{"id":"c995d0fd-a09c-4c9b-9986-c4d8b492caad","arxiv_id":"2506.03799","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ConText is the first visual in-context learning model for text removal and segmentation, chaining the two tasks and using self-prompting to reach new state-of-the-art scores.","lead":"ConText is a vision model trained to erase text and find text in images by learning from a single example prompt, chaining the two tasks together. It reports large gains over existing OCR specialists and over generic in-context models, and the authors release code.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The generalized in-context learnability claim in Section 5.3 is confounded: the 'GT demonstration' appears to be the query's own label, which is exactly what self-prompting trains on, while the baselines were not trained with self-prompting.","rationale":"The paper has real strengths: it releases code, reports broad comparisons, and the random-demonstration gains are large and consistent across datasets. I do not object to the SOTA claim per se. My concern targets the distinct claim that self-prompting preserves generalized in-context learnability. The evaluation condition used to demonstrate that claim is the same condition used by the training technique, namely an identical demonstration-query pair. That makes the RS-versus-GT gap a poor instrument for measuring generalization, and the ablation confirms that self-prompting, not task-chaining or context-aware aggregation, is what creates the gap. This is not an accusation of data fabrication; it is an uncontrolled confound in the protocol, and it is fixable with a controlled baseline and a non-self GT evaluation. The reader's pseudo-label point partially overlaps because it also questions the evidence base for the headline results. I recommend keeping the reader's CONDITIONAL verdict unchanged, with the added condition that the learnability experiment be re-run with a self-prompting-treated baseline and with a GT demonstration that is not the query image. If those tests show the gap persists only for ConText, the in-context learnability claim would be substantially strengthened.","tokens_in":24337,"tokens_out":12075,"duration_ms":112806,"concrete_test":"Add the SP-0.2 recipe alone to the SegGPT-to-Removal+Segmentation baseline and run the same RS-versus-GT protocol; simultaneously re-run ConText with a GT demonstration drawn from a different image (not the query) whose labels are ground truth. If the baseline's RS-versus-GT gap rises to ConText's level, or if ConText's gap collapses when the GT demonstration is not the query itself, then the Figure 5 and Table 5 evidence is an artifact of the self-prompting protocol rather than evidence of generalized in-context learnability. The released evaluation code should also state explicitly whether 'GT demonstration' was the query's own image.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Self-prompting (Section 4.3) trains ConText on two identical input-output pairs with probability 0.2. The learnability protocol in Section 5.3, Figure 5, and Table 5 measures performance with a 'ground-truth (GT) demonstration'. Appendix B.2, Table 11 treats this as an upper bound. If 'GT demonstration' means the demonstration pair is the query image with its own ground-truth label, then this is exactly the self-prompting input distribution on which ConText was trained, but it is out-of-distribution for the Painter and SegGPT baselines. Table 5 shows that this is the operative factor: without self-prompting the RS-versus-GT gap is only +0.65 fgIoU / +0.37 PSNR, while adding SP-0.2 raises it to +3.98 fgIoU / +1.42 PSNR. Because the paper claims self-prompting prevents the model from 'devolving into specialist-like, context-free inference', the large gap is the central evidence for generalized in-context learning, but it may instead reflect memorization of the self-prompting condition. This does not invalidate the random-demonstration SOTA numbers, but it does undermine the ICL-specific claim. The pseudo-label concern identified by the reader is real and is even acknowledged in the paper's Impact Statement, but it is secondary: those labels affect training data, not the learnability measurement protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ConText, a visual in-context learning (V-ICL) framework for text removal and segmentation. Three components are introduced: a task-chaining prompt that combines image, removal, and segmentation labels; a context-aware aggregation (CAA) module that fuses cross-demonstration information; and a self-prompting (SP) training strategy that occasionally feeds the model two identical input-output pairs. Experiments on multiple text removal and segmentation benchmarks report state-of-the-art results for random-demonstration inference (average PSNR 38.36 over three removal benchmarks; average fgIoU 76.77 over four segmentation benchmarks) and further claim generalized in-context learnability, evidenced by a large performance gap when the demonstration is the ground-truth label instead of a randomly selected sample.","tokens_in":24696,"tokens_out":7067,"duration_ms":71800,"significance":"If the central claims hold, this is a useful contribution to V-ICL for OCR: it is the first OCR-focused V-ICL generalist, it releases code, and it reports strong improvements over both generalist baselines and task-specific specialists on standard benchmarks. The random-demonstration SOTA numbers are supported by extensive comparisons and appear independent of the learnability-protocol confound discussed below. However, the paper's broader claim that self-prompting preserves generalized in-context learnability rests on a GT-demonstration measurement that may be in-distribution for ConText