acceptodds
Under review as a conference paper at ICLR 2027

ATTNSEG: Attention as the Interface Between Language Models and Segmenters

Abstract

Vision-language models for segmentation must communicate where a referred object lies to a pixel decoder. Existing approaches do so by either serializing the mask into the token stream or compressing spatial information into the hidden state of a single [SEG] token. We argue that both overlook a spatial representation the model has already computed: the [SEG] token's attention over image tokens, which is natively defined on the image patch grid and can directly encode the referred region. Probing a frozen segmentation VLM shows that attention contains substantially more directly readable localization information than the [SEG] hidden state at matched capacity. This signal captures object shape rather than merely pointing to the referent and is concentrated in a small subset of late attention heads, while standard uniform aggregation obscures much of this spatial structure. Motivated by these findings, we introduce ATTNSEG, an attention-based interface that makes the VLM the primary localizer and leaves the segmentation decoder to refine its prediction. ATTNSEG learns to aggregate the [SEG] token's attention maps into a dense spatial prompt and directly supervises it with an attention-guidance objective. A two-stage training strategy further enforces this division of labor by first learning localization in the VLM and then training the decoder for mask refinement. Applied to Sa2VA, ATTNSEG improves ReasonSeg by gIoU and GS-Eval by at 4B, and by and at 14B, respectively. Gains reach on object parts, multiple objects, and stuff regions, where identifying where to segment rather than refining boundaries is the primary challenge.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision Language Models, Image segmentation, Referring, Grounding, LLMs

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Yqb3OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2(+1)
Soundness
1
Read review
Paper summary

The authors propose a new finetuning method for visual grounding. They try to use the attention distributions of the [SEG] token from the multiple layers and heads in VLM as the prompt for decoder for mask prediction.

Critical strengths

The proposed method achieves substantial performance improvements over existing baselines, demonstrating its potential effectiveness. Moreover, the proposed method is relatively straightforward. Since existing methods tend to produce high-confidence false-positive predictions, the authors introduce negative categories during training to suppress these incorrect predictions.

Critical weaknesses

The writing quality is poor, making it difficult to follow the motivation and methodology. I am not confident that I have fully understood the proposed approach. Furthermore, the training details do not clearly specify whether novel categories are included during training as it needs to use VLM to expand the absolutely absent categories. If so, the performance comparison would be unfair, as it would violate the fundamental principle of open-vocabulary detection (OVD), which aims to detect categories unseen during training. Furthermore, the selection of the 40 absolutely absent categories for each sample requires further explanation. The authors should clarify how these categories are selected and whether they include novel categories reserved for testing.

Questions for authors
  1. What is the motivation. I can just guess about that the authors believe single [SEG] token cannot encode spatial information. However, the spatial information is important for segmentation prediction. Then, the authors find that the attention map of the [SEG] token can represent the spatial information since the high values represents the [SEG] token are more related to that image token. Thus, the authors try to use the attention map of the [SEG] token as a prompt for decoder to predict the segmentation mask.
  2. For figure 2, could you please explain how to derive the different diagrams with detailed implementations. Did you just use the row attention of [SEG] token to replace the original [SEG] token in the baseline?
  3. In my opinion, the [SEG] token may encode certain spatial information, as each image token is associated with positional encoding. Therefore, it would be helpful for the authors to clarify how the proposed method distinguishes or extracts spatial information from the [SEG] token.
  4. The equations lack sufficient definitions of the mathematical notations used. The authors should clearly define all variables and symbols to improve readability and ensure that readers can fully understand the proposed method. For example, there is no explanation on S in line 240. What is w_k^{(0}} in Eq. (2). What is c and the symbol of a circle with center dot in Eq. (3)
  5. a is an attention map, m is the mask. How to realize the cosine similarity in Eq. (3).
  6. The network output appears to be an attention map, which is typically normalized using a softmax function, resulting in values between 0 and 1. However, unlike a ground-truth (GT) binary mask, which contains multiple pixels with values of 1 and 0, a softmax-normalized attention map represents a probability distribution, where the values are generally continuous and sum to 1 along the normalized dimension. Therefore, it is unclear how the proposed method addresses the discrepancy between the prediction and the GT mask in terms of their value distributions. The authors should clarify how the attention map is converted into a segmentation prediction and how the loss function is designed to accommodate this difference.
6 comments
AuthorsOpenReview ↗

Authors Response 1, Part 1


We thank the reviewer for their time and for recognizing the substantial gains over existing baselines and the simplicity of the method. We are grateful for this rebuttal, as it gives us the chance to correct a misunderstanding of the paper's premise. AttnSeg is not an open-vocabulary detection method. It uses no category vocabulary: no novel or held-out categories, no "absolutely absent" categories, and no negative categories during training. AttnSeg is a referring and reasoning segmentation method that changes how a vision-language model (VLM) passes the location of the referent to its mask decoder. Instead of a single [SEG] vector, AttnSeg passes the [SEG] token's own attention map, and trains the VLM so that this map localizes the referent. The VLM thus does the heavy lifting of localization, and the decoder only refines. Q1 explains this step by step. We have also rewritten large parts of the paper, the method section in particular, and we provide an anonymous page with an animation of the method: https://anonymous00440055.github.io/attseg_project_page/.


Q1) What is AttnSeg, and what motivates it?

We restate the core of the paper here, from the question it asks to the method that answers it.

The research question. Segmentation VLMs such as Sa2VA [1] mark the referent with a [SEG] token and pass the token's final hidden state, a single vector, to a mask decoder such as SAM2 [2]. The decoder never sees the text, so this vector must carry both where the referent is and how much of it is meant: "the left ear of the dog" must be told apart from "the dog". We ask two questions:

  • Which internal signal should be the interface between the VLM and the decoder? We argue it is the [SEG] token's attention over the image tokens, which is a map over the image by construction, rather than its hidden state.
  • Which component should do the heavy lifting of localization? We argue it is the VLM, which reads the expression, and that the decoder should only refine. To enforce this, we supervise the VLM's attention directly and train in two stages. Stage one trains the VLM to localize, with no decoder involved. Stage two freezes the VLM and trains only SAM, so SAM can only refine what the VLM has already localized.

