acceptodds
Under review as a conference paper at ICLR 2027

Two Vectors Replace In-Context Demos: Structured Task Adaptation via Embeddings

Abstract

In-context learning (ICL) adapts frozen large multimodal models (LMMs) to new tasks from a few demonstrations (demos), but re-encodes them at every query, where each demo image adds up to hundreds of visual tokens. Demo-free methods remove this cost with a compact task state. However, they add it at locations searched per task or at every decoder layer, where task parameters grow with depth. Moreover, inserted tokens or keys cannot change how the original prompt divides its attention within a layer. To address these issues, we propose Structured Task Adaptation via Embeddings (**STAVE**), which replaces demos with two task-specific vectors added to existing input embeddings. Specifically, a readout vector updates the answer-producing tokens and a context vector updates the other structural token groups. Both are trained with answer labels on prompts with and without demos. We justify these design choices theoretically using a first-order analysis of the loss and a margin bound. Extensive experiments on six LMMs and five large language models show that STAVE matches or outperforms state-of-the-art methods on multimodal tasks with far fewer task parameters and surpasses 15-shot ICL and prior task vectors on 18 text tasks, all at zero-shot inference cost.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
In-context learning, task vectors, parameter-efficient adaptation, large multimodal models, large language models

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_Yu4NOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

STAVE learns two vectors that are added to the existing prompt embeddings of a frozen model: one changes the tokens that produce the answer, and the other changes the remaining context tokens. The vectors are trained on labeled examples with and without demonstrations, then used without demonstrations at test time. Experiments cover six multimodal and five language models, with ablations examining the token split and training objective.

Critical strengths
  1. Storing a task in just two vectors is useful because it avoids repeating demonstrations at inference, and Table 6 shows the same prompt length and measured cost as zero-shot inference.
  2. The gains also appear on held-out data across several backbones, and the token-split and training ablations help explain why the method works (Tables 2, 3, and 10; Appendix H).
Critical weaknesses
  1. Choosing the best epoch on the evaluation set makes Table 1 an optimistic estimate, and evaluating 16/32-shot ICL on fewer queries makes those comparisons less direct, although the epoch check partly addresses the first concern (Appendices F.2 and F.5).
  2. PT is trained for 500 steps using only the first answer token, while STAVE can train for 4,000 steps using the full answer, so the gap may partly reflect how they were trained (Appendix F.3).
Questions for authors
  1. How do the Table 1 results change when checkpoints are chosen without using the evaluation set and all methods are tested on the same queries?
  2. Does STAVE still outperform PT when PT is trained on the full answer and given a comparable training and tuning budget?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Yu4N (1/2)

We sincerely thank the reviewer for the positive assessment and the constructive questions. We answer both below.

W1/Q1. Choosing the best epoch on the evaluation set makes Table 1 optimistic, and 16/32-shot ICL uses fewer queries. How do the results change when checkpoints are chosen without the evaluation set and all methods use the same queries?

A1. STAVE changes by at most 0.34 points when its checkpoints are chosen on a separate validation split, and 16/32-shot ICL scored on the full evaluation sets stays far below STAVE in every column.

(1) Already in the paper. Appendix F.2 states: "MimIC and HiFICL report results from the best-performing epoch on the evaluation split of these benchmarks, and we follow this protocol", and "Tables 2 and 10 instead select the checkpoint on validation samples". The protocol of Table 1 sets aside no validation split. Tables 2 and 10 score only the selected checkpoint on test, and Table 3 makes every per-task choice on a development split, never on the test split (Appendix F.4). There, STAVE has the best average on both Table 2 backbones, the best mean on five of the six Table 10 columns, and the highest Table 3 average.

