{"id":"f8e585ca-60f4-45fe-b62e-883be46bf7a9","arxiv_id":"2607.25106","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Web-image-enriched queries improve open-vocabulary map grounding and ObjectNav success, especially for fine-grained categories, with gains limited by downstream detectors.","lead":"IMPRINT improves zero-shot object-goal navigation by enriching text queries with web images matched against semantic maps. It also releases HSSD-rare, a long-tail ObjectNav benchmark, and shows detection quality is the main bottleneck after better grounding.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The headline grounding gains are averaged over a plausibly misconfigured SigLIP text baseline; per-encoder gains on the well-tuned baselines are small (2–4 SR) and reported without any statistical uncertainty, so \"consistent improvement\" rests on thin deltas.","rationale":"The reader identified retrieval quality as the weakest assumption — a real concern, but the paper partially self-mitigates it (metadata gating, dedup, n-image ablation in Fig. 5 showing saturation rather than collapse, and a limitations paragraph). I find a different, more load-bearing soft spot: the quantitative headline itself. Averaging across encoders hides that the large gains come from a SigLIP text-only baseline that performs so poorly (2–8 SR) it is likely misconfigured rather than representative, and the gains on the two credibly tuned encoders are 2–7 SR on OVON-syn and only ~2 SR on HSSD-rare, with no episode counts or uncertainty estimates anywhere in the paper. This does not break the paper — the direction of the effect is consistent across six rows, the detector-intervention analysis (Table III) is honestly reported, and the authors disclose the modest HSSD-rare online gains — but it does mean the central quantitative claim is weaker and less general than the abstract's phrasing. This is exactly the kind of issue the reader's CONDITIONAL verdict already anticipates (\"statistical uncertainty is unreported... addressable issues\"), so my read reinforces rather than revises that verdict. The concrete test is cheap: one re-run of the SigLIP baseline with standard prompting plus bootstrap CIs, and the claim's true strength becomes clear.","tokens_in":13948,"tokens_out":2335,"duration_ms":23827,"concrete_test":"Re-run Table I with two changes: (a) SigLIP text queries embedded using the standard prompt ensemble (e.g., \"a photo of a {category}\" templates averaged) rather than the raw category string; (b) report episode counts per cell and bootstrap 95% CIs on the image-minus-text SR delta per encoder. If tuned SigLIP text-only rises well above 7.74/2.15, the headline average gains shrink accordingly; and if the BLIP2 and SED deltas on HSSD-rare (+2.58, +2.18 SR) have CIs crossing zero at the actual n, the claim of consistent grounding improvement holds only for specific encoders, not in general.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim cites average gains of +10.54 SR (OVON-syn) and +6.37 SR (HSSD-rare) for image- vs text-only queries (Table I). But those averages are dominated by the SigLIP rows: text-only SigLIP scores 7.74 SR on OVON-syn and 2.15 on HSSD-rare — absurdly low next to BLIP2's 38.71/17.11 on the identical maps and protocol. A text-only SigLIP grounding pipeline that fails on >90% of episodes where BLIP2 succeeds strongly suggests the text baseline is under-configured (e.g., raw category string without the prompt ensembling SigLIP-style encoders typically require), not that text is intrinsically uninformative. Excluding SigLIP, the gains are +3.55 (BLIP2) and +7.42 (SED) on OVON-syn, and only +2.58 (BLIP2) and +2.18 (SED) on HSSD-rare. The long-tail claim therefore rests on ~2-point deltas with no reported episode counts, no variance, and no significance testing — and Sec. VI-A says episodes are \"balanced across categories\" without giving n. Meanwhile the ablation's own OVON-seen row shows image conditioning adds just +1.01 SR, so the method's benefit is clearly regime-dependent. The claim \"consistently improves isolated object grounding\" survives for image-vs-that-text-baseline, but the magnitude and generality of the improvement are inflated by one anomalous baseline and the residual deltas are within plausible noise. This is a baseline-validity and statistical-power concern, distinct from the reader's retrieval-noise concern.","agreement_with_reader":"partial"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The paper introduces IMPRINT, a training-free, plug-and-play module for open-vocabulary ObjectNav built on queryable semantic maps: given a text goal, it retrieves N web images (gated by a head-noun metadata check), encodes them with the same VLM used for mapping, computes per-image cosine-similarity maps over the map, and averages them for goal localization. It is evaluated in a static grounding phase (BLIP2, SigLIP, SED encoders; OVON-syn and the new HSSD-rare benchmark) and an online navigation phase (ZSON, VLFM, OneMap), with ablations over aggregation rule, image count, and OVON splits; a failure analysis; a controlled detector intervention (IMPRINT*); and a 