The analysis behind each premise (Sec. 3.2). We test these premises with probing, a standard way to measure what a frozen network already knows. We freeze the model and train a small readout to predict the mask from one of its internal signals; if the small readout succeeds, the signal already contains the mask. On a frozen Sa2VA-4B:

  • the same readout recovers 52.1 / 53.2 gIoU (ReasonSeg / GS-Eval) from the [SEG] attention, but only 30.6 / 35.0 from the [SEG] hidden state (Finding 1);
  • giving the readout the image helps the hidden state but not the attention, which already encodes the referent's extent (Finding 2);
  • the localization is concentrated in the late layers (Finding 3).

AttnSeg in four steps, with the tensor shapes for Sa2VA-4B (L=36L = 36 layers, H=16H = 16 heads, SS tokens in the sequence, an h×wh \times w grid of image tokens):

  1. Read (Eq. 2). From the attention tensor A∈RL×H×S×SA \in \mathbb{R}^{L \times H \times S \times S}, take the [SEG] token's row in every layer and head, restricted to the hwhw image tokens (L×H×hwL \times H \times hw). Rescale each row to mean one and fold it onto the image grid, giving LH=576LH = 576 maps of size h×wh \times w.
  2. Combine (Eq. 3). A 1×11\times1 convolution with 576 signed weights sums them into one map a^∈Rh×w\hat{a} \in \mathbb{R}^{h \times w}.
  3. Prompt (Eq. 7). The map is normalized, resized to 256×256256 \times 256 and fed to SAM2's dense mask-prompt channel, in place of the [SEG] vector hSEG∈R2048h_{\text{SEG}} \in \mathbb{R}^{2048}. SAM2 outputs the mask at full image resolution.
  4. Supervise (Eqs. 4–5). The ground-truth mask is downsampled to the same grid, m∈0,1h×wm \in \\{0,1\\}^{h \times w}, and Latt\mathcal{L}_{att} compares it with a^\hat{a}.

This gives +6.4 / +10.2 gIoU at 4B and +8.8 / +11.3 at 14B. The largest gains are on object parts, multiple objects and single objects, where the VLM must single out the referent.

AuthorsOpenReview ↗

Authors Response 1, Part 2


Q2) Are novel or "absent" categories used in training?

No. AttnSeg is a referring and reasoning segmentation method, not an open-vocabulary detector, and it uses no category vocabulary. Each training sample is a triplet: an image, a free-form expression, and the ground-truth mask of the region it refers to. There is no list of absent, novel or negative categories, so no category selection is involved. We train on the thirteen-dataset mixture released with Sa2VA (Appendix D, Table 6). We evaluate on ReasonSeg and GS-Eval, whose training data is not in that mixture; our ReasonSeg results use no ReasonSeg training data at all.

False positives are suppressed by an emergent property, not by negative categories. Figure 3 shows the learned head weights on a real example, "the door with number 8":

  • The 1×11\times1 convolution gives positive weights to heads on the correct door and on the number 8.
  • It gives negative weights to a head locked onto the other door (a distractor), to a head collapsed onto an attention sink, and to a diffuse head. These heads are therefore subtracted from the map.

Nothing supervises these roles. The convolution is trained only to make the combined map match the ground-truth mask, as in any segmentation task, and the ability to suppress distractors emerges from that objective. This is the VLM doing the heavy lifting: its attention becomes discriminative and separates the referent from look-alikes before the decoder sees it. Across the trained model, about 40% of the weights are negative (Sec. 3.3).


Q3) How are the diagrams in Figure 2 derived, and does AttnSeg simply replace the [SEG] token with its attention row?

Figure 2 is our probing analysis (Sec. 3.2), on a frozen Sa2VA-4B. Each point is a small readout trained to predict the mask from one internal signal of the frozen model, scored in gIoU:

  • Panel (a) compares two signals, the [SEG] token's attention maps (purple) and its hidden state (green). Each is read by three readouts of increasing capacity: a single 3×33\times3 convolution, a U-Net, and a U-Net that also sees the RGB image. The two signals differ only in the input; the readout and its training are identical (Appendix B).
  • Panel (b) repeats the probe on each window of five consecutive layers, to find where in the network the localization lives.

The per-head maps and their learned weights are shown in Figure 3 (Q2).

The replacement is not arbitrary; it is what the analysis prescribes. Figure 2 shows three things:

  • the attention carries 18 to 22 gIoU more localization than the hidden state (Finding 1);
  • it already encodes the referent's extent (Finding 2);
  • it lives in the late layers (Finding 3).

AttnSeg therefore prompts the decoder with the [SEG] token's attention instead of its hidden state. This interface is the first of three components. The other two, supervising the attention with Latt\mathcal{L}_{att} and the two-stage training, make the VLM do the heavy lifting of localization. Our ablation (Table 3) shows that the gain needs all three. The interface alone adds +0.26 / +1.09 gIoU over the same model finetuned on the same data, while the full method adds +4.19 / +9.56.

AuthorsOpenReview ↗

Authors Response 1, Part 3


Q4) Doesn't the [SEG] token already encode spatial information, and how does AttnSeg extract it?

The [SEG] hidden state does carry some spatial information, and measuring how much is exactly what our analysis does. We use probing (Sec. 3.2): we freeze the model, train a small readout to predict the mask from one internal signal, and take the readout's accuracy as a measure of what that signal already contains. The hidden state gives 30.6 / 35.0 gIoU (ReasonSeg / GS-Eval). Its spatial information is real but limited: it is a single vector of 2048 numbers with no spatial index, so it must be decoded back onto the image.

AttnSeg instead reads the attention weights that the [SEG] token assigns to the image tokens. This row has one entry per image token, so it is indexed by image position by construction and folds directly onto the h×wh \times w grid, with nothing to decode (Eq. 2). Read by the same readout, it gives 52.1 / 53.2, 18 to 22 gIoU more than the hidden state (Finding 1). The localization concentrates in the late layers (Finding 3). The learned 1×11\times1 combination (Eq. 3) selects and weights the heads automatically, so no per-model analysis is needed.


Q5) Can the notation be defined?

