{"id":"dd2c79b5-ace3-4f88-ab25-486c1c7591c8","arxiv_id":"2501.02268","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"G-Prune prunes visual tokens for MLLMs via graph-based information propagation, cutting LLaVA-NeXT FLOPs by about 63% with small accuracy loss.","lead":"This paper introduces G-Prune, a way to speed up multimodal AI models by removing redundant image tokens before they enter the language model. It builds a similarity graph over image patches, lets importance scores flow across the graph, and keeps the most representative patches, which can be in the foreground or background.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 5 degree normalization can erase the score signal inside connected components, so G-Prune's claimed 'one representative per object' mechanism is not established and may rest on early stopping.","rationale":"The reader's weakest assumption correctly focuses on the l2-norm proxy and degree normalization, but I find a more specific and more damaging problem: degree normalization is not merely an uncontrolled heuristic, it is the inverse of the stationary distribution of the row-stochastic propagation in the uniform-weight limit. Thus within a large, tightly connected object, the normalized score becomes flat and the top-k choice is arbitrary. The paper's own ablation shows that the graph component, not the l2-norm initialization, carries the performance, so the concern targets the actual source of the method's success. I do not recommend changing the verdict because the empirical claim -- that G-Prune reduces LLM FLOPs with modest accuracy loss on LLaVA-NeXT benchmarks -- is plausible and is supported by the reported comparisons and ablations. The concern is about the mechanism and the generality of the 'foreground and background are both critical' conclusion, not about the headline benchmark numbers. Hence the reader's CONDITIONAL verdict stands unchanged, with the added condition that the graph-normalization mechanism be either justified analytically or probed with the proposed component-level diagnostic.","tokens_in":16000,"tokens_out":11694,"duration_ms":125096,"concrete_test":"Run Algorithm 1 with s=0.5, t=5 on the actual 2880 LLaVA-NeXT visual tokens from a sample of TextVQA images; compute the connected components of the thresholded adjacency and the within-component coefficient of variation (CV) of S'_i. Then rerun selection with t=50 (near convergence) and, separately, with degree normalization removed. If the largest components have near-zero CV at t=5, or if t=50 retains the reported accuracy, then the normalized propagation is not selecting object representatives and the paper's explanation is unsupported; if CV is substantial and t=50 collapses accuracy, the early-stopped diffusion is the real mechanism and needs explicit justification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central narrative is that Eq. 3 propagates information and Eq. 5 normalizes by degree to reveal the most representative token of each object. This is mathematically suspect. A' in Eq. 3 is row-stochastic, so in a connected component that has converged, S(t) approaches the stationary distribution of the random walk. For the intended 'same object' case of near-uniform connection weights, that stationary distribution is proportional to degree D_i, making S'(t)_i = S(t)_i / D_i constant within the component. Top-k selection inside that object is then arbitrary, and the claimed mechanism cannot be what produces the reported gains. The method only avoids this collapse by stopping at t=5, before convergence, but the paper does not justify why this particular early-stopped diffusion should identify object representatives. The ablation in Table 2 sharpens the issue: graph-only propagation (initial scores all 1) already gives TextVQA 64.01 versus 64.05 for the full method, so the l2-norm initialization is not the driver; the unexplained graph/normalization heuristic is. This is load-bearing because the novelty of G-Prune rests on the graph-selection rule, not merely on the empirical numbers.