{"id":"bd4a154c-5796-4d50-b621-0d14efde9b13","arxiv_id":"2608.07088","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"RoRA allocates a fixed visual-token budget into semantic core, context, and detail roles guided by attention-anchored regions, improving pruned MLLM accuracy and speed without training.","lead":"This paper introduces RoRA, a training-free method for pruning visual tokens in multimodal LLMs by assigning retained tokens to three roles: a protected semantic core, complementary context, and fine-grained detail. It reports consistent accuracy gains over prior pruning baselines and lower selection overhead across LLaVA and Qwen-VL models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim that role-oriented allocation drives the gains is not cleanly isolated from the new object prior; ablations compare against FastV, not against D2Pruner without role allocation.","rationale":"The paper presents a well-motivated, training-free method with code and consistent small-to-moderate gains across models. The strongest claim, however, is that role-oriented allocation causes the improvement. The ablations isolate the core, context, and detail components relative to FastV, but they do not isolate the object prior from the role allocation, nor do they compare against D2Pruner with the role stages removed. Since D2Pruner is the strongest baseline and already performs debiased attention plus diversity, the marginal gain of RoRA could come from the object prior alone. The Qwen3-VL results are the largest wins and have no ablation at all. The reader's GQA-calibration concern is legitimate but is partially mitigated by the Qwen benchmarks excluding GQA; the missing ablation is therefore more load-bearing. Given the absent appendix with hyperparameters, the result is best treated as CONDITIONAL until the ablation and reproducibility details are provided. No ad hominem; this is an internal-validity gap.","tokens_in":15505,"tokens_out":12700,"duration_ms":103758,"concrete_test":"Run the official code to ablate the object prior by setting lambda_obj=0 and to ablate the role allocation by setting Kc=Kd=0 (core-only) on Qwen3-VL-8B at 75% and 90% pruning, then compare the normalized averages in Table 2 against D2Pruner. If the core-only or no-object-prior variant drops from 98.0%/87.9% toward D2Pruner's 92.6%/82.6%, the role-allocation claim is unsupported; if the gap persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that role-oriented regional allocation (protected core, complementary context, AAR-guided detail) is what enables RoRA's improvements over training-free baselines. The method differs from D2Pruner in two ways: it adds a prompt-calibrated object prior (lambda_obj in Eq. 3) and it replaces D2Pruner's dense diversity graph with AAR-guided context/detail allocation (Eqs. 5-10). The ablations in Figure 3 compare FastV, Core, and Core+Context on RefCOCO, and with/without detail on TextVQA/VizWiz. They show the core alone accounts for almost all of the gain (RefCOCO normalized average from 35.44% to 74.44% at K=144), while adding AAR context improves by only about 0.6 points and detail by about 0.3 points. No ablation varies lambda_obj, and no ablation compares against D2Pruner with and without the role allocation stages. Therefore the large Qwen3-VL gain over D2Pruner (5.4 and 5.3 normalized points at 75% and 90% pruning in Table 2) could be driven by the object prior rather than by role-oriented allocation. Since the paper's claim is causal, this missing isolation is the most load-bearing gap. The reader's GQA-calibration concern is real but less decisive because the Qwen benchmarks do not include GQA.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RoRA, a training-free visual token pruning method for multimodal LLMs. The method operates at an early LLM layer, debiases text-conditioned attention with a positional prior and a prompt-calibrated object prior, selects a protected semantic core, then builds Attention-Anchored Regions (AARs) from high-confidence anchors. The remaining budget is allocated to complementary context outside AARs and to fine-grained detail inside AARs, with pairwise similarity used only as a redundancy filter in the context stage. Experiments are reported on LLaVA-1.5, LLaVA-NeXT, Qwen2.5-VL, and Qwen3-VL across many benchmarks, with normalized averages under matched token budgets. The