The revision defines every symbol where it is introduced (Sec. 3.1 to 3.3), and Figure 3 annotates the shape of every tensor. The symbols the reviewer lists are:

  • SS (Eq. 2): the length of the token sequence, which comprises the image tokens, the text tokens and the generated answer. The attention weights of one forward pass form a tensor of shape L×H×S×SL \times H \times S \times S, for LL layers and HH heads.
  • w(0)w^{(0)} (Eq. 3): the initial value of the learned weight wl,kw_{l,k} of layer ll and head kk. It is set to the uniform average 1/(LH)1/(LH), so training starts from the plain mean of the heads.
  • cc (Eq. 4): per-patch weights, cp=3c_p = 3 on patches inside the ground-truth mask and cp=1c_p = 1 elsewhere, which up-weight the foreground.
  • ⊙\odot and the cosine (Eq. 4): ⊙\odot is the elementwise product. Let aa be the combined attention map and mm the ground-truth mask on the same h×wh \times w grid, both flattened to vectors. The weighted inner product is ⟨a,m⟩c=∑pcp,ap,mp\langle a, m \rangle_c = \sum_p c_p\\, a_p\\, m_p, the weighted norm is ∥a∥c=⟨a,a⟩c\lVert a \rVert_c = \sqrt{\langle a, a \rangle_c}, and the loss is ℓ(a,m)=1−⟨a,m⟩c/(∥a∥c∥m∥c)\ell(a, m) = 1 - \langle a, m \rangle_c / (\lVert a \rVert_c \lVert m \rVert_c). In code, this is an elementwise product with cc followed by sums.

Q6) How is a softmax attention map compared with a binary mask, and how does it become a segmentation?

This difference is exactly what our objective is designed for. As Sec. 3.3 states, "since only where the map attends carries information, and not how much attention reaches the image, any such objective must be invariant to the scale of a^\hat{a}." The method handles it in three places:

  1. Normalization (Eq. 2). Each head's map is rescaled to spatial mean one. A head is then characterized by where it attends, not by how much of its mass reaches the image, which varies with the sequence length.
  2. A scale-invariant loss (Eq. 4). The cosine compares the shape of the map with that of the mask, not their values, so the map does not need to be binary or to sum to one. After the signed combination, the map is no longer a probability distribution anyway. The loss is minimized when the map is uniform over the target and flat elsewhere. This is why we chose a cosine rather than a loss that compares values, such as binary cross-entropy (Sec. 3.3).
  3. The decoder, not the map, makes the mask. The combined map is min-max normalized, resized and converted to logits (Appendix C, Eq. 7). It then enters SAM2's dense mask-prompt channel without thresholding. SAM2, trained with the standard mask loss in stage two, outputs the binary mask at full resolution.

We hope this clarifies the approach. The rewritten paper and the animation (https://anonymous00440055.github.io/attseg_project_page/) walk through it step by step, and we would be glad to answer any further question.


References:

[1] Sa2VA: Marrying SAM2 with MLLM for Dense Grounded Understanding of Images and Videos

[2] SAM 2: Segment Anything in Images and Videos

Reviewer_YV3LOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that segmentation VLMs should pass the [SEG] token's attention over image tokens, rather than its hidden state, to the mask decoder. Probing a frozen Sa2VA-4B, it reports that a readout on the [SEG] attention recovers 18.2/21.5 more gIoU than the same readout on the [SEG] embedding, and that the signal sits in a few late heads. ATTNSEG aggregates the per-layer/per-head attention with a learned signed 1×1 convolution and supervises the resulting map with an area-weighted, correctness-gated cosine loss (L_att). The map is fed to SAM's dense mask-prompt channel. Training has two stages: LoRA on the VLM under L_att, then decoder-only training. On Sa2VA-4B/14B, the method improves ReasonSeg by +6.4/+8.8 gIoU and GS-Eval by +10.2/+11.3, with the largest gains on parts, multiple objects and stuff.

Critical strengths
  1. The empirical gains over the released Sa2VA checkpoints are consistent at both scales. They remain substantial against the fine-tuning-only control (Table 2, row 2).
  2. The ablations are candid. Table 2 isolates fine-tuning, the interface and the loss, and Table 3 re-probes the trained model.
  3. The recomputation of only the [SEG] attention rows alongside fused attention kernels (Sec. 4.4) is a useful practical detail.
Critical weaknesses
  1. The core interface has clear prior work, and the closest work is not cited. Reading a frozen LMM's attention over image tokens, stacking all layers and heads, decoding it with a small U-Net, and feeding the result to SAM as a dense mask prompt is the approach of F-LMM (Wu et al., CVPR 2025):

    • F-LMM stacks word-to-image attention maps from all layers and heads of a frozen LMM.
    • It decodes them with a three-stage U-Net mask head.
    • It refines the output with a SAM-based refiner that takes the logits as a dense mask prompt, plus a box prompt.
    • It evaluates on referring and reasoning segmentation.

    This anticipates both the probing setup of Sec. 3.1 (a three-stage U-Net reading attention from a frozen model; App. B, Table 4) and the dense, unthresholded "Interface III". F-LMM is neither cited nor compared. Two statements in the paper are therefore inaccurate: that attention-based systems "all pass through a sparse, discretized bottleneck" (Sec. 2), and that "almost every multimodal segmenter leaves [SAM's dense channel] at its learned no_mask_embed default" (App. A). Cao et al. (2025), which the paper cites, also evaluated prompting SAM with the LMM attention map as a mask prompt (their Table 4). Kang et al. (CVPR 2025) showed that a few localization heads carry the grounding signal; the learned head weighting here is a trainable variant of that observation. Relative to these works, the remaining delta is limited: using the [SEG] token's own row instead of a word or answer token, signed rather than selected head weights, and an auxiliary loss on the map.

  2. The paper's own ablation does not support its central claim. The title and Sec. 3 present the attention interface as the key idea. Yet Table 2 shows that replacing the [SEG] embedding with the attention map adds only +0.26 (ReasonSeg) and +1.09 (GS-Eval). Nearly all of the improvement (+3.93/+8.47, 61–83% of the total) comes from L_att. Sec. 4.3 argues that the interface is a "precondition" because "only the attention map is an object a spatial objective can act on". However, the [SEG] attention exists whether or not it is forwarded to SAM. The decisive control is missing: train the VLM with L_att (stage 1), then keep the original [SEG]-embedding interface for SAM. Without this row, it is unclear whether the proposed interface matters at all, or whether the method reduces to auxiliary attention supervision.

