{"id":"07068d47-a490-4465-865a-1764bc2fa075","arxiv_id":"2607.10796","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Routing CV experts into atomic evidence then Bloom-staged verbalization improves LVLM benchmarks and yields measurable query-conditioned reasoning traces.","lead":"The paper adds specialized vision experts and a Bloom-taxonomy staged reasoning protocol to LVLMs so answers rest on atomic, citeable evidence. The design improves perception and reasoning scores while exposing how different queries enter different cognitive levels.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Ablation does not isolate Bloom synthesis from expert-pool gains; the central claim that the hierarchical protocol itself improves verifiable multimodal behavior is under-supported.","rationale":"The reader's weakest assumption (training-free template verbalizers + hand-designed Bloom steps remain non-inferential and reliably consumable) is real but secondary. The more load-bearing gap is causal attribution: the paper never shows that the ordered Bloom scaffold, rather than the expert pool or extra context, drives the reported gains and the verifiability story. Table 2 is the only controlled comparison and it is incomplete (MM-Vet flat; no matched unstructured-context control; main results vs non-expert baselines). The reader's CONDITIONAL verdict and call for tighter expert-augmented controls already point in the right direction; the stress test sharpens that the missing control is specifically Experts-only vs equal-budget unstructured evidence vs Bloom, not merely code release. No change of verdict category is required—still CONDITIONAL—but the condition should be stated as isolation of the hierarchical protocol, not only reproducibility of the training-free pipeline. The architecture is clear and the diagnostics (start-level distribution, citation rates) are useful; the soft spot is simply that the central causal claim is not yet tightly supported by the reported ablations.","tokens_in":15252,"tokens_out":724,"duration_ms":8536,"concrete_test":"Re-run the full Table 1 / Fig. 4 suite with three matched arms on the same backbone and expert pool: (A) Experts Only (no Bloom draft), (B) Experts + unstructured long evidence dump of equal token budget to y_Bloom, (C) Experts + full Bloom verbalization. If (C) does not beat (A) and (B) by a clear margin on MMBench L2 (CP/FP-S/LR/RR) and POPE adversarial, the hierarchical-protocol claim does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is that imposing an evidence-grounded Bloom hierarchy on multi-expert LVLM inference yields more verifiable/analyzable behavior and improves perception/reasoning benchmarks. The paper's own ablation (Table 2, §4.3) is the load-bearing support for attributing gains to Bloom rather than to the vision experts alone. That table shows Vision Experts Only already at MMB 70.4 / MM-Vet 49.3 / SEED-Text 47.06, versus Experts+Bloom at 75.7 / 49.0 / 51.76. MM-Vet is flat-to-worse; the MMB and SEED-Text lifts are modest and come with higher latency, GFLOPs, and token length. Main Table 1 and Fig. 4 compare MoCE only against non-expert baselines (InstructBLIP, Qwen-VL, LLaVA), not against the expert-augmented MoVA-style pool the method builds on (§3.1, §3.5). The Reasoning Trace Module (Table 3) shows most traces start at LIST (83.82%) with sparse step presence (~14–16%), so the hierarchy is only partially exercised. Thus the claim that the Bloom protocol (not the experts or longer context) is what improves verifiability and performance rests on a weak isolation of the synthesis stage.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes MoCE, a training-free LVLM inference framework that routes specialized vision experts (detection, segmentation, OCR/chart parsers), converts their outputs into a typed, citeable Literal Evidence Summary of atomic statements, then applies a Bloom-inspired staged verbalizer (LIST→EXPLAIN→ILLUSTRATE→COMPARE→HYPOTHESIZE→FINAL) to produce an evidence-cited draft that is fed with the image and query to a LLaVA-style Vicuna-7B backbone. A lightweight Reasoning Trace Module parses the draft to recover query-conditioned entry levels, step-wise citations, and evidence-type usage. Empirically, MoCE reports gains over InstructBLIP, Qwen-VL, and