but out-of-distribution for the baselines, so the ICL-specific conclusion needs additional controls or a reinterpretation before the contribution is fully established.","major_comments":[{"comment":"The ground-truth demonstration protocol is not precisely defined, and the reported numbers indicate a confound. If the 'ground-truth-based demonstration' is the query image paired with its own ground-truth label, then this input is exactly the self-prompting condition described in Section 4.3 ('two identical input-output pairs'), which ConText was trained on with probability 0.2 but the Painter and SegGPT baselines were not. Table 5 shows that the RS-versus-GT gap on TotalText segmentation and SCUT-EnsText removal grows from +0.65 fgIoU / +0.37 PSNR without SP to +3.98 fgIoU / +1.42 PSNR with SP-0.2. This pattern is consistent with the model memorizing the identical-pair training condition rather than with a general ability to reason from arbitrary GT demonstrations. Please (a) define the protocol explicitly, (b) add a control with a non-identical but ground-truth demonstration, and (c) either train baselines under the same self-prompting schedule or reinterpret the gap as an upper-bound measurement rather than evidence against 'context-free inference'.","section":"Section 5.3, Figure 5, Table 5, Appendix B.2"},{"comment":"The training targets for ConTextV include pseudo-labels: removal labels for TextSeg are generated by ViT-Eraser, and segmentation masks for SCUT-EnsText and SCUT-Syn are produced by gray-scale distance between original and erased images. The Impact Statement itself acknowledges that these labels can introduce noise and bias. Because ConTextV is trained on these pseudo-labeled datasets and then compared with specialists on TextSeg segmentation (Table 2) and SCUT-EnsText removal (Table 3), a systematic bias in the pseudo-labels could inflate the reported specialist-level results. Please quantify the quality of the pseudo-labels (for example, IoU against human masks on a subset) or provide an ablation that removes the pseudo-labeled datasets; without such evidence, the specialist comparison does not fully isolate the effect of the model design from the effect of the training labels.","section":"Appendix B.1 and Impact Statement"},{"comment":"The claim that PromptText is 'not training-involved' is inaccurate for ConTextV and for any baseline trained on TextSeg, because PromptText is constructed from the TextSeg validation set. Since ConTextV is trained on TextSeg training data, the evaluation is not fully held-out for that model. The HierText-trained ConText row provides a partially clean check, but the section should either restrict the claim to ConText or evaluate PromptText on a dataset that is disjoint from all training sets; the current wording overstates the generalization evidence.","section":"Section 5.3, PromptText, Appendix B.4"}],"minor_comments":[{"comment":"The bullet lists '+4.50PNSR'; 'PNSR' should be 'PSNR.'","section":"Section 1 (contribution bullet)"},{"comment":"The table headers 'RS / GT' are ambiguous: clarify that the first value is the random-selected score and the second value is the delta from the GT-demonstration score, not an absolute GT score.","section":"Section 5.4, Table 5"},{"comment":"The rows reporting ground-truth-demonstration results are not visually labeled; add a caption note or row labels (e.g., 'GT demo') so readers can distinguish them from random-demonstration rows.","section":"Appendix B.2, Tables 11-13"},{"comment":"The phrase 'a sounding upper and lower bounds' should be 'a sound upper and lower bound' (or 'sound upper and lower bounds').","section":"Figure 5 caption"},{"comment":"The sentence 'the intuitivelinear fusion (eF1) yields a significant improvement (+3.61%) fgIoU for segmentation and +1.33 PSNR for removal)' has a mismatched parenthesis and should be reworded for clarity.","section":"Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"The random-demonstration SOTA results are credible and well supported, and the released code is a plus. The main issue is the interpretation of the GT-demonstration gap in Section 5.3, which currently conflates in-distribution self-prompting behavior with generalized in-context learning. I recommend major revision rather than rejection, because the central claim can be repaired with additional controls or a more careful claim, and the primary empirical contribution appears sound."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before reading further. First, the headline results are real: with randomly selected demonstrations, ConText beats the prior generalists and the task-specific specialists on text removal and segmentation across several benchmarks. Second, the paper's central claim about preserving generalized in-context learning is weakened by a measurement confound that the authors don't address.\n\nWhat's genuinely new: this is the first OCR-specific visual in-context learning framework. The task-chaining prompt (image-removal-segmentation), the context-aware aggregation module, and the self-prompting training strategy together form a coherent and empirically successful recipe. The ablation study is thorough, and the comparisons to both fine-tuned generalists and specialists are broad and fair. The random-demonstration improvements (+4.50 PSNR, +3.34 fgIoU over the best generalist baselines) are substantial and worth taking seriously.\n\nThe soft spots, in proportion. The in-context learnability evidence in Section 5.3 is the load-bearing problem. The ground-truth demonstration they use as an upper bound is the query image paired with its own label. Self-prompting trains the model on exactly this distribution 20% of the time, while the baselines never see it. Table 5 shows the RS-vs-GT gap jumps from +0.65 fgIoU / +0.37 PSNR without self-prompting to +3.98 / +1.42 with SP-0.2. So the large gap they cite as evidence of in-context reasoning mostly reflects the model memorizing the self-prompting condition, not a clean measure of learnability. The random-demonstration SOTA numbers survive this criticism; the ICL-specific claim does not.