(2) New: Table 1 with validation-chosen checkpoints. For each dataset we built a validation split from the training split that shares no image with the training samples or the evaluation set (1,000 items for VQAv2 and OK-VQA, and 2,000 images for COCO). For every run, we scored all saved STAVE epochs on this split, kept the epoch with the highest validation score, and reported only that epoch on the full evaluation set. VQA accuracy in %, COCO CIDEr:

RowLLaVA VQAv2LLaVA OK-VQALLaVA COCOIdefics2 VQAv2Idefics2 OK-VQAIdefics2 COCO
STAVE, validation-chosen75.81 ± 0.4255.47 ± 0.36130.44 ± 1.2073.39 ± 0.0660.91 ± 0.44134.04 ± 1.06
Change vs. Table 1 STAVE-0.26-0.34-0.09-0.04-0.09-0.14
Strongest two other methods in Table 1HiFICL 75.36, MimIC 75.19MimIC 55.74, HiFICL 54.50HiFICL 130.24, LIVE 129.75PT-Pre 72.75, HiFICL 71.61HiFICL 60.29, MimIC 59.69LoRA 132.84, MimIC 132.42

The selection rule changes STAVE by at most 0.34 points. Even with STAVE chosen on the validation split and other methods still chosen on the evaluation set, STAVE outperforms or matches the strongest methods of Table 1 in every column.

(3) New: 16/32-shot ICL on the full evaluation sets. The paper scored these two rows on the first 1,000 items because of their prompt length. We now score them on the same queries as every other row (10,000 VQAv2, 5,046 OK-VQA and 5,000 COCO items) with the demos of the same three runs.

Column16-shot ICL, full set32-shot ICL, full setSTAVE
LLaVA VQAv266.24 ± 0.0462.91 ± 0.3975.81 ± 0.42
LLaVA OK-VQA46.61 ± 0.2141.81 ± 0.2655.47 ± 0.36
LLaVA COCO117.74 ± 0.72116.28 ± 0.48130.44 ± 1.20
Idefics2 VQAv266.85 ± 0.1353.67 ± 0.1973.39 ± 0.06
Idefics2 OK-VQA57.55 ± 0.1653.42 ± 0.2060.91 ± 0.44
Idefics2 COCO122.52 ± 0.69114.72 ± 0.34134.04 ± 1.06

STAVE is above both ICL rows in every column, by 7.49 to 19.72 points over 32-shot ICL, with no demos in its prompt. In every column 32 demos score lower than 16, so adding demos does not close the gap. This is a weakness of ICL itself: beyond 16 demos, more demos lower its score instead of raising it, by 1.46 to 13.18 points across the six columns.

W2/Q2. PT trains for 500 steps on the first answer token, while STAVE can train for 4,000 steps on the full answer. Does STAVE still outperform PT with the full answer and a comparable training and tuning budget?

A2. Yes. Even with the full-answer loss and 4,000 steps, 4 times as many as STAVE uses on VizWiz and OK-VQA, all three PT modes stay below STAVE in every VQA cell. On Flowers, the only dataset where STAVE itself trains for 4,000 steps, STAVE is above every PT mode on Idefics2 and comparable to PT on Qwen-VL. Across all five Table 2 datasets, STAVE keeps the best average on both backbones. We also did more than tune PT's hyperparameters: PT is the only method whose learning rate was adjusted for individual runs, so its scores are in fact optimistic (see (3)). Points (2) to (5) continue in 2/2.

(1) Scope. The difference applies to Table 2 only. In Table 1, PT and STAVE already train with the same full-answer loss, for the same number of epochs, on the same 1,000 items, and PT's learning rate is tuned (Appendix F.2). There, STAVE is above all three PT modes in all six columns.

AuthorsOpenReview ↗

Response to Reviewer Yu4N (2/2)

A2 (continued).

