We organized the Second Workshop on Sign Language Processing (WSLP-2026), co-located with EMNLP 2026 in Budapest, Hungary, on 29 October 2026. The workshop continues the programme begun at WSLP 2025: to give under-resourced sign languages a dedicated computational venue, and to treat sign language processing as a set of related problems rather than a single translation benchmark. We received 40 submissions and accepted 22 papers. This overview concentrates on the 15 accepted main-workshop research papers. Those papers cover new corpora for Catalan, Uruguayan, Icelandic, Russian, and Bangla sign languages; compact gloss-free translation; isolated recognition and fingerspelling; and a cluster of diagnostic studies on what actually transfers when a model is moved from a high-resource sign language to a low-resource one. As part of the workshop, we hosted three shared tasks on Indian Sign Language, namely translation, isolated recognition, and word/sign presence prediction, which together attracted 58 registered teams, and 19 teams actively participated. The open questions have shifted from whether sign languages can be modeled at all to how annotation budgets should be allocated, which parts of a pretrained checkpoint are worth transferring, and how to evaluate systems when automatic metrics remain only weakly correlated with grammatical correctness.

WSLP-2026 hosted three shared tasks on Indian Sign Language: sentence-level translation from video or pose into English (SLT), isolated recognition over a large one-shot vocabulary (ISLR), and word/sign presence prediction in continuous sentence video (WPP). Shared tasks were hosted on Codabench. Twenty-three teams registered for translation and five submitted; thirteen registered for WPP and five submitted; twenty-two registered for recognition and nine submitted. On the lexical tracks, the systems that report usable official scores treat the problem as retrieval or temporal alignment within a pose-embedding space, often with an external exemplar bank. Translation submissions almost uniformly map pose sequences into a pretrained T5 decoder. They obtain fluent English more readily than scores that can be trusted at the current accuracy level: official BLEU can sit at zero, while chrF remains small and sensitive to output length. The edition shows that ISL can be measured by automatic evaluation metrics, but that current scores still say more about retrieval recipes and decoder fluency than about signed grammar. A later edition should mark pose-tracker failures and place native-signer judgments alongside BLEU and chrF.

Sign Language Translation (SLT) is a key accessibility technology for Deaf communities. Although recent years have seen rapid progress in SLT, Indian Sign Language (ISL) and other sign languages of under-resourced regions remain under-researched and overshadowed by sign languages with a longer research history and stronger infrastructure. In this direction, we present ISLiM (ISL in Miniature), a compact end-to-end keypoint-based gloss-free SLT architecture that maps body-pose landmark sequences directly to spoken-language text. At its core, a stack of temporal convolutional layers captures low-level pose motion patterns, a Transformer-based encoder models higher-order kinematic interactions and sign-level temporal ordering, and a pretrained T5-small decoder with a prior grasp of the target language performs the semantic heavy-lifting of generating fluent text. We further explore complementary loss functions that sustain training on small datasets, and alternative qualitative evaluation metrics tailored to the particular challenges of SLT. ISLiM achieves a BLEU-4 score of 5.85 on iSign, establishing state-of-the-art results among reported keypoint-based models on the only large-scale ISL dataset. We also show promising transferability to German Sign Language on Phoenix-2014-T through modular decoder swap, pointing toward cross-lingual extensibility where both sign and spoken languages can be swapped independently. A preliminary mobile deployment study confirms that the full keypoint-based model runs on a consumer Android smartphone with ~384ms mean inference latency on CPU, demonstrating a practical path toward offline, accessible assistive tools for the Deaf community in India and the world. Code is available at https://github.com/ahmedshoaib/ISLiM

We describe our submission to the WSLP 2026 shared task on Isolated Sign Language Recognition. The task is one-shot retrieval over 4,361 Indian Sign Language word classes with roughly one training clip per class. Our final system placed 3rd on top-1 accuracy at 0.6825 and 2nd on top-10 at 0.7357. The system is deliberately simple. It combines MediaPipe keypoints, a graph attention backbone pretrained on a different ISL corpus, PCA whitening, and nearest-exemplar cosine retrieval against an enriched gallery. Our main finding concerns how two very differently priced interventions relate. PCA whitening of the frozen cross-corpus encoder is worth +18.85 points on a leave-one-out gallery evaluation. Fine-tuning that encoder end-to-end is worth a comparable amount. The two do not compose. Once the encoder has been fine-tuned, whitening adds +0.38 points, an interaction of −18.5. Whitening and fine-tuning act as substitutes because both correct the same embedding anisotropy, and a transform that is fitted in seconds captures most of what a day of GPU training buys. We also report fifteen interventions that produced no usable gain: twelve indistinguishable from zero and three actively harmful. These were evaluated with McNemar tests on a 2,030-query held-out split with a minimum detectable effect of 1.23 points. We argue that the 526-clip official test set is too small to support the comparisons participants routinely make with it.

