{"id":"7ebeb353-8697-4d0c-80f9-20ede64cb520","arxiv_id":"2603.18558","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free hierarchical multimodal selector turns a video question into a fuzzy-logic tree over vision and audio experts and beats prior frame selectors under a 16-frame budget.","lead":"HiMu picks the few most useful video frames for answering hard questions by turning the question into a logic tree and scoring frames with cheap vision and audio experts. It can match much denser sampling while staying training-free and avoiding expensive multi-round model loops.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The central claim rests on unmeasured fidelity of the single-shot LLM logic-tree parse for compositional queries.","rationale":"The reader correctly isolates the single-shot LLM parse as the weakest assumption supporting the claim that HiMu solves the “flattening” problem of global-embedding selectors. Controlled K=16 results, Flat-Fusion ablation (–5.5 pp), multi-LVLM transfer, and the 16-vs-64 frame comparison remain solid empirical evidence; the remaining risk is precisely that those gains may shrink once tree quality is measured rather than assumed. Because the paper already flags the limitation and the reader already assigns CONDITIONAL, no verdict change is warranted—only the same tightening the reader requested (tree audits or oracle bounds) before the Pareto-front language can be treated as fully settled.","tokens_in":21763,"tokens_out":574,"duration_ms":26716,"concrete_test":"On the LongVideoBench referring-query subset (or Video-MME questions containing explicit temporal/cross-modal cues), collect the generated trees, have two annotators label each for (i) correct operator choice and nesting, (ii) correct expert routing. Replace every imperfect tree with a corrected version, re-run expert scoring + PASS + Qwen3-VL-8B QA at K=16. If accuracy rises by more than ~2–3 pp relative to the automatic trees, the parse-fidelity assumption is load-bearing and currently under-supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"HiMu's advantage over flat similarity selectors is that a single text-only LLM call produces a hierarchical tree whose leaves route to modality experts and whose internal nodes apply fuzzy And/Or/Seq/RightAfter (Sec. 3.1, Eqs. 3–8). The satisfaction curve T(t) and the subsequent PASS selection therefore inherit any errors in that parse. The paper itself states that “malformed or overly shallow trees can misroute predicates and degrade selection quality” (Limitations). Supplementary parser swaps (Table 8) show only ~1 pp variance across four LLMs, yet that measures end-to-end accuracy, not tree correctness. There is no human audit of operator choice, expert routing, or nesting depth, nor an oracle-tree upper bound on the compositional subsets (temporal-order or cross-modal questions) that the method claims to unlock. If a non-negligible fraction of trees collapse to near-flat structure, the reported gains over BOLT/AKS/T* and the 4× frame-budget claim would be partly attributable to multi-expert scoring rather than true hierarchical composition—the very distinction the paper advertises.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"HiMu is a training-free frame selector for long-form VideoQA. A single text-only LLM call decomposes the query into a hierarchical logic tree whose leaves are atomic predicates routed to lightweight vision (CLIP, OVD, OCR) and audio (ASR, CLAP) experts; signals are median/MAD-normalized, bandwidth-smoothed, and composed bottom-up with fuzzy And/Or/Seq/RightAfter operators into a satisfaction curve T(t), from which PASS selects K frames. Under a fixed K=16 budget the method reports SOTA accuracy among frame selectors on Video-MME, LongVideoBench val and HERBench-Lite with Qwen3-VL-8B, improves over uniform sampling across seven LVLMs as a drop-in module, and matches uniform sampling at 4× the frame budget, while remaining cheaper than agentic multi-call systems.","tokens_in":22089,"tokens_out":950,"duration_ms":9391,"significance":"If the results hold, HiMu meaningfully advances the efficiency–accuracy Pareto front for query-aware frame selection. The design is training-free, caches expert features, incorporates audio as first-class selection evidence, and supplies per-leaf attribution for interpretability—practical strengths for multi-query long-video pipelines. Controlled K=16 comparisons against Uniform/BOLT/AKS/T*, multi-model generalization, expert/composition ablations, frame-budget curves, and FLOPs plots constitute a solid empirical package that other groups can build on or combine with token-compression methods.","major_comments":[{"comment":"The central claim that hierarchical composition (not merely multi-expert scoring) drives the gains is only partially isolated. Table 3 shows Flat Fusion drops overall accuracy by 5.49 pp, which is the largest ablation, but Flat Fusion still aggregates the same multi-modal leaves; there is no oracle-tree upper bound, no human audit of operator