paper claims consistent gains over prior training-free baselines, including 96.5% of full LLaVA-1.5 performance at 88.9% pruning, a 5.4-point normalized average gain over D2Pruner on Qwen3-VL at 75% pruning, and a selector overhead of 0.7 ms at 66.7% pruning on an H800.","tokens_in":15799,"tokens_out":4777,"duration_ms":42877,"significance":"If the central claim is supported, RoRA is a practically useful contribution: it is training-free, has low selection overhead, and its component-wise ablations show that each of the three roles contributes. The paper also evaluates on four model families spanning different visual encoders and token budgets, which is a broad and appropriate experimental scope. The main weakness is that the causal claim 'role-oriented regional allocation drives the gains' is not cleanly isolated from the new object prior, and the prior calibration source overlaps with an evaluation benchmark. These issues are fixable with additional ablations and validation details, so the work is potentially publishable, but the current evidence is not fully conclusive for the headline causal interpretation.","major_comments":[{"comment":"The central claim that role-oriented regional allocation is what enables RoRA's gains is not isolated from the prompt-calibrated object prior. RoRA differs from D2Pruner in two ways: it adds the λ_obj o_i term in Eq. (3) and it replaces D2Pruner's dense diversity graph with AAR-guided context/detail allocation in Eqs. (5)-(10). The ablations in Fig. 3 vary core/context/detail composition but never vary λ_obj, and no experiment compares D2Pruner with and without the object prior. Thus the 5.4- and 5.3-point gains over D2Pruner in Table 2 could be driven entirely by the object prior rather than by role allocation. Please add an ablation with λ_obj = 0 and a D2Pruner-plus-object-prior variant, and report the effect of λ_obj on each model family.","section":"Methods, Eq. (3), and Fig. 3"},{"comment":"The positional prior b_i and object prior o_i are estimated from 1,000 images sampled from the GQA training split, and GQA is one of the nine evaluation benchmarks in Table 1. Because the priors are fixed spatial distributions derived from the same image distribution as the evaluation set, the GQA column (e.g., 57.5 vs. 56.9 at K=64 for LLaVA-1.5) and the overall average may be inflated by construction. The Qwen experiments are less affected because GQA is not in Table 2, but the LLaVA tables still support part of the paper's claims. Please validate with priors estimated from an unrelated source (e.g., COCO or a mixed collection) and also report GQA-excluded averages.","section":"Methods, calibration paragraph; Table 1"},{"comment":"The paper states that 'RoRA uses one fixed configuration for each backbone–pruning-ratio pair' and that the configuration is shared across tasks, but it does not disclose how these configurations were selected. The method has many free parameters: λ_obj, λ_ctx, λ_det, ω_a, ω_m, ω_l, τ_sem, Kp/Kc/Kd, anchor count M and radius r_a, and pruning layer ℓ_p. If these were tuned on the evaluation benchmarks, the matched-budget comparisons would be biased in RoRA's favor. Please specify the validation set and selection criterion used to fix these hyperparameters, and include a sensitivity analysis over the most influential parameters, particularly λ_obj and the budget split.","section":"Experiments, Implementation details"}],"minor_comments":[{"comment":"Several normalized averages in Table 2 exceed 100% of the unpruned model (e.g., FastV 101.4%, D2Pruner 105.6%, and RoRA 106.9% on Qwen2.5-VL at 75% pruning). Since these numbers are presented as percentages of full performance, the paper should explain how pruning can improve the normalized average and should include the raw benchmark scores to aid interpretation.","section":"Table 2"},{"comment":"The complexity claim 'RoRA has complexity O(N + N_t^2)' is not fully supported by the algorithm description. The context-stage selection in Eq. (7) compares each candidate only against the currently retained tokens, which gives O(N_t * (Kp + Kc)) in the worst case rather than O(N_t^2). Please either correct the complexity expression or clarify the assumed implementation.","section":"Conclusion, complexity discussion"},{"comment":"The ablation text refers to a 'standard detail configuration' and to 'allocating 18 detail tokens', but