Released sign-to-text checkpoints combine visual and text-side parameters, so intact-checkpoint comparisons cannot identify what drives transfer. We study this confound in low-resource Taiwan Sign Language (TSL) by separating and recombining the visual and mT5 components of Uni-Sign checkpoints trained on Chinese Sign Language (CSL)-to-Chinese and American Sign Language (ASL)-to-English data, then fine-tuning the combinations identically on TSL. The CSL-derived visual initialization improves content-word accuracy by 15.0 points with its released mT5 parameters, but changes accuracy by only -0.5 points with untouched mT5-base, showing that the measured benefit of visual pretraining depends on the text-side initialization. We assess video dependence by replacing target clips with duration-matched held-out clips at evaluation. Further interventions localize much of the accuracy and video-dependence effects to the untied output projection: replacing an ASL-to-English model’s projection with mT5-base’s improves accuracy by 25.0 points and increases sensitivity to clip replacement. Permutation and timing controls show that the effect depends on pretrained row-to-token correspondence and on the projection being present during fine-tuning. Absolute translation quality remains low, so we interpret these experiments diagnostically. The results show that released checkpoints should be analyzed as structured initializations whose components can interact, rather than treated as indivisible evidence of cross-language transfer.

Deaf signers face persistent communication barriers in healthcare settings, where reliance on third-party interpreters can compromise privacy precisely when confidentiality matters most. We present LSC-CARE, the first annotated dataset of Catalan Sign Language (LSC) for the medical domain, comprising 2,784 signed segments (approximately 4.5 hours) produced by 11 deaf native signers across ten emergency-care topics. Each segment is paired with time-aligned ID-gloss annotations produced by a deaf native linguist and written Catalan translations produced by an LSC interpreter, following conventions consistent with the LSC SignBank. To assess the dataset’s utility for Sign Language Translation, we benchmark three settings: a gloss-free end-to-end video-to-text model, a cascade of sign language recognition and gloss-to-text translation, and an oracle gloss-to-text upper bound. The cascade consistently outperforms the end-to-end approach despite a high gloss error rate, and, unlike it, is statistically invariant to signer identity, while the 21-BLEU gap between cascade and oracle identifies visual sign recognition as the main bottleneck for low-resource medical SLT. We release the dataset to support accessible healthcare technologies for the Catalan Deaf community.

We address cross-signer Isolated Sign Language Recognition (ISLR) for Indian Sign Language as part of the WSLP-2026 shared task, where most target words have exactly one available training video and a system must recognize a query video performed by a signer never seen for that class during training. We approach this with a metric-learning encoder trained in two phases: pretraining on a large, multi-signer American Sign Language corpus to learn general signer-invariance, then fine-tuning on the target vocabulary using every real cross-signer recording we can responsibly source. Our preprocessing detects and extracts multiple performances of a word within a single recording as separate training samples, a technique not previously reported in SLR literature, and adapts a published missing-keypoint interpolation technique into a retrieval rather than classification pipeline. Our final system achieves 67.11% Top-1, 69.20% Top-5, and 70.34% Top-10 accuracy on the WSLP-2026 test-phase evaluation.

Sign spotting, the task of deciding whether an isolated query sign occurs inside a clip of continuous signing, enables sign language video retrieval without gloss or text annotation. We describe our system for the WSLP@EMNLP-2026 Word/Sign Presence Prediction shared task on Indian Sign Language: a lightweight (2.86M-parameter) pose-based Siamese encoder that embeds both clips in one space, forms an explicit frame-to-frame cosine similarity map, and reduces it with multiple-instance pooling over temporal windows. Training combines binary cross-entropy with a multiple-instance contrastive objective (MIL-NCE). The system reaches 72.69 accuracy and 71.95 macro-F1 on the blind test set, against 61.66 and 58.42 for the published pose baseline on the same data. Ablations identify contrastive alignment as the component that makes the architecture work: the MIL-NCE objective is what places the two clips in a comparable metric space, and training does not leave chance without it. We further document a label-balance property of the released splits that fixes 76.3% of blind test labels without a model, and recommend a change for future editions.

