{"id":"06585d3c-ba03-498d-a964-25687f96a034","arxiv_id":"2507.15428","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"EgoPrune prunes egomotion video tokens by homography-based frame alignment and MMR selection, keeping accuracy close to the full-token baseline while reducing compute.","lead":"EgoPrune is a training-free token pruning method that uses camera geometry to remove redundant visual tokens before feeding egomotion video into a vision-language model. It claims to keep over 99% of full-input accuracy while cutting FLOPs, memory, and latency, and it runs on a Jetson Orin NX edge device.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claims are not yet established end-to-end: TFLOPs and latency are reported at fixed retention without stating whether PARF and MMR overhead are included, so the measured savings may reflect only the LLM forward pass.","rationale":"The reader's verdict is already CONDITIONAL, and the identified weaknesses—no error bars, post-hoc 90s filter, missing keyframe-selector ablation, author overlap—are real. My concern differs from the reader's weakest_assumption: I do not think the homography failure mode is the most load-bearing issue. Empirically, EgoPrune preserves high accuracy on VSI-Bench, which contains indoor parallax-heavy scenes, so the homography assumption appears to hold well enough in practice. The more vulnerable part of the central claim is the efficiency half: the paper promises large FLOPs, memory, and latency reductions, but the main efficiency figure is collected with DeepSpeed Profiler at a fixed token retention, making it unclear whether the reported TFLOPs include the overhead of PARF and MMR. Since MMR is O(n^2) in the number of visual tokens and PARF requires multiple feature-matching operations per keyframe pair, this overhead is not obviously negligible. The paper does report on-device end-to-end latency on Jetson Orin NX, which is valuable independent support and suggests the pipeline can be fast in practice, but it uses a different backbone and quantization and does not provide a stage-by-stage breakdown. Therefore the efficiency claim needs a direct end-to-end measurement before it can be accepted, which is consistent with keeping the reader's CONDITIONAL verdict unchanged. The recommended test is a single profiled run that includes all preprocessing stages for all methods, which would settle whether the promised compute savings survive full accounting.","tokens_in":13851,"tokens_out":9605,"duration_ms":101260,"concrete_test":"Instrument the full EgoPrune inference path—including ORB keypoint extraction, FLANN/RANSAC homography estimation, perspective warp, PARF similarity filtering, and MMR selection—and the full DivPrune and PACT preprocessing paths, using a profiler that captures wall-clock time and FLOPs for all stages on the same ≤90s filtered VSI-Bench and UrbanVideo-Bench videos at 50% retention. Report the breakdown (preprocessing vs. LLM forward) and verify whether EgoPrune's end-to-end latency and FLOPs remain below baselines by the claimed margin; if the pruning overhead dominates, the central efficiency claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central efficiency claim—'significantly reducing FLOPs, memory usage, and latency'—is supported by Section 5.2.2, but the reported TFLOPs are collected with DeepSpeed Profiler, which profiles model execution, not the surrounding Python/pybind stages. PARF (Section 4.2.1) runs ORB detection, FLANN matching, RANSAC, and a perspective warp, while the MMR selector (Section 4.2.2) has O(n^2) complexity with a reported window w=10 and token dimension d of the VLM; with T≈90 frames and N=196 tokens per frame, n≈17,640, making MMR non-trivial. The paper's own algorithm (Algorithm 1) scans all previously selected tokens, not just the most recent w, so the stated complexity is ambiguous. Moreover, if 'all methods retain 50% of tokens' means an equal number of tokens reaches the LLM, then the LLM forward-pass TFLOPs should be identical across methods; any reported TFLOPs difference must come either from token counts not actually being equal or from pruning-stage FLOPs that are not separately reported. The efficiency figure therefore does not yet establish that EgoPrune's end-to-end cost is lower, which is the central claim. On-device latency in Section 5.4 likely includes the pipeline, but no breakdown is given, and it uses a different model (VILA-1.5 3B) and INT4 quantization, so it does not resolve the TFLOPs accounting. This is more load-bearing than the homography concern alone: PARF failure would degrade accuracy, but the experiments already show high accuracy on parallax-heavy indoor video, whereas the efficiency claim is unsupported without an end-to-end measurement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EgoPrune, a training-free token pruning pipeline for egomotion video reasoning in vision-language models. It combines a keyframe selector adapted from EmbodiedR, a Perspective-Aware Redundancy Filtering (PARF) stage that uses homography-based alignment to drop temporally redundant tokens, and an MMR-based