{"id":"601831b2-66ff-4d54-991a-6092978b532a","arxiv_id":"2605.29657","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"OccamToken replaces absolute token ranking with register-anchored relative evidence testing to enable adaptive, high-ratio visual token pruning in VLMs while preserving most accuracy.","lead":"OccamToken is a training-free method for pruning visual tokens in vision-language models by using register tokens as a reference for relative evidence testing instead of fixed top-K ranking. Smart generalists might read it to understand potential efficiency gains for running multimodal AI models with less computation and memory.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Register tokens as stable low-info reference lacks isolation from other attention effects in the pruning logic.","rationale":"Reader's weakest assumption directly matches the load-bearing step. Full text expands the method description but does not add an isolating experiment, so the concern remains unchanged. This moves the verdict from UNVERDICTED to CONDITIONAL pending the proposed check.","tokens_in":1770,"tokens_out":278,"duration_ms":15499,"concrete_test":"On LLaVA-NeXT with the 40-token budget, replace the register-derived threshold with an equivalent dynamic threshold computed from mean attention across all visual tokens; if accuracy falls below 90% of the reported OccamToken result, the register-anchoring step is necessary for the headline performance.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on replacing absolute ranking with register-anchored relative testing. This requires that register attention specifically captures low-information patterns in a way that makes the derived thresholds query-adaptive and robust at 1.4% retention. The paper states this as the key insight but provides no ablation that holds all other components fixed while swapping the register reference for a non-register baseline (e.g., mean visual attention or a learned sink). Without that isolation, observed gains could stem from dynamic thresholding alone rather than the register property.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes OccamToken, a training-free token pruning framework for VLMs that replaces absolute importance ranking with register-anchored relative evidence testing. It claims that register tokens serve as a stable low-information reference, enabling image-adaptive redundancy pruning and query-adaptive relevance pruning via dynamic thresholds; experiments on LLaVA-NeXT, LLaVA-v1.5 and Qwen3-VL reportedly reduce 2880 visual tokens to ~40 while retaining >93% accuracy even at 1.4% retention.","tokens_in":1861,"tokens_out":490,"duration_ms":14668,"significance":"If the central claims hold after proper isolation of the register effect, the method would provide a practical, training-free route to extreme visual-token compression that adapts to both image redundancy and query dependence, addressing known brittleness of fixed top-K pruning in the presence of attention sinks.","major_comments":[{"comment":"Abstract (key insight paragraph): the assertion that register tokens 'naturally absorb low-information attention patterns' and thereby supply a stable reference is load-bearing for the entire framework, yet the manuscript supplies no ablation that holds all other components fixed while replacing the register reference with a non-register baseline (e.g., mean visual attention or a learned sink token). Without this isolation, observed gains at 1.4% retention cannot be attributed to the register property rather than dynamic thresholding alone.","section":"Abstract"},{"comment":"Method section (register-anchored relative testing): the derivation of query-adaptive thresholds from register attention is described only at a high level; the paper must supply the precise formula (including any scaling or normalization constants) and demonstrate that the resulting thresholds remain stable across the reported retention regimes.","section":"Method"}],"minor_comments":[{"comment":"Abstract: quantitative claims (2,880 tokens → ~40, >93% accuracy) are presented without reference to the exact evaluation protocol, datasets, or number of runs; these details belong in the abstract or a dedicated experimental-setup paragraph.","section":"Abstract"},{"comment":"Throughout: notation for register tokens versus visual tokens should be introduced once with consistent symbols rather than relying on prose descriptions.","section":null}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which help clarify the presentation of our core claims. We address each major point below and will revise the manuscript accordingly to strengthen the evidence and exposition.","responses":[{"response":"We agree that an explicit ablation isolating the register reference is necessary to attribute performance gains specifically to the register property rather than dynamic thresholding in isolation. In the revised manuscript we will add this ablation (holding all other components fixed) comparing register-anchored relative testing against non-register baselines such as mean visual attention and a fixed sink token, across the same retention regimes and models.","revision_made":"yes","referee_comment":"[Abstract] Abstract (key insight paragraph): the assertion that register tokens 'naturally absorb low-information attention patterns' and thereby supply a stable reference is load-bearing for the entire framework, yet the manuscript supplies no ablation that holds all other components fixed while replacing the register reference with a non-register baseline (e.g., mean visual attention or a learned sink token). Without this isolation, observed gains at 1.4% retention cannot be attributed to the register property rather than dynamic thresholding alone."},{"response":"We will expand the Method section to include the exact mathematical formulation of the query-adaptive thresholds (including all scaling and normalization constants) derived from register attention. We will also add a dedicated analysis subsection demonstrating threshold stability across the full range of reported retention ratios (50% down to 1.4%) on the evaluated models.","revision_made":"yes","referee_comment":"[Method] Method section (register-anchored relative testing): the derivation of query-adaptive thresholds from register attention is described only at a high level; the paper must supply the precise formula (including any scaling or normalization constants) and demonstrate that the resulting thresholds remain stable across the reported retention regimes."