We present the first experiments on isolated sign language recognition (ISLR) for Icelandic Sign Language (ÍTM). We use ÍTM SignWiki, a dataset derived from a bilingual Icelandic–ÍTM online dictionary. It is genuinely low-resource: 1,845 videos cover 849 classes, 86% of which have only two examples, making the full task effectively one-shot recognition across signers. We compare two open-source ISLR frameworks, OpenHands and SPOTER, on three tasks of increasing vocabulary size (22, 117 and 849 classes), and evaluate three pose estimators and two forms of cross-lingual transfer. With ÍTM data alone, SPOTER outperforms OpenHands on all three tasks, and MediaPipe poses give better results than AlphaPose or SDPose. Cross-lingual transfer brings the largest gains: pretraining SPOTER on American Sign Language data before finetuning on ÍTM raises accuracy by 14–24 percentage points, to 72.7%, 47.9% and 22.6% on the three tasks, and multilingual training with data from six other sign languages lifts OpenHands from 1.41% to 28.86% on the full task. Although far from practical use, the results suggest that transfer from better-resourced sign languages is promising for very low-resource ones. We release our adapted versions of both frameworks.

Pose-based sign language recognizers represent the two hands as separate halves of a skeleton graph, so a left-dominant and a right-dominant signer produce the same sign as mirror images, and a model must learn every sign in both orientations. Which hand leads, however, is a fact about the signer, not the sign. We therefore canonicalize dominance, mapping the hands onto sign phonology’s dominant/non-dominant roles: any training clip whose left hand carries clearly more motion energy is mirrored – a lossless operation, since the mirror is an exact automorphism of the skeleton. Canonicalized donors trained on Turkish Sign Language transfer to six other sign languages +6.01 accuracy percentage points better than standard donors under full fine-tuning, and +5.19 with frozen features; every canonicalized donor outscores every standard one (exact p = 1/70 over n = 8 donors), the gain is positive on all six recipients, and most of the gain remains even when recipient data is not canonicalized. To locate the source of the gain, we train nine donor-training variants that contain identical data and differ only in which clips are mirrored. Every other variant beats the natural donor, and three that keep the natural left/right proportion recover 57 to 79% of the gain (so a global orientation change is not the whole effect); this narrows the mechanism without identifying it. One variant, a random 20% re-mirroring after canonicalization, outscores the operator itself, so it is sufficient rather than optimal.

The motion-energy comparison itself is fragile: across seven corpora, 0.12–65.1% of clips never have a hand detected, so up to 80.7% of apparently one-handed clips are tracking artifacts rather than signing. Restricting the transfer comparison to clips in which both hands were tracked shrinks the fine-tune gain from +5.72 to +4.77 with full separation (exact p = 1/15 using four standard and two canonical donors). We also release relevant per-corpus validity checks.

Uruguayan Sign Language (LSU) is a low-resource language with a lack of computational models. Building on the first gloss-annotated corpus of LSU interpreter videos, we release LSU-446, a curated dataset of 17,793 sign instances over 446 glosses. Unlike standard benchmarks recorded under controlled conditions, LSU-446 consists of interpreted discourse, with an average sign duration of 403 ms, and heavy coarticulation. To the best of our knowledge, this work establishes the first isolated sign recognition baselines for LSU at scale, comparing a Transformer (SPOTER) and a graph convolutional network (MSG3D); MSG3D reaches 32.7% Top-1 over 446 classes and, unlike SPOTER, degrades only marginally as the vocabulary grows. We then use both architectures for a sign spotting experiment on a held-out video from an unseen signer and domain (33.8% out-of-vocabulary), comparing a segmenter-based pipeline against a sliding window, under an evaluation protocol that decouples temporal detection from lexical recognition.

Expert annotation is a central bottleneck in sign-language corpus construction. A natural assumption is that annotation budgets should prioritize diverse or underrepresented signing patterns. We test this in a controlled retrospective study on PHOENIX14T (German Sign Language) and Isharah (Saudi Sign Language), across gloss recognition, translation, multiple budgets, and three seeds, comparing random sampling against geometric, active-learning, pose-derived articulatory, and privileged selectors. Articulatory coverage is generally weak across all four corpus–task settings and rarely outperforms random sampling. A density analysis shows it selects locally atypical clips even after controlling for duration; removing the sparsest 10–20% helps at larger budgets but never enough to beat random. A privileged gloss-diversity oracle also performs poorly despite selecting comparatively typical clips, showing that atypicality is not the whole explanation. Facility-location representativeness is the strongest feasible selector in three of four largest-budget settings, with lower WER than random in 16 of 18 seed-matched recognition comparisons, compared with only 2 of 18 for articulatory-only selection. The results suggest pose-derived articulatory coverage and gloss-type coverage, as operationalized here, are unreliable proxies for annotation utility in low-resource sign-language corpus construction. The source code is available at https://github.com/gyan-shrestha/WSLP-2026.