  3. Auxiliary attention supervision is itself not new. Supervising a VLM's attention over visual tokens with ground-truth spatial maps has been shown to improve grounding; see KLAL (Esmaeilkhani et al., "Direct Visual Grounding by Directing Attention of Visual Tokens", WACV 2026), which aligns answer-token attention across all LLM layers to GT maps with a KL loss. Since L_att accounts for most of the gain, the paper should position this component against such work. The specific refinements (area weighting with ρ = 0.30 and a [0.25, 12] clamp, the correctness gate) are heuristics, and the paper does not show their sensitivity.

  4. The motivating probe compares unlike representations. In Sec. 3.1, the embedding probe must map a spatially unindexed vector to a 16×16 grid through a learned projection, without access to image features. The embedding is never used that way in practice: SAM decodes it jointly with dense image features. The 18–22 gIoU gap (Finding 1) therefore says little about which representation is the better interface. Sec. 4.3 itself concedes that the gap largely vanishes once a real decoder is trained (Table 2). Findings 1–2 thus do not establish the design premise.

Questions for authors
  1. How does ATTNSEG compare with F-LMM, conceptually and empirically? What does it offer beyond stacking LMM attention maps into a mask head and feeding SAM's dense prompt?
  2. Please report the missing row of Table 2: stage-1 L_att training followed by the original [SEG]-embedding interface.
  3. How sensitive are the results to the area-weighting exponent ρ, the clamp range and the correctness gate? Can you report variance over seeds?
  4. Do the findings and gains transfer to other [SEG]-based models, e.g., LISA or GLaMM built on different backbones?
9 comments
AuthorsOpenReview ↗

Authors Response 1, Part 1

We thank the reviewer for the careful review. We are glad they found the gains consistent at both scales and substantial against the finetuning-only control, the ablations holistic, and the recomputation of the [SEG] rows a useful practical detail. The main concern is how much is new relative to prior work. That work either reads attention that it leaves unchanged or supervises attention as an auxiliary objective. We show that in a VLM trained to segment, neither the attention interface nor attention supervision alone delivers the gain. It comes from combining them, so that the VLM does the heavy lifting of localization through its own [SEG] attention (Q3).

Section and table numbers refer to the revision. We also provide an anonymous project page whose video demo animates how AttnSeg works: https://anonymous00440055.github.io/attseg_project_page/


Q1) Do the findings and gains transfer to other [SEG]-based models?

Beyond the two Sa2VA-InternVL models [4] in the submission, we applied AttnSeg to a second VLM family, Sa2VA with Qwen2.5-VL-7B [9] at two input resolutions, and to two other frameworks, LISA-7B [2] and UniPixel-7B [10]. Each result cell gives gIoU as base → AttnSeg, where the base is the same model finetuned on the same data without AttnSeg. "Map" scores the thresholded raw attention map itself. Table 1 of the revision includes the Sa2VA-Qwen2.5-VL and UniPixel results at 1,204 px, per split. Appendix H (Table 11) reports the full study.

ModelInput (H, W)Image tokensAtt. map (H, W)ReasonSegGS-EvalMap, ReasonSegMap, GS-Eval
LISA-7B(224, 224)256(16, 16)50.0 → 49.960.8 → 60.41.40 → 31.441.53 → 33.62
Sa2VA-Qwen2.5-VL-7B(448, 588)336(16, 21)61.32 → 61.7966.55 → 74.639.56 → 31.279.78 → 38.99
Sa2VA-Qwen2.5-VL-7B(896, 1204)1,376(32, 43)60.77 → 64.2868.81 → 74.045.95 → 40.666.79 → 50.89
UniPixel-7B(896, 1204)1,376(32, 43)65.21 → 64.9765.63 → 69.754.60 → 47.424.20 → 54.75
Sa2VA-InternVL2.5-4B(1344, 1792)3,328(48, 64)59.92 → 64.1162.55 → 72.110.03 → 49.600.16 → 59.09
Sa2VA-InternVL3-14B(1344, 1792)3,328(48, 64)60.88 → 69.7163.05 → 74.310.10 → 53.570.32 → 60.18
  • On LISA, AttnSeg brings no gain, which we attribute to its small attention map. AttnSeg lifts LISA's map from 1.4 to 31.4 gIoU on ReasonSeg, as much as on Sa2VA-Qwen at 600 px, yet the mask does not improve. LISA's CLIP encoder takes a 224 px image, so its map is only 16×16, the coarsest of all settings, and a better map gives the decoder little extra spatial detail to work with.
  • A controlled experiment supports this. We ran the same model, Sa2VA-Qwen2.5-VL-7B, at two input resolutions, so the only change is the size of the image the VLM sees and hence of its map (16×21 vs. 32×43). The larger map raises the ReasonSeg gain from +0.47 to +3.51, and the map itself improves on both benchmarks (31.3 → 40.7 and 39.0 → 50.9). LISA does not allow a clean version of this control. Its CLIP ViT-L/14 encoder is pretrained at a fixed 224 px, and its LLaMA-based language model has a 2,048-token context. A larger input would therefore also move the encoder off its training resolution and crowd out the text. Qwen2.5-VL is trained natively at variable resolution.
  • Across the five Sa2VA and LISA settings, the gain grows with the map. On ReasonSeg, it rises from −0.1 at 16×16 to +0.5 at 16×21, +3.5 at 32×43 and +4.2 at 48×64 (+8.8 for the 14B model). Over the same settings, the map rises from under 10 gIoU in every base model to 31.4, 31.3, 40.7 and 49.6. On GS-Eval, every setting larger than 16×16 gains 5 to 11 points. As in Table 1, these gains concentrate on object parts (+20.83 and +12.73 for Sa2VA-Qwen), while stuff moves by less than one point.
  • A large map is not an extra requirement. Current VLMs encode images at high or native resolution by design: InternVL tiles each image into 448 px crops, and Qwen2.5-VL encodes it at native resolution. Our main results use Sa2VA-InternVL's default input, which already yields a 48×64 map. Only LISA's 224 px CLIP encoder falls below this range, and on GS-Eval AttnSeg gains at every larger map.
  • UniPixel [10] confirms the transfer to a second framework on GS-Eval. UniPixel couples Qwen2.5-VL-7B with SAM 2 through its own segmentation token, and at 1,204 px it has the same 32×43 map as Sa2VA-Qwen. AttnSeg lifts its map from 4.6 to 47.4 gIoU on ReasonSeg and from 4.2 to 54.8 on GS-Eval, and improves its masks on GS-Eval by +4.1, with a gain on every split (multiple objects +8.0, object parts +5.2, single objects +3.9, stuff +1.0). On ReasonSeg, the mask stays on par (65.21 → 64.97); UniPixel's base is already the strongest of the six settings on this benchmark.
AuthorsOpenReview ↗

Authors Response 1, Part 2


Q2) How does AttnSeg compare with F-LMM, conceptually and empirically?