5-episode LoCoBot demo. The authors claim consistent grounding and navigation gains, larger under fine-grained specificity, and identify object detection as the key bottleneck in the long-tail regime. HSSD-rare contributes 1000 episodes over 20 parent categories / 559 subcategories in 17 scenes, with a documented viewpoint-generation pipeline.","tokens_in":24506,"tokens_out":1991,"duration_ms":1196575,"significance":"If the numbers hold after baseline cleanup, IMPRINT is a useful, low-overhead addition to open-vocabulary ObjectNav: training-free, no policy modification, demonstrated across three encoders and three navigation stacks, with gains consistent in direction on two benchmarks. Two elements have value beyond the method itself: HSSD-rare, a genuinely new subcategory-level ObjectNav benchmark (1000 episodes, 559 fine-grained subcategories, documented viewpoint generation, OOV and LLM-scored rarity analyses), and the controlled detector intervention (IMPRINT*) that cleanly separates grounding gains from detection bottlenecks — an informative systems result. Regime-dependence is reported honestly (OVON-seen +1.01 SR). The method itself (retrieve–encode–average) is simple, so the benchmark and the grounding-vs-detection decomposition are the lasting contributions.","major_comments":[{"comment":"The pooled gains quoted in the text (+10.54 SR on OVON-syn, +6.37 on HSSD-rare) are dominated by the SigLIP text-only rows (7.74 and 2.15 SR), which trail BLIP2 text-only (38.71/17.11) on identical maps and protocol by >30 points. SigLIP-style encoders are normally strong zero-shot classifiers when standard prompt ensembling is used, so this pattern suggests an under-configured text baseline (e.g., raw category strings without prompt templates) rather than an intrinsic deficit of text queries. Excluding SigLIP, per-encoder gains are +3.55 (BLIP2) and +7.42 (SED) on OVON-syn, and +2.58/+2.18 on HSSD-rare — positive but much smaller than the headline. Please disclose the per-encoder text-prompting protocol, re-report SigLIP with standard prompt ensembling, and present per-encoder deltas in the text rather than pooled averages that hinge on one anomalous row.","section":"§VI-C, Table I"},{"comment":"No evaluation reports episode counts or statistical uncertainty. The OVON-syn static evaluation is described only as 'balanced across categories' (§VI-A) without n, and several online deltas are small enough to be sampling noise at typical ObjectNav episode counts (e.g., ZSON +0.47 SR and OneMap +0.47 SR on HSSD-rare, Table II). Please state n for every table (static and online, including the subsampling protocol for OVON splits) and report paired confidence intervals (e.g., bootstrap over episodes) or per-category success rates, at least for Tables I–III. The 'consistent improvement' claim currently rests partly on deltas within plausible noise.","section":"§VI-A/C, Tables I–III"},{"comment":"Two online configurations are not reproducible from the text. (a) IMPRINT* 'conditions YOLOWorld with the top-3 retrieved reference images,' but YOLOWorld is a text-prompted detector and §IV describes only map-query enrichment; the image-conditioning mechanism, detection thresholds, and top-3 selection rule are never specified, yet Table III carries the 'detection bottleneck' contribution. (b) ZSON builds no semantic map, so the similarity-map aggregation of §IV-B does not apply; please describe how image embeddings condition the ZSON policy (e.g., an averaged goal embedding) and whether N and the gating rule match the map-based setting.","section":"§VI-C (Table III), §IV-D (Table II)"},{"comment":"The method's premise — that metadata-gated web retrieval yields sufficiently relevant exemplars — is never measured directly. The n-ablation (saturation then decline) is consistent with retrieval noise, and the gating step itself is not ablated. A small relevance audit (e.g., human or VLM relevance labels on retrieved sets for a sample of categories) plus a gated-vs-ungated comparison would substantiate the premise. Relatedly, please release the cached retrieved image sets: search-engine drift otherwise makes the headline numbers difficult for others to reproduce.","section":"§IV-A, Fig. 5"}],"minor_comments":[{"comment":"The caption states the encoder is 'fixed as BLIP2,' but the legend shows BLIP2, SigLIP, and SED; please reconcile. Also state whether n>15 was tested — the HSSD-rare optimum sits at the edge of the swept range (n=15).","section":"Fig. 5"},{"comment":"The harmonic mean is computed over cosine similarities, which can be negative; the harmonic mean is well-defined only for positive values. Please clarify any shift/clamping — otherwise the harmonic/hybrid collapse (e.g., SigLIP HSSD-rare SR 1.23) may be a numerical artifact rather than evidence about aggregation quality.","section":"Table V"},{"comment":"'github.io/IMPRINT' is not a resolvable repository URL. Please provide the actual link and specify the release scope (HSSD-rare episodes, generated viewpoints, cached retrievals, evaluation configs) and license.","section":"Abstract / §I"},{"comment":"With five episodes, 1/5 vs 4/5 is not a meaningful