token selector that balances prompt relevance and intra-frame diversity. The method is evaluated on VSI-Bench and UrbanVideo-Bench with LLaVA-OneVision-7B against DivPrune and PACT at 70%, 50%, and 30% token retention, with additional efficiency measurements and an on-device deployment on a Jetson Orin NX 16GB. The abstract claims over 99% task accuracy preservation while significantly reducing FLOPs, memory, and latency.","tokens_in":14239,"tokens_out":5701,"duration_ms":58011,"significance":"If the efficiency claims hold, EgoPrune would be a useful contribution: it is training-free, geometry-aware, and plausibly deployable on embodied platforms. The appendix ablations of PARF and MMR are a strength, and the on-device evaluation adds practical evidence. However, the significance is currently tempered by unresolved accounting of pruning overhead, the absence of a keyframe-selector ablation, and the lack of statistical confidence in small accuracy differences. The core idea is promising and worth pursuing, but the empirical support needs to be tightened.","major_comments":[{"comment":"The efficiency claim is not yet established. Figure 5 reports TFLOPs, latency, and memory with all methods retaining 50% of tokens, but the paper does not state whether the PARF (ORB+FLANN+RANSAC+warp) and MMR selection costs are included; DeepSpeed Profiler profiles model execution, not surrounding Python/pybind stages. If all methods feed the same number of tokens to the LLM, LLM-forward TFLOPs should be identical, so the reported differences must come from unequal final token counts or from un-reported pruning-stage costs. The paper needs an end-to-end accounting with and without pruning overhead, plus a clear statement of the actual token counts entering the LLM. In addition, Algorithm 1 line 8 computes max over all previously selected tokens j in S, not only the most recent w, so the claimed O(rwn^2) complexity with w=10 does not match the described procedure.","section":"Section 5.2.2 and Algorithm 1"},{"comment":"The keyframe selector is never ablated. The pipeline's first stage is 'Embodied-R's overlap-aware keyframe selection', but no experiment compares EgoPrune with and without this stage or replaces it with uniform sampling. Consequently the reported gains over DivPrune and PACT cannot be attributed to PARF/MMR; they may largely stem from the keyframe selection. The appendix ablates PARF and MMR but holds the keyframe selector fixed. Please add a keyframe ablation and ensure baselines use the same frame-sampling policy.","section":"Section 4.2.1 and Section 5.3"},{"comment":"The headline accuracy-preservation claim rests on very small differences. For example, VSI-Bench at 50% retention gives 35.43 for EgoPrune versus 35.45 for full tokens, and UrbanVideo-Bench at 30% gives 45.66 versus 46.23 for full tokens; the improvements over DivPrune and PACT are typically 0.2 to 1.0 points. No error bars, number of evaluation repeats, or significance tests are provided. Since the abstract's 'over 99%' claim is computed from these differences, the paper should report variance (e.g., bootstrap or multiple runs) and show that the ranking is stable.","section":"Tables 1 and 2"},{"comment":"Excluding videos longer than 90 seconds removes the regime where token pruning matters most. At 1 FPS and 196 tokens per frame, 90 seconds yields about 17.6K visual tokens, which is moderate for a 7B VLM. The abstract's motivation is 'long, redundant video inputs', but the evaluation truncates long videos. Please report results on the full benchmarks, and/or analyze accuracy and efficiency as a function of video length, to justify the cutoff and the long-video claim.","section":"Section 5.1.3"},{"comment":"PARF relies on a single homography per consecutive frame pair, which is exact only for planar scenes or pure rotation. Real egomotion includes translation with depth variation, so many tokens will be misaligned; the 75% cosine-similarity threshold could then discard informative tokens. The paper provides no alignment-quality metric or failure analysis. Please add quantitative evidence (e.g., fraction of frames where RANSAC inliers are below a threshold, or comparison to optical-flow or essential-matrix alignment) to show that misalignment does not undermine pruning decisions.","section":"Sections 3.2 and 4.2.1"}],"minor_comments":[{"comment":"The caption contains a typo: 'token retation rate' should be 'token retention rate'.","section":"Figure 7 caption"},{"comment":"The RANSAC reference is given as Cantzler 1981; the standard reference is Fischler and Bolles (1981). Please correct this citation.","section":"References"},{"comment":"The text repeatedly prints 'LLaV A' with a space; use 'LLaVA' consistently.","section":"Throughout"},{"comment":"The paper states three NVIDIA A6000 GPUs with 50GB, but the A6000 has 48GB memory; please correct or clarify the hardware configuration.","section":"Section 5.1.3"},{"comment":"Algorithm 1 does not define the similarity function used in line 8; Section 4.2.2 defines sim after the algorithm, so the algorithm should refer explicitly to that