{"id":"0df555a5-7aac-4153-800a-44782c47053a","arxiv_id":"2607.12503","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free pipeline that feeds MLLMs reprojected motion arrows plus a structured trace graph lifts 4D spatio-temporal QA accuracy on three benchmarks.","lead":"DynTrace is a training-free add-on that helps video-capable AI models track moving objects in 3D and keep that information while answering questions about motion and distance. On three 4D-reasoning benchmarks it lifted open-source models by roughly 3-6 points, though the main benchmark was built by the same research group.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncontrolled input substitution: DTG supplies answer-relevant metric values, so improved accuracy may reflect reading text rather than tracking dynamic evidence.","rationale":"The reader's weakest_assumption targets geometric fidelity: monocular depth, pose, and the affine background model feed directly into DTV and DTG. That is a genuine risk, but Table 7 partially mitigates it by comparing GT-based and predicted evidence (55.9% vs 54.2%). The untested confound I focus on is input substitution. The reader lists this as one reason for CONDITIONAL, but does not make it the weakest assumption. In my reading it is more load-bearing because it directly threatens the interpretation of every reported gain: if DTG text contains the answer-relevant metric values, the method's success may not demonstrate that MLLMs reason better with tracked dynamic evidence, only that they can read explicit numbers. The paper's ablations show DTG is the main contributor, which strengthens this concern. The proposed test is straightforward and would settle whether the gains survive when raw frames remain available or when DTG values are perturbed. Since the reader already issued CONDITIONAL and my concern reinforces that, the verdict should remain UNCHANGED rather than moving to ACCEPT or REJECT.","tokens_in":45106,"tokens_out":9000,"duration_ms":102610,"concrete_test":"Run a single controlled condition on Dyn-Bench with Qwen3-VL-8B: feed the baseline model the raw 16 frames plus the exact DTG text (without DTV images), using the same final prompt. If accuracy is within ~1 point of the full DynTrace (65.8%), the gain is primarily from DTG text, not from DTV or from replacing frames. Second, perturb DTG numeric values (multiply distances/speeds by 1.5, swap left/right) while keeping DTV unchanged; a large accuracy drop would show the model is reading text values directly. Also rerun with raw frames added to the DynTrace input; if accuracy does not drop, the 'instead of raw frames' design is not the source of benefit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical comparison is not controlled for the information content of the input. In the DynTrace condition the answering MLLM receives, in place of raw frames, DTV overlays plus DTG serialized text (Eqs. 7-9, Eq. 10). DTG explicitly reports end-of-window positions, speeds, headings, distances, bearings, and key moments such as CLOSEST_POINT with timestamps. Many Dyn-Bench, VLM4D, and DSI-Bench questions ask exactly for these quantities (e.g., approximate distance at 1.5 s, direction of motion, speed). The baseline sees raw frames and must estimate them. Table 3 shows DTG alone accounts for most of the gain (+2.2 vs +0.9 for DTV), consistent with a read-the-answer-from-text effect rather than improved tracking or reasoning. The paper does not run a control where raw frames remain visible alongside DTV/DTG, nor a control where an equally informative but non-tracking textual summary is given to the baseline. If DTG values are directly answer-relevant, the headline claim that tracking dynamic object evidence improves MLLM reasoning is underdetermined: the gain might simply reflect that the test-time input now contains the answer. This is distinct from geometric accuracy; even with perfect monocular estimates (Table 7 shows only a 1.7-point gap between GT-based and predicted evidence), the method's advantage could be an artifact of giving away metric cues.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DynTrace, a training-free inference-time framework that augments MLLM inputs for 4D spatio-temporal reasoning. It extracts query-relevant dynamic objects, reconstructs their world-coordinate trajectories from monocular depth and pose, and provides two forms of evidence: DTV (Dynamic Trajectory Visualization), which reprojects world trajectories onto the image plane as directional overlays, and DTG (Dynamic Trace Graph), which serializes per-window object/relation tokens containing dynamic cues, trace evolution, and key moments. The enhanced MLLM receives DTV images plus DTG text instead of raw frames. Experiments on Dyn-Bench, VLM4D, and DSI-Bench report consistent improvements over four open-source MLLM backbones, with ablations attributing +2.2 of the +5.0 Dyn-Bench gain to DTG and +0.9 to DTV.","tokens_in":45276,"tokens_out":5010,"duration_ms":55374,"significance":"If the central claim is established, DynTrace would be a useful, inexpensive, model-agnostic way to inject explicit dynamic evidence into