F-LMM [1] answers a different question. Finetuning a VLM to segment through a [SEG] token, as in LISA [2] and GLaMM [3], erodes its conversational ability. F-LMM avoids this by keeping the LMM frozen: it reads the attention of the phrase's word tokens and trains a U-Net and a SAM-based refiner on top. Sa2VA [4] has since shown that freezing is unnecessary: it trains the VLM for segmentation together with conversational data and largely maintains the chat performance of its base model. The VLM therefore no longer has to be frozen and limited.

We ask a different question: which internal signal should reach the decoder of such a VLM, and which of the two should do the heavy lifting of localization, the VLM or the decoder? Table 3 of the revision gives the answer (SAM2 columns, gIoU):

FinetuningInterfaceLatt\mathcal{L}_{att}ReasonSegGS-Eval
A✗Embeddings✗57.7061.91
B✓Embeddings✗59.9262.55
C✓Embeddings✓59.8167.99
D✓Attention✗60.1863.64
E✓Attention✓64.1172.11

Our answer differs from F-LMM's design as follows:

  1. Reading the attention is not enough; supervising it is. Passing the [SEG] attention to a trained SAM2 [5] without supervising it (row D) adds only +0.26 / +1.09 gIoU over finetuning. Training the VLM with Latt\mathcal{L}_{att}, which uses only the masks already provided by the segmentation data, adds the remaining +3.93 / +8.47, 61% and 83% of the total gain (row E). Supervising the attention without passing it to the decoder (row C) is not enough either. F-LMM keeps the attention fixed by design.

  2. We analyze and shape the attention, and the shaping is emergent. Our probes place the localization in a late layer band (Finding 3). The learned, signed 1×11\times1 combination recovers that band without a per-backbone search, and it learns to separate the referent from look-alikes. Figure 3 shows this on a real example, "the door with number 8":

    • heads on the correct door and on the number 8 receive positive weights;
    • a head locked onto the other door (a distractor), a head collapsed onto an attention sink, and a diffuse head receive negative weights, and are subtracted.

    Nothing supervises these roles. The combination is trained only to match the ground-truth mask, and the ability to suppress distractors emerges from that objective. This is the VLM doing the heavy lifting: its attention becomes discriminative before the decoder sees it. Across the trained model, about 40% of the weights are negative. The revision also compares three attention objectives (Sec. 4.3, Table 2) and finds that what misleads the decoder is concentrated off-target attention, not its total amount.

Empirically, we integrated Localization Heads [6] and F-LMM [1] into Sa2VA-InternVL2.5-4B, with Localization Heads applied to the same [SEG]-token attention that AttnSeg reads, and compare at three levels of decoding (Appendix F, Table 9). F-LMM and Localization Heads are also in Table 1. gIoU:

MethodBackboneDecoderRS Sent.RS PhraseRS AllGS StuffGS PartsGS MultiGS SingleGS All
Attention map only
Localization HeadsSa2VA-InternVL2.5-4Bnone15.7016.4015.9013.707.4030.3014.0016.50
AttnSegSa2VA-InternVL2.5-4Bnone50.6146.4349.6079.2029.0872.6147.5659.09
Off-the-shelf SAM, no training
Localization HeadsLLaVA-1.5-7BSAM25.0921.7424.2845.469.2613.4812.7321.37
Localization HeadsSa2VA-InternVL2.5-4BSAM25.3020.4024.1038.707.2040.1017.2026.60
AttnSegSa2VA-InternVL2.5-4BSAMRefiner [7] + SAM61.4759.4760.9991.1336.5187.9162.1972.24
Trained decoder
F-LMMLLaVA-1.5-7BU-Net + SAM27.931.028.780.512.471.337.553.1
F-LMMSa2VA-InternVL2.5-4BU-Net + SAM53.262.355.492.038.985.363.172.6
AttnSegSa2VA-InternVL2.5-4BSAM264.8061.9364.1191.2035.2986.3762.9772.11

Without a decoder, or with an off-the-shelf SAM, AttnSeg leads by a wide margin: our map reaches 49.60 / 59.09 on its own and 60.99 / 72.24 through SAM, against 15.90 / 16.50 and 24.10 / 26.60 for Localization Heads on the same model and attention. With trained decoders on the same backbone, AttnSeg exceeds F-LMM by 8.71 gIoU on ReasonSeg (64.11 vs. 55.4) and is on par on GS-Eval (72.11 vs. 72.6). Even with no decoder at all, our map exceeds F-LMM's full pipeline on LLaVA-1.5-7B (49.60 / 59.09 vs. 28.7 / 53.1), as well as the 46.2 on ReasonSeg that F-LMM reports with DeepSeek-VL-7B.

AuthorsOpenReview ↗

Authors Response 1, Part 3


Q3) Is your loss just an auxiliary loss? Please report the missing row of Table 2: stage-1 Latt\mathcal{L}_{att} training followed by the original [SEG]-embedding interface.

We added this row to Table 3 (row C), which now holds the full 2×2. The row shares the stage-one VLM of AttnSeg, trained with Latt\mathcal{L}_{att}, and prompts the decoder with the [SEG] embedding. Per split, in gIoU:

RowDecoder promptLatt\mathcal{L}_{att}RS Sent.RS PhraseRS AllGS StuffGS PartsGS MultiGS SingleGS All
B[SEG] embedding✗60.7857.2159.9290.7619.0374.4650.4762.55
C[SEG] embedding✓60.8956.3759.8190.2928.6682.9157.0967.99
Eattention map (AttnSeg)✓64.8061.9364.1191.2035.2986.3762.9772.11

Used as an auxiliary loss behind the [SEG] embedding (row C), Latt\mathcal{L}_{att} leaves ReasonSeg unchanged (−0.11) and adds 5.44 on GS-Eval. Read through the attention map (row E), the same VLM gains +4.19 / +9.56. Prompting the decoder with the embedding instead of the map costs 4.30 / 4.12 gIoU, mostly on object parts (−6.6) and single objects (−5.9). The map that Latt\mathcal{L}_{att} shapes has to be the map the decoder receives.