the corresponding number of detail tokens for the standard configuration is not defined in the main text or figure caption. Please state the exact Kd values and the hyperparameter settings used for each ablation point.","section":"Figure 3 and ablation text"},{"comment":"The implementation details and the RefCOCO evaluation protocol are deferred to the appendix, but the provided manuscript does not include the appendix. Please ensure the appendix is part of the submission so that the hyperparameter tables and the localization evaluation protocol are available to reviewers.","section":"Appendix availability"}],"recommendation":"major_revision","confidential_remarks":"The manuscript carries a AAAI copyright line at the bottom of the first page. This is unusual for a journal submission and should be checked for prior submission or publication status. The overlap between the GQA-derived priors and the GQA evaluation set is a genuine correctness-risk issue, and I recommend requiring the alternative-calibration-source experiment before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"RoRA is a well-executed token-pruning paper with one genuinely new idea—splitting the retained budget into core, context, and detail roles guided by attention-anchored regions—and a broad, honest evaluation. The code is released, the baselines are matched, and the ablations in Figure 3 show each component gives some improvement. If you work on training-free MLLM compression, this is worth reading.\n\nThe soft spot is that the central causal claim—that role-oriented allocation drives the gains over D2Pruner—is not cleanly isolated. The 'Core' variant in the ablation already contains the prompt-calibrated object prior (lambda_obj in Eq. 3), not just the protected-semantic-core rule. Moving from FastV to Core gives +39 points on RefCOCO at K=144, while adding the AAR-guided context adds ~0.6 and detail ~0.3. So the large Qwen3-VL margins over D2Pruner (5.4 and 5.3 normalized points) could plausibly be driven mostly by the object prior, with role allocation contributing a small increment. No ablation varies lambda_obj, and no comparison runs D2Pruner with vs. without the role-allocation stages. That is the main gap.\n\nThe GQA-calibration concern is real but secondary. Priors learned on GQA training images while GQA is an evaluation benchmark could inflate that column, but the Qwen results don't involve GQA and still show gains. An alternative calibration source would be a cheap robustness check.\n\nMinor items: no error bars or repeated runs; margins on LLaVA-1.5 are within a point or so and could be noise. The per-backbone-per-ratio hyperparameters are deferred to an appendix that isn't included; the paper should disclose how they were chosen to rule out benchmark-specific tuning.\n\nBottom line: a solid engineering paper with a useful method, but the abstraction needs to be disentangled from the object prior before the causal claim is accepted. It deserves a serious referee—send it out, with a request to add the missing ablation and a validation protocol.","headline":"RoRA is a genuine engineering contribution with a clean evaluation, but the paper's causal claim that role-oriented allocation drives its gains is not isolated from the object prior.","tokens_in":16393,"tokens_out":3367,"would_cite":true,"duration_ms":29358,"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":"RoRA claims that visual token pruning should allocate the retained budget across three complementary evidence roles—protected semantic core, complementary context, and fine-grained detail—and that this role-oriented regional allocation…","keywords":["visual token pruning","multimodal large language models","training-free acceleration","attention debiasing","Attention-Anchored Regions","role-aware allocation","KV-cache reduction","LLaVA-1.5"],"falsifier":"Re-fit the two calibration priors on a source with no overlap with the evaluation benchmarks, for example 1,000 unlabeled images from COCO or Flickr, keep every other hyperparameter fixed, and re-run the LLaVA-1.5 and Qwen tables; a meaningful drop in normalized average accuracy on GQA, or in the overall average, would show that the reported gains depend on the particular calibration corpus rather than on generic role allocation.","tokens_in":15277,"feed_emoji":"✂️","tokens_out":8706,"duration_ms":64144,"temperature":0.7,"pith_summary":"This