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes G-Prune, a training-free visual token pruning method for multimodal large language models (MLLMs). G-Prune treats visual tokens as nodes in a graph, connects nodes by thresholded cosine similarity, propagates importance scores over the graph via a row-stochastic transition matrix, and retains the top-k tokens by a degree-normalized score. The method is evaluated on LLaVA-NeXT-8B across eight benchmarks, reporting about 63.5% FLOPs reduction with small accuracy drops on general VQA and text-oriented VQA, and favorable comparisons against Random pruning, ToMe, and FastV. Ablations show that the graph-based propagation is the main contributor, with the l2-norm initialization providing only a small additional gain. The paper releases code at a public repository.","tokens_in":16224,"tokens_out":11373,"duration_ms":107178,"significance":"If the empirical results hold, G-Prune is a practical, training-free acceleration tool for MLLMs: it is plug-and-play, does not require intervention in the LLM reasoning process, and retains strong accuracy even at high pruning ratios, especially on text-oriented benchmarks where existing baselines degrade sharply. The release of code and the breadth of benchmarks (general VQA, MLLM benchmarks, and text-oriented VQA) are strengths, and the reported FLOPs savings are consistent across tables. However, the conceptual contribution is currently undermined by an unverified mechanism: the paper claims that the graph propagation and degree normalization identify 'the most representative token of each object,' yet this claim is not supported by the mathematics of the proposed update, and the algorithm as stated has several inconsistencies. The paper is best viewed as an empirical heuristic paper whose central narrative needs substantial correction.","major_comments":[{"comment":"The claim that degree normalization in Eq. (5) reveals 'the most representative token of each object' is not supported by the mathematics of the proposed propagation. For the row-softmax matrix A' defined after Eq. (1), the stationary distribution pi of the Markov chain in Eq. (3) satisfies pi_i proportional to the row sum R_i = sum_j exp(A_ij) (by detailed balance, since A is symmetric). In an object that is a clique with near-uniform edge weight w, R_i is identical for all nodes in the object, so S'(t)_i = S(t)_i / D_i is constant within that component and top-k selection inside the object is arbitrary. In an object that is not a clique, R_i = N - 1 + (e^w - 1) D_i, so S'(t)_i is approximately (N-1)/D_i + (e^w - 1), which assigns the highest score to the lowest-degree (most peripheral) token, not to a representative or central token. The paper does not explain why stopping at t=5 avoids this behavior; Table 3 shows that t=1 already gives essentially the same accuracy (71.94 vs. 71.98 at t=5), so the result is not tied to convergence to a representative stationary distribution. The authors should either provide a correct characterization of what Eq. (5) selects or explicitly present the method as an empirically motivated heuristic without the representativeness claim.","section":"Method, Eq. (5)"},{"comment":"The degree D_i defined in Eq. (4) can be zero for any token whose cosine similarities to all other tokens are below the threshold s. In that case, Eq. (5) divides by zero and the normalized score S'(t)_i is undefined. The paper does not specify how isolated nodes are handled (e.g., setting D_i to a small constant, adding self-loops, or excluding such tokens). With the default s=0.5 and 2880 high-dimensional visual tokens, isolated nodes are a realistic possibility, yet the reported results are clean, suggesting the implementation silently applies a rule that is not described. This is a correctness gap in the algorithm as stated and should be addressed explicitly.","section":"Algorithm 1, Eqs. (4)-(5)"},{"comment":"The algorithm selects the global top-k tokens by S'(t)_i via Eq. (6), but the text immediately after states that 'only the most representative token is retained to represent each object.' This description is only accurate if k equals the number of connected components. In general, if k is smaller than the number of objects, some objects receive no token; if k is larger, multiple tokens from a single object are retained. The visualizations in Fig. 6 also show multiple tokens kept in each textured region at various pruning ratios. The paper should either change the selection rule to enforce one token per connected component (which would require knowing the component decomposition) or correct the narrative to describe the actual global top-k behavior.","section":"Method, Eq. (6) and following text"},{"comment":"The hyperparameters s=0.5 and t=5 appear to be chosen as the best configuration from the grid in Table 3, evaluated on the same benchmarks (GQA, POPE, TextVQA) used later for the final reported results. This is a form of selection on the test benchmarks; the paper should state whether a separate validation split was used or report the full grid of results so the reader can assess the sensitivity. In addition, all benchmark numbers are single-run without error bars or a statement of