Q4) How does AttnSeg relate to attention-prompted SAM (Cao et al.) and localization heads (Kang et al.)?

We agree that neither prompting SAM with an attention map nor the observation that a few heads localize is ours; the paper cites both [6, 8]. Our contribution is what makes them work inside a segmentation VLM.

  • Cao et al. [8] expose the problem we solve. Their attention comes from an LMM whose attention is never supervised, is averaged over all layers and heads, and goes to a frozen SAM. In their Table 4, this map as a mask prompt scores below a single point prompt (51.7 vs. 63.3 mIoU on GCG). AttnSeg changes both sides: Latt\mathcal{L}_{att} trains the VLM so that its [SEG] attention localizes, and stage two trains the decoder to read the dense map. Table 3 (reproduced in Q2) shows that the dense interface pays off only with both. Without Latt\mathcal{L}_{att} it adds +0.26 / +1.09 over finetuning; with it, +4.19 / +9.56.

  • Kang et al. [6] motivate our head weighting, which is a component, not the contribution. Their heads are selected offline by spatial entropy and then frozen, so a head can only be kept or dropped. Our 1×1 weighting is learned jointly with Latt\mathcal{L}_{att}, so it needs no per-backbone selection. It is also signed: about 40% of its weights are negative and subtract sink and distractor heads (Fig. 3), which a selection cannot express. Most of the gain comes from Latt\mathcal{L}_{att}, not from how the heads are combined: without Latt\mathcal{L}_{att}, the interface adds only +0.26 / +1.09 (Table 3).

Reviewer_BVLtOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Existing reasoning segmentation approaches either predict segmentation geometry in text form or compress spatial information into a single [SEG] token representation passed to a segmentation decoder. This paper proposes ATTNSEG, which extracts the attention between the [SEG] token and image tokens to generate a dense spatial prompt for segmentation. The authors introduce an attention-guidance loss to improve the localization information captured by these attention maps, followed by a two-stage training strategy that first fine-tunes the vision-language model and then trains the segmentation decoder.

Critical strengths

The paper is well-written and easy to follow. It presents an interesting insight into how the attention between [SEG] and image tokens already contains rich spatial information that can be exploited for dense segmentation. The experimental results demonstrate meaningful improvements over the baseline, particularly for challenging segmentation categories such as object parts and multiple objects. Ablation studies also largely help isolate the contributions of different components of the proposed approach.

Critical weaknesses

However, the paper has a few critical weaknesses:

  • The idea of mining segmentation information or spatial localization from attention maps has been explored in previous methods such as PnP-OVSS (Luo et al., CVPR 2024), Segmentation From Attention (Hossain et al., TMLR 2026), and Open-vocabulary Object Segmentation with Diffusion Models (Li et al., ICCV 2023). In particular, Segmentation From Attention (TMLR 2026) also explores few-shot fine-tuning of attention maps to obtain more accurate segmentation. In my opinion, the main contribution of the paper is using and fine-tuning the attention between the [SEG] token and image tokens. However, the [SEG] token already contains substantial spatial information through the segmentation fine-tuning of such models. Therefore, I find the overall technical novelty a bit limited.

  • The authors haven't cited any of the aforementioned works, PnP-OVSS (CVPR 2024), Segmentation From Attention (TMLR 2026), and Open-vocabulary Object Segmentation with Diffusion Models (ICCV 2023), as prior literature on extracting segmentation information from attention maps. The authors should acknowledge and discuss these approaches.

  • The authors only evaluate their approach on the InternVL family of backbones using the Sa2VA framework. From the motivation and overall framing of the paper, it appears that attention can serve as a general interface between language models and segmenters. However, whether it generalizes to other LLM-based segmentation models that use [SEG] tokens, such as LISA and GLaMM, is unclear. I feel there should be an experiment to determine whether the proposed approach generalizes to any of these frameworks.

  • One of the key arguments of the paper is that the attention maps between the [SEG] token and image tokens already contain substantial spatial information, and the segmentation decoder primarily performs refinement. Although the authors evaluate lightweight decoders in their probing experiments, they do not perform any experiments replacing SAM2 with such lightweight decoders in their full framework. I feel this experiment would help validate their motivation better. They could also try simple post-processing techniques on the attention maps, such as applying DenseCRF.

Questions for authors

Please check the weakness section. In particular:

  • Does the proposed approach generalize to other segmentation frameworks beyond Sa2VA?
  • How does the proposed approach perform with a lightweight decoder instead of SAM2?
  • Can the authors clarify how their method differs from existing attention-based segmentation approaches and better establish its novelty?
7 comments
AuthorsOpenReview ↗

Authors Response 1, Part 1

We thank the reviewer for the careful review. We are glad they found the paper well written, the insight into the [SEG] attention interesting, the improvements meaningful (especially on object parts and multiple objects), and the ablations informative. We also provide an anonymous project page whose video demo animates how AttnSeg works: https://anonymous00440055.github.io/attseg_project_page/

We address the three questions below.


Q1) Does the proposed approach generalize to other segmentation frameworks beyond Sa2VA?

Beyond the two Sa2VA-InternVL models [1] in the submission, we applied AttnSeg to a second VLM family, Sa2VA with Qwen2.5-VL-7B [2] at two input resolutions, and to two other frameworks, LISA-7B [3] and UniPixel-7B [4]. Each result cell gives gIoU as base → AttnSeg, where the base is the same model finetuned on the same data without AttnSeg. "Map" scores the thresholded raw attention map itself. Table 1 of the revision includes the Sa2VA-Qwen2.5-VL and UniPixel results at 1,204 px, per split. Appendix H (Table 11) reports the full study.

