{"id":"98b34b4f-3dfd-4a89-87a7-b17b895077ef","arxiv_id":"2508.13672","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"ITL-LIME replaces random perturbations with real instances selected by clustering and contrastive learning to improve LIME explanations in low-resource settings.","lead":"The paper introduces ITL-LIME, a method that uses real data from a related source domain to make LIME explanations more reliable when training data is scarce. It aims to reduce the instability and unrealistic perturbations that standard LIME produces on small datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Source-prototype similarity may not transfer across domains; surrogate weighting by contrastive proximity is unvalidated as a proxy for local model behavior, undermining the core fidelity claim.","rationale":"The reader's verdict is UNVERDICTED because only the abstract is available. Our stress-test identifies the same core assumption as the reader's weakest_assumption: that source-cluster prototype similarity and contrastive-learning proximity are reliable indicators that retrieved source instances are locally informative for the target model. We agree this is the most load-bearing assumption because if it fails, the entire method collapses to a potentially harmful form of negative transfer, producing explanations that are less faithful than standard LIME. However, our concern does not change the verdict: the abstract provides no evidence either way, so the appropriate status remains UNVERDICTED. We are not claiming the method is wrong; we are specifying a concrete test that would validate the transfer assumption and thus the central claim. The reader's weakest_assumption captures the same issue, so our agreement is 'agree.'","tokens_in":784,"tokens_out":3457,"duration_ms":37086,"concrete_test":"Construct a synthetic binary classification task with known ground truth, e.g., two interleaving moons, and train a black-box model (e.g., a neural network) on a small target sample. Create a related but shifted source domain (e.g., rotate the moons by 30° or shift the center). For a fixed set of target instances near the decision boundary, compute ITL-LIME explanations using source clusters/prototypes and contrastive weighting, and compare against (a) LIME with random perturbations and (b) LIME with only target neighbors. Measure explanation fidelity by the agreement of the surrogate's local gradient with the true model gradient (or by deletion/insertion AUC on the target instance). If ITL-LIME does not outperform both baselines on fidelity and stability across replicates, the prototype-transfer assumption fails. Run the same experiment with several shift magnitudes to see when transfer","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ITL-LIME improves explanation fidelity and stability in low-resource settings by replacing random perturbations with real source instances. The load-bearing step is the retrieval rule: the source domain is partitioned into clusters, and for a given target instance the method retrieves instances from the cluster whose prototype is most similar to the target. This presupposes that prototype similarity in the clustering representation space (which is presumably the raw feature space or a learned representation) is a sufficiently reliable indicator that the retrieved source instances lie close to the target's local data manifold and that the target black-box model's local decision boundary there approximates the behavior around the target. Two unstated conditions must hold: (1) cross-domain transfer is valid—i.e., the source domain's feature-label relationship aligns with the target domain's in the retrieved local region; and (2) the contrastive weighting mechanism assigns high weight to exactly those instances that are informative for the surrogate, rather than merely those that are feature-proximate. If the source domain has even mild covariate shift or the contrastive encoder ignores decision-relevant dimensions, the retrieved instances can be off-manifold or from a different mode, and the surrogate may be fit to misleading data. In that case, ITL-LIME's fidelity could be worse than standard LIME with target-only neighbor sampling. The abstract provides no experiment or theoretical argument to rule this out. Because the entire method hinges on this transfer assumption, it is the most load-bearing concern. This is not a claim of internal inconsistency; it is a request for evidence that the proposed retrieval actually improves local surrogates under the stated low-resource conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ITL-LIME, an extension of LIME for low-resource data settings. Instead of generating random perturbations, it retrieves real source-domain instances from the source cluster whose prototype is most similar to the target instance, combines them with neighboring real target instances, weights the combined set with a contrastive-learning encoder, and trains the LIME surrogate on this weighted set. The abstract claims that this approach enhances explanation fidelity and stability in data-constrained environments, but no experimental results, baseline comparisons, or methodological details are presented in the available text.","tokens_in":1110,"tokens_out":3222,"duration_ms":35242,"significance":"If validated, the core idea addresses a well-known limitation of LIME: random perturbations can produce unrealistic, off-manifold neighbors, especially when training data are scarce. Using real source instances and instance-based transfer learning is a plausible and timely direction for XAI in small-data domains. The proposed pipeline is concrete and potentially reproducible, and a systematic evaluation would be of value to the community. However, the abstract alone does not establish the central claim; the contribution currently rests on the promise of an empirical improvement rather than on demonstrated evidence.","major_comments":[{"comment":"The central claim that ITL-LIME 'enhances explanation fidelity and stability' is an empirical assertion, but the manuscript reports no experiments, baselines, datasets, metrics, or statistical comparisons. As written, the claim is unsupported. The paper should include a controlled evaluation against standard LIME and, ideally, other local explanation methods, with