definition.","section":"Algorithm 1"},{"comment":"The connection between 'over 99% of task accuracy' in the abstract and the numbers in Tables 1 and 2 should be made explicit, since the comparison is to the full-token baseline and not to a task-specific ceiling.","section":"Abstract and Section 5.2.1"},{"comment":"The keyframe selector from EmbodiedR is not described sufficiently in this paper; a short algorithmic description or pseudocode would improve reproducibility.","section":"Section 4.2.1"}],"recommendation":"major_revision","confidential_remarks":"The overlap between the authors and the EmbodiedR/UrbanVideo-Bench papers creates a potential disclosure concern; the paper should state this overlap explicitly in the text. In addition, the efficiency comparison needs a clear explanation of why TFLOPs differ when all methods are said to retain the same percentage of tokens."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is PARF: aligning consecutive frames with a homography before cosine-similarity pruning. That is a sensible, well-motivated fix for egomotion video, where fixed-position token comparisons fail. The rest of the pipeline is an integration of known parts—the keyframe selector from EmbodiedR and the MMR selector from CDPruner—but the combination is new, and the paper is honest about that lineage. Credit where due: the problem is real, the method is simple and plausible, and the evaluation on two egomotion benchmarks at three retention ratios is more thorough than most pruning papers. The on-device deployment on a Jetson Orin NX is also a concrete step beyond synthetic FLOP counts. The reported accuracy numbers—near or slightly above the full-token baseline at 70%, 50%, and 30% retention—are believable, and the appendix ablations show both PARF and MMR contribute.\n\nWhere the paper is soft is the efficiency claim, which is the load-bearing one. If all methods retain 50% of tokens, the LLM forward pass should see the same token count, so identical TFLOPs—yet the paper reports differences. That means the reported TFLOPs either come from uneven token counts or are measuring only a slice of the pipeline. The pruning overhead (ORB, FLANN, RANSAC, homography warp, and the O(n^2) MMR scan) is never included in the TFLOPs figure, and the DeepSpeed Profiler typically measures model execution, not Python-side preprocessing. So the central promise of \"significantly reducing FLOPs\" is not established. The latency measurements on the edge device look more end-to-end, but without a breakdown they don't settle the accounting either.\n\nOther issues are smaller. The 90-second video filter is a post hoc exclusion, though at least the paper states it as an assumption. The keyframe selector is never ablated, so you cannot tell how much of the gain comes from that borrowed component versus the new ones. The accuracy differences are a few tenths with no error bars, which makes the \"consistently outperforms\" language a bit strong. The homography assumption (planar scenes or pure rotation) is acknowledged, and the experiments on parallax-heavy indoor video suggest the method tolerates misalignment better than you'd fear, but the paper doesn't analyze when it breaks.\n\nBottom line: the accuracy-preservation story is credible and the PARF idea is worth pursuing. The efficiency claim needs an honest end-to-end measurement that includes pruning overhead and reconciles the token-count accounting. I would send this to review, and ask for a revised efficiency section and a keyframe ablation. A serious referee can get value out of this paper.","headline":"EgoPrune's accuracy story mostly holds; its headline efficiency claim is not yet backed by end-to-end measurement.","tokens_in":760,"tokens_out":944,"would_cite":true,"duration_ms":29480,"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":"EgoPrune claims that a training-free, geometry-aware token-pruning pipeline—homography alignment, a 75% cosine-similarity cutoff, then prompt-balanced MMR selection—keeps over 99% of egomotion video-reasoning accuracy while cutting FLOPs…","keywords":["egomotion video reasoning","training-free token pruning","homography alignment","visual token redundancy","vision-language models","embodied AI","edge deployment","maximal marginal relevance"],"falsifier":"Run EgoPrune at 30% token retention on egomotion videos where translation dominates and depth varies sharply—a handheld or drone camera moving straight through a cluttered room—and compare answer accuracy against the full-token baseline; the published gaps at 30% retention are about 0.3–0.6 points, so a drop of more than a point or two would falsify the alignment premise. A more direct check is to compute the per-token reprojection error of the RANSAC homography and verify that the tokens pruned as redundant are precisely the aligned ones, rather than a sizeable fraction of misaligned pixels.","tokens_in":13672,"feed_emoji":"🎥","tokens_out":18087,"duration_ms":165066,"temperature":0.7,"pith_summary":"EgoPrune is a training-free token-pruning method aimed at a specific problem: videos recorded by a moving agent, where