ModelInput (H, W)Image tokensAtt. map (H, W)ReasonSegGS-EvalMap, ReasonSegMap, GS-Eval
LISA-7B(224, 224)256(16, 16)50.0 → 49.960.8 → 60.41.40 → 31.441.53 → 33.62
Sa2VA-Qwen2.5-VL-7B(448, 588)336(16, 21)61.32 → 61.7966.55 → 74.639.56 → 31.279.78 → 38.99
Sa2VA-Qwen2.5-VL-7B(896, 1204)1,376(32, 43)60.77 → 64.2868.81 → 74.045.95 → 40.666.79 → 50.89
UniPixel-7B(896, 1204)1,376(32, 43)65.21 → 64.9765.63 → 69.754.60 → 47.424.20 → 54.75
Sa2VA-InternVL2.5-4B(1344, 1792)3,328(48, 64)59.92 → 64.1162.55 → 72.110.03 → 49.600.16 → 59.09
Sa2VA-InternVL3-14B(1344, 1792)3,328(48, 64)60.88 → 69.7163.05 → 74.310.10 → 53.570.32 → 60.18
  • On LISA, AttnSeg brings no gain, which we attribute to its small attention map. AttnSeg lifts LISA's map from 1.4 to 31.4 gIoU on ReasonSeg, as much as on Sa2VA-Qwen at 600 px, yet the mask does not improve. LISA's CLIP encoder takes a 224 px image, so its map is only 16×16, the coarsest of all settings, and a better map gives the decoder little extra spatial detail to work with.
  • A controlled experiment supports this. We ran the same model, Sa2VA-Qwen2.5-VL-7B, at two input resolutions, so the only change is the size of the image the VLM sees and hence of its map (16×21 vs. 32×43). The larger map raises the ReasonSeg gain from +0.47 to +3.51, and the map itself improves on both benchmarks (31.3 → 40.7 and 39.0 → 50.9). LISA does not allow a clean version of this control. Its CLIP ViT-L/14 encoder is pretrained at a fixed 224 px, and its LLaMA-based language model has a 2,048-token context. A larger input would therefore also move the encoder off its training resolution and crowd out the text. Qwen2.5-VL is trained natively at variable resolution.
  • Across the five Sa2VA and LISA settings, the gain grows with the map. On ReasonSeg, it rises from −0.1 at 16×16 to +0.5 at 16×21, +3.5 at 32×43 and +4.2 at 48×64 (+8.8 for the 14B model). Over the same settings, the map rises from under 10 gIoU in every base model to 31.4, 31.3, 40.7 and 49.6. On GS-Eval, every setting larger than 16×16 gains 5 to 11 points. As in Table 1, these gains concentrate on object parts (+20.83 and +12.73 for Sa2VA-Qwen), while stuff moves by less than one point.
  • A large map is not an extra requirement. Current VLMs encode images at high or native resolution by design: InternVL tiles each image into 448 px crops, and Qwen2.5-VL encodes it at native resolution. Our main results use Sa2VA-InternVL's default input, which already yields a 48×64 map. Only LISA's 224 px CLIP encoder falls below this range, and on GS-Eval AttnSeg gains at every larger map.
  • UniPixel [4] confirms the transfer to a second framework on GS-Eval. UniPixel couples Qwen2.5-VL-7B with SAM 2 through its own segmentation token, and at 1,204 px it has the same 32×43 map as Sa2VA-Qwen. AttnSeg lifts its map from 4.6 to 47.4 gIoU on ReasonSeg and from 4.2 to 54.8 on GS-Eval, and improves its masks on GS-Eval by +4.1, with a gain on every split (multiple objects +8.0, object parts +5.2, single objects +3.9, stuff +1.0). On ReasonSeg, the mask stays on par (65.21 → 64.97); UniPixel's base is already the strongest of the six settings on this benchmark.
AuthorsOpenReview ↗

Authors Response 1, Part 2


Q2) How does the proposed approach perform with a lightweight decoder instead of SAM2?

We froze the VLM and its attention map and replaced only the decoder, so every decoder receives the same cached map. The learned decoders are trained on the same data for the same budget and selected on a held-out split of the training mixture. DenseCRF, the guided filter and SAMRefiner need no training. All are scored once on ReasonSeg and GS-Eval (gIoU). The revision reports this study in Appendix G (Table 10).

DecoderImageParamsReasonSegGS-Eval
ARaw attention map (the decoder's input)☐049.659.1
BDenseCRF [5]☑049.260.5
CGuided filter / JBU [6, 7]☑051.163.5
DLearned upsampler☐0.033M54.860.7
ESAM2 [8] decoder, image features zeroed☐4.2M54.559.5
FUNet [9]☐0.471M55.660.5
GSegFormer-B0 [10]☐3.711M56.160.5
HSAM2 decoder + conv encoder trained from scratch☑4.7M57.364.4
ISAM-S☑46M57.563.5
JSegFormer-B0☑3.716M59.664.2
KUNet☑0.478M60.664.8
LMobileSAM [11] (distilled encoder + SAM decoder)☑10.1M61.868.6
MSAMRefiner [12] (frozen SAM ViT-H, no training)☑636M61.072.2
NSAM2☑224M64.172.1

B and C learn no weights; their hyperparameters are hand-set (DenseCRF: Gaussian kernel sxy = 3, compat = 3; bilateral kernel sxy = 50, srgb = 5, compat = 20. Guided filter: radius = 16, ε = 1e-4).

A 0.48M decoder already beats the original interface with the full SAM2. The UNet with the image (K) achieves 60.6/64.8. That is above the released Sa2VA-4B with its 224M SAM2 decoder (57.70 / 61.91), and above the same model finetuned on our data (59.92 / 62.55). MobileSAM (L, 10.1M) reaches 61.8 / 68.6. Between the raw map (A) and full SAM2 (N), they close 76% and 84% of the ReasonSeg gap, with 1/470 and 1/22 of the parameters. This is our premise at work: when the VLM does the heavy lifting of localization, the decoder only has to refine, and a 0.48M one suffices.

An off-the-shelf SAM needs no finetuning at all. SAMRefiner (M) prompts a frozen SAM ViT-H from a coarse mask without any training. Applied to our raw map, it reaches 61.0 / 72.2, on par with our trained SAM2 on GS-Eval (72.1) and above the released Sa2VA-4B with its trained decoder. SAMRefiner is not a lightweight decoder, but it shows from another angle that our map is a usable prompt for an unmodified SAM.

Classical post-processing recovers little. DenseCRF (B) slightly lowers ReasonSeg (49.6 → 49.2) and adds 1.4 on GS-Eval, and the guided filter (C) adds 1.5 / 4.4. Snapping the map to the image edges is not enough; closing the gap requires learned image features.