determinism. For claims of small accuracy differences, such as the 0.95% drop on VQA2.0 at 70% pruning, the absence of variance information makes it difficult to judge significance.","section":"Experiments, Table 3 and Implementation Details"},{"comment":"The analysis leading to the conclusion that 'both foreground and background tokens are critical' and that l2-Norm distributions of foreground and background overlap is based on segmenting the image into foreground and background areas, but the segmentation method is not specified anywhere in the paper. Without a description of how the segmentation was obtained (e.g., object detector, saliency map, manual annotation), the observation in Fig. 1(b) is not reproducible, and it is the paper's first stated contribution. The authors should specify the segmentation procedure or weaken the claim.","section":"Introduction, Fig. 1(b)"}],"minor_comments":[{"comment":"The text reads 'we evaluate OncePrue for text-oriented VQA benchmarks'; 'OncePrue' should be 'G-Prune'.","section":"Datasets and Metrics"},{"comment":"The abstract uses 'front' instead of 'foreground' in the sentence about retaining tokens from front or background.","section":"Abstract"},{"comment":"The paper refers to 'ChartVQA' but the benchmark is consistently named ChartQA elsewhere; please unify.","section":"Experiments, comparison with ToMe"},{"comment":"The sentence 'the difference between the highest and the lowest is only 0.32%' does not match Table 3, where the average ranges from 71.75 (t=50) to 71.98 (t=5), a difference of 0.23 points. Please correct the number.","section":"Experiments, Table 3 discussion"},{"comment":"The reported drop of '40.35%' for ToMe on TextVQA at 90% pruning differs from Table 1 (65.41 to 38.36, which is about 41.4% relative drop); please reconcile the percentage.","section":"Experiments, comparison with ToMe"},{"comment":"The model name is typeset inconsistently as 'LLaVA-NeXT', 'LLaV A-NeXT', and 'LLaV A-NeXT'; a single consistent form should be used.","section":"Throughout"},{"comment":"Eq. (6) uses lowercase n for the number of tokens while the rest of the paper uses uppercase N; please unify the notation.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an empirical efficiency paper with a simple and potentially useful idea, and the benchmark results are consistently in its favor. However, the paper currently overclaims the mechanism: the specific stationary-distribution argument in the stress-test note is not exactly correct (the stationary distribution of a row-softmax chain is proportional to the row sum of exp(A), not to the degree), but the corrected analysis is even more problematic for the 'representative token' story, since it yields inverse-degree selection in non-clique objects. The authors can address this either by deriving what the normalized score actually selects or by repositioning the method as a heuristic without representativeness claims. I also note the lack of comparison with other graph-based token pruning methods and the missing segmentation details in Fig. 1(b); both are within the scope of a major revision. The novelty is moderate, but the practical contribution and code release are valuable enough to warrant a revise-and-resubmit rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: G-Prune is a training-free token pruning heuristic that gets good numbers on LLaVA-NeXT, especially at high pruning ratios, and the paper is worth a referee. But the explanation of why it works—degree normalization revealing one representative token per object—doesn't survive contact with the math.\n\nWhat's new: The combination of cosine-similarity graph construction, l2-norm initialization, iterative propagation, and degree normalization applied to MLLM visual tokens is not in the cited prior work. The paper also makes a reasonable empirical case that both foreground and background tokens matter, and the benchmark coverage is broad. The code is public.\n\nWhere it bends: The stress-test note is right. Equation 3 row-normalizes the adjacency, so as t grows S(t) approaches the stationary distribution, which for a symmetric weight matrix is proportional to weighted degree. Dividing by binary degree (Eq. 5) then makes S' approximately constant inside a connected component with homogeneous edge weights. So the \"most representative token of each object\" story is not what the selection is doing; the method only avoids collapse by stopping at t=5, and the paper never justifies that early stopping. The ablation in Table 2 makes this worse: graph-only propagation with uniform initialization already gets TextVQA 64.01 vs 64.05 for the full method, so l2-norm initialization is not the driver either. The gains come from the graph/normalization heuristic, but not in the way the paper claims.