paper argues that visual token pruning in multimodal language models should allocate the retained budget across three complementary evidence roles—a protected semantic core, complementary context, and fine-grained detail—instead of ranking every token with one homogeneous importance score. Its central claim is that role-oriented regional allocation, guided by Attention-Anchored Regions (AARs), retains more of the unpruned model's accuracy than prior training-free pruning methods at the same token budget. Under matched budgets, RoRA consistently leads across the LLaVA and Qwen-VL families, holding 96.5% of full LLaVA-1.5 performance while dropping 88.9% of visual tokens, and improving on the strongest baseline by about five normalized points on Qwen3-VL at 75% pruning. A sympathetic reader would care because the method is training-free, adds only about 0.7 ms of selection cost, and preserves object and detail evidence that attention-only or crop-based schemes miss. The method treats pruning as evidence allocation rather than as a single importance ranking, which is why it remains accurate at very aggressive retention ratios.","feed_headline":"Token pruning by role keeps 96.5% of accuracy at 88.9% pruning","feed_subtitle":"RoRA splits visual tokens into core, context, and detail, beating matched-budget baselines with 0.7 ms overhead","key_machinery":"The load-bearing object is the Attention-Anchored Region (AAR): the union $\\mathcal{R}_{\\mathrm{AAR}} = \\bigcup_{a \\in A} \\mathcal{N}_{r_a}(a)$ of local grid neighborhoods centered at the $M_a$ highest-calibrated-attention tokens. An AAR is not an object mask; it is a spatial proxy for 'object support already covered.' It carries the argument by making coverage explicit, so the residual budget can be spent asymmetrically: context tokens get a boost outside AARs ($c_i = p_i + \\lambda_{\\mathrm{ctx}} \\mathbb{1}[i \\notin \\mathcal{R}_{\\mathrm{AAR}}]$), while detail tokens get a bonus inside AARs ($d_i = \\omega_a \\hat{a}_i + \\omega_m \\hat{m}_i + \\omega_\\ell \\hat{\\ell}_i + \\lambda_{\\mathrm{det}} \\eta_i$). Redundancy filtering then compares candidates only against retained evidence instead of building a full $N \\times N$ similarity matrix, which is why the selection overhead stays near 0.7 ms per image.","core_discovery":"RoRA's central claim is that the reason training-free pruning fails at high ratios is not a weak importance signal but a missing model of what each retained token is for. It partitions the retained set into a disjoint union $S = P \\cup C \\cup D$: protected semantic-core tokens lock onto the queried object; complementary-context tokens cover secondary objects and scene structure outside core regions; detail tokens recover text, boundaries, and small parts. The paper proposes Attention-Anchored Regions (AARs), unions of local 2D-grid neighborhoods around high-confidence anchor tokens, as lightweight proxies for object support already covered by the core. Context candidates outside AARs receive an exploration bonus, detail candidates inside AARs receive a repair bonus, and token-token similarity is used only when filtering context redundancy against already-retained tokens, not as a global pairwise graph. With this allocation, the paper reports the best normalized averages among training-free baselines at every matched budget on LLaVA-1.5-7B, LLaVA-NeXT-7B, Qwen2.5-VL-7B, and Qwen3-VL-8B, including 96.5% of unpruned performance at 88.9% pruning on LLaVA-1.5.","pith_inferences":["A consequence the paper leaves implicit is that at extreme pruning ratios, accuracy is limited less by the quality of the importance signal than by whether the retained set contains at least one anchor per semantically distinct region; this suggests budgets should scale with the number of salient regions, not with image resolution.","The transferability of the GQA-fit priors is directly testable: fitting them on a non-GQA corpus such as COCO or Flickr and re-running the same tables would reveal whether the GQA column is inflated by prior leakage.","The same anchor neighborhoods could seed region-level grounding or KV-cache eviction policies, a use the paper does not explore.","A natural extension to video would treat frame-local anchors as AARs and allocate spatiotemporal context outside them."],"forward_implications":["Because coverage is made explicit by