choice/routing/nesting depth, and no stratified breakdown on the temporal-order or cross-modal subsets that the introduction and abstract advertise as the motivating failure mode of global embeddings. Without that evidence it remains possible that a non-negligible fraction of trees are near-flat and that reported gains over BOLT/AKS/T* partly reflect multi-expert scoring rather than true hierarchical temporal logic.","section":null},{"comment":"Default experiments use the same LLM family for tree parsing and answering (Sec. 4.1). Supplementary Table 8 shows only ~1 pp end-to-end variance across four parsers, yet that measures final QA accuracy, not tree fidelity. The Limitations section itself states that malformed or shallow trees misroute predicates; a modest human audit or oracle-tree experiment on a compositional subset would substantially strengthen the causal link between the neuro-symbolic design and the SOTA numbers.","section":null},{"comment":"Asymmetric comparisons to agentic systems (VideoZoomer 128 frames, VideoChat-A1 512 frames, VSLS 32 frames) in Table 2 are informative for the Pareto narrative but place HiMu at a numerical advantage that is not fully controlled. The paper should either re-run those systems at K=16 with the same backbone or more carefully qualify the claim that 16 compositionally selected frames are “more effective” than hundreds of densely sampled frames.","section":null}],"minor_comments":[{"comment":"Figure 2 FLOPs axis and Table 1 latency numbers for baselines are estimated rather than re-measured on the same hardware; a short note on estimation methodology would improve reproducibility.","section":null},{"comment":"PASS parameters (Np, Nn, w, Δ) are set to floor(sqrt(K)) schedules; a one-sentence justification or sensitivity note in the main text (beyond the supplement) would help readers.","section":null},{"comment":"Abstract and introduction use both MLLM and LVLM; pick one acronym for consistency.","section":null},{"comment":"Eqs. (5)–(8) for Seq and RightAfter are clear, but a short worked numerical example of a two-step Seq would aid readers unfamiliar with fuzzy temporal operators.","section":null},{"comment":"HERBench-Lite absolute numbers remain low (~43 %); a brief discussion of the “fusion deficit” relative to oracle frame sets would contextualize the modest absolute gains.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is a solid systems contribution with careful ablations and multi-benchmark coverage. The main risk is over-claiming hierarchical composition without direct tree-fidelity evidence; that is fixable with a modest audit or stratified analysis and does not require rejecting the work. Fit for a top CV/ML venue after minor revision."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The punchline is simple: under a fixed 16-frame budget, HiMu beats the usual similarity selectors and improves a stack of LVLMs as a drop-in, while staying far cheaper than agent loops. That is a practical result people will actually use.\n\nWhat is new is not any single expert or fuzzy operator. It is the full training-free pipeline: one text-only parse into a hierarchical tree, leaves routed to vision and audio experts (CLIP, OVD, OCR, ASR, CLAP), bandwidth-matched smoothing, fuzzy And/Or/Seq/RightAfter into a satisfaction curve, then PASS. Related work already has flat similarity, detector zoom, temporal-logic specs, and multi-call agents. HiMu sits in the gap and measures it carefully. The controlled Qwen3-VL-8B / K=16 comparisons, multi-benchmark coverage, leave-one-out and flat-fusion ablations, frame-budget curves (16 frames beating uniform at 64), multi-model transfer, and FLOPs Pareto plot are the real strength. Flat fusion dropping harder than removing any single expert is the cleanest evidence that composition is doing work, not just multi-signal scoring. Audio as a first-class selection signal is also well motivated and ablated.\n\nSoft spots, in proportion. The load-bearing assumption is that the single-shot LLM tree is faithful enough. The paper admits malformed trees hurt; the supplement shows end-to-end accuracy is stable across parsers, but that is not a human audit of operator choice, nesting, or routing, and there is no oracle-tree upper bound on the temporal/cross-modal subset. That is a genuine gap for the “hierarchical composition” claim, not a fatal one. Other ordinary issues: free hyperparameters (sensitivity is mild), proprietary models on a 25% subset, and some literature comparisons at much larger frame budgets. Math is standard continuous fuzzy logic; citations cover the space fairly, including their own HERBench.