(2) New: PT retrained with the full-answer loss. We retrained all three PT modes with the full-answer cross-entropy, and set PT's training to 4,000 steps on VizWiz, OK-VQA and Flowers. STAVE itself trains for 4,000 steps only on Flowers and for only 250 to 1,000 steps on the other datasets. On DTD and CUB, its steps are no more than PT's 500 in Table 2, so there PT keeps its 500 steps and only its loss changes. Accuracy (%):

CellSTAVEPT-PrePT-AppPT-BL
Idefics2 VizWiz69.63 ± 0.3769.12 ± 0.5568.00 ± 0.6164.63 ± 7.07
Idefics2 OK-VQA59.59 ± 0.5059.00 ± 0.4558.92 ± 0.5159.23 ± 0.33
Idefics2 DTD95.36 ± 0.4795.04 ± 0.2094.30 ± 0.3894.66 ± 0.49
Idefics2 Flowers95.67 ± 0.4292.53 ± 5.9793.20 ± 1.0093.40 ± 0.28
Idefics2 CUB93.50 ± 0.7278.88 ± 25.3395.00 ± 0.3995.35 ± 0.31
Qwen-VL VizWiz68.60 ± 0.5567.48 ± 0.6866.58 ± 0.4566.92 ± 0.86
Qwen-VL OK-VQA62.39 ± 0.1259.83 ± 0.3558.92 ± 0.5359.58 ± 0.62
Qwen-VL DTD95.20 ± 0.3295.09 ± 0.3293.40 ± 0.9279.96 ± 24.53
Qwen-VL Flowers96.30 ± 0.2396.59 ± 0.5495.42 ± 0.3195.73 ± 0.15
Qwen-VL CUB95.56 ± 0.2695.39 ± 0.4194.07 ± 0.7094.68 ± 0.10

Even with many more training steps than STAVE, PT remains below it in all 12 VizWiz and OK-VQA comparisons. Over the five datasets, STAVE has the best average on both backbones (82.75 and 83.61 vs at most 81.88 and 82.88 for PT), and no PT mode is above STAVE in more than one of the 10 cells: PT-Pre only on Qwen-VL Flowers, and PT-App and PT-BL only on Idefics2 CUB, where PT-Pre falls to 78.88. Moreover, PT is less practical than STAVE:

  • Its score depends on where the virtual tokens are placed. PT-Pre is the best mode in seven cells and PT-BL in three, and the three modes differ by up to 16.47 points within a cell. Each task therefore needs a search over placements, which STAVE does not need.
  • Its training is unstable across runs. With the same data and number of steps, two Idefics2 VizWiz PT-BL runs reach 68.63 and 68.80, but the third ends at 56.47. One run of Idefics2 CUB PT-Pre and one of Qwen-VL DTD PT-BL end near 50% (49.63 and 51.63), while the other two runs are above 93.
  • It costs more at inference. PT adds 20 virtual tokens to every prompt and has 10 times as many parameters, while STAVE keeps the zero-shot prompt length and cost (Table 7).

(3) Tuning budget: PT's scores are optimistic. PT is very sensitive to its learning rate, one more major drawback: a rate that trains well in one run can collapse in another on the same model and dataset (Appendix F.2). So in Table 2, we retrained every failed PT run at a retuned learning rate. No other method received such a per-run adjustment, so PT's scores are optimistic compared with those of STAVE and the task vector methods. Examples on Idefics2 (Table 2 setting):

PT entryLearning rateAccuracy (%)
VizWiz PT-BL3×10⁻³66.25 in one run, 0.01 in another (66.59 after changing to 5×10⁻³)
OK-VQA PT-BL10⁻²55.50 in one run, 0.11 in another (56.84 after changing to 10⁻³)
VizWiz PT-Pre, one run10⁻² / 3×10⁻³ / 10⁻³ / 3×10⁻⁴43.39 / 0.00 / 0.00 / 64.74

A collapsed run outputs a repeated word (for example, "un un un"). In Table 2, a rate that was stable in one run failed in another run in 8 of the 15 Idefics2 PT entries, while no STAVE run on these cells diverged.

(4) Beyond the mean score. STAVE is also more stable than PT. Across repeated runs with the same data and