{"id":"3dd6c913-7e32-46ae-b6c3-569e44d56273","arxiv_id":"2607.09080","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GeoTrace cuts video tokens to 10% with ~99.1% performance retention (12.99x fewer TFLOPs on LLaVA-OneVision) via farthest-point skeleton selection plus trajectory-based residual condensation.","lead":"GeoTrace is a new training-free method that shrinks a video's visual tokens to 10% before a video large language model processes them, cutting compute ~13x while keeping ~99% of original performance. It works by picking a small, spread-out set of 'skeleton' tokens and merging the rest along the paths they follow across frames.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Efficiency claim hinges on FLOPs accounting; compression overhead and memory-bound decoding may erase the 12.99x reduction.","rationale":"The reader's weakest-assumption identification matches our own: the headline efficiency number is only meaningful if FLOPs are accounted correctly and if they correspond to practical speedups. Since the full text is unavailable, we cannot resolve whether the paper already addresses this (e.g., by reporting wall-clock timing and including compression overhead). Therefore the appropriate verdict remains UNVERDICTED, consistent with the reader's assessment. If the concrete test were to fail, the verdict would move to CONDITIONAL or REJECT depending on magnitude; but we have no evidence yet that it fails. The concern is technically substantive, not a manufactured objection, and it directly targets the central claim's validity.","tokens_in":1014,"tokens_out":1926,"duration_ms":22320,"concrete_test":"Reproduce the efficiency claim on LLaVA-OneVision using a public FLOPs profiler (e.g., fvcore or calflops) that counts all operations in the instrumented model, including the CFPA/TCRC compression pass. Compare total inference FLOPs with GeoTrace (10% tokens) vs vanilla. Then measure wall-clock end-to-end latency on the same GPU (e.g., A100) with identical batch size and video input, including the compression step. If the total-FLOPs reduction is <5x or the wall-clock speedup is <3x, the 12.99x headline overstates the real gain. Also require reporting the specific benchmark metric and its standard deviation for the 99.1% retention claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result (10% tokens, 12.99x TFLOPs reduction, 99.1% performance retained) rests on a FLOPs-accounting premise that is unverified in the abstract. The 12.99x ratio is plausible only if (a) compute is dominated by attention over visual tokens, (b) the compression pass (CFPA/TCRC) costs negligible FLOPs, and (c) FLOPs translate directly to wall-clock speedup. None of these is secure. CFPA performs farthest-point sampling over per-frame pixel clouds, which is O(T*H*W) in distance computations with a significant constant, and TCRC builds temporal trajectories and performs condensation, adding further cost. If that overhead is netted out, the real reduction shrinks. More importantly, Video LLM inference on GPUs is often memory-bandwidth bound rather than FLOPs bound, especially for long-context decoding; cutting attention FLOPs 12.99x may yield a far smaller latency reduction. The abstract also does not name the evaluation metric, so '99.1% of vanilla performance' is not interpretable (e.g., a 0.9-point drop on Video-MME accuracy is different from a 0.9% relative drop). This is a structural concern about the central efficiency claim, not about performance retention per se.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GeoTrace, a training-free spatiotemporal token compression framework for Video LLMs. The method decomposes visual tokens into 'skeleton tokens' selected by Contextual Farthest-Point Anchoring (CFPA) and 'traceable residual event tokens' produced by Trajectory-Constrained Residual Condensation (TCRC). The abstract claims that on LLaVA-OneVision, retaining only 10% of visual tokens yields a 12.99x TFLOPs reduction while preserving 99.1% of vanilla performance, and that the method generalizes across four Video LLMs and four video understanding benchmarks. This report is based exclusively on the abstract, as the full text was not made available to the referee.","tokens_in":1221,"tokens_out":3372,"duration_ms":36069,"significance":"If the headline results are substantiated, GeoTrace would be a practically important contribution: a training-free, architecture-agnostic compression method that reduces Video LLM inference compute by more than an order of magnitude with under 1% measured performance loss would have broad applicability. The abstract's positive features include a clear problem motivation, a method that appears to combine geometric sampling with trajectory constraints, and a commitment to release code. However, the significance currently rests on a small number of point estimates with no named metric, benchmark, variance, baseline, or efficiency-accounting details. The central claims are plausible but not verifiable from the abstract alone.","major_comments":[{"comment":"The statement 'preserving 99.1% of the vanilla performance' is not interpretable without naming the evaluation metric and benchmark. Is 99.1% a ratio of scores, a relative retention percentage, or an average across tasks? A 0.9% relative drop on an accuracy metric is different from a 0.9-point absolute drop. Additionally, no variance or number of runs is reported. As this is the central accuracy-preservation claim, the metric and per-benchmark breakdown must be specified, along with error bars or at least run-to-run variability.","section":"Abstract, headline accuracy claim"},{"comment":"The 12.99x TFLOPs reduction is the headline efficiency result. The text does not state whether the CFPA pass (farthest-point sampling over per-frame pixel clouds) and the TCRC pass (temporal trajectory construction and condensation) are included in the FLOPs