LLaVA-1.5/NeXT on TextVQA, MMBench, MMVet, POPE, and SEED, with ability breakdowns favoring perception and relational reasoning; an experts-only vs. experts+Bloom ablation and N=5000 trace diagnostics are used to argue that the hierarchical protocol improves verifiability and performance.","tokens_in":15594,"tokens_out":1509,"duration_ms":36636,"significance":"If the hierarchical, evidence-grounded protocol is shown to be the operative ingredient—not merely the expert pool, longer intermediate context, or adapter fusion—the work would offer a practical, training-free way to make multi-expert LVLM reasoning more inspectable and less hallucination-prone, with a reusable diagnostic (the Reasoning Trace Module) that quantifies cognitive entry levels and evidence use. The explicit typed evidence store with citation IDs and the deterministic trace parser are concrete engineering contributions that go beyond typical CoT prompting. Significance is currently limited by incomplete isolation of the Bloom stage from the MoVA-style expert stack the method reuses, so the architectural claim is promising but not yet firmly established as a cognitive-protocol result rather than a multi-expert system result.","major_comments":[{"comment":"§4.3, Table 2 is the load-bearing isolation of Bloom synthesis from the expert pool, yet it is incomplete for the central claim. Experts Only already reaches MMB 70.4 / MM-Vet 49.3 / SEED-Text 47.06; adding Bloom lifts MMB (+5.3) and SEED-Text (+4.7) but leaves MM-Vet flat-to-worse (49.3→49.0) while increasing latency (10.24s→13.40s), GFLOPs (+13.75%), and average tokens (63.70→76.87). There is no control that appends unstructured intermediate text or free-form CoT of matched length/token budget, so gains cannot yet be attributed to the ordered Bloom hierarchy rather than extra context or more verbose generation. A length-matched non-hierarchical scaffold is needed before concluding that the protocol itself improves verifiable multimodal behavior.","section":null},{"comment":"§4.2 Table 1 and Fig. 4 compare MoCE only to non-expert LVLMs (InstructBLIP, Qwen-VL, LLaVA-1.5/NeXT). The method reuses the expert pool and adapter design from MoVA (§3.1, §3.5, ref. [55]), and Vision Experts Only is already a strong intermediate system in Table 2. Without a main-table comparison to MoVA (or an equivalently expert-augmented baseline under the same backbone and inference budget), the performance narrative over-attributes gains to the cognitive verbalization framework relative to prior multi-encoder expert integration. Reporting MoVA / Experts Only alongside MoCE in Table 1 and Fig. 4 is necessary for the contribution claim.","section":null},{"comment":"§4.3 Table 3 reports step-present rates of only ~14–16% per Bloom step over N=5000, while among N=952 detectable traces 83.82% start at LIST. These figures undercut the claim that the hierarchy is systematically operationalized at inference time (§3.2, conclusion). The manuscript should clarify the relationship between “step present,” “detected traces,” and hierarchy compliance (skips, out-of-order headers, missing FINAL), quantify how often y_Bloom is malformed or empty of structure, and analyze whether performance gains concentrate on the minority of examples that actually execute multi-step traces. Without that, the Reasoning Trace Module’s diagnostics do not yet support “query-conditioned cognitive trajectories” as a reliable property of the system.","section":null},{"comment":"§3.2–3.4 leave the implementation of f_Bloom and the router under-specified for a training-free, reproducibility-critical pipeline. It is unclear whether Bloom verbalization and expert selection are produced by the same Vicuna-7B via fixed prompts, a separate LLM, or templates; how ℓ0 is chosen from q; and what the exact routing prompt and top-K are. Because the paper’s free parameters include router top-K, adapter depth L, and entry-level logic, and because atomic verbalization is claimed to be non-inferential, these details (and sensitivity to them) are load-bearing for the training-free and fidelity claims in §4.1.","section":null}],"minor_comments":[{"comment":"Several typographical and formatting issues: “V erbalization” / “T race” with stray spaces in headings; “strenghten” (Related Works); “L VLMs”; “Course