{"id":"090ad2ea-1eab-4c06-8565-83764c87f3dc","arxiv_id":"2608.05673","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"INTraJ decomposes social influence into a socially shaped reference trajectory and a gated residual correction, yielding small but consistent ADE/FDE improvements across Argoverse 2, ETH/UCY, and SDD.","lead":"This paper introduces INTraJ, a trajectory prediction framework that separates social influence into a planning stage and a reaction stage. The framework is tested on four benchmarks and reports modest accuracy gains, with code released on GitHub.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The planning reference is a smoothed-future estimate, and the final prediction is a gate-weighted blend of that estimate with the backbone output; reported gains may stem from low-pass regularization rather than from the claimed planning-before-reaction social decomposition.","rationale":"The reader's conditional verdict is appropriate. My stress pass identifies the same weakest point and makes it sharper by exhibiting the algebraic structure of the final prediction: because T_final is a convex combination of T_full and a network supervised by smoothed GT, all reported gains could arise from generic low-pass regularization of the backbone rather than from a socially shaped planning stage. The ablation table is the key evidence: ideal-only is not better than the backbone, and ideal+fusion without the gate is exactly the backbone; only adding the learned gate moves the metrics. This does not disprove the planning-before-reaction interpretation, but it removes the empirical support for the strong abstract claim that staged social modeling is critical. A non-social smoothed-reference control and a smoothing-strength sweep would settle the issue. Because this is a missing-comparison and sensitivity problem rather than an internal inconsistency, the verdict should remain conditional; I would not change the reader's verdict.","tokens_in":16902,"tokens_out":4301,"duration_ms":47203,"concrete_test":"On the same QCNet/AV2 protocol, run two ablations with all losses and the gate unchanged: (1) Replace the supervision target T_smooth in Eq. (12) with raw GT and with several smoothing strengths (e.g., no smoothing, mild, strong); if the FDE gain disappears with raw GT or tracks smoothing strength, the improvement is temporal smoothing rather than social planning. (2) Replace the future social carrier z in Eq. (7)/(9) with a constant or empty carrier so T_ideal is generated from pure intent alone while keeping smoothed-GT supervision; if the full model's gain is preserved, the social decomposition is not load-bearing. Repeating each run with multiple seeds would also settle whether the 0.06–0.16 metric gaps exceed run-to-run noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (12) supervises T_ideal with T_smooth, a smoothed version of the future ground truth, and Eq. (13) gives T_final = T_ideal + α·(T_full − T_ideal) = α·T_full + (1−α)·T_ideal. Thus the final output is a per-step weighted blend of the backbone's full-context prediction and a network trained to regress a low-pass-filtered version of the future trajectory. The planning-stage social carrier is only one way to construct such a smoothing reference; a non-social low-pass estimate of the target trajectory would enter the same interpolation. Table 5 shows the full gain appears only when the gate α is learned (row ✓✓✓); without the gate, T_final collapses to T_full and gives no improvement, and the ideal-only branch (✓✗✗) actually worsens minADE1. Therefore the empirical gain is attributable to the gated blending mechanism under smoothed-future supervision, not necessarily to the claimed planning-before-reaction decomposition. The author response concedes that T_ideal and T_full−T_ideal should not be interpreted as identifiable causal or psychological variables, which is consistent with this concern. No smoothing-strength sensitivity, no non-social reference control, and no seed-level variance analysis are reported, so the central claim that staged social modeling is critical is not isolated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes INTraJ, a plug-in framework that decomposes social influence in trajectory prediction into a planning stage and a reaction stage. The planning stage constructs a reference trajectory T_ideal from a pure-intent seed plus a future social carrier, and the reaction stage forms the final prediction as T_final = T_ideal + α·(T_full − T_ideal), where T_full is the backbone prediction and α is a learned gate. The method is instantiated on multi-target (QCNet) and single-target (DeMo) backbones for Argoverse 2, on Resonance for ETH/UCY and SDD, and on a lightweight baseline for AV2-ped. The paper reports consistent improvements over the backbones, with state-of-the-art results on some crowd benchmarks, and provides ablation and efficiency analyses. The manuscript also contains an embedded \"Author Response\" section that qualifies the interpretation of T_ideal and T_full − T_ideal.","tokens_in":17234,"tokens_out":3836,"duration_ms":41602,"significance":"The two-stage planning/reaction formulation is potentially useful as a structured modeling principle, and the plug-in nature across two prediction paradigms and four benchmarks is an attractive feature. The paper provides public code and a broad empirical evaluation. However, the experiments as reported do not isolate the claimed mechanism from a simpler explanation: the final prediction is a gate-weighted blend of the backbone output and a network trained