AARs, the residual budget can be spent asymmetrically—context outside the AAR, detail inside it—so the method avoids the dense pairwise redundancy graph that dominates D2Pruner's selection cost.","At 66.7% pruning on LLaVA-1.5-7B, RoRA keeps 99.8% of the normalized average performance across nine benchmarks, and 96.5% at 88.9% pruning, the best among training-free baselines at each matched budget.","On Qwen3-VL-8B at 75% and 90% pruning, RoRA beats D2Pruner by 5.4 and 5.3 normalized average points, showing the allocation transfers to newer dynamic-resolution backbones.","The full POPE run with 192 retained tokens finishes in 5:43 on an H800 with batch size 1, giving a 24.6% end-to-end runtime reduction and 0.7 ms selection overhead while retaining 98.4% of unpruned accuracy.","Localization ablations on RefCOCO-family splits show the protected semantic core is the dominant component: normalized accuracy rises from 35.44% to 74.44% at 144 retained tokens once core protection is added, with AAR-guided context adding a further improvement."],"supporting_citations":[{"why":"It supplies the attention-based FastV baseline and the early-LLM-layer pruning setting that RoRA builds on.","marker":"Chen et al. 2024"},{"why":"It supplies the attention-debiasing principle and the D2Pruner baseline that RoRA compares against and improves.","marker":"Zhang et al. 2026"},{"why":"It provides the HoloV crop-based region allocation baseline whose fixed quotas RoRA contrasts with AAR-guided allocation.","marker":"Zou et al. 2026"},{"why":"It supplies DART, the diversity-based selection baseline that RoRA outperforms under matched budgets.","marker":"Wen et al. 2025"},{"why":"It provides the GQA training images used to estimate the positional and object priors, and GQA as an evaluation benchmark.","marker":"Hudson and Manning 2019"},{"why":"It defines the LLaVA-1.5 backbone on which the main fixed-grid pruning results and efficiency measurements are reported.","marker":"Liu et al. 2024a"},{"why":"It defines the LLaVA-NeXT dynamic-resolution backbone used for the long-sequence experiments.","marker":"Liu et al. 2024b"},{"why":"It defines the Qwen2.5-VL backbone used in the Qwen evaluation block.","marker":"Bai et al. 2025b"},{"why":"It defines the Qwen3-VL backbone where RoRA surpasses D2Pruner by about five normalized points.","marker":"Bai et al. 2025a"}],"fun_headline_variants":["Role-based token pruning keeps 96.5% at 88.9% pruning","RoRA prunes by role: core, context, detail","Prune tokens by role, not just importance","Role-aware pruning: 96.5% accuracy at 89% cut","Token roles enable high-ratio pruning in MLLMs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two calibration statistics learned from 1,000 unlabeled GQA images—where the model tends to look and where objects tend to appear—are generic enough to transfer to every test benchmark, including GQA itself.","fun_headline_variants_meta":{"raw":{"variants":["Role-based token pruning keeps 96.5% at 88.9% pruning","RoRA prunes by role: core, context, detail","Prune tokens by role, not just importance","Role-aware pruning: 96.5% accuracy at 89% cut","Token roles enable high-ratio pruning in MLLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000332,"raw_usage":{"total_tokens":1939,"prompt_tokens":1128,"completion_tokens":811,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":721}},"tokens_in":744,"tokens_out":811,"duration_ms":7094,"temperature":1.0,"reasoning_tokens":721,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:28:35.358266+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-fit the two calibration priors on a source with no overlap with the evaluation benchmarks, for example 1,000 unlabeled images from COCO or Flickr, keep every other hyperparameter fixed, and re-run the LLaVA-1.5 and Qwen tables; a meaningful drop in normalized average accuracy on GQA, or in the overall average, would show that the reported gains depend on the particular calibration corpus rather than on generic role allocation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the attention-debiasing principle and the D2Pruner baseline that RoRA compares against and improves."},{"cited_title":"Highlighted Tokens","cited_arxiv_id":null,"evidence_quote":"It provides the HoloV crop-based region allocation baseline whose fixed quotas RoRA contrasts with AAR-guided allocation."}],"review_version":2}