{"id":"212131cb-fa2d-4b37-a88f-d31940b2be31","arxiv_id":"2603.12824","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A 69M text-only student encodes queries for visual document retrieval while a frozen 2B VLM indexes documents offline, retaining ~95% teacher quality via cosine distillation and translated queries.","lead":"NanoVDR distills a 2B vision-language document indexer into a ~70M text-only query encoder for visual document retrieval. It keeps most teacher quality while cutting parameters and CPU query latency by large factors, making VDR practical without GPUs at query time.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Abstract-only review cannot verify the load-bearing claim that pointwise cosine alignment on pre-cached teacher query embeddings is sufficient for a text-only student to retrieve against VLM-indexed visual documents.","rationale":"The Reader correctly identified the load-bearing design choice (pointwise cosine on pre-cached teacher query embeddings, no document processing) and correctly set CONDITIONAL / LOW confidence because only the abstract is available. No stronger internal inconsistency is visible from the abstract; the concern is missing verification of that choice and of the reported numbers, not a contradiction with consensus. Full-text tables, the exact objective definitions, and a reproducible training script would settle it. Until then the Reader's verdict stands: useful if true, but not yet checkable. No adjustment is warranted.","tokens_in":2098,"tokens_out":580,"duration_ms":4673,"concrete_test":"Obtain the full paper (or code) and re-run the six-objective ablation on the same three backbones and 22 ViDoRe sets, confirming that (1) only pre-cached teacher query embeddings are used as targets, (2) no document images or OCR text enter student training, and (3) evaluation is against the frozen VLM visual document index. If pointwise cosine still wins and the 95.1% retention / DSE-Qwen2 outperformance hold under those constraints, the concern is discharged; if any objective that uses documents or ranking closes the gap, or if the numbers drop when the index is strictly visual, the headline claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency result (NanoVDR-S-Multi retains 95.1% of the 2B teacher and beats DSE-Qwen2 on ViDoRe v2/v3) rests on one design choice: that pointwise cosine alignment of student query embeddings to pre-cached teacher query embeddings (text-only path of the frozen VLM) is a sufficient and stable distillation target, without any document-side processing or ranking/contrastive loss during student training. The abstract asserts this after comparing six objectives, but supplies no numbers, no description of how the six objectives were defined or controlled, no confirmation that teacher embeddings were frozen and text-only, and no evidence that the student actually retrieves against the teacher's visual document index rather than a text-only proxy. Without those details, the asymmetry argument and the 95.1%/outperformance numbers cannot be checked for hidden leakage (e.g., student seeing document text, teacher query path not matching the document path, or MT augmentation contaminating the evaluation splits). This is exactly the weakest assumption the Reader flagged; the abstract alone does not secure it.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes NanoVDR, which exploits query–document asymmetry in visual document retrieval: a frozen 2B VLM teacher indexes documents offline, while a distilled text-only student (as small as DistilBERT, 69M) encodes queries at inference. The abstract reports a systematic comparison of six distillation objectives across three backbones and 22 ViDoRe datasets, concluding that pointwise cosine alignment of student query embeddings to pre-cached teacher query embeddings is superior to ranking-based and contrastive alternatives and requires no document processing during student training. Cross-lingual transfer is identified as the main bottleneck and addressed via machine-translated query augmentation. The resulting NanoVDR-S-Multi is claimed to retain 95.1% of teacher quality, outperform DSE-Qwen2 (2B) on ViDoRe v2 and v3, use 32× fewer parameters and 50× lower CPU query latency, and train in under 13 GPU-hours.","tokens_in":2411,"tokens_out":1008,"duration_ms":17276,"significance":"If the reported retention, outperformance, and efficiency numbers hold under full experimental scrutiny, the work is a practically significant contribution to visual document retrieval. Decoupling a strong visual document indexer from a tiny text-only query encoder directly targets a real deployment cost (multi-billion-parameter query encoding for plain-text queries). The systematic objective comparison, the low training budget, and the concrete efficiency claims (parameter count, CPU latency, GPU-hours) are strengths that would make the result useful and falsifiable for the community. The design is outside the usual symmetric dual-encoder template but is not, on its face, internally