The WSLP 2026 shared task studies sentence-level translation from Indian Sign Language pose sequences to English, with training on iSign and evaluation on a disjoint collection of source videos. We address two challenges in this setting: unstable pose scale caused by landmark errors and weak grounding between pose inputs and generated text. We introduce robust clip-level normalization and shared-encoder contrastive grounding, which aligns pose and target-text representations in a pretrained T5 encoder during training while retaining standard autoregressive decoding at inference. On a video-held-out iSign development split, contrastive grounding improves chrF from 0.102 to 0.121, while clip-level normalization further improves it to 0.128; the normalization gain is consistent across three matched seeds. A retrieval probe provides complementary evidence of improved pose–text alignment. Our final shared-task submission achieves chrF 0.16 on the official test set. However, input-independent English controls reach chrF 0.15, showing that small differences in the official score are difficult to interpret at the current level of translation accuracy. The source code is available at https://github.com/gyan-shrestha/WSLP-2026-SLT.

Isolated sign language recognition is commonly formulated as supervised classification, but this formulation is challenging when nearly every target gloss has only one labeled video. We study the Isolated Sign Language Recognition track of the WSLP 2026 Shared Task on Indian Sign Language Processing, a predominantly one-shot target setting with 4,398 training videos spanning 4,361 glosses, over 99% of which are singletons. We introduce CARE, a Coverage-Aware Retrieval Ensemble that augments a complete local visual dictionary with a CISLR reference bank covering a subset of the target vocabulary. A frozen Video-Swin encoder ranks all target labels locally, while frozen I3D and SignRep encoders provide external retrieval evidence for lexically matched labels. CARE estimates orthogonal Procrustes mappings from shared-class centroids, retains both raw and aligned representation geometries, and preserves external videos as individual nearest-neighbor references. Query-wise score standardization and a deterministic coverage mask restrict external evidence to supported labels. The method requires neither target-specific backbone fine-tuning nor a large-vocabulary classification head. CARE obtains 74.31%/76.15%/76.15% Top-1/5/10 accuracy on the WSLP 2026 validation split and 65.59%/67.11%/68.25% on the official test split. The 8.72% Top-1 gap warrants caution when interpreting performance on the small 109-query validation split used for model selection.

This paper investigates the recognition of the Russian fingerspelling alphabet, also known as the Russian Sign Language (RSL) dactyl. Dactyl is a component of sign languages where distinct hand movements represent individual letters of a written language. This method is used to spell words without specific signs, such as proper nouns or technical terms. An alphabet learning simulator is one meaningful application of isolated dactyl recognition. There is a notable issue of data shortage in isolated dactyl recognition: existing Russian dactyl datasets lack subject heterogeneity, contain insufficient samples, or cover only static signs. We provide Bukva, the first open-source video dataset for RSL dactyl recognition that is free of all these shortcomings. It contains 3,757 videos with more than 100 samples for each RSL alphabet sign, including dynamic ones. We utilized crowdsourcing platforms to increase the subjects’ heterogeneity, resulting in the participation of 155 deaf and hard-of-hearing experts in the dataset creation. We use a TSM (Temporal Shift Module) block to handle static and dynamic signs effectively, achieving 83.6% top-1 accuracy with real-time inference on CPU only. The dataset, demo code, and pre-trained models will be made publicly available.

Sign Language Translation (SLT) is especially difficult for low-resource languages such as Turkish Sign Language (TID), where aligned sign-text data are scarce. Self-supervised encoders pretrained on large sign corpora can reduce this dependence, but existing models are developed and evaluated almost exclusively on high-resource sign languages. One such model is SHuBERT (Sign Hidden-Unit BERT), a publicly released self-supervised contextual representation model trained on American Sign Language (ASL). We ask whether SHuBERT can be adapted to TID translation on the open-domain BUTID benchmark. Pretrained SHuBERT is incorporated into an encoder-decoder based translation model and evaluated under both frozen and fine-tuned settings. This is compared against two references trained under the same recipe: a pose baseline, and a raw-feature setting using DINOv2-encoded hand/face crops and pose. All variants outperform the pose baseline, frozen SHuBERT performs on par with raw-feature setting, and fine-tuning yields the strongest result, reaching 7.15 BLEU-4 on the test set (+1.62 over the pose baseline, +0.77 over the raw-feature setting). Our results suggest that reusing a high resource, ASL-pretrained self-supervised encoder can benefit low-resource SLT, but this benefit emerges through fine-tuning rather than frozen transfer alone.

