); please verify the reported numbers.","section":"Table 8, row d2"},{"comment":"The paper uses the term 'randomnesses' in a non-standard plural form; consider rephrasing to 'random components' or 'stochasticities' for clarity.","section":"§2, Related Works and Abstract"},{"comment":"The natural frequency f0 is mentioned in the resonance analogy but is never mapped to a concrete quantity in the model; define how the model's spectral features relate to a 'natural frequency' or remove the term to avoid confusion.","section":"Fig. 2 caption and §1"},{"comment":"The 'two-player cooperative game' explanation is presented as intuition, not proof, but the manuscript does not clearly mark it as speculative; consider labeling it as an analogy to prevent readers from mistaking it for a theoretical result.","section":"Appendix C.2"}],"recommendation":"major_revision","confidential_remarks":"The paper has strong empirical results and a publicly available implementation, but the abstract and introduction overstate the interpretability claims. The main quantitative contribution — competitive or SOTA performance on multiple benchmarks — is solid and would survive a revision. However, the central 'decoupled randomness' claim needs either architectural enforcement or a much more cautious presentation. If the authors reframe the paper as a competitive architecture with spectral resonance-based social features and move the perpendicular-vibration material to an explicitly exploratory section, the paper may become acceptable. I also note a heavy reliance on the authors' own prior work (SocialCircle, SocialCircle+) in comparisons; this is not inappropriate, but the novelty relative to those papers should be clarified. No concerns about the validity of the benchmark numbers were found."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this upfront: the paper is a competent empirical contribution with a genuinely new framing, but the central interpretability claim is a lot weaker than the text suggests.\n\nWhat is actually new: treating a predicted trajectory as a superposition of a linear base plus a self-sourced vibration (intention/self randomness) and a social-sourced vibration (resonance-bias, social randomness) is a fresh decomposition, and the spectral element-wise product for pairwise resonance is a reasonable encoding. The method is well-specified, the ablations are extensive (Tables 4, 6–9, 11), and results on ETH-UCY, SDD, NBA, and nuScenes are competitive. Code is linked, and the appendix makes a good-faith effort to discuss limitations, including explicitly noting the perpendicular-vibration story is “not strict proofs.” That honesty deserves credit.\n\nThe soft spot is load-bearing for the paper’s narrative: the re-bias branch has direct access to ego-only features (Δf_e) and its own noise z_r, so the architecture does not enforce that it captures only social randomness. Nothing in the min-over-K loss penalizes overlap between the two biases or requires the re-bias to vanish when no neighbors are present. The “causality validation” is really a sensitivity analysis, not causal evidence. The appendix itself admits the perpendicular-vibration phenomenon fails on NBA, which further weakens the decoupling claim. Also minor: the linear base assumption is undercut by the model’s own ablations on NBA and nuScenes, where adding the linear base hurts performance. And there are no error bars or multi-seed results, which matters given that gains over strong baselines are often single-digit percentage improvements.\n\nNone of this makes the empirical work useless. The model performs well, and the ablations show both bias terms contribute. But the paper oversells the decoupling and the causal language. For someone working on interpretable trajectory prediction, this is a worthwhile read for the framing and the careful benchmarking, but it should not be cited as evidence that self and social randomness can be cleanly separated without architectural enforcement.\n\nRecommendation: send it to peer review. A serious referee should engage with it, but the default should be major revision—tone down the claims, correct the causal language, and require variability estimates. The core idea is publishable; the current presentation is not.","headline":"A solid, clearly specified trajectory prediction paper whose headline claim—decoupling self vs. social randomness—is not actually enforced by the architecture; worth reviewing but needs major claim revisions.","tokens_in":28771,"tokens_out":1964,"would_cite":true,"duration_ms":21316,"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":"This paper claims that pedestrian trajectories can be forecast as the superposition of a linear base and two cause-specific vibrations, with social influence read from the spectral similarity