Perception” in Fig. 4 should be “Coarse Perception.”","section":null},{"comment":"§3.2 Eq. (4)–(5): notation for ˜T and the Reasoning Structure/Trace Module is slightly inconsistent across text and figure captions; unify naming.","section":null},{"comment":"Fig. 1 and Fig. 3 are helpful but do not show failure cases (wrong expert routing, hallucinated citations, hierarchy skips); one negative qualitative example would strengthen the interpretability narrative.","section":null},{"comment":"Related work positions MoAI/MoVA well but could more sharply state what is new relative to those systems beyond the Bloom scaffold (e.g., typed evidence store + deterministic trace metrics).","section":null},{"comment":"Appendix tables on TEXT evidence atomicity are useful; a brief pointer from §3.2 would help readers find them.","section":null},{"comment":"All benchmark numbers are point estimates with no seeds, variance, or significance tests; even for a training-free method, multi-run or bootstrap intervals on the main deltas would improve credibility.","section":null}],"recommendation":"major_revision","confidential_remarks":"The manuscript is closest in spirit to MoVA/MoAI-style multi-expert LVLMs with an added cognitive scaffold. Fit for a solid CV/ML venue is reasonable after major revision, but novelty will be judged largely on whether Bloom+trace is more than a prompting wrapper around an existing expert stack. I would not reject on novelty alone if the isolation experiments are strengthened. No integrity concerns noted."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit here is the concrete inference-time stack: route a few specialized CV experts (DETR/SAM/OCR/chart parsers), turn their outputs into short typed atomic evidence items with IDs, run a Bloom-ordered verbalizer that produces a staged draft with citations, feed that draft plus image/query to the LLM, and parse the result with a lightweight Reasoning Trace Module. That combination is not in MoVA/MoAI/LLaVA as written, and the start-level / citation diagnostics (N=5000, mostly LIST at ~84%) are a genuine addition for inspectability.\n\nWhat works: the pipeline is training-free, the evidence store is explicit and citeable, and MoCE beats the usual non-expert baselines on TextVQA, MMBench, MMVet, POPE, SEED, with clearer lifts on the perception and relation slices. The experts-only vs experts+Bloom ablation (Table 2) does show a real MMB and SEED-text bump from the staged synthesis, and the trace table makes hierarchy compliance and evidence use measurable instead of hand-wavy.\n\nSoft spots, in proportion: the isolation is incomplete. Experts alone already sit at strong numbers; Bloom adds modest accuracy at higher latency/tokens/GFLOPs, and MM-Vet is flat-to-worse. Main Table 1 and the ability plots compare only against InstructBLIP/Qwen/LLaVA, not against the MoVA-style expert pool the method actually builds on. Step headers appear only ~14–16% of the time, so the full LIST→…→FINAL ladder is lightly exercised. Bloom is a convenient scaffold, not a deep cognitive model, and free choices (top-K, adapter depth, entry logic) are not stress-tested. No error bars, no released prompts/code in the manuscript.\n\nThis is for people building tool-augmented or interpretable VL systems who want a practical pattern and a diagnostic, not a new capability regime. Math is light (just the formalization of the store and steps), data are standard public benchmarks, citations cover the right MoE/VLM and cognitive-science neighbors without obvious gaps. I would send it to peer review; a referee can demand the missing expert-augmented baselines and code. Worth a look if you care about grounded multi-expert routing; not urgent otherwise.","headline":"Clean training-free recipe that makes multi-expert LVLM evidence citeable and traces Bloom-style steps, with real but modest gains that mostly ride the expert pool rather than the hierarchy itself.","tokens_in":16174,"tokens_out":592,"would_cite":false,"duration_ms":19428,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A Bloom-inspired hierarchy turns multi-expert vision outputs into staged, citeable reasoning that LVLMs can use at inference time.","keywords":["vision-language models","mixture of experts","Bloom taxonomy","cognitive verbalization","evidence