Sign language translation (SLT) can help bridge the communication gap between the Deaf and hard-of-hearing community and the general population. Yet progress has been driven largely by a handful of high-resource sign languages, such as German or American Sign Language, supported by large annotated corpora and large computing budgets. Many sign languages, including Bangla Sign Language (BdSL), remain underrepresented because annotated data is scarce and much research assumes access to powerful hardware. In this paper, we share an early-stage adaptation of the gloss-free SpaMo framework for Bangla, called SpaMo-BdSL, and reflect on what we learned while building it under real resource limits. We report two findings that we believe matter for low-resource and underrepresented SLT. First, a small language-specific model (BanglaT5, 247M parameters) matches the translation quality of a much larger multilingual model (mT0-XL, 3.7B parameters) on our Bangla dataset, while being far cheaper to run; for example, it translates the test set in 40 seconds versus 2 minutes 43 seconds for mT0-XL. Second, a set of simple training changes suggested by a large language model (Claude Opus 4.6) allowed us to train the whole system on a single consumer-grade GPU. We argue that the common assumption ”bigger and more multilingual is better” can quietly exclude underrepresented sign languages and the researchers who work on them, and we use our experience to open a discussion on more inclusive, multilingual, and culturally aware SLT.

Sign language processing advances rapidly for high-resource languages such as American Sign Language (ASL), yet most of the world’s sign languages lack the phonological annotations new methods require. We present the first zero-shot cross-lingual framework for handshape recognition, transferring from ASL to Catalan Sign Language (LSC). Our approach leverages the decomposition of handshapes into five phonological features — selected fingers, flexion, spread, thumb position, and thumb contact — shared across both languages, to decode LSC handshapes from predicted features via a composite phonological distance metric. We evaluate three architectures (MLP, SL-GCN, SHuBERT) trained on two ASL corpora (PopSign, Sem-Lex) against a 37-handshape, single-signer LSC benchmark. Zero-shot transfer proves viable once recording-format disparities are harmonized, reaching 80.0% phonological feature accuracy and 54.5% expected handshape accuracy. Phonological decomposition thus offers a bridge for extending sign language technologies to low-resource languages without any target-language video training labels.

Fingerspelling is a significant component of Sign Language (SL), allowing the interpretation of proper names, characterized by fast hand movements during signing. Although previous works on fingerspelling recognition have focused on processing the temporal dimension of videos, there remains room for improving the accuracy of these approaches. This paper introduces HandReader, a group of three architectures designed to address the fingerspelling recognition task. HandReader_RGB employs the novel Temporal Shift-Adaptive Module (TSAM) to process RGB features from videos of varying lengths while preserving important sequential information. HandReader_KP is built on the proposed Temporal Pose Encoder (TPE) operating on keypoints as tensors. Such keypoint composition in a batch allows the encoder to pass them through 2D and 3D convolution layers, utilizing temporal and spatial information and accumulating keypoint coordinates. We also introduce HandReader_RGB+KP — architecture with a joint encoder to benefit from RGB and keypoint modalities. Each HandReader model possesses distinct advantages: on the ChicagoFSWild and ChicagoFSWild+ datasets, HandReader models achieve the best results among methods trained on a single training set, including a new state-of-the-art result on ChicagoFSWild+.

Fingerspelling — the sequential signing of letters that compose proper names, technical terms, and out-of-vocabulary words — is an integral component of any sign language. Despite Russian Sign Language (RSL) being used by more than 150,000 native signers, no public fingerspelling dataset for RSL exists, in contrast to American, British, German, and several other sign languages. We introduce Znaki the first open RSL fingerspelling dataset: 37,252 video recordings (35.4 hours) of 1,593 unique sequences — proper names sampled from 16 domain categories — recorded by 68 native RSL signers (37 Deaf, 29 hard-of-hearing, and 2 certified interpreters) through a crowdsourcing pipeline with a two-stage signer qualification procedure and expert validation. The dataset covers all 33 Cyrillic handshapes, and its letter-frequency distribution matches the Russian National Corpus. As reference points for future research, we establish two baselines adapted from American Sign Language fingerspelling recognition: a Squeezeformer-based encoder-decoder, reaching a character-level accuracy (1-CER) of 0.932 and a sequence accuracy of 0.596, and Fingerspelling-PoseNet, reaching 0.886 and 0.545, respectively.