ablations that isolate the contribution of each component (clustering-based retrieval, source-instance combination, and contrastive weighting).","section":"Abstract (overall claim)"},{"comment":"The retrieval rule assumes that source instances from the cluster whose prototype is most similar to the target instance are useful for approximating the target model's local decision boundary. This is a non-trivial cross-domain transfer assumption: source and target may have different feature-label relationships, and prototype similarity in the clustering representation may not track local model behavior. The authors should provide evidence, either theoretical or empirical, that this selection yields useful perturbations under covariate shift or other domain shifts.","section":"Abstract (retrieval rule)"},{"comment":"The weighting mechanism is underspecified. If the contrastive encoder is trained using the same target instances that later define the neighborhood or are used in evaluation, the weighting can be self-fulfilling and inflate apparent fidelity or stability. The authors must specify the training data, the contrastive objective, and the evaluation protocol, and should validate that contrastive proximity is a proxy for local model behavior rather than merely feature proximity, e.g., by comparing against simple kernel or distance-based weighting.","section":"Abstract (contrastive weighting)"},{"comment":"The method as described depends on several free parameters and unspecified design choices: the number of source clusters k, the prototype similarity metric, the contrastive learning hyperparameters, the size of the combined set, and the proportion of source versus target instances. The abstract also does not define how 'neighboring real instances' are selected or what constitutes a 'compact locality'. These choices must be made explicit for the method to be reproducible and for the claimed benefit to be attributable to the mechanism rather than to parameter tuning.","section":"Abstract (reproducibility and parameters)"}],"minor_comments":[{"comment":"'Prototype' is not defined; it could mean a cluster centroid, medoid, or a learned representation. Please clarify.","section":"Abstract (terminology)"},{"comment":"'Fidelity' and 'stability' are used without formal definitions. The full paper should state the exact metrics used to measure these quantities.","section":"Abstract (metrics)"},{"comment":"The abstract does not position ITL-LIME relative to existing LIME variants or transfer-learning XAI methods. The full paper should include a discussion of prior work on stabilizing LIME and on cross-domain explanations.","section":"Abstract (related work)"}],"recommendation":"major_revision","confidential_remarks":"The present review is based only on the abstract; no full text was available. The central claim is empirical and currently unsupported. I recommend that the authors be asked to provide the full manuscript with experimental results before a final decision is made. If the full paper already contains experiments, this report should be updated accordingly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a methods paper with a sensible idea—use real instances from a related source domain instead of random perturbations for LIME's surrogate, with a contrastive encoder to weight locality. If the full experiments hold up, it's a useful contribution to the low-resource XAI niche. But the abstract asserts the payoff without showing any numbers, so don't take the headline claim at face value.\n\nWhat's genuinely new: most LIME variants focus on better perturbation schemes or kernel weighting within the target domain; borrowing real instances from a source domain via clustering and prototype similarity is not something I've seen as a standard baseline. The motivation is also fair—LIME's random perturbations do go off-manifold when data is scarce, and that's a real failure mode worth attacking.\n\nWhat the paper does well, from the abstract: it identifies a concrete problem, proposes a mechanism that addresses it directly, and the pipeline (cluster source domain, retrieve by prototype similarity, combine with target neighbors, weight by a contrastive encoder) is coherent. There's no obvious internal contradiction.\n\nNow the soft spots. The big one is evidence: no results, no baselines, no datasets, no code mentioned in the abstract. The claim that ITL-LIME 'enhances explanation fidelity and stability' is exactly the thing that needs support. The stress-test concern about transfer is well placed: prototype similarity in raw feature space is not the same as being near the target's local decision boundary in the black-box model. If the source domain has mild covariate shift, the retrieved instances could be off-manifold in the target, and the contrastive weighting might reward feature proximity rather than surrogate usefulness. That could make fidelity worse than standard LIME. The abstract gives no argument or experiment to rule that out. It's not a fatal flaw, but it's the load-bearing assumption and it needs direct testing.\n\nA second, minor worry: the 'combined set' construction and contrastive encoder could be fitted in a way that makes the evaluation circular—if the same instances inform both the encoder and the surrogate assessment, the fidelity numbers could be self-fulfilling. Again, a request for experimental hygiene, not a confirmed problem.\n\nWho's this for? Researchers working on LIME, especially in medical or sensor domains with small labeled sets. It's a niche contribution, but honest progress. I'd send it to peer review, not desk reject it, because the idea is clear and the failure mode it targets is real. The referee should ask for strong empirical comparison against standard LIME, random perturbation with manifold constraints, and other transfer-based baselines, plus ablations on the clustering and contrastive components.