\n\nOther soft spots: hyperparameters s and t are tuned on the same benchmarks used for the headline results (Table 3), results are single-run with no error bars, and the segmentation method behind Figure 1 is unspecified. These are fixable.\n\nWho this is for: people working on MLLM inference efficiency. It's a useful empirical baseline, not a conceptual breakthrough. The reviewer should focus on whether the early-stopped diffusion is a principled selection rule or just a heuristic that happens to work—right now it's the latter. I'd encourage you to engage with it; the empirical results are solid enough to deserve a serious referee, and the mechanism question is worth resolving.","headline":"G-Prune is a plausible training-free token pruning heuristic with solid benchmark results, but its 'one representative per object' explanation is mathematically shaky and the real mechanism is an unexplained early-stopped diffusion.","tokens_in":16762,"tokens_out":3202,"would_cite":false,"duration_ms":30828,"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":"Visual tokens can be pruned by graph-based information propagation, keeping representative tokens from foreground and background, so that LLaVA-NeXT runs at about 63.5% fewer FLOPs with accuracy drops of 0.95% and 2.34% on VQA2.0 and…","keywords":["visual token pruning","multimodal large language models","training-free acceleration","graph information propagation","foreground and background tokens","LLaVA-NeXT","token redundancy","vision-language benchmarks"],"falsifier":"Replace the $\\ell^2$-norm initialization in Eq. 2 with uniform scores, keeping the same graph and iterations; if TextVQA accuracy at 90% pruning stays within about one point of G-Prune's 59.31, the norm prior is not doing the work. Similarly, a variant that clusters tokens and keeps one per cluster at controlled cluster sizes would test whether degree normalization is really compensating for component-size bias.","tokens_in":3208,"feed_emoji":"✂️","tokens_out":4848,"duration_ms":108989,"temperature":0.7,"pith_summary":"This paper argues that the large stacks of visual tokens fed to multimodal large language models (MLLMs) are highly redundant, and that a training-free pruning method can remove most of them without destroying performance. Its central finding is that both foreground and background tokens matter, so pruning must keep representative tokens from every semantic region rather than only the salient objects. On this basis the paper proposes G-Prune, which builds a similarity graph over tokens, propagates importance scores through the graph, and keeps the top-scoring normalized tokens. Applied to LLaVA-NeXT, G-Prune reports about 63.5% fewer FLOPs at 70% pruning with a 0.95% accuracy drop on VQA2.0 and a 2.34% drop on TextVQA.","feed_headline":"Graph-based token pruning cuts MLLM compute 63% with small loss","feed_subtitle":"Keeps foreground and background tokens; VQA2.0 drops 0.95%, TextVQA 2.34%, at 63.57% fewer FLOPs.","key_machinery":"The central object is a token-similarity graph with an information-propagation score. Given $N$ visual tokens, edges are set by cosine similarity above a threshold $s$; after softmax row-normalization, scores begin as each token's $\\ell^2$-norm and evolve by repeated multiplication with the adjacency, so information flows from high-norm tokens to their semantic neighbors. Degree normalization divides each final score by the number of connected neighbors, making scores comparable across objects and regions of different sizes, and the top-$k$ tokens by normalized score are retained. This machinery does the work of turning local token statistics into a global representativeness ranking that is not biased toward the foreground.","core_discovery":"The paper's central claim is that a simple graph-diffusion score can identify the visual tokens an MLLM actually needs, and that these tokens are spread across both foreground and background. G-Prune treats each visual token as a graph node, connects nodes whose cosine similarity exceeds a threshold, and iterates the update $S_t = S_0 (A')^t$, where $S_0$ is each token's $\\ell^2$-norm and $A'$ is a row-normalized adjacency matrix; dividing the final score by node degree removes the bias of larger connected components. Keeping the top-$k$ normalized scores retains one or a few representative tokens per object or background region. The experiments show that random pruning collapses text-oriented VQA accuracy at 90% pruning, while G-Prune keeps