accounting. If the compression overhead is excluded, the net reduction is overstated. Moreover, TFLOPs reductions do not necessarily translate to wall-clock speedups on memory-bandwidth-bound video-LLM inference. The authors should report measured throughput/latency on the same hardware and state clearly whether the FLOPs figure is a theoretical attention-only calculation or an end-to-end accounting.","section":"Abstract, efficiency claim"},{"comment":"The abstract states that GeoTrace is evaluated on 'four Video LLMs across four video understanding benchmarks,' but names neither the models nor the benchmarks. Without this information, the 'generalization across different model architectures and scenarios' claim cannot be assessed. The main text must include a table of per-model, per-benchmark results, and the abstract should at least identify the models and benchmarks or clearly point to such a table.","section":"Abstract, generalization claim"},{"comment":"No baseline method or ablation is visible in the abstract. The contribution of CFPA and TCRC relative to simpler alternatives is unclear. For instance, random or uniform token subsampling at the same 10% retention might already capture part of the efficiency gain, and existing token merging or pruning methods should be compared. The authors need to provide a comparison against such baselines and ablations that remove or replace each component to support the claim that GeoTrace is effective beyond token subsampling.","section":"Abstract, baselines and ablations"}],"minor_comments":[{"comment":"The term 'exact skeleton tokens' is ambiguous. 'Exact' suggests a formal guarantee or deterministic definition that is not explained in the abstract; consider using 'geometrically selected' or defining what exactness means.","section":"Abstract, terminology"},{"comment":"The phrase 'traceable residual event tokens' uses 'traceable' without a definition. It would help to state what traceability means and why it reduces ambiguity.","section":"Abstract, terminology"},{"comment":"The phrases 'one-to-one temporal trajectories' and 'near-manifold condensation' are underspecified. The full text should give precise definitions and pseudocode, but the abstract should at least give a one-sentence operational description.","section":"Abstract, method description"},{"comment":"The code link is provided, but no license, dependency list, or reproducibility instructions are mentioned. If the repository is public, ensure it includes version-stamped model configurations and evaluation scripts.","section":"Software, code availability"}],"recommendation":"uncertain","confidential_remarks":"The reviewer received only the abstract, not the full text. Based on the abstract alone, I cannot make a confident decision: the headline numbers are plausible but lack the specificity needed to verify the central claims. I recommend an editorial decision to request the full manuscript and then evaluate whether the missing details (metric definition, variance, FLOPs accounting, baselines, ablations) are adequately addressed. If the full text does not provide these, the paper would require major revision before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — quick take on GeoTrace. The abstract describes a training-free spatiotemporal compression scheme that splits video tokens into skeleton tokens kept via farthest-point-style anchoring (CFPA) and residual event tokens folded along temporal trajectories with a near-manifold condensation step (TCRC). The central figure is 10% tokens retained on LLaVA-OneVision, a 12.99x TFLOPs reduction, and 99.1% of vanilla performance kept, claimed to generalize across four Video LLMs and four benchmarks. If that holds up to a full experimental read, it's a genuinely useful efficiency result for the video-LLM crowd — not a new paradigm, but a practical lever that could make long-video inference cheaper.\n\nWhat's new here is the two-part decomposition and the explicit attempt to avoid the usual frame-wise saliency or merging artifacts. The framing is sensible, the method is training-free, and they promise code. I believe the paper is an honest empirical heuristic evaluated on external benchmarks; nothing about the construction forces the result. The building blocks aren't exotic — farthest-point sampling and trajectory-based condensation are known moves — but the particular combination and the across-model evaluation are worth credit.\n\nThe soft spots are mostly about what the abstract doesn't say. The '99.1%' is not interpretable without the metric named; is it a 0.9-point drop on Video-MME accuracy or a relative 0.9% change? And the 12.99x TFLOPs reduction is only meaningful if the compression pass is included in the ledger and if the model is actually FLOPs-bound. Video-LLM decoding can be memory-bandwidth bound, so the real wall-clock gain could be smaller. That doesn't make the claim wrong; it makes the abstract incomplete. I'd want the full paper to show the accounting, plus error bars or at least multiple runs.\n\nI also want to see ablations for the two components — how much CFPA alone vs TCRC alone buys — and whether the token budget or condensation tolerance was tuned per benchmark. That's the standard worry for this genre, not a fatal one.