between observed trajectories.","keywords":["trajectory prediction","pedestrian motion forecasting","social interaction","resonance","spectral analysis","vibration decomposition","explainable randomness","Haar transform"],"falsifier":"Take a real ego-neighbor pair, keep both current positions fixed, and replace the neighbor's observed path with a straight-line version that has the same endpoints but no spectral structure. If the predicted trajectory changes materially, spectral shape is doing causal work; if it does not change, only position matters and the resonance mechanism is not the driver of social prediction.","tokens_in":27686,"feed_emoji":"🚶","tokens_out":4631,"duration_ms":49608,"temperature":0.7,"pith_summary":"This paper proposes that human trajectory prediction be treated as a vibration problem: each future path is the sum of a linear baseline, a self-sourced vibration capturing the agent's own intention randomness, and a social-sourced vibration capturing interaction randomness. The central claim is that these two kinds of randomness can be learned separately, with social influence read from the spectral overlap between the ego agent's observed trajectory and each neighbor's observed trajectory, in analogy to resonance. If true, predicted paths become decomposable and explainable, letting one see which part of the randomness comes from the agent's own choices and which from social pressure, rather than sampling from a single black-box noise distribution. The paper reports that this decomposition improves accuracy on pedestrian benchmarks and shows competitive results on a vehicle dataset.","feed_headline":"Predict pedestrian paths as summed vibrations","feed_subtitle":"Self-sourced and social-sourced vibration terms decompose randomness into explainable parts.","key_machinery":"The machinery is a vibration decomposition of trajectories. A linear least-squares fit provides equilibrium points, called the linear base; two noise-conditioned Transformer networks, working in the discrete Haar transform domain, produce a self-bias and a social (resonance) bias; the prediction is their sum. The specifically named object is the resonance feature $f^{i \\leftarrow j} = N_{r2}(\\mathrm{Flatten}(f^i_r \\odot f^j_r))$, an element-wise product of the ego's and a neighbor's spectral embeddings, gathered into an angle-partitioned resonance matrix. This product is the mechanism that turns 'similar spectrum' into 'social influence.'","core_discovery":"The core discovery, on the paper's own terms, is that trajectory randomness can be forecast as the superposition of independent vibrations whose states are set by sampled noise variables: $\\hat{Y}^i = \\hat{Y}^i_l + \\Delta\\hat{Y}^i_s + \\Delta\\hat{Y}^i_r$. The self-bias $\\Delta\\hat{Y}^i_s$ is generated by comparing the observed trajectory with its linear least-squares fit in the frequency domain and decoding a Transformer feature conditioned on noise $z_s$; the re-bias $\\Delta\\hat{Y}^i_r$ is generated by a second Transformer that ingests a resonance matrix built from element-wise products of the ego's and each neighbor's spectral embeddings, together with angle-based position partitions. The paper argues that this structure separates self-randomness from social-randomness, that the resonance features cluster neighbors into behaviorally meaningful groups without manual annotation, and that visualized biases show the two vibrations growing in nearly orthogonal directions. On the strength of these components, the Resonance model reports state-of-the-art or competitive ADE/FDE numbers across the tested datasets.","pith_inferences":["If the spectral-similarity mechanism is right, then perturbing a neighbor's trajectory in the frequency domain is a principled way to generate counterfactual social situations for training safety-critical predictors; this is a testable extension the paper does not explore.","The emergent near-orthogonality of the two biases suggests the network learns a coordinate system for 'intention versus interaction' that could be probed against human judgments of which motion aspects are intentional versus reactive.","The finding that the fixed linear base hurts on NBA and nuScenes points toward a natural variant with a learned or scene-adaptive reference trajectory, which would generalize the vibration idea beyond near-linear pedestrian motion.","The two-bias design also enables controlled interventions, such as holding the ego trajectory fixed while replacing a neighbor's path with its spectrally flattened version, to isolate whether spectral shape or position drives a particular social prediction."],"forward_implications":["Separating randomness by cause means a predictor can be tuned to over- or under-weight intention noise versus social noise, which is not possible with