Gloss-free sign language translation (SLT) aims to translate sign videos directly into spoken-language sentences without relying on costly gloss annotations. However, when supervision is available only at the sentence level, the correspondence between visual dynamics and linguistic content remains weak. We introduce TransAlign-SLT, a lightweight fully gloss-free framework that derives complementary temporal and semantic supervision from the available translations. Auxiliary TranslationCTC provides coarse temporal regularization, while Masked Entropic Optimal Transport (MOT) aligns visual temporal features with contextualized word-level representations from a frozen, corpus-adapted BERT encoder. Since the semantic branch is used only during training, the inference pipeline remains lightweight and autoregressive. TransAlign-SLT achieves 23.54 BLEU-4 / 48.46 ROUGE-L on PHOENIX14T and 16.39 BLEU-4 / 44.68 ROUGE-L on CSL-Daily, with 45.6M inference parameters and 123 GFLOPs. Ablation studies show that temporal modeling, semantic alignment, and decayed TranslationCTC supervision provide complementary benefits. Overall, the results show that useful visual-language structure can be learned directly from translation supervision without glosses, pseudo-glosses, or a large pretrained language decoder at inference.

Cross-lingual transfer is a standard approach to low-resource sign language translation, but it is unclear what the transferred model actually provides. We fine-tune a pose-based model from American Sign Language (ASL) to Indian Sign Language (ISL) and decompose the resulting transfer across its encoder, decoder, and articulator streams. The matched ASL encoder-decoder pair reaches 3.61 BLEU-4 on ISL, while either component alone performs close to training from scratch. Content-word recall reveals a sharp frequency effect: frequently observed content improves with training and additional data, whereas rare and unseen content remains near zero, even after twenty epochs and a 49% increase in training data. Finally, comparing ASL and Chinese Sign Language (CSL) sources shows that the apparent source-language gap is substantially reduced when the output language is controlled, with remaining differences concentrated differently across articulators. Because the source checkpoints also differ in pretraining domain, we treat this final pattern as suggestive rather than causal. Together, the results show that transfer provides a coupled encoder-decoder starting point, but does not by itself solve the long tail of target-language content.

We present our submissions to the WSLP 2026 Isolated Sign Language Recognition shared task. The task requires predicting the correct English gloss from a short Indian Sign Language video.

Our approach first establishes a strong primary prediction using vision features and identifies uncertain cases using model disagreement and prediction confidence. Rather than allowing all available experts to influence every prediction, we selectively invoke auxiliary models for difficult cases and incorporate their evidence through constrained late-fusion strategies. This enables complementary appearance, pose, and vision-language information to be used when the primary prediction is uncertain while preserving reliable predictions for confident cases. We introduce a disagreement-gated Qwen2.5-VL model as an auxiliary decision-making module and further incorporate complementary appearance and pose experts through several constrained fusion strategies.

The final system combines multiple fusion variants using reciprocal rank fusion and achieves 66.16% Top-1 accuracy, 72.43% Top-5 accuracy, and 73.19% Top-10 accuracy on the blind test set

Recent benchmarks show that multimodal large language models (MLLMs) can answer questions about sign-language video, but accuracy alone does not reveal which visual cues support those answers. We test behavioral sensitivity to hand-localized visual detail using 150 front-view isolated signs from NationalCSL-DP. For each item, a hand-target condition (T) blurs tracked hand regions and a same-arm control (C) applies the same Gaussian operator, nominal σ, and translated mask geometry to nearby arm regions; the underlying regional content is not information-matched. The same four-choice question is used across conditions and four video MLLMs. All 1,800 final Primary responses are uniquely scorable, with 1,200 T/C responses entering the primary contrast. Mean accuracy is 38.17% on original videos (O), 33.50% on T, and 36.67% on C. The pre-specified C–T contrast is +3.17 percentage points (95% bootstrap CI [−0.17, 6.50]; item-cluster sign-flip p = 0.08172). The four model-level point estimates are descriptively positive, but two models are floor-limited. The evidence is therefore directional rather than confirmatory. The contribution is a matched-control diagnostic that complements aggregate benchmark accuracy while making its residual anatomical-content mismatch explicit.