MLLMs, and its decomposition into Dynamic Source Confusion and Dynamic Trace Loss would provide a principled taxonomy for 4D reasoning failures. The paper has several strengths: the pipeline is clearly described and training-free; gains are reported on three benchmarks, including two externally-authored ones; ablations separate the contributions of DTV, DTG, and DT-Token components; Table 7 gives a GT-versus-predicted robustness check; and the appendix contains honest limitation cases. However, the headline interpretation—that tracking dynamic object evidence improves MLLM reasoning—is currently underdetermined because the experimental setup changes both the information content and the input modality, not just the reasoning scaffold.","major_comments":[{"comment":"The DynTrace condition is not controlled for the information content of the input. The enhanced model receives DTG serialized text (Eqs. 7–9) that explicitly lists end-of-window positions, speeds, headings, distances, bearings, and key moments such as CLOSEST_POINT with timestamps (e.g., Fig. 7; Fig. 12), while the baseline receives raw frames and must estimate these quantities. Many questions in Dyn-Bench, VLM4D, and DSI-Bench ask for exactly these values (e.g., distance at a given time, average speed). Table 3 shows that DTG alone accounts for +2.2 of the +5.0 gain, versus only +0.9 for DTV, which is consistent with the model reading the answer directly from the text. Please add controls: (a) keep raw frames visible in the DynTrace condition; (b) give the baseline an equally informative but non-tracking textual summary (e.g., per-window positions without trajectory-trend or key-moment","section":"§3.3, Eq. (10); §4.1; Table 3"},{"comment":"Dyn-Bench is used as the primary benchmark and is also cited in the introduction as empirical validation of the paper's own premise: 'Dyn-Bench [10] empirically validates this coupled mechanism.' Ref. [10] shares at least four authors with this paper (Yuzhi Huang, Yue Huang, Xinghao Ding, Jingyan Jiang). Using a same-group benchmark to validate a central premise of the same group creates a circularity risk, and the paper does not disclose the overlap. This concern is mitigated by the external VLM4D and DSI-Bench results, but the SOTA claim on Dyn-Bench and the framing of Dyn-Bench as the main support need revision. Please disclose the author overlap and restructure the evidence so that externally-authored benchmarks carry at least equal weight.","section":"§1, §4.1; Ref. [10]"},{"comment":"The world-coordinate trajectories, DTV arrows, and DTG metric values are all built on monocular depth and pose estimates plus a 2D affine model of camera-induced background motion. No accuracy analysis of these intermediate estimates is reported, and no error bars or significance tests are given for any of the headline results. Since the DynTrace input replaces the raw evidence with these estimates, errors in depth scale, pose drift, or parallax that the affine model cannot capture propagate directly into answers. Table 7 is a useful start but only reports one averaged subset. Please report sensitivity of final accuracy to plausible perturbations of the intermediate geometry (e.g., depth scaled by ±10%, ±20%; pose jitter) and provide at least bootstrap confidence intervals for the main gains, especially for category-level gains as small as +0.5 and +0.8 in Table 1.","section":"§3.2, Eqs. (5)/(22), (12)–(14); Tables 1–7"}],"minor_comments":[{"comment":"The legend is difficult to read because the same marker style is used for both benchmarks and DynTrace/baseline pairs. Please use distinct line styles and label each series directly.","section":"Fig. 4"},{"comment":"The notation alternates between D/DTV and G/DTG in the text and figure. Please define the symbols once in a notation table to avoid confusion.","section":"§3.3 and Fig. 2"},{"comment":"The final reasoning prompt asks for 'ONLY the letter (A, B, C, or D)', but the qualitative examples show long reasoning chains. Please clarify whether the reported numbers come from the letter-only protocol or from a separate reasoning-then-answer setting.","section":"Fig. 18 prompt vs. Fig. 3/14/15 examples"},{"comment":"The two limitation cases—egocentric frame transformation and fine-grained body-part relations—are honestly reported and align with the method's design. It would strengthen the paper to add one sentence in the main text referencing these boundaries.","section":"Appendix B, 'Current limitations'"}],"recommendation":"major_revision","confidential_remarks":"The author overlap between Dyn-Bench and this submission is significant and should be disclosed. The paper currently cites Dyn-Bench as external validation of its own premise, which is a circularity risk. The uncontrolled input substitution is the more serious technical issue: DTG supplies answer-relevant metric values, so the main empirical claim needs additional controls. If those controls confirm the gains, the paper would be suitable for publication; without them, the central interpretation is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a clean, well-documented training-free pipeline that lifts open-source MLLMs by 3-6 points on three 4D spatio-temporal QA benchmarks, with consistent gains across four backbones. Second, I don't trust the central interpretation: the experimental setup lets DTG hand the model the numeric answers, so the gains may be a reading-comprehension effect rather than evidence that tracking dynamic evidence helps reasoning.