every frame is a different viewpoint of the same environment. The paper claims that this egomotion redundancy can be removed with the camera's own geometry—estimate a homography between consecutive frames, warp one onto the other, and drop tokens whose aligned counterparts exceed a 75% cosine-similarity cutoff—followed by a Maximal Marginal Relevance selector that keeps tokens both relevant to the question and visually diverse. On the VSI-Bench and UrbanVideo-Bench egomotion benchmarks, retaining only 30–70% of visual tokens, the method reports average accuracy within a small fraction of the full-token baseline (35.16–35.43 vs. 35.45 on VSI-Bench; 45.66–46.25 vs. 46.23 on UrbanVideo-Bench) while lowering FLOPs, peak memory, and end-to-end latency. A deployment on a Jetson Orin NX 16GB edge device with an INT4-quantized 3B model is offered as evidence that the savings survive real embodied constraints. If the claim holds, spatially grounded video reasoning can run on-device for drones and mobile robots, and pruning methods for video no longer have to assume a fixed camera.","feed_headline":"Camera-aware pruning keeps 99% of video-reasoning accuracy","feed_subtitle":"Homography alignment plus relevance-diversity selection cuts FLOPs, memory, and latency on egomotion benchmarks.","key_machinery":"The load-bearing object is the homography warp inside Perspective-Aware Redundancy Filtering (PARF): a $3\\times3$ matrix $\\mathbf{H}$ mapping each frame onto its successor, estimated from ORB matches via RANSAC, which turns viewpoint-shifted frames into a common coordinate system so that the 75% cosine-similarity threshold is a meaningful redundancy test. Because a single homography is exact only for planar scenes or pure rotation, the alignment is approximate in general egomotion, and the method's accuracy rests on that approximation being good enough. The second mechanism is the MMR token selector, which greedily maximizes $\\lambda\\cdot\\mathrm{rel}(v_i)-(1-\\lambda)\\cdot\\max_{j\\in S}\\mathrm{sim}(v_i,v_j)$, where relevance is the cosine similarity to the mean prompt-token vector and diversity is measured against a short window of earlier picks; $\\lambda=0.5$ balances the two. Because neither stage reads attention scores, the pipeline composes with FlashAttention-2 and with quantized edge runtimes, which is what makes the on-device deployment possible.","core_discovery":"The discovery is that egomotion videos carry their redundancy in the camera's perspective, not in fixed pixel positions: after EmbodiedR's overlap-aware keyframe selection leaves roughly 50–60% overlap between consecutive frames, PARF estimates a homography from ORB keypoint matches filtered by FLANN and Lowe's ratio test, with RANSAC rejecting outliers; warping the previous frame onto the current one makes local token pairs comparable, and tokens with cosine similarity above 75% are declared redundant and dropped. The surviving tokens are then chosen by an MMR selector that scores each candidate by cosine similarity to the averaged prompt tokens, minus a penalty for similarity to already-selected tokens (with a window of ten selections keeping the cost at $O(n^2)$ in token count). The paper argues that this two-stage design answers the three failure modes of prior methods: fixed-position cosine similarity breaks under viewpoint shift, attention-based pruning collides with positional bias and FlashAttention, and diversity-only selection ignores the question being asked. On both benchmarks, pruned models at 70%, 50%, and 30% retention match or slightly beat the full-token average, with the clearest gains on perception-heavy tasks such as object appearance, size estimation, and counting.","pith_inferences":["The single-homography model is the fragile point: for egomotion with strong translation through depth-varying scenes (walking past close objects, a drone flying toward buildings), one planar warp cannot align all pixels, so the 75% cutoff will misclassify some informative tokens as redundant; a depth-aware or multi-plane extension would be the natural stress test.","If the finding generalizes, the geometry-first principle—estimate the motion, prune in the aligned frame, then query-select—should transfer to other moving-camera settings such as driving videos, handheld AR capture, and robot manipulation, where the fixed-camera assumption of prior pruning methods fails just as it does in egomotion video.","The observation that pruned models sometimes beat the full-token baseline suggests the MMR stage acts as a soft query-grounded filter; a testable consequence is that adding the same MMR selection to other pruning methods (or to third-person video) would improve their accuracy at low retention independently of the egomotion setting."],"forward_implications":["At 70%, 50%, and 30% token retention on both benchmarks, EgoPrune's average accuracy stays within about a point of the full-token baseline (on VSI-Bench at 50% retention it is 35.43 vs. 35.45), so egomotion video redundancy can be largely removed without any fine-tuning.","FLOPs, peak memory, and end-to-end latency all decrease