TextVQA at 59.31 against ToMe's 39.02, supporting the claim that graph-selected tokens preserve fine-grained information.","pith_inferences":["The propagation step resembles a random-walk centrality measure on the similarity graph; if that reading is right, the method could be replaced by a one-step spectral or PageRank-like score with similar behavior and less iteration.","Since the ablation shows graph structure alone accounts for most of the gain, a testable extension is to replace the l2-norm prior with other saliency cues, such as CLS-attention or text-prompt similarity, and check whether the gains stack.","The degree-normalized top-k selection effectively picks one representative per clustered region; a natural extension is to make k adaptive per image or per region based on entropy or task type.","The method is demonstrated on a single MLLM; if the graph-score mechanism is truly model-agnostic, it should transfer to other high-resolution MLLMs, which is a direct test a reader could run."],"forward_implications":["At 70% pruning, LLaVA-NeXT runs at about 6.76 TFLOPs instead of 18.52 TFLOPs, a 63.57% reduction, with VQA2.0 dropping from 82.70 to 81.91 and TextVQA from 65.41 to 63.87.","At 90% pruning, G-Prune keeps text-oriented benchmarks usable: TextVQA at 59.31 and DocVQA at 48.94, compared with 38.36 for ToMe and 53.53 for FastV on TextVQA.","Because the method is training-free and does not modify the MLLM's weights or attention, it can be dropped into an existing model at inference time.","The graph score selects tokens from background and textured regions as well as foreground objects, which the paper argues is necessary for fine-grained tasks such as reading text.","The method keeps performance close to baseline across general VQA, MLLM-specific benchmarks, and text-oriented VQA at moderate pruning ratios."],"supporting_citations":[{"why":"Provides the baseline MLLM, its 2880 visual tokens, and the 18.52 TFLOPs reference used for all pruning comparisons.","marker":"Liu et al. 2024a"},{"why":"Token merging baseline; the high-pruning-ratio comparison shows G-Prune's margin, such as TextVQA 59.31 vs 39.02 at 90% pruning.","marker":"Bolya et al. 2022"},{"why":"Attention-based pruning baseline compared in Table 1 and Figure 3.","marker":"Chen et al. 2024a"},{"why":"Supplies the norm-based token scoring idea behind the l2-norm initialization in Eq. 2 and its ablation.","marker":"Lu and Zhang 2022"},{"why":"Representative foreground-oriented token pruning work used to motivate why background tokens also matter for MLLMs.","marker":"Rao et al. 2021"},{"why":"VQA2.0 benchmark; source of the headline 0.95% accuracy-drop figure.","marker":"Goyal et al. 2017"},{"why":"TextVQA benchmark; the fine-grained task where G-Prune shows its largest advantage at high pruning ratios.","marker":"Singh et al. 2019"}],"fun_headline_variants":["Graph diffusion picks the right visual tokens for MLLMs","G-Prune: 63% fewer FLOPs with graph-based token selection","Visual tokens: both foreground and background matter - G-Prune","Training-free token pruning keeps MLLM accuracy via graphs","G-Prune trims 63.57% MLLM compute, drops ≤2.34% accuracy"],"cache_read_input_tokens":18944,"weakest_assumption_plain":"The load-bearing premise is that a token's $\\ell^2$-norm measures how much information it carries, and that dividing scores by node degree makes them comparable across regions of different sizes; if either assumption fails, the top-k selection is arbitrary.","fun_headline_variants_meta":{"raw":{"variants":["Graph diffusion picks the right visual tokens for MLLMs","G-Prune: 63% fewer FLOPs with graph-based token selection","Visual tokens: both foreground and background matter - G-Prune","Training-free token pruning keeps MLLM accuracy via graphs","G-Prune trims 63.57% MLLM compute, drops ≤2.34% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1352,"prompt_tokens":994,"completion_tokens":358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":260}},"tokens_in":610,"tokens_out":358,"duration_ms":3941,"temperature":1.0,"reasoning_tokens":260,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:13:51.404982+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the $\\ell^2$-norm initialization in Eq. 2 with uniform scores, keeping the same graph and iterations; if TextVQA accuracy at 90% pruning stays within about one point of G-Prune's 59.31, the norm prior is not doing the work. Similarly, a variant that clusters tokens and keeps one per cluster at controlled cluster sizes would test whether degree normalization is really compensating for component-size bias.","supporting_citations":[],"review_version":1}