\n\nWhat's actually new is the coupling: DTV reprojects world-coordinate trajectories onto the image plane as motion arrows, and DTG serializes object/relation cues, trace evolution, and key moments into a text graph. Each piece is established separately, but this particular two-branch test-time augmentation is new. The geometry is simple and correctly applied; I found no derivation gaps. The ablation structure is sensible (DTG +2.2, DTV +0.9, combined +5.0), and the authors report external benchmarks (VLM4D, DSI-Bench), a GT-vs-predicted robustness check, occlusion robustness, and an honest limitations section. That's more than most papers in this area.\n\nThe soft spot is load-bearing. The DynTrace condition replaces raw frames with DTV and DTG. DTG explicitly reports end-of-window positions, speeds, headings, distances, bearings, and key moments like CLOSEST_POINT with timestamps. Many questions on all three benchmarks ask for exactly these quantities (distance at a given time, direction of motion, speed). The baseline must estimate them from frames; DynTrace can read them from text. Table 3 shows DTG alone accounts for most of the gain, which is consistent with a read-the-answer effect. The paper never runs a control where raw frames remain visible, nor a control where the baseline receives an equally informative but non-tracking textual summary. The GT-based check only addresses geometric accuracy, not this confound. There's also a secondary circularity concern: the primary benchmark Dyn-Bench has overlapping authors with this paper, and the failure-mode subsets are defined by a 15-rule prompt that mines exactly the two failure modes the method targets. External benchmarks mitigate the circularity but not the input-substitution issue.\n\nBottom line: this is a serious, useful paper for people building training-free augmentations for video/4D MLLMs. The method is plausible and the engineering is solid. But the headline claim about tracking dynamic object evidence is not yet established. It deserves peer review, not desk rejection. The referee should demand controlled experiments (raw frames + DTV+DTG; equally informative text baseline) and code/data release. If the gains survive those controls, this becomes a real contribution.","headline":"Useful training-free pipeline for 4D MLLM reasoning, but the headline claim is undercut by an uncontrolled input-substitution confound: DTG hands the model the metric answers.","tokens_in":45981,"tokens_out":3743,"would_cite":false,"duration_ms":37777,"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":"DynTrace claims that dynamic 4D video reasoning fails because models confuse object motion with camera motion and lose continuous object traces, and that a training-free infusion of geometry-grounded arrows and structured text traces fixes","keywords":["4D spatio-temporal reasoning","multimodal large language models","dynamic object tracking","camera motion disentanglement","trajectory visualization","visual question answering","training-free inference","dynamic trace graph"],"falsifier":"Use synthetic videos with known ground-truth 3D trajectories and a camera path engineered to produce parallax (depth discontinuities moving across the frame). If DynTrace's accuracy on static foreground objects falls below the same MLLM seeing raw frames, the claim that tracked dynamic evidence—rather than lucky geometry—carries the result is falsified; equivalently, deliberately scaling the estimated depth should change answers in proportion to the depth error.","tokens_in":44822,"feed_emoji":"🎯","tokens_out":6885,"duration_ms":64886,"temperature":0.7,"pith_summary":"The paper sets out to explain why large multimodal models (MLLMs) are bad at dynamic 4D spatio-temporal reasoning—questions about how objects move, turn, approach, or separate over time in video. It argues that the failures come from two evidence deficits: models confuse real object motion with camera-induced apparent motion, and sparse frame sampling breaks the continuous trail of each object. DynTrace is a training-free fix: before the answering model sees anything, it extracts query-relevant moving objects, reconstructs their 3D trajectories, and presents them to the MLLM twice—as annotated arrows overlaid on frames and as a structured text graph of cues, trends, and key moments. The paper reports that this consistently improves four open-source MLLMs across three 4D benchmarks, with the fused evidence beating each component alone and an 8B model with DynTrace outscoring a raw 32B model. If right, the practical lesson is that explicit, continuously tracked dynamic evidence can substitute for much larger models or additional training.","feed_headline":"Motion arrows plus text traces lift 4D video answers up to 