\n\nBottom line: worth a serious look, but the abstract alone doesn't earn the central claim.","headline":"Plausible transfer-learning twist on LIME, but the abstract alone can't support the fidelity/stability claim; worth sending to review if the full paper has real experiments.","tokens_in":1606,"tokens_out":2027,"would_cite":false,"duration_ms":20482,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ITL-LIME replaces LIME's random perturbations with real source-domain instances to keep local explanations stable when data is scarce.","keywords":["explainable AI","LIME","local explanations","transfer learning","instance-based transfer","data scarcity","contrastive learning","surrogate models"],"falsifier":"Run ITL-LIME on a target dataset with a few hundred labeled instances and a source domain deliberately unrelated to the target (e.g., images vs. tabular health records). If explanation fidelity, measured as the surrogate's agreement with the black-box model on held-out target neighbors, is no better than classic LIME with random perturbations, then the transfer premise fails.","tokens_in":725,"feed_emoji":"🔍","tokens_out":1842,"duration_ms":21796,"temperature":0.7,"pith_summary":"The paper proposes ITL-LIME, a variant of LIME that addresses the instability and poor fidelity of local explanations when the target training set is small. Instead of sampling random perturbations that can fall off the true data manifold, ITL-LIME retrieves real instances from a related source domain, weights them by proximity to the target instance using a contrastively trained encoder, and trains the local surrogate on those combined instances. The authors aim to show that this instance-based transfer-learning approach yields more faithful and stable explanations in data-constrained settings. If it works, it would make LIME usable where data scarcity currently undermines its core promise of locally faithful explanation.","feed_headline":"Real instances replace random noise to stabilize LIME on small data","feed_subtitle":"A transfer-learning variant retrieves nearby source-domain samples to train LIME's surrogate, improving fidelity and stability.","key_machinery":"The central machinery is the instance-retrieval and weighting pipeline: clustering the source domain into prototypes, selecting the cluster closest to the target instance, pulling real instances from that cluster, combining them with target-side neighbors, and weighting the combined set with a contrastive-learning encoder. The contrastive encoder is the component that defines a compact locality, assigning higher weight to instances whose learned representation is close to the target. This weighted real-instance set is what replaces the random perturbation neighborhood in classic LIME.","core_discovery":"ITL-LIME introduces instance transfer learning into the LIME framework to improve explanation fidelity and stability in low-resource data settings. The method partitions the source domain into clusters with representative prototypes, then, for a target instance, retrieves real source instances from the cluster whose prototype is most similar to that target instance. These source instances are combined with the target instance's neighboring real instances, and a contrastive-learning-based encoder assigns weights to all instances based on their proximity to the target. The weighted set trains the surrogate model that produces the explanation. The core claim is that using real, weighted instanc","pith_inferences":["A natural test is whether prototype similarity alone is a reliable guide to local model behavior; source clusters may be close in representation space yet far apart in the black-box function's decision boundary, which would weaken the transfer.","The method could be extended to actively select which source instances to include, rather than taking a whole cluster, by adding a model-fidelity criterion to prune irrelevant instances.","A direct comparison against other LIME stabilization strategies, such as sampling from a learned density model of the target data, would clarify whether the benefit comes from transfer or simply from using real points instead of random ones."],"forward_implications":["LIME's explanations could become reproducible and trustworthy in applications with small datasets, such as medical or financial settings where labeled examples are scarce.","The method decouples the explanation process from the need to generate synthetic perturbations, so the surrogate is trained only on data points that actually exist in a related domain.","If the contrastive encoder indeed captures local behavioral similarity, the weighted locality could generalize to other local surrogate explanation methods beyond LIME.","Source-domain data becomes a reusable asset for explanation quality, not just for model training."],"supporting_citations":[],"fun_headline_variants":["Real data, not noise, steadies LIME on small samples","Transfer learning swaps LIME's random noise for real instances","LIME gets real: source-domain samples boost small-data explanations","Replace LIME's random perturbations with real neighbor instances","ITL-LIME: borrow real instances to stabilize LIME in low-resource settings"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The method assumes that source-domain instances pulled from the cluster whose prototype is closest to the target instance lie near the target's local data manifold in a way that makes them useful perturbations for training the local surrogate, and that this cross-domain proximity transfers without explicit validation.","fun_headline_variants_meta":{"raw":{"variants":["Real data, not noise, steadies LIME on small samples","Transfer learning swaps LIME's random noise for real instances","LIME gets real: source-domain samples boost small-data explanations","Replace LIME's random perturbations with real neighbor instances","ITL-LIME: borrow real instances to stabilize LIME in low-resource settings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1412,"prompt_tokens":784,"completion_tokens":628,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":539}},"tokens_in":528,"tokens_out":628,"duration_ms":6503,"temperature":1.0,"reasoning_tokens":539,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:56:36.815146+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ITL-LIME on a target dataset with a few hundred labeled instances and a source domain deliberately unrelated to the target (e.g., images vs. tabular health records). If explanation fidelity, measured as the surrogate's agreement with the black-box model on held-out target neighbors, is no better than classic LIME with random perturbations, then the transfer premise fails.","supporting_citations":[],"review_version":1}