and scale more smoothly with input length than the DivPrune and PACT baselines, which is the property that makes long egomotion video usable in time-constrained embodied settings.","Ablations show the two stages are complementary: removing PARF hurts geometry-sensitive tasks (absolute distance, size estimation, route planning), while removing MMR hurts object-centric and directional tasks (object appearance, relative direction, cognitive map construction).","Because the method needs no attention scores, it can sit on top of FlashAttention-style kernels, and the Jetson Orin NX deployment with an INT4-quantized VILA-1.5 3B model shows the efficiency gains transfer to an edge runtime rather than existing only on server GPUs.","Accuracy staying flat from 70% down to 30% retention implies the token budget for egomotion reasoning can be chosen by the latency budget, since pruning strength does not trade away task accuracy in the tested range."],"supporting_citations":[{"why":"EmbodiedR supplies the overlap-aware geometric keyframe selector that EgoPrune builds on, leaving the 50–60% inter-frame overlap that PARF then prunes.","marker":"(Zhao et al. 2025b)"},{"why":"RANSAC is the outlier-rejecting estimator used to compute the homography matrix from ORB keypoint matches inside PARF.","marker":"(Cantzler 1981)"},{"why":"MMR is the maximal marginal relevance greedy selection algorithm that the second-stage token selector instantiates.","marker":"(Wu, Wang, and Ma 2023)"},{"why":"CDPruner provides the average-prompt-token representation that EgoPrune uses as the relevance target in the MMR selector.","marker":"(Zhang et al. 2025)"},{"why":"DivPrune is the diversity-based, training-free pruning baseline that EgoPrune must outperform across retention ratios.","marker":"(Alvar et al. 2025)"},{"why":"PACT is the clustering-and-importance pruning baseline that EgoPrune must outperform, and the FlashAttention-compatibility target.","marker":"(Dhouib et al. 2025)"},{"why":"LLaVA-OneVision-7B is the vision-language model on which all benchmark comparisons and ablations are run.","marker":"(Li et al. 2024)"},{"why":"VILA-1.5 3B is the model used for the Jetson Orin NX edge deployment, establishing that the pruning works on quantized on-device runtimes.","marker":"(Lin et al. 2024)"}],"fun_headline_variants":["Perspective-aware pruning cuts compute, keeps accuracy","Homography alignment drops redundant tokens in egomotion video","EgoPrune: train-free token pruning for on-device video reasoning","Camera-aware token pruning for efficient embodied AI"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single homography can align consecutive egomotion frames well enough for the 75% cosine-similarity cutoff to separate genuinely redundant tokens from informative ones; the homography assumption is exact only for planar scenes or purely rotational camera motion, so under real translation and depth variation the warp is wrong for many pixels and the filter can discard tokens the model still needs.","fun_headline_variants_meta":{"raw":{"variants":["Perspective-aware pruning cuts compute, keeps accuracy","Homography alignment drops redundant tokens in egomotion video","EgoPrune: train-free token pruning for on-device video reasoning","Camera-aware token pruning for efficient embodied AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000274,"raw_usage":{"total_tokens":1696,"prompt_tokens":1058,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":674,"completion_tokens_details":{"reasoning_tokens":572}},"tokens_in":674,"tokens_out":638,"duration_ms":6844,"temperature":1.0,"reasoning_tokens":572,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:32:18.506078+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run EgoPrune at 30% token retention on egomotion videos where translation dominates and depth varies sharply—a handheld or drone camera moving straight through a cluttered room—and compare answer accuracy against the full-token baseline; the published gaps at 30% retention are about 0.3–0.6 points, so a drop of more than a point or two would falsify the alignment premise. A more direct check is to compute the per-token reprojection error of the RANSAC homography and verify that the tokens pruned as redundant are precisely the aligned ones, rather than a sizeable fraction of misaligned pixels.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RANSAC is the outlier-rejecting estimator used to compute the homography matrix from ORB keypoint matches inside PARF."},{"cited_title":"R.; Singh, G.; Akbari, M.; and Zhang, Y","cited_arxiv_id":null,"evidence_quote":"DivPrune is the diversity-based, training-free pruning baseline that EgoPrune must outperform across retention ratios."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PACT is the clustering-and-importance pruning baseline that EgoPrune must outperform, and the FlashAttention-compatibility target."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"VILA-1.5 3B is the model used for the Jetson Orin NX edge deployment, establishing that the pruning works on quantized on-device runtimes."}],"review_version":1}