}],"tokens_in":1412,"tokens_out":402,"duration_ms":14381,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that OccamToken replaces absolute token ranking with register-anchored relative testing and reports strong compression numbers without any training. On LLaVA-NeXT it drops 2880 visual tokens to roughly 40 while keeping over 93 percent accuracy, and it shows similar patterns on LLaVA-v1.5 and Qwen3-VL.\n\nWhat is new is the framing that register tokens absorb low-information attention and therefore serve as a stable baseline for deciding whether a visual token adds real evidence. The method then derives image-adaptive and query-adaptive thresholds from that reference. This is positioned against the brittleness of fixed top-K selection under attention sinks and varying queries.\n\nThe paper does well at stating a clear problem with existing absolute-ranking approaches and at giving concrete retention numbers at the extreme 1.4 percent regime. Those numbers, if they hold under the full experimental protocol, would be useful for anyone trying to cut prefill cost on long visual sequences.\n\nThe soft spot is the missing isolation for the register claim. The abstract presents registers as naturally low-information sinks, yet there is no sign of an ablation that keeps the dynamic-threshold logic fixed and swaps the register reference for a simple mean or non-register baseline. Without that, the gains could come from adaptivity alone rather than the specific register property. The abstract also omits dataset details, baselines, and variance, so the full paper must supply those for the accuracy claims to be evaluable.\n\nThis is for practitioners who need to reduce VLM inference cost on visual inputs. A reader working on deployment or token compression would get value from trying the method even if they later add their own controls.\n\nIt deserves peer review because the empirical target is concrete and the core idea is simple enough to test.","headline":"OccamToken shows practical training-free pruning that hits high compression with retained accuracy, but the register reference needs an ablation to separate it from dynamic thresholding.","tokens_in":2372,"tokens_out":439,"would_cite":false,"duration_ms":20702,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"OccamToken replaces absolute visual token ranking with register-anchored relative testing to prune tokens down to 1.4% retention while keeping over 93% accuracy.","keywords":["token pruning","vision-language models","efficient inference","attention mechanisms","register tokens","adaptive pruning","training-free methods"],"falsifier":"On LLaVA-NeXT, accuracy at a 40-token budget with register-anchored relative testing falls below the accuracy obtained by absolute top-K ranking at the same budget.","tokens_in":2669,"feed_emoji":"⚡","tokens_out":651,"duration_ms":20362,"temperature":0.7,"pith_summary":"Vision-language models incur high prefill costs from long visual token sequences. Absolute ranking methods that pick a fixed top-K set prove brittle because attention sinks skew scores and because image redundancy plus query dependence make any fixed budget unreliable. OccamToken instead anchors pruning on register tokens that absorb low-information attention patterns and uses them as a reference to test whether each visual token adds genuine evidence. Dynamic thresholds derived from register attention then drive both image-adaptive redundancy removal and query-adaptive relevance removal. The approach runs without training and improves the accuracy-efficiency curve on LLaVA-NeXT, LLaVA-v1.5, and Qwen3-VL.","feed_headline":"OccamToken prunes 2880 visual tokens to 40 while retaining 93% accuracy","feed_subtitle":"Training-free method anchors pruning on register tokens to adaptively remove low-evidence visual tokens in VLMs.","key_machinery":"Register-anchored relative evidence testing, which compares each visual token's attention contribution against a register-based reference to decide retention via dynamic thresholds.","core_discovery":"The paper claims that register tokens naturally absorb low-information attention patterns and therefore supply a stable reference for identifying genuinely informative visual evidence. Replacing absolute importance ranking with register-anchored relative evidence testing yields both image-adaptive redundancy pruning and query-adaptive relevance pruning through dynamic thresholds derived from register attention, delivering consistent gains in the accuracy-efficiency trade-off without any additional training.","pith_inferences":["The same register-reference idea could be tested in other attention-heavy multimodal architectures where certain tokens act as sinks.","Memory footprint during inference would drop roughly in proportion to the token reduction if the accuracy claim holds.","Future work could examine whether the register tokens themselves can be further compressed once they have served as the reference."],"forward_implications":["Accuracy-efficiency trade-off improves across LLaVA-NeXT, LLaVA-v1.5, and Qwen3-VL without training.","Stable compression remains possible even at the 1.4% retention regime.","Both image-adaptive redundancy pruning and query-adaptive relevance pruning are achieved through the same register-derived thresholds.","The method eliminates the need for per-input fixed token budgets."],"fun_headline_variants":["OccamToken anchors pruning on register tokens for adaptive VLM inference","Register tokens supply stable reference for relative evidence testing","OccamToken performs image and query adaptive pruning training free","Relative testing on registers replaces absolute token ranking in VLMs"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Register tokens naturally absorb low-information attention patterns, making them a stable reference for identifying genuinely informative visual evidence.","fun_headline_variants_meta":{"raw":{"variants":["OccamToken anchors pruning on register tokens for adaptive VLM inference","Register tokens supply stable reference for relative evidence testing","OccamToken performs image and query adaptive pruning training free","Relative testing on registers replaces absolute token ranking in VLMs"]},"model":"grok-4.3","cost_usd":0.007396,"raw_usage":{"total_tokens":3420,"prompt_tokens":708,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":73962000,"prompt_tokens_details":{"text_tokens":708,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2647,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":708,"tokens_out":65,"duration_ms":26455,"temperature":1.0,"reasoning_tokens":2647,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T08:44:09.345972+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"On LLaVA-NeXT, accuracy at a 40-token budget with register-anchored relative testing falls below the accuracy obtained by absolute top-K ranking at the same budget.","supporting_citations":[],"review_version":1}