to regress a smoothed version of the ground-truth future. Because the central claim that \"staged social modeling is critical\" rests on this isolation, the significance is conditional on additional controls and claim revisions.","major_comments":[{"comment":"The ablation does not support the claim that both the planning stage and the reaction stage are necessary. Compared with the QCNet baseline (minADE1 1.69, minFDE1 4.30), the ideal-only branch degrades minADE1 to 1.75, the ideal-plus-fusion branch without the learned gate matches the baseline (1.69/4.30), and only the full gated model improves both metrics (1.63/4.14). Thus, the entire observed gain on this benchmark comes from the learned gate, and the planning reference alone is harmful. The text's assertion that removing either stage causes a performance drop is contradicted by the table's numbers.","section":"§4.5, Table 5"},{"comment":"Because T_ideal is trained to regress a smoothed ground-truth trajectory T_smooth and T_final = α·T_full + (1−α)·T_ideal, the reported gains could in principle be obtained by any low-pass-filtered reference blended with the backbone, without a planning-before-reaction mechanism. The paper does not provide a control experiment with a non-social smoothing reference, nor a sensitivity analysis over the smoothing strength used to define T_smooth. Such a control is necessary to attribute the improvement to the proposed social planning decomposition rather than to low-pass regularization.","section":"§3.5, Eq. (12) and Eq. (13)"},{"comment":"The manuscript's own author response concedes that \"T_ideal and T_full−T_ideal should not be interpreted as uniquely identifiable psychological or causal variables.\" This concession is in tension with the abstract and conclusion, which state that the framework \"validates that staged social modeling is critical\" and that the residual \"recovers local reactive adjustments.\" The claims in the abstract, Section 1, and Section 5 should be revised to the more limited formulation of a planning-oriented reference and a reaction-oriented correction, with the non-identifiability limitation stated in the main text.","section":"Embedded Author Response"},{"comment":"All reported results are single numbers with no error bars, confidence intervals, or significance tests, and many of the gains are small (for example, Ours (DeMo) vs. DeMo: minADE1 1.48 vs. 1.49, minFDE1 3.71 vs. 3.74; Ours (QCNet) vs. QCNet: minADE6 0.64 vs. 0.65). Given that the central claim is \"consistent improvements\" across benchmarks, the paper should report variance across at least three seeds or paired significance tests to establish that the improvements are not noise.","section":"§4.2–§4.3, Tables 1–3"}],"minor_comments":[{"comment":"The method name is typeset inconsistently as \"INTraJ\" in the abstract and as \"INTraJ\" elsewhere; please standardize the spelling.","section":"Abstract and throughout"},{"comment":"The robustness subsection is interrupted by the embedded author response block, with Figure 5 appearing after that block. The figure and the surrounding discussion should be moved into Section 4.7, and the author-response text should be removed from the camera-ready version or integrated into the main paper if it is meant to be part of the submission.","section":"§4.7 and Figure 5"},{"comment":"The caption says \"the proposed benchmark test split\" without identifying the benchmark; please specify that this is the AV2-ped split described in §4.1.","section":"Table 3 caption"},{"comment":"The hyperparameter λ_end appears in the planning reference loss but is not listed among the balancing coefficients or given a value; please report its value and whether it is tuned.","section":"§3.5, Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the reported gains may reduce to a gate-weighted smoothing blend rather than to the claimed planning-before-reaction decomposition. The authors' own response already concedes non-identifiability of the components. I recommend requiring the non-social control experiment, smoothing-strength sensitivity analysis, and seed-level variance reporting before acceptance; these are within the scope of a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read INTraJ. It's a plug-in module that decomposes social influence into a planning reference and a reaction residual, then blends them with a learned gate. The packaging is new—planning-before-reaction—and the authors evaluate it across four benchmarks and several backbones, with code released. That is real work, and the gains, though modest, are consistent: QCNet minADE1 1.69→1.63, FDE 4.30→4.14; SDD FDE 10.02→9.87; ETH/UCY average FDE 0.28→0.26.\n\nThe soft spot is the interpretation. Eq. (13) is T_final = (1−α)T_ideal + αT_full, and T_ideal is trained against a smoothed version of the future ground truth. So the final output is a per-step blend of the backbone prediction and a low-pass-filtered future estimate. The ablation in Table 5 shows that without the learned gate the framework gives no improvement over the backbone (row ✓✓✗ is exactly T_full), and the ideal-only branch actually worsens minADE1 (1.75 vs 1.69). So the empirical gain is attributable to the gated blending under smoothed-future supervision, not to the decomposition per se. No sensitivity to smoothing strength and no non-social control are reported, so the claim that \"staged social modeling is critical\" is not isolated.\n\nThe author response is honest about this: they concede T_ideal and T_full−T_ideal are not identifiable psychological or causal variables. Good. But then the abstract and conclusion should not say \"validating that staged social modeling is critical.