{"id":"3d7c80e1-b921-4d54-9dc9-ea24e4d3268b","arxiv_id":"2606.11576","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"AVIS is an adaptive policy that jointly scales visual context via key-based token pruning and reasoning via difficulty-predicted self-consistency to improve the accuracy-compute curve on image and video tasks.","lead":"AVIS adapts how much visual detail and reasoning effort a vision-language model uses for each query using token pruning and a difficulty predictor. This targets better accuracy per unit of compute in models that currently scale both axes inefficiently.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Learned difficulty predictor's generalization and overhead not validated enough to guarantee reliable VRS allocation","rationale":"The reader's weakest assumption directly identifies the same load-bearing component. Because the full manuscript was not supplied in the initial query, no stronger internal inconsistency can be diagnosed; the concern remains the unverified reliability of the learned policy rather than any contradiction in the stated method.","tokens_in":1728,"tokens_out":310,"duration_ms":18601,"concrete_test":"Split the difficulty-predictor training queries from one benchmark (e.g., the image set) and evaluate predictor accuracy plus end-to-end accuracy-compute curve on a held-out video benchmark; if the adaptive curve falls below the best fixed-rollout baseline on the held-out set, the claimed joint scaling advantage does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"AVIS's central claim rests on the learned difficulty predictor adaptively choosing rollout count (VRS) per query to improve the joint accuracy-compute frontier over fixed VCS-only or VRS-only baselines. This requires the predictor to (a) incur negligible extra cost, (b) avoid systematic misallocation on a non-trivial fraction of inputs, and (c) generalize across the reported image/video benchmarks and RL-post-trained VLMs. The abstract provides no details on training data, loss, architecture, or cross-benchmark hold-out performance of this component, so the adaptive benefit could be an artifact of in-distribution tuning rather than a robust policy.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript introduces AVIS for adaptive test-time scaling in Vision-Language Models. It decomposes inference cost into Visual Context Scaling (VCS) via training-free Key Diversity Visual (KDV) token pruning and Visual Reasoning Scaling (VRS) via a learned difficulty predictor that selects the number of self-consistency rollouts per query. The central claim is that this joint adaptive policy improves the accuracy-compute frontier over VCS-only and VRS-only baselines across image and video reasoning benchmarks, remains effective on RL post-trained VLMs, and is compatible with shared-prefill inference.","tokens_in":1843,"tokens_out":411,"duration_ms":18543,"significance":"If the empirical claims hold with proper validation, the work would be significant for practical VLM deployment by showing how to jointly allocate compute across visual evidence and reasoning search without prohibitive overhead. The training-free KDV pruning and shared-prefill design are practical strengths that could generalize beyond the reported setting.","major_comments":[{"comment":"Abstract: The abstract states performance gains but supplies no quantitative results, error bars, dataset details, or validation of the difficulty predictor; the central claim cannot be assessed from the provided text.","section":"Abstract"},{"comment":"Method section on the difficulty predictor: The learned predictor is load-bearing for the adaptive VRS benefit, yet no details are supplied on its training data, loss, architecture, inference overhead, or cross-benchmark hold-out accuracy; without these, it is impossible to verify that the predictor avoids systematic misallocation or adds negligible cost on a non-trivial fraction of inputs.","section":"Method (difficulty predictor)"}],"minor_comments":[{"comment":"Abstract: Consider naming the specific image and video benchmarks and reporting at least one headline accuracy-compute number to allow readers to gauge the magnitude of the claimed improvement.","section":"Abstract"}],"recommendation":"uncertain","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We appreciate the referee's comments highlighting areas where the manuscript can be improved for clarity and completeness. We address each point below.","responses":[{"response":"We agree that the abstract would be strengthened by including quantitative results. In the revised version, we will add specific performance metrics, dataset details, and references to the validation of the difficulty predictor to make the central claim more assessable from the abstract.","revision_made":"yes","referee_comment":"[Abstract] Abstract: The abstract states performance gains but supplies no quantitative results, error bars, dataset details, or validation of the difficulty predictor; the central claim cannot be assessed from the provided text."},{"response":"We acknowledge that more details are needed on the difficulty predictor. We will expand the method section in the revised manuscript to provide information on its training data, loss function, architecture, inference overhead, and cross-benchmark hold-out accuracy to allow proper verification.","revision_made":"yes","referee_comment":"[Method (difficulty predictor)] Method section on the difficulty predictor: The learned predictor is load-bearing for the adaptive VRS benefit, yet no details are supplied on its training data, loss, architecture, inference overhead, or cross-benchmark hold-out accuracy; without these, it is impossible to verify that the predictor avoids systematic misallocation or adds negligible cost on a non-trivial fraction of inputs."}],"tokens_in":1372,"tokens_out":311,"duration_ms":24599,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"AVIS combines key-diversity visual pruning with an adaptive difficulty predictor to jointly tune visual context and reasoning depth at test time. The shared prefill for multiple rollouts is a useful engineering choice.