grounding","hallucination reduction","reasoning trace","multimodal reasoning"],"falsifier":"On a held-out suite of perception- and reasoning-intensive VQA items, remove or scramble the Bloom stages while keeping the same expert pool and measure whether the accuracy gains and the measured start-level/evidence-citation statistics disappear.","tokens_in":16155,"feed_emoji":"👁️","tokens_out":612,"duration_ms":7422,"temperature":0.7,"pith_summary":"Large vision-language models still struggle to ground answers in fine visual detail and often apply the same reasoning path to every query. This paper claims that specialized computer-vision models already supply the missing perceptual pieces—object detection, masks, OCR, charts—and that the missing architectural step is an explicit, hierarchical protocol for turning those outputs into evidence and then into staged reasoning. The proposed system first decomposes expert outputs into short, non-inferential statements stored with type and grounding references. It next runs a Bloom-inspired verbalizer that starts at a query-determined cognitive level and walks through list–explain–illustrate–compare–hypothesize–final, citing evidence IDs at each step. A lightweight parser recovers the start level, executed steps, and evidence-use statistics. The authors show that this training-free pipeline improves scores on perception- and reasoning-heavy benchmarks and that different questions really do enter the hierarchy at different levels, making the model’s reasoning trajectory measurable.","feed_headline":"Bloom steps turn vision experts into citeable LVLM reasoning","feed_subtitle":"Training-free pipeline lifts perception and reasoning scores and makes evidence use measurable","key_machinery":"Two-stage cognitive verbalization: a Literal Evidence Summary that atomizes expert outputs into typed, citeable statements, followed by Bloom Verbalization that converts those statements into an ordered, evidence-cited reasoning trace consumed by the final LLM; a Reasoning Trace Module then recovers start level, step set, and evidence-use statistics.","core_discovery":"Imposing an evidence-grounded, hierarchical (Bloom-inspired) reasoning protocol on multi-expert LVLM inference produces more verifiable and analyzable multimodal behavior and yields measurable gains on perception- and reasoning-intensive real-world scene-understanding benchmarks, all without additional training.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Bloom hierarchy turns vision experts into verifiable LVLM traces","Evidence-grounded Bloom steps make LVLM reasoning citeable","Multi-expert outputs become analyzable via Bloom verbalization","Training-free Bloom protocol lifts perception and reasoning scores","Cognitive experts plus hierarchical traces reduce LVLM hallucination"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That the hand-designed Bloom step order and template-based atomic verbalizers produce faithfully non-inferential evidence that a frozen base language model can reliably use at inference time without any further training.","fun_headline_variants_meta":{"raw":{"variants":["Bloom hierarchy turns vision experts into verifiable LVLM traces","Evidence-grounded Bloom steps make LVLM reasoning citeable","Multi-expert outputs become analyzable via Bloom verbalization","Training-free Bloom protocol lifts perception and reasoning scores","Cognitive experts plus hierarchical traces reduce LVLM hallucination"]},"model":"grok-4.5","effort":"low","cost_usd":0.003908,"raw_usage":{"total_tokens":1224,"prompt_tokens":763,"num_sources_used":0,"completion_tokens":62,"cost_in_usd_ticks":39080000,"prompt_tokens_details":{"text_tokens":763,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":399,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":763,"tokens_out":62,"duration_ms":5468,"temperature":1.0,"reasoning_tokens":399,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T09:11:37.745545+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out suite of perception- and reasoning-intensive VQA items, remove or scramble the Bloom stages while keeping the same expert pool and measure whether the accuracy gains and the measured start-level/evidence-citation statistics disappear.","supporting_citations":[],"review_version":1}