\n\nThe pseudo-label issue is real but secondary. Removal labels for TextSeg come from ViT-Eraser, and segmentation masks for SCUT-EnsText/SCUT-Syn come from gray-scale differencing. The Impact Statement acknowledges potential noise and bias, but the authors don't quantify how this affects the SOTA numbers. I'd treat the specialists comparison as slightly optimistic.\n\nWho this is for: people working on OCR or on visual in-context learning more broadly. The framework is interesting, and the results will likely be reproduced and built upon. The paper deserves a serious referee, but the authors need to re-frame the ICL claim: either drop the 'upper bound' framing or run a cleaner test where the GT demonstration is not the query's own label (e.g., a held-out oracle pair from the training set). My recommendation is to send it to peer review, with the expectation of a major revision on that specific issue.","headline":"Solid SOTA numbers for OCR text removal and segmentation, but the in-context learnability evidence is confounded by the self-prompting training trick.","tokens_in":25157,"tokens_out":1594,"would_cite":true,"duration_ms":16038,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"ConText claims that adapting visual in-context learning to OCR by chaining text removal and segmentation into a single prompt yields state-of-the-art results on both task families.","keywords":["visual in-context learning","text removal","text segmentation","task chaining","self-prompting","context-aware aggregation","scene text","masked autoencoder"],"falsifier":"Train ConText with the same recipe but replace every pseudo-label with human-annotated removal and segmentation targets on the same images, then re-run the same benchmarks; if the reported margins collapse, the gains were partly fitting the pseudo-label teacher's bias. A second check: evaluate ConText with a demonstration drawn from a different script or style than the query, and if performance barely moves, the model has lost the demonstration sensitivity that self-prompting is claimed to preserve.","tokens_in":24183,"feed_emoji":"🔤","tokens_out":9558,"duration_ms":85802,"temperature":0.7,"pith_summary":"This paper tries to establish that visual in-context learning—solving a task from example image-label pairs placed in the prompt—can be made to work for fine-grained OCR tasks such as removing text from photographs and segmenting text into foreground masks. Earlier in-context vision models used a single image-label pair and forced the model to jump straight from input to output in one reconstruction step. ConText instead chains two related tasks into one prompt—original image, text-removed image, and segmentation mask—so the model reasons through an intermediate representation, much as chain-of-thought prompting helps language models. On top of that prompt design, the paper adds a cross-attention module that lets the two demonstrations inform each other and a self-prompting training trick that keeps the model sensitive to the demonstration. The paper reports state-of-the-art numbers on both tasks and takes the large gap between ground-truth and random demonstrations as evidence that genuine in-context learning, rather than context-free specialization, is happening.","feed_headline":"Chained in-context prompts beat OCR specialists on both tasks","feed_subtitle":"One chained prompt lets a single generalist model outdo task-specific specialists on text removal and segmentation.","key_machinery":"The load-bearing object is the task-chaining compositor: the prompt is built as a 2×3 grid of [original image, removed image, segmentation mask] covering both the demonstration pair and the query. During training both the removal and the segmentation labels are partially masked with spatially aligned masks, and a weight-shared decoder reconstructs both, so the model can exploit the fact that both outputs describe the same text regions. Two supporting mechanisms carry the rest: context-aware aggregation (CAA), a lightweight shared cross-attention module that enriches each label representation with features drawn from the other demonstration, and self-prompting, which with probability 0.2 builds the input from two identical pairs so the model keeps relying on the demonstration rather than becoming a task specialist. The backbone is a ViT-L masked autoencoder initialized from a pre-trained in-context segmentation checkpoint.","core_discovery":"The central claim is that restructuring the in-context prompt from a single image-label pair into a task chain—original image, removed image, segmentation mask—lets one masked-autoencoder-style model perform text removal and segmentation in a single forward pass, and do so better than models fine-tuned for either task alone. The argument leans on a simple identity: the segmentation mask is exactly the visual difference between the original image and its erased version, so chaining the tasks gives the model an intermediate to reason through instead of demanding a direct one-step mapping. The paper attributes the remaining margin to context-aware aggregation, a shared cross-attention module that injects each demonstration's features into the other's label representation, and to self-prompting, which with probability 0.2 feeds two identical input-output pairs so the model cannot collapse into a demonstration-independent specialist. On the paper's numbers, ConText reaches a PSNR of 38.36 averaged