inconsistent.","major_comments":[{"comment":"Abstract (central design claim): The load-bearing assertion that pointwise cosine alignment on pre-cached teacher query embeddings is a sufficient training signal for a text-only student to retrieve against a VLM’s visual document index cannot be verified from the abstract alone. The manuscript must define the six objectives, state how they were controlled (same data, same backbones, same evaluation protocol), confirm that teacher embeddings are frozen and come from the text-only query path, and show that evaluation is against the teacher’s visual document index rather than a text-only proxy. Without those details, the 95.1% retention and outperformance numbers rest on an unchecked assumption.","section":null},{"comment":"Abstract (results claim): The reported 95.1% teacher retention and outperformance of DSE-Qwen2 (2B) on ViDoRe v2 and v3 are the paper’s main empirical claims. The full manuscript must supply per-dataset or per-version tables, the exact metric (e.g., nDCG@k), error bars or multi-seed variance, and a clear statement that evaluation splits are free of machine-translation leakage. Absent that evidence, the efficiency narrative (32× parameters, 50× CPU latency, <13 GPU-hours) cannot be accepted as established.","section":null},{"comment":"Abstract (cross-lingual bottleneck): Identifying cross-lingual transfer as the primary bottleneck and resolving it with machine-translated query augmentation is presented as decisive. The paper must document the MT policy (languages, volume, source of translations), show ablations with vs. without MT, and demonstrate that translated queries do not contaminate held-out ViDoRe evaluation sets. This is load-bearing for the Multi variant’s claimed gains.","section":null}],"minor_comments":[{"comment":"Abstract: The six objectives are named only generically (pointwise cosine vs. ranking-based and contrastive). Naming them explicitly in the abstract or early methods would help readers assess the comparison scope before the full tables.","section":null},{"comment":"Abstract: “retains 95.1% of teacher quality” should specify the aggregate metric and whether it is micro- or macro-averaged over the 22 datasets.","section":null},{"comment":"Abstract: “50× lower CPU query latency” should state the hardware and batch size used for the latency measurement so the figure is reproducible.","section":null}],"recommendation":"uncertain","confidential_remarks":"Only the abstract was available for this review; the full text was not provided. The recommendation is therefore uncertain rather than a substantive accept/revise/reject on the completed manuscript. The design is interesting and the efficiency claims, if substantiated, would fit a strong IR/systems venue. Once the full paper is in hand, the three major points above (objective comparison protocol, evaluation against the visual index, MT contamination controls) should be checked first; if they hold with transparent tables and ablations, the bar for acceptance is primarily presentation and completeness rather than conceptual repair."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing: this paper claims you can freeze a 2B VLM as a document indexer and distill a 69M text-only query encoder that keeps 95.1% of teacher quality on ViDoRe, beats a 2B baseline, and runs 50× faster on CPU, all for under 13 GPU-hours. If that holds, it is a practical systems win for multimodal IR.\n\nWhat looks new is not the asymmetry idea itself—query/document decoupling and retriever distillation are familiar—but the systematic finding that pointwise cosine alignment to pre-cached teacher query embeddings beats ranking and contrastive objectives for this setting, plus cheap MT query augmentation to fix cross-lingual transfer. The abstract reports that comparison across six objectives, three backbones, and 22 datasets, and the efficiency numbers are concrete. That is the right kind of engineering result if the tables back it up.\n\nThe soft spot is exactly what the stress-test flags, and it is real because we only have the abstract. The whole story rests on the claim that matching teacher query embeddings (text path only) is enough for the student to retrieve against the teacher’s visual document index, with no document-side training and no ranking loss. The abstract asserts this after the six-objective sweep, but we cannot see the objective definitions, controls, whether the teacher path was truly frozen and text-only, or whether evaluation is cleanly against the visual index rather than a text proxy. MT augmentation also needs a clean train/eval split story. None of that is circularity; it is missing evidence. Free parameters (student size, hyperparams, MT policy) are normal for this kind of paper and not a red flag by themselves.