\n\nThe paper does a solid job defining the two scaling axes and showing how existing work handles them separately. The pruning rule is training-free and efficient, which is a plus, and the compatibility with RL post-trained models broadens its applicability. These elements make the method feel grounded in deployment realities.\n\nThe main soft spot is that the abstract contains no quantitative results, no description of the predictor's training or accuracy, and no specifics on benchmarks or baselines. This means we cannot yet tell if the adaptive allocation improves the trade-off or simply incurs extra cost without benefit. The stress-test concern about generalization and overhead stands until the full paper provides validation data.\n\nThis work is aimed at researchers and engineers optimizing inference for vision-language models in production settings. A reader looking for practical test-time scaling techniques would get value from the method description and the joint policy framing.\n\nIt deserves serious referee attention because the core idea is coherent and targets a real efficiency bottleneck, even if the current evidence is thin. I would recommend sending it to peer review so the authors can supply the missing experimental details and ablations.","headline":"AVIS pairs a simple training-free visual pruning rule with a learned difficulty predictor for joint scaling, but the abstract supplies no numbers or predictor validation so the claimed trade-off gains cannot be assessed.","tokens_in":2353,"tokens_out":346,"would_cite":false,"duration_ms":28635,"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":"AVIS jointly adapts visual context and reasoning scaling per query to improve accuracy-compute trade-off in vision-language models.","keywords":["vision-language models","test-time scaling","adaptive inference","token pruning","self-consistency","visual reasoning","chain-of-thought"],"falsifier":"If, on a held-out set of image and video reasoning tasks, AVIS reaches the same accuracy levels only by spending more total tokens or latency than a well-tuned fixed VCS or VRS baseline, the claimed improvement in the accuracy-compute trade-off would be refuted.","tokens_in":2641,"feed_emoji":"⚖️","tokens_out":712,"duration_ms":21385,"temperature":0.7,"pith_summary":"The paper introduces AVIS, a lightweight policy that decides per query how much visual evidence to keep and how many reasoning rollouts to run. It handles the first choice with a training-free rule that drops redundant visual tokens and the second with a small predictor that picks the rollout count from query difficulty. Existing work tunes only one of these two axes at a time; AVIS tunes both together while reusing a single prefill pass across rollouts. A reader would care because large visual contexts and long decoding chains make test-time scaling expensive, so smarter per-query allocation can preserve gains at lower total cost. The same policy stays useful after the base model has been reinforced.","feed_headline":"AVIS adapts visual pruning and rollout count per query","feed_subtitle":"The policy improves accuracy per compute token on image and video tasks by choosing both context size and reasoning depth on the fly.","key_machinery":"Adaptive Visual Inference Scaling (AVIS) policy that couples Key Diversity Visual (KDV) pruning for Visual Context Scaling with a learned difficulty predictor for Visual Reasoning Scaling.","core_discovery":"AVIS realizes Visual Context Scaling through Key Diversity Visual pruning, an O(N) rule that removes redundant visual tokens before the language model sees them, and realizes Visual Reasoning Scaling through adaptive self-consistency driven by a learned difficulty predictor that selects the number of rollouts. The design keeps all rollouts on a shared KV cache. Across image and video reasoning benchmarks this joint adaptation produces a better accuracy-compute curve than either VCS-only or VRS-only baselines and continues to help on RL post-trained VLMs.","pith_inferences":["Similar predictor-plus-pruning logic could be tested on long-context language models where context length and search depth are the two main cost drivers.","The difficulty signal might be reused to gate other test-time techniques such as tool use or external retrieval.","If the predictor is small enough, it could be distilled into the base VLM to remove the separate forward pass entirely."],"forward_implications":["AVIS yields higher accuracy at the same compute budget than methods that scale only visual context or only reasoning steps.","The gains hold on top of models that have already received reinforcement-learning post-training.","Shared-prefill compatibility keeps added latency and memory low because every rollout reuses one KV cache.","KDV pruning is training-free and linear in the number of visual tokens, so it adds negligible cost."],"fun_headline_variants":["AVIS tunes visual pruning and rollout count per query","AVIS balances token pruning and adaptive rollouts dynamically","Policy adapts visual context size and reasoning depth per query","AVIS jointly scales VCS and VRS via shared KV cache"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The learned difficulty predictor chooses rollout counts that match query needs without adding large overhead or wasting compute on a noticeable fraction of inputs.","fun_headline_variants_meta":{"raw":{"variants":["AVIS tunes visual pruning and rollout count per query","AVIS balances token pruning and adaptive rollouts dynamically","Policy adapts visual context size and reasoning depth per query","AVIS jointly scales VCS and VRS via shared KV cache"]},"model":"grok-4.3","cost_usd":0.003287,"raw_usage":{"total_tokens":1778,"prompt_tokens":711,"num_sources_used":0,"completion_tokens":63,"cost_in_usd_ticks":32874500,"prompt_tokens_details":{"text_tokens":711,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1004,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":711,"tokens_out":63,"duration_ms":8770,"temperature":1.0,"reasoning_tokens":1004,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T10:45:30.614453+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"If, on a held-out set of image and video reasoning tasks, AVIS reaches the same accuracy levels only by spending more total tokens or latency than a well-tuned fixed VCS or VRS baseline, the claimed improvement in the accuracy-compute trade-off would be refuted.","supporting_citations":[],"review_version":1}