\n\nThis is for people building long-video QA systems under context limits, not for pure theory. I would bring it to reading group, cite it when discussing frame selection or multimodal evidence routing, and send it to peer review. Tighten the tree-fidelity analysis and the asymmetric baselines; the core result still stands.","headline":"A real systems win on compositional multimodal frame selection under tight budgets; the LLM-tree fidelity gap is real but does not erase the empirical case.","tokens_in":22679,"tokens_out":563,"would_cite":true,"duration_ms":10909,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A single text-only parse into a logic tree of vision and audio experts can pick the right 16 frames for long-video QA without iterative model calls.","keywords":["Video Question Answering","Frame Selection","Neuro-Symbolic Reasoning","Multimodal Understanding","Long-form Video","Fuzzy Logic Composition","Training-free Selection"],"falsifier":"Hold the answering model and the 16-frame budget fixed, replace HiMu's tree with deliberately shallow or misrouted trees (or with flat leaf-score summation) on the same Video-MME and LongVideoBench questions, and check whether the reported accuracy gap over uniform sampling and over similarity selectors disappears.","tokens_in":22661,"feed_emoji":"🎬","tokens_out":722,"duration_ms":9125,"temperature":0.7,"pith_summary":"Long-video question answering is bottlenecked by which frames a model is allowed to see under a tight context budget. Similarity-based selectors collapse compositional, cross-modal questions into one embedding and lose order and modality bindings; agentic selectors recover structure only by expensive iterative vision-language calls. HiMu claims that a single text-only language-model call can decompose the question into a hierarchical logic tree of atomic predicates, each scored by a cheap expert (vision, OCR, speech, non-speech audio), then composed with fuzzy temporal operators into a continuous satisfaction curve from which a small set of frames is chosen. Under a fixed 16-frame budget it reports state-of-the-art accuracy among frame-selection methods on three benchmarks, lifts seven different answering models as a drop-in module, and matches the accuracy of uniform sampling that uses four times as many frames, all without retraining or multi-round selection loops. A sympathetic reader cares because the result says compositional multimodal evidence can be recovered at single-shot cost rather than being traded against latency.","feed_headline":"16 frames beat 64 when selection is a logic tree","feed_subtitle":"One text-only parse plus cheap vision and audio experts matches uniform sampling at four times the budget.","key_machinery":"The hierarchical multimodal logic tree: leaves are atomic expert predicates (CLIP, open-vocabulary detection, OCR, ASR, CLAP); internal nodes are continuous fuzzy operators (And, Or, Seq, RightAfter) that produce a satisfaction curve T(t); PASS then spreads selection across peaks of that curve.","core_discovery":"Compositional multimodal frame selection for long-form video QA does not require iterative large vision-language inference. A single text-only parse of the query into a hierarchical logic tree, leaves routed to lightweight vision and audio experts, signals normalized and bandwidth-matched, then composed bottom-up by fuzzy operators that enforce co-occurrence, sequence and tight adjacency, yields a per-frame satisfaction curve from which a fixed budget of frames can be selected. Under that regime HiMu is the strongest reported selector on Video-MME, LongVideoBench and HERBench-Lite and matches uniform sampling at four times the frame budget.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Logic-tree frame pick: 16 beats uniform 64","Single parse plus experts match 4× frame budget","HiMu logic tree tops uniform sampling at 16 frames","Fuzzy hierarchical selectors beat 64-frame uniform","16-frame multimodal logic tree matches 4× budget"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The whole pipeline assumes that one text-only language-model parse produces a faithful hierarchical tree of atomic expert predicates; if the tree is shallow or misrouted, the satisfaction curve and selected frames collapse.","fun_headline_variants_meta":{"raw":{"variants":["Logic-tree frame pick: 16 beats uniform 64","Single parse plus experts match 4× frame budget","HiMu logic tree tops uniform sampling at 16 frames","Fuzzy hierarchical selectors beat 64-frame uniform","16-frame multimodal logic tree matches 4× budget"]},"model":"grok-4.5","effort":"low","cost_usd":0.005636,"raw_usage":{"total_tokens":1584,"prompt_tokens":871,"num_sources_used":0,"completion_tokens":81,"cost_in_usd_ticks":56360000,"prompt_tokens_details":{"text_tokens":871,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":632,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":871,"tokens_out":81,"duration_ms":5455,"temperature":1.0,"reasoning_tokens":632,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T22:31:02.656801+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Hold the answering model and the 16-frame budget fixed, replace HiMu's tree with deliberately shallow or misrouted trees (or with flat leaf-score summation) on the same Video-MME and LongVideoBench questions, and check whether the reported accuracy gap over uniform sampling and over similarity selectors disappears.","supporting_citations":[],"review_version":1}