\n\nWho it is for: people building production visual document retrieval who care about query latency and CPU deployment. A serious referee should see the full paper, tables, and preferably code. I would not desk-reject on the abstract; the claimed recipe and measured outcome are specific enough to deserve a real review. I would bring it to reading group only after the full text is up, and I would not cite it yet.","headline":"Useful asymmetric distillation recipe for visual document retrieval, but abstract-only so the load-bearing cosine-alignment claim is still unverified.","tokens_in":2995,"tokens_out":525,"would_cite":false,"duration_ms":5377,"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 69M text-only query encoder retains 95% of a 2B vision-language retriever by distilling only on text queries.","keywords":["visual document retrieval","vision-language models","knowledge distillation","query-document asymmetry","pointwise cosine alignment","cross-lingual transfer","ViDoRe","efficient retrieval"],"falsifier":"Train the same 69M DistilBERT student with the identical pointwise cosine objective but without machine-translated query augmentation, then measure nDCG on the multilingual ViDoRe splits; a large drop relative to the multi-lingual student would show that the reported quality depends critically on the translation fix rather than the distillation objective itself.","tokens_in":2980,"feed_emoji":"🔍","tokens_out":654,"duration_ms":5953,"temperature":0.7,"pith_summary":"This paper argues that visual document retrieval does not need a multi-billion-parameter vision-language model for both document indexing and query encoding. Documents are visually complex and benefit from a large frozen VLM teacher that indexes them offline, but queries are short text strings that can be handled by a much smaller text-only student. The authors show that the right distillation target is simply pointwise cosine alignment of student query embeddings to pre-cached teacher query embeddings, without any document processing or ranking losses during student training. Cross-lingual gaps are the main remaining bottleneck and are closed cheaply by machine-translating the training queries. The resulting 69M DistilBERT student keeps 95.1% of the 2B teacher’s quality, beats a 2B baseline on ViDoRe v2 and v3, and delivers roughly 50× lower CPU query latency at under 13 GPU-hours of total training cost.","feed_headline":"69M text encoder keeps 95% of a 2B visual document retriever","feed_subtitle":"Pointwise cosine distillation plus translated queries cuts parameters 32× and CPU latency 50×","key_machinery":"Pointwise cosine alignment on query text alone: the student is trained only to match the teacher’s text-query embeddings (pre-cached, no document images or ranking/contrastive losses needed during student training). This objective consistently beat five ranking- and contrastive-based alternatives across three backbones and 22 ViDoRe datasets.","core_discovery":"A text-only student as small as 69M parameters, trained solely by pointwise cosine alignment to a frozen 2B VLM teacher’s pre-cached query embeddings and augmented with machine-translated queries, retains 95.1% of teacher quality on visual document retrieval and outperforms a 2B VLM baseline while cutting parameters by 32× and CPU query latency by 50×.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["69M text encoder retains 95% of 2B visual document retriever","Text-only 69M student matches 95% teacher quality on VDR","Pointwise cosine distillation yields 69M encoder with 95% retention","NanoVDR cuts parameters 32× while keeping 95% of 2B quality","69M DistilBERT student beats 2B VLM baseline on visual retrieval"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"That matching a large VLM’s text-only query embeddings is enough for a tiny text encoder to retrieve against documents that were indexed with full visual understanding, without any document-side training signal.","fun_headline_variants_meta":{"raw":{"variants":["69M text encoder retains 95% of 2B visual document retriever","Text-only 69M student matches 95% teacher quality on VDR","Pointwise cosine distillation yields 69M encoder with 95% retention","NanoVDR cuts parameters 32× while keeping 95% of 2B quality","69M DistilBERT student beats 2B VLM baseline on visual retrieval"]},"model":"grok-4.5","effort":"low","cost_usd":0.004572,"raw_usage":{"total_tokens":1358,"prompt_tokens":846,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":45720000,"prompt_tokens_details":{"text_tokens":846,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":423,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":846,"tokens_out":89,"duration_ms":3806,"temperature":1.0,"reasoning_tokens":423,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T22:05:35.824708+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the same 69M DistilBERT student with the identical pointwise cosine objective but without machine-translated query augmentation, then measure nDCG on the multilingual ViDoRe splits; a large drop relative to the multi-lingual student would show that the reported quality depends critically on the translation fix rather than the distillation objective itself.","supporting_citations":[],"review_version":1}