quantitative comparison; please frame the LoCoBot study strictly as a qualitative proof-of-concept and, if possible, include per-episode details (goal, outcome, failure cause).","section":"Appendix C"},{"comment":"On OVON-syn, IMPRINT* raises SR from 13.00 to 18.17 but worsens DTG (4.53 → 4.98). A one-sentence explanation (e.g., additional successes on farther goals) would preempt confusion.","section":"Table III"},{"comment":"SR is defined via the stop action (§III), which never occurs in the static phase; please define static-phase SR explicitly (top-ranked map cell within 1 m of the target) and state the map grid resolution, since 1 m success depends on cell size.","section":"§VI-A / §IV-B"},{"comment":"Terminology: 'OVON-syn,' 'OVON-synon.' (Table IV), and 'val seen synonyms' are used interchangeably; please unify. IMPRINT* is first used in Table III without a defining sentence in the text.","section":"§VI-A/C"},{"comment":"Category-wise gains (treadmill, clothing, bottle, table lamp) are described but not tabulated; a per-category table in the appendix would make the 'broadly distributed' claim checkable.","section":"§VI-C"},{"comment":"Formatting artifacts: '2,m', '1,m', '0.5,m' in Appendix A and several broken inline math expressions (e.g., observation dimensions in §IV-B) should be repaired in the final PDF.","section":"Appendix A / throughout"}],"recommendation":"major_revision","confidential_remarks":"The core method is a straightforward retrieve–encode–average scheme, so novelty on the algorithmic side is modest; the real contributions are the HSSD-rare benchmark and the grounding-vs-detection decomposition, both of which I expect to be used by others. My main reservation is evaluation hygiene rather than correctness: the headline pooled averages lean on a likely under-configured SigLIP text baseline, and no table carries uncertainty estimates. All issues raised look fixable within the manuscript's scope (clarified protocols, a SigLIP re-run, CIs, and method-detail paragraphs), and I see no reason to doubt the direction of the reported effects. The code link in the abstract is a placeholder; the editor may wish to verify the actual artifact release before final acceptance."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a clean systems paper. Image-conditioned web exemplars as a drop-in map query help long-tail ObjectNav grounding, they ship HSSD-rare with a real viewpoint pipeline, and they correctly flag detection—not the map query—as the thing that kills end-to-end gains. That package is useful even if the magnitude story is softer than the abstract implies.\n\nWhat is new is the combination, not the ingredients. Web retrieval for long-tail recognition and queryable VLM maps both exist. Running multi-image cosine maps through existing ObjectNav stacks (VLFM, OneMap, ZSON) with no policy change, plus building subcategory episodes from HSSD with a careful visibility-checked viewpoint sampler, is the actual contribution. Static tables across three encoders, aggregation and image-count ablations, failure breakdown, the YOLOWorld conditioning intervention (IMPRINT*), and a small LoCoBot run are the right experiments. Code and data are promised. Citation pattern is normal.\n\nSoft spots, in proportion. The stress-test is right on the averages: +10.5 / +6.4 SR are dominated by SigLIP text-only at 7.7 and 2.2 SR, which is not a credible text baseline next to BLIP2/SED on the same protocol. Drop SigLIP and you still see consistent but small lifts (~2–7 SR static; online HSSD-rare gains are a point or two until the detector is also image-conditioned). No episode counts, no variance, no significance. Web gating is brittle and N is free. None of that breaks the claim that images help and that detection gates transfer; it does mean “consistently improves” should be read as directionally true, not as a large effect size.\n\nWho it is for: people building open-vocab ObjectNav stacks and anyone who needs a finer-grained Habitat benchmark than OVON-syn. Worth a serious referee. I would engage—use the benchmark, try the plug-in, and quote the detection bottleneck—without treating the averaged SR deltas as gospel.","headline":"Solid plug-and-play engineering plus a real long-tail benchmark; headline grounding averages are inflated by a broken SigLIP text baseline, but the direction of the result and the detection-bottleneck diagnosis still hold.","tokens_in":15474,"tokens_out":540,"would_cite":true,"duration_ms":18798,"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":"Adding web images to text queries improves long-tail object finding in mapped environments without retraining the navigator.","keywords":["Object Goal Navigation","queryable semantic maps","vision-language models","image-conditioned queries","long-tail recognition","zero-shot navigation","HSSD-rare","web image retrieval"],"falsifier":"On HSSD-rare-style subcategory episodes, replace retrieved images with random or off-topic images (or disable gating) and check whether success rate and distance-to-goal still beat text-only; if they do not fall, the claimed benefit of image conditioning is not doing the