over three text removal benchmarks and an fgIoU of 76.77% averaged over four segmentation datasets, outperforming both in-context generalists and task-specific specialists, and it retains a visible performance gap between ground-truth and randomly sampled demonstrations.","pith_inferences":["If the intermediate-reasoning story is right, then adding a third linked task to the chain—say, text detection or recognition—should further improve removal and segmentation; this is a direct and cheap test of the task-chaining mechanism.","Because the removal labels for TextSeg come from another eraser model and the segmentation masks for SCUT-EnsText and SCUT-Syn come from gray-scale differencing, the reported margins may partly encode that teacher's biases; retraining with purely human annotations would measure that contamination.","The fixed self-prompting probability of 0.2 trades off in-context sensitivity against task accuracy, so a schedule that anneals the probability during training, or adapts it per sample, might dominate the fixed setting.","The watermark-removal result suggests the same chain composer transfers beyond OCR to any foreground-erasure task pair, such as shadow or reflection removal, where an erasure output and a foreground mask are naturally paired."],"forward_implications":["Any pair of visual tasks joined by a shared logical identity—here, that the segmentation mask equals the difference between an image and its erased version—can be chained into a single prompt to lift performance on both tasks.","In-context models trained on visually heterogeneous domains need explicit pressure, such as self-prompting, to stay demonstration-sensitive; without it, the model degrades into a context-free specialist.","The growing gap between ground-truth and random demonstration performance as training data scales indicates that the model's gains come partly from genuine in-context inference, and that this capability improves with more data.","A single randomly chosen demonstration suffices for the model to beat task-specific specialists, suggesting that costly demonstration retrieval is not needed for OCR-style in-context tasks.","The model can follow explicit human-style visual prompts—colored circles, boxes, and strokes drawn on the query image—without ever training on such annotations, which the paper presents as evidence of generalized in-context understanding."],"supporting_citations":[{"why":"Defines the composited-prompting MAE inpainting paradigm that ConText starts from as its baseline.","marker":"(Bar et al., 2022)"},{"why":"Supplies the masked-autoencoder mask-reconstruction backbone for the whole pipeline.","marker":"(He et al., 2022)"},{"why":"Painter, the single-task in-context generalist that serves as a no-fine-tuning and fine-tuned baseline across benchmarks.","marker":"(Wang et al., 2023b)"},{"why":"SegGPT, both the main generalist baseline and the source of the pre-trained initialization for ConText.","marker":"(Wang et al., 2023d)"},{"why":"Provides the label-anchor account of in-context learning that motivates the context-aware aggregation module.","marker":"(Wang et al., 2023a)"},{"why":"Supports the claim that label positions absorb demonstration information, the design principle behind the CAA module.","marker":"(Yu & Ananiadou, 2024)"},{"why":"Chain-of-thought prompting, the language-side analogue that motivates chaining related visual tasks.","marker":"(Wei et al., 2022)"},{"why":"ViT-Eraser, used to generate removal pseudo-labels for TextSeg and as a task-specific specialist baseline.","marker":"(Peng et al., 2024a)"},{"why":"Supplies the removal ground-truth labels for HierText and TotalText used in training.","marker":"(Zhu et al., 2024)"}],"fun_headline_variants":["Chained in-context prompts beat OCR specialists","Task-chaining prompts for text removal and segmentation","ConText: Chained prompts yield state-of-the-art OCR","One model, two OCR tasks: chain prompts to win"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the training labels are trustworthy, and several are not human-annotated: as disclosed in Appendix B.1, removal images for TextSeg are produced by an external eraser model and segmentation masks for SCUT-EnsText and SCUT-Syn are computed by gray-scale differencing, so part of the reported gains could be fitting those pseudo-labels rather than genuine in-context reasoning.","fun_headline_variants_meta":{"raw":{"variants":["Chained in-context prompts beat OCR specialists","Task-chaining prompts for text removal and segmentation","ConText: Chained prompts yield state-of-the-art OCR","One model, two OCR tasks: chain prompts to win"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000543,"raw_usage":{"total_tokens":2635,"prompt_tokens":1015,"completion_tokens":1620,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":1556}},"tokens_in":631,"tokens_out":1620,"duration_ms":13672,"temperature":1.0,"reasoning_tokens":1556,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:54:58.757693+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train ConText with the same recipe but replace every pseudo-label with human-annotated removal and segmentation targets on the same images, then re-run the same benchmarks; if the reported margins collapse, the gains were partly fitting the pseudo-label teacher's bias. A second check: evaluate ConText with a demonstration drawn from a different script or style than the query, and if performance barely moves, the model has lost the demonstration sensitivity that self-prompting is claimed to preserve.","supporting_citations":[{"cited_title":"Visual Text Generation in the Wild","cited_arxiv_id":"2407.14138","evidence_quote":"Supplies the removal ground-truth labels for HierText and TotalText used in training."}],"review_version":1}