\n\nBottom line: this is a paper with a real potential contribution. It deserves a serious referee, but the referee should push on the efficiency accounting and the metric definition. I'd bring it to a reading group if the full text were available; from the abstract alone, maybe.","headline":"A plausible training-free token-compression scheme for Video LLMs with a headline number that the abstract alone can't validate; worth a serious look at the full experiments.","tokens_in":1852,"tokens_out":1799,"would_cite":true,"duration_ms":19694,"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":"GeoTrace cuts video-LLM compute 12.99x while preserving 99.1% of performance.","keywords":["video LLM","token compression","visual tokens","efficiency","farthest-point sampling","trajectory constraints","skeleton tokens","without training"],"falsifier":"Time end-to-end inference on the same GPU for LLaVA-OneVision with and without GeoTrace at 10% token retention; if the wall-clock speedup is far below the 12.99x FLOPs reduction—say, under 2x—then the headline efficiency number does not translate to practical savings.","tokens_in":789,"feed_emoji":"⚡","tokens_out":3568,"duration_ms":32578,"temperature":0.7,"pith_summary":"GeoTrace is a training-free compression method that reduces the visual token count in Video Large Language Models by roughly tenfold—from full video evidence to a sparse skeleton plus condensed residual events. On LLaVA-OneVision, keeping only 10% of tokens delivers a 12.99x reduction in TFLOPs while preserving 99.1% of the uncompressed model's benchmark performance. The same method holds across four Video LLMs and four video-understanding benchmarks, suggesting it is a general lever for making video reasoning cheaper. If this holds, video LLM inference could become an order of magnitude more affordable without retraining, opening room for longer videos or higher-resolution inputs on the same hardware.","feed_headline":"Video LLMs run 12.99x cheaper with 90% fewer tokens","feed_subtitle":"GeoTrace keeps 99.1% of video understanding using only 10% of visual tokens, no retraining.","key_machinery":"CFPA (Contextual Farthest-Point Anchoring) selects skeleton tokens as a diverse, context-aware subset via farthest-point sampling in an embedding space. TCRC (Trajectory-Constrained Residual Condensation) compresses the remaining residual tokens along one-to-one temporal trajectories and constrains the condensed features to lie near a data manifold. Together they yield a compact, traceable token representation that replaces frame-wise saliency heuristics and ambiguous feature merging.","core_discovery":"The paper's central claim is that video evidence can be decomposed into two complementary token types: 'exact skeleton tokens' that are salient, context-consistent, and high-coverage, selected by Contextual Farthest-Point Anchoring (CFPA), and 'traceable residual event tokens' produced by Trajectory-Constrained Residual Condensation (TCRC), which follows one-to-one temporal trajectories and condenses residual features near a manifold to avoid ambiguity. With only 10% of visual tokens retained, GeoTrace achieves a 12.99x TFLOPs reduction on LLaVA-OneVision while preserving 99.1% of vanilla performance, and the approach generalizes across four Video LLMs and four benchmarks.","pith_inferences":["The abstract's 'exact' skeleton tokens likely mean geometrically well-spread rather than literally unique; the reported 99.1% preservation is the measurable guarantee.","The 12.99x figure is a FLOPs reduction, not a promise of wall-clock speedup; in memory-bound decoding the end-to-end gain may be smaller, and a fair comparison should include the cost of the compression pass itself.","The trajectory-constrained condensation idea could transfer to other temporally ordered data, such as audio or point-cloud sequences, wherever a token budget limit binds.","A useful stress test: push the token retention below 10% on long-video benchmarks to find where the skeleton/residual decomposition begins to lose coverage."],"forward_implications":["Video LLM inference can be made over an order of magnitude cheaper in FLOPs with less than 1% performance loss, purely by input token selection, no training required.","Because the method is architecture-agnostic, any current or future Video LLM can adopt GeoTrace as a drop-in preprocessing step.","With the token budget cut to 10%, the freed compute can be spent on longer videos, more frames, or higher resolution inputs, extending the practical range of video understanding.","The traceable event-token decomposition may give partial interpretability: residual tokens correspond to specific trajectories, so attention can be traced to concrete motion evidence."],"fun_headline_variants":["GeoTrace: 90% fewer video tokens, 99.1% accuracy retained","Training-free GeoTrace cuts video tokens 90%, keeps 99.1% perf","Video LLMs: 12.99x compute reduction with GeoTrace's 10% tokens","GeoTrace: 90% less tokens, 99.1% video understanding, no training","GeoTrace slashes video LLM tokens by 90% while preserving 99.1%"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The efficiency claim assumes that the attention computation over tokens dominates total model cost and that the CFPA/TCRC compression pass is cheap enough to be ignored in the FLOPs ledger; if compression overhead is significant or decoding is memory-bound, the real speedup will be smaller than 12.99x.","fun_headline_variants_meta":{"raw":{"variants":["GeoTrace: 90% fewer video tokens, 99.1% accuracy retained","Training-free GeoTrace cuts video tokens 90%, keeps 99.1% perf","Video LLMs: 12.99x compute reduction with GeoTrace's 10% tokens","GeoTrace: 90% less tokens, 99.1% video understanding, no training","GeoTrace slashes video LLM tokens by 90% while preserving 99.1%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000908,"raw_usage":{"total_tokens":3774,"prompt_tokens":811,"completion_tokens":2963,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":2841}},"tokens_in":555,"tokens_out":2963,"duration_ms":21352,"temperature":1.0,"reasoning_tokens":2841,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T07:41:56.174937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Time end-to-end inference on the same GPU for LLaVA-OneVision with and without GeoTrace at 10% token retention; if the wall-clock speedup is far below the 12.99x FLOPs reduction—say, under 2x—then the headline efficiency number does not translate to practical savings.","supporting_citations":[],"review_version":2}