work.","tokens_in":15277,"feed_emoji":"🗺️","tokens_out":920,"duration_ms":19275,"temperature":0.7,"pith_summary":"Queryable semantic maps let robots find objects from open-vocabulary text, but a category name alone often fails when the target is a fine-grained or product-level subcategory. IMPRINT enriches that text query with a handful of web-retrieved images, encodes them with the same vision-language model used for the map, and averages the resulting similarity maps to pick a better place to go. The method is zero-shot and plug-and-play: it does not train anything or change the navigation policy. On a synonym-heavy split and on a new long-tail benchmark built from richly annotated synthetic scenes, image-conditioned queries raise isolated grounding success and produce end-to-end navigation gains. The paper further shows that those localization gains only fully become navigation success when the downstream object detector is also strong, identifying detection quality as a systems bottleneck for long-tail embodied search.","feed_headline":"Web images lift long-tail object finding without retraining","feed_subtitle":"Averaging map similarity from retrieved photos beats text-only queries; detection remains the bottleneck.","key_machinery":"IMPRINT: retrieve N web images for the query (with head-noun metadata gating), encode them, produce one cosine-similarity map per image over the stored feature map, and average those maps (optionally with the text map) to obtain the target localization used by an unchanged navigation stack.","core_discovery":"Enriching a text object query with web-sourced image embeddings, matching each against a queryable semantic map, and aggregating the similarity maps consistently improves zero-shot object grounding and yields navigation gains on both synonym-level and true subcategory long-tail ObjectNav, without any training or policy change; converting those gains fully into success still depends on downstream detection quality.","pith_inferences":["Joint image conditioning of both the map query and the detector may be the practical default for long-tail embodied search, not an optional ablation.","Retrieval quality and domain shift of web images could become a first-class failure mode when the same method moves from synthetic scenes to messy real homes.","Patch-level encoders help on common categories but may need different aggregation or depth filtering when targets are rare product variants.","Viewpoint-generation pipelines like the one used for HSSD-rare could be reused to stress-test other open-vocabulary navigation stacks on the same subcategory ontology."],"forward_implications":["Text-only queryable maps underperform as soon as targets move from broad classes to fine-grained subcategories; image exemplars close much of that gap.","The same plug-in can sit on map-based pipelines (frontier similarity plus PointGoal) and on direct policy conditioning without redesigning the controller.","Isolated grounding gains do not automatically equal navigation gains; detector quality becomes the binding constraint under long-tail conditions.","Conditioning the open-vocabulary detector on the same retrieved images unlocks further end-to-end success on both OVON-syn and HSSD-rare.","HSSD-rare supplies a reusable testbed of co-occurring subcategory targets for measuring true long-tail ObjectNav rather than synonym rewording alone."],"fun_headline_variants":["Web images enrich queries for long-tail ObjectNav gains","Image-conditioned maps beat text-only in zero-shot ObjectNav","IMPRINT aggregates photo similarities for rarer object grounding","No-train image queries improve long-tail navigation success","Detection quality limits how map gains become nav wins"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Web search plus simple metadata filtering must return visual examples that are relevant enough that averaging their embeddings against the map is a trustworthy location signal rather than noise.","fun_headline_variants_meta":{"raw":{"variants":["Web images enrich queries for long-tail ObjectNav gains","Image-conditioned maps beat text-only in zero-shot ObjectNav","IMPRINT aggregates photo similarities for rarer object grounding","No-train image queries improve long-tail navigation success","Detection quality limits how map gains become nav wins"]},"model":"grok-4.5","effort":"low","cost_usd":0.002171,"raw_usage":{"total_tokens":933,"prompt_tokens":748,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":21708000,"prompt_tokens_details":{"text_tokens":748,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":121,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":748,"tokens_out":64,"duration_ms":3855,"temperature":1.0,"reasoning_tokens":121,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T01:01:08.404846+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On HSSD-rare-style subcategory episodes, replace retrieved images with random or off-topic images (or disable gating) and check whether success rate and distance-to-goal still beat text-only; if they do not fall, the claimed benefit of image conditioning is not doing the work.","supporting_citations":[],"review_version":1}