\" That is an overclaim relative to the evidence.\n\nAlso, no error bars or significance tests anywhere. For a paper whose headline is consistent gains, a few seed-level standard deviations would help. That is a standard ask for this field, not a fatal flaw.\n\nWho is this for? People working on plug-in refinement or structured trajectory prediction. The method is simple, cheap, and likely useful as a drop-in on top of existing backbones. The conceptual framing may or may not survive contact with a proper control experiment; right now it is an attractive story, not a demonstrated mechanism.\n\nRecommendation: send to peer review. It deserves referee time: the experiments are broad, the code is out, and the core question—does a smoothing-reference blend explain the gains—can be settled by a revision that adds the missing controls. But the authors should be pushed to either defend the decomposition or drop the causal language.","headline":"A useful plug-in refinement method whose planning-reaction story is not actually isolated by the experiments; the gain comes from a learned gate around a smoothed-future reference.","tokens_in":17713,"tokens_out":2779,"would_cite":true,"duration_ms":28459,"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":"INTraJ claims that trajectory prediction improves when social influence is split into a planning stage, which builds a smooth reference path from future social cues, and a reaction stage, which adds gated local corrections.","keywords":["trajectory prediction","motion forecasting","social interaction modeling","planning and reaction decomposition","multi-target prediction","single-target prediction","gated residual fusion","pedestrian trajectory prediction"],"falsifier":"Run INTraJ with several smoothing strengths for $T_\\mathrm{ideal}$; if the FDE improvement is insensitive to smoothing strength or appears even when $T_\\mathrm{ideal}$ is replaced by an unfiltered copy of the ground truth, the two-stage interpretation loses support. Equally, train a matched model with identical parameters but no explicit planning reference, predicting $T_\\mathrm{final}$ directly from the same socialized intent, and compare FDE: if the gains vanish, the specific decomposition is not the mechanism.","tokens_in":16742,"feed_emoji":"🚗","tokens_out":7795,"duration_ms":67077,"temperature":0.7,"pith_summary":"The paper tries to establish that social influence in trajectory prediction should be modeled as a two-stage process rather than a single fused latent factor. It claims that an agent first forms a socially shaped overall plan by anticipating other agents' future motions, and then makes local reactive adjustments to that plan. INTraJ implements this idea by decoding a smoothed planning reference from a pure intent seed augmented with future social information, keeping the backbone's full-context prediction, and combining the two through a gated residual. The paper reports consistent improvements in final displacement error and long-horizon consistency across four benchmarks, including autonomous driving and pedestrian crowds, suggesting that the planning-before-reaction decomposition is a transferable modeling principle.","feed_headline":"Planning-before-reaction model improves trajectory forecasts","feed_subtitle":"Splitting social cues into a smooth plan plus local corrections lifts FDE across driving and crowd data.","key_machinery":"The Plan-React module carries the argument. It defines the identity $T_\\mathrm{final} = T_\\mathrm{ideal} + \\alpha \\cdot S_\\mathrm{high}$, where $T_\\mathrm{ideal}$ is a planning reference decoded from a socially shaped intent representation, $T_\\mathrm{full}$ is the backbone's full-context prediction, and $S_\\mathrm{high} = T_\\mathrm{full} - T_\\mathrm{ideal}$ is the reaction residual. The planning reference is produced by cross-attending a pure intent seed (derived from the focal agent's history and local context) to a future social carrier $z$ — predicted trajectories of other agents in the multi-target paradigm, or a counterfactual latent response from a static-ego branch in the single-target paradigm. A learnable gate $\\alpha$ controls how much of the residual is retained per sample, mode, and time step, and a variance-based auxiliary loss prevents the gate from collapsing to a constant.","core_discovery":"The central claim is that trajectory formation follows a planning-before-reaction order: before acting, an agent anticipates the future behavior of surrounding agents and forms a socially shaped plan, and only then makes local reactive adjustments when executing that plan. Concretely, INTraJ computes a planning-level reference trajectory $T_\\mathrm{ideal}$ from a pure intent seed updated through cross-attention with a future social carrier (predicted futures of others in multi-target settings, a counterfactual latent response in single-target settings). It keeps the backbone's full-context prediction $T_\\mathrm{full}$ and treats the difference $S_\\mathrm{high} = T_\\mathrm{full} - T_\\mathrm{ideal}$ as reaction-stage adjustment, fused by $T_\\mathrm{final} = T_\\mathrm{ideal} + \\alpha S_\\mathrm{high}$ with a learned gate $\\alpha$. The paper argues that this explicit separation of global passing tendency from local interaction response yields more stable, coherent predictions, and evidences it with gains in FDE across AV2, AV2-ped, ETH/UCY, and SDD.","pith_inferences":["An untested consequence of the paper's view is that the