a single latent sampling.","The resonance representation yields an interpretable social feature: each neighbor contributes a resonance energy, and the paper shows these energies cluster neighbors into groups consistent with real walking groups.","The near-orthogonality of self- and social-biases gives a coordinate decomposition in which self-bias randomizes direction while re-bias randomizes speed along the motion, clarifying what kind of uncertainty each sampled prediction carries.","The method transfers to vehicle trajectories with strong results, although the paper notes that the fixed linear base is less suited to fast-changing scenes such as NBA and nuScenes.","Because the two Transformer branches are mirrored and additive, the architecture admits cheap ablated versions that retain reasonable performance while using roughly one third of the full model's inference time."],"supporting_citations":[{"why":"Supplies the ETH-UCY leave-one-out protocol and the social-pooling baseline against which trajectory prediction methods, including Re, are measured.","marker":"[1]"},{"why":"Introduces the best-of-K training and evaluation scheme that Re uses to sample and score multiple predicted trajectories.","marker":"[19]"},{"why":"Provides the waypoint-conditioned prediction idea that Re adopts when forecasting self-biases on a limited set of interpolated waypoints.","marker":"[48]"},{"why":"Defines the Transformer encoder-decoder architecture that serves as the backbone for both the self-sourced and social-sourced vibration branches.","marker":"[72]"},{"why":"Introduces the angle-based social interaction representation that Re adapts into its resonance gathering mechanism.","marker":"[77]"},{"why":"Provides the SocialCircle+ interaction representation used as a main comparison and as the 'SC' ablation variant when learning the social vibration.","marker":"[78]"},{"why":"Motivates the use of discrete Haar transform spectrums for trajectories, which Re adopts as the transform T for all spectral computations.","marker":"[79]"}],"fun_headline_variants":["Pedestrian paths as co-vibrations: decompose randomness","Forecast trajectories by summing independent vibrations","Vibrational model separates social and self randomness","Resonance: trajectories as superposed vibrations","Decompose pedestrian randomness into vibrations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that social influence between two agents can be measured by comparing the shapes of their observed paths in the frequency domain, specifically by multiplying their frequency representations together.","fun_headline_variants_meta":{"raw":{"variants":["Pedestrian paths as co-vibrations: decompose randomness","Forecast trajectories by summing independent vibrations","Vibrational model separates social and self randomness","Resonance: trajectories as superposed vibrations","Decompose pedestrian randomness into vibrations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1425,"prompt_tokens":899,"completion_tokens":526,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":515,"completion_tokens_details":{"reasoning_tokens":457}},"tokens_in":515,"tokens_out":526,"duration_ms":5313,"temperature":1.0,"reasoning_tokens":457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:26:03.442080+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real ego-neighbor pair, keep both current positions fixed, and replace the neighbor's observed path with a straight-line version that has the same endpoints but no spectral structure. If the predicted trajectory changes materially, spectral shape is doing causal work; if it does not change, only position matters and the resonance mechanism is not the driver of social prediction.","supporting_citations":[{"cited_title":"Social gan: Socially acceptable tra- jectories with generative adversarial networks","cited_arxiv_id":null,"evidence_quote":"Introduces the best-of-K training and evaluation scheme that Re uses to sample and score multiple predicted trajectories."},{"cited_title":"From goals, waypoints & paths to long term hu- man trajectory forecasting","cited_arxiv_id":null,"evidence_quote":"Provides the waypoint-conditioned prediction idea that Re adopts when forecasting self-biases on a limited set of interpolated waypoints."},{"cited_title":"Attention is all you need","cited_arxiv_id":null,"evidence_quote":"Defines the Transformer encoder-decoder architecture that serves as the backbone for both the self-sourced and social-sourced vibration branches."},{"cited_title":"Socialcircle: Learning the angle-based social in- teraction representation for pedestrian trajectory prediction","cited_arxiv_id":null,"evidence_quote":"Introduces the angle-based social interaction representation that Re adapts into its resonance gathering mechanism."}],"review_version":1}