11 points","feed_subtitle":"Training-free visual arrows and text traces fix two failure modes in video-question-answering AI models.","key_machinery":"At the core are two complementary evidence channels built from the same metric reconstruction. DTV (Dynamic Trajectory Visualization) takes the world-coordinate trajectory of each tracked object, projects it through the camera model onto the frame, and draws a direction arrow on the object's mask; because the arrow comes from world motion, not image displacement, it stays truthful when the camera itself moves. DTG (Dynamic Trace Graph) partitions the video into temporal windows and serializes, per window, a camera summary, object tokens, and relation tokens; each token records the state at the window's end (cues), how it changed through the window (trace), and sparse decisive events such as","core_discovery":"DynTrace's central claim is that current MLLMs fail at continuous dynamic scene reasoning not from a lack of visual content but from how the evidence is delivered: sparse, untracked, camera-uncompensated frames. The method builds Geometry-Grounded Dynamic Evidence: object trajectories lifted into a shared world frame via monocular depth and pose, camera behavior summarized separately, and pair relations (distance, bearing, interaction moments) computed from those trajectories. It then renders this evidence twice—Dynamic Trajectory Visualization (DTV), which reprojects world-space motion as arrows on the image plane, and Dynamic Trace Graph (DTG), a windowed text graph of object and relation","pith_inferences":["The authors' framing suggests a cheap adaptive deployment rule: for questions about camera-relative direction, feed only DTV; for questions about distance evolution or event order, feed only DTG; a router could halve token cost without losing much accuracy.","The reported gains probably upper-bound the value of tracking; the real bottleneck is the metric quality of the monocular depth and pose estimates and the assumption that camera-induced flow is a global affine transform. Videos with strong parallax or scale drift would stress-test this.","The failure on part-level and first-person-frame questions implies the next advance is not better object tracking but per-part trajectories and viewpoint transformation—two concrete extensions suggested by the paper's own limitations.","Because raw frames are never shown, the answering model cannot verify the arrows or numbers against pixels; keeping a few original frames as a cross-check channel might catch residual geometric errors."],"forward_implications":["Any open-source MLLM can be upgraded at inference time by prepending DynTrace-style evidence; the paper reports gains of +4.3 to +5.0 on its main benchmark and +3.0 to +6.1 on the other two.","Explicit dynamic evidence can partially substitute for scale: an 8B model with DynTrace beats a raw 32B model.","The two components are complementary, not redundant: DTG alone adds +2.2 and DTV alone +0.9 over the raw 8B backbone, while together they add +5.0.","The benefit persists under temporary occlusion (overall 58.1% to 62.0% on 75 such clips), because the text trace survives gaps in visual evidence.","Both failure modes improve separately: gains appear on question subsets classified as Dynamic Source Confusion and on subsets classified as Dynamic Trace Loss."],"fun_headline_variants":["Track dynamic cues to boost 4D video reasoning in MLLMs","Arrows and trace tokens fix 4D video QA with training-free method","Training-free motion tracking lifts 4D video AI answers by 11 pts","Dynamic trace graph and visual arrows improve 4D video reasoning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole evidence chain treats monocular depth, camera pose, and a global affine camera-motion model as reliable enough that their errors never need visual verification, while raw frames are withheld from the answering model.","fun_headline_variants_meta":{"raw":{"variants":["Track dynamic cues to boost 4D video reasoning in MLLMs","Arrows and trace tokens fix 4D video QA with training-free method","Training-free motion tracking lifts 4D video AI answers by 11 pts","Dynamic trace graph and visual arrows improve 4D video reasoning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001161,"raw_usage":{"total_tokens":4682,"prompt_tokens":822,"completion_tokens":3860,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":3790}},"tokens_in":566,"tokens_out":3860,"duration_ms":21865,"temperature":1.0,"reasoning_tokens":3790,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T06:28:12.718587+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use synthetic videos with known ground-truth 3D trajectories and a camera path engineered to produce parallax (depth discontinuities moving across the frame). If DynTrace's accuracy on static foreground objects falls below the same MLLM seeing raw frames, the claim that tracked dynamic evidence—rather than lucky geometry—carries the result is falsified; equivalently, deliberately scaling the estimated depth should change answers in proportion to the depth error.","supporting_citations":[],"review_version":2}