reaction residual $S_\\mathrm{high}$ should spike at concrete local interaction events, such as yielding or avoiding a collision; one could verify this by checking whether high-residual moments align with time-to-collision minima in the data.","Because the planning reference is supervised by a smoothed ground truth, the paper has not ruled out the alternative that the planning stage is merely a low-pass filter; a testable extension would replace the smoothed target with a scene-derived planning signal, such as lane or goal conditioning, and compare.","If the decomposition is genuinely causal, perturbing the future social carrier should alter $T_\\mathrm{ideal}$ more than $S_\\mathrm{high}$; this could be tested by selectively ablating the carrier and measuring each component's change."],"forward_implications":["If the two-stage view is right, then the same decomposition should keep improving any future backbone it is plugged into, since it adds a structured planning head without altering the base predictor.","The framework's gains in FDE and long-horizon consistency imply that explicitly separating global passing tendency from local reaction is a useful inductive bias for trajectory stability.","Because INTraJ works in both multi-target and single-target paradigms, the planning-before-reaction principle is not tied to a particular architecture or input representation.","The consistent improvements on both autonomous driving and pedestrian crowd benchmarks suggest that staged social modeling generalizes across application domains."],"supporting_citations":[{"why":"Provides the multi-target backbone and the full-context prediction $T_\\mathrm{full}$ that anchors the residual reaction stage.","marker":"[46]"},{"why":"Serves as the single-target backbone instantiation, providing the decoupled design that yields the counterfactual social carrier.","marker":"[43]"},{"why":"Serves as the crowd backbone onto which the Plan-React module is plugged, and its state-of-the-art results are surpassed.","marker":"[39]"},{"why":"Serves as the plug-in refinement baseline that INTraJ outperforms on the QCNet backbone, enabling direct comparison with post-processing approaches.","marker":"[45]"},{"why":"Provides the Argoverse 2 driving benchmark and the official metrics (minADE, minFDE, MR) used for evaluation.","marker":"[37]"},{"why":"Supplies the ETH/UCY crowd trajectories that serve as the primary benchmark for evaluating the planning-before-reaction decomposition.","marker":"[1]"},{"why":"Defines the Stanford Drone Dataset, the second crowd benchmark used to measure FDE gains.","marker":"[2]"},{"why":"Provides the Argoverse 2-ped pedestrian benchmark and the lightweight baseline setup on which the framework achieves state-of-the-art performance.","marker":"[34]"}],"fun_headline_variants":["Two-stage planning-reaction model sharpens trajectory forecasts","INTraJ splits social cues into plan plus local fixes","Frame trajectories as plan-then-react, cut FDE errors","Staged social modeling lifts long-horizon trajectory accuracy","Planning stage uses future cues before reactive adjustment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The planning reference is trained against a mildly smoothed ground-truth trajectory; if that smoothed target is not what an agent would actually plan before reacting, the planning/reaction split may be an artifact of low-pass filtering rather than a real behavioral mechanism.","fun_headline_variants_meta":{"raw":{"variants":["Two-stage planning-reaction model sharpens trajectory forecasts","INTraJ splits social cues into plan plus local fixes","Frame trajectories as plan-then-react, cut FDE errors","Staged social modeling lifts long-horizon trajectory accuracy","Planning stage uses future cues before reactive adjustment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00047,"raw_usage":{"total_tokens":2352,"prompt_tokens":968,"completion_tokens":1384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":1306}},"tokens_in":584,"tokens_out":1384,"duration_ms":10972,"temperature":1.0,"reasoning_tokens":1306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T04:22:11.304150+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run INTraJ with several smoothing strengths for $T_\\mathrm{ideal}$; if the FDE improvement is insensitive to smoothing strength or appears even when $T_\\mathrm{ideal}$ is replaced by an unfiltered copy of the ground truth, the two-stage interpretation loses support. Equally, train a matched model with identical parameters but no explicit planning reference, predicting $T_\\mathrm{final}$ directly from the same socialized intent, and compare FDE: if the gains vanish, the specific decomposition is not the mechanism.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the single-target backbone instantiation, providing the decoupled design that yields the counterfactual social carrier."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Serves as the crowd backbone onto which the Plan-React module is plugged, and its state-of-the-art results are surpassed."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Argoverse 2 driving benchmark and the official metrics (minADE, minFDE, MR) used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the ETH/UCY crowd trajectories that serve as the primary benchmark for evaluating the planning-before-reaction decomposition."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Argoverse 2-ped pedestrian benchmark and the lightweight baseline setup on which the framework achieves state-of-the-art performance."}],"review_version":1}