The remaining gap to SAM2:

  1. The image matters more than the decoder. Adding the image to the UNet (F → K) gains +5.0 on ReasonSeg for 0.007M parameters. Growing a map-only decoder 112× (D → G) gains +1.3.
  2. Capacity saturates. UNets with 0.472M, 1.884M and 4.236M parameters score 60.5, 60.3 and 60.7 on ReasonSeg, so the table lists only the smallest UNet.
  3. SAM2's edge is its pretrained image encoder, not its mask decoder. SAM2's pretrained decoder with the image features zeroed (E, 4.2M) scores 54.5, below a 0.471M UNet trained from scratch (F, 55.6). With its encoder attached (N), the same decoder gains +9.6. A frozen-SAM2 ablation agrees: zeroing its image features drops held-out gIoU from 63.7 to 34.4.
AuthorsOpenReview ↗

Authors Response 1, Part 3


Q3a) How does AttnSeg differ from existing attention-based segmentation approaches?

We thank the reviewer and agree with the premise: reading localization out of attention is well established. We do not claim it, nor to be the first to supervise attention with masks (Appendix A, "Supervising attention"). We have added all three works to the revised Appendix A, together with the two closest works on multimodal LLMs, Localization Heads [13] and F-LMM [14]. They are covered in the paragraph "Segmentation from the attention of models not trained to segment", and in Table 4.

The difference is the core research question. These works ask whether a model never trained to segment can be made to segment from its attention, and four of the five never change the attention. We ask a different question: in a VLM already trained to segment through a dedicated decoder, what should cross the interface to that decoder, and which of the two should do the heavy lifting of localization: the VLM or the decoder? The table below condenses Table 4 of the revised appendix.

WorkAttention trained?Role of the attention
Li et al. [15]✗not used for the mask: a grounding head trained on frozen diffusion features
PnP-OVSS [16]✗the mask itself (class-name tokens, GradCAM, CRF)
Segmentation From Attention [17]one-shotthe mask itself (class-name tokens, ConvCRF)
Localization Heads [13]✗a box prompt for a frozen SAM
F-LMM [14]✗input to a trained U-Net and SAM
AttnSeg (ours)✓the decoder's only prompt, read from the [SEG] token's own attention, so the VLM does the heavy lifting

"The [SEG] token already contains substantial spatial information." It does, but not in a form the decoder can use, and that gap is what we address. In the frozen Sa2VA-4B, the same probe reads 30.6 / 35.0 gIoU (ReasonSeg / GS-Eval) from the [SEG] hidden state and 52.1 / 53.2 from its attention (Sec. 3.2). Yet even after finetuning with the attention interface, a 560-parameter gate reads only 24.0 / 26.8 from it, rising to 42.0 / 44.7 with Latt\mathcal{L}_{att} (Table 3, rows D and E). Under the mask loss alone, the decoder ends up doing the heavy lifting: its gradient into the map is about 49× that of Latt\mathcal{L}_{att} (Sec. 3.3), and the released model scores 17.8 gIoU on object parts.

Latt\mathcal{L}_{att} is not an auxiliary loss (Table 3). Every row finetunes the released Sa2VA-4B with the same data and seed. Rows C and E share the same stage-1 VLM and differ only in what the decoder reads.

RowDecoder promptLatt\mathcal{L}_{att}RS Sent.RS PhraseRS AllGS StuffGS PartsGS MultiGS SingleGS All
B[SEG] embedding✗60.7857.2159.9290.7619.0374.4650.4762.55
C[SEG] embedding✓60.8956.3759.8190.2928.6682.9157.0967.99
Eattention map (AttnSeg)✓64.8061.9364.1191.2035.2986.3762.9772.11

Used as an auxiliary loss behind the [SEG] embedding (row C), Latt\mathcal{L}_{att} leaves ReasonSeg unchanged (−0.11) and adds 5.44 on GS-Eval. Read through the attention map (row E), the same VLM gains +4.19 / +9.56. Prompting the decoder with the embedding instead of the map costs 4.30 / 4.12 gIoU, mostly on object parts (−6.6) and single objects (−5.9). The map that Latt\mathcal{L}_{att} shapes has to be the map the decoder receives.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

We thank the reviewers for their constructive feedback.

AttnSeg passes the [SEG] token's attention map to the mask decoder instead of its embedding and supervises that map, so the VLM does the heavy lifting of localization and the decoder only refines. It improves ReasonSeg / GS-Eval by +6.4 / +10.2 gIoU at 4B and +8.8 / +11.3 at 14B.


The revision addresses every concern:

  • Generality (BVLt, YV3L): AttnSeg transfers to a second VLM family (Sa2VA-Qwen2.5-VL-7B) and to a second framework (UniPixel-7B), improving GS-Eval in every setting whose attention map exceeds 16×16 (+4.1 to +11.3; Table 1, App. H).

  • Lightweight decoders (BVLt): a 0.48M U-Net on our map beats the released Sa2VA-4B with its full 224M SAM2, and a frozen off-the-shelf SAM reaches 61.0 / 72.2 with no training (App. G).

  • Novelty (BVLt, YV3L): App. A positions AttnSeg against attention mining (PnP-OVSS, Segmentation From Attention, Localization Heads, F-LMM) and attention supervision (KLAL). On the same backbone and [SEG] attention, our map reaches 61.0 / 72.2 through a zero-shot SAM, against 24.1 / 26.6 for Localization Heads (App. F).

  • Not an auxiliary loss (YV3L): The full 2×2 (Table 3) shows that our loss behind the [SEG] embedding gives −0.1 / +5.4, but through the attention map +4.2 / +9.6. The map the loss shapes must be the map the decoder reads.

  • Probe fairness (YV3L): Table 3 now reads every ablation model both by SAM2 and by the probes. Our loss makes the attention readable by a 560-parameter gate (24.0 → 42.0 gIoU).

  • Clarity (Yqb3): AttnSeg is a referring and reasoning segmentation method and uses no category vocabulary. The method section is rewritten with every symbol defined, and an anonymous page animates the method: https://anonymous00440055.github.io/attseg_project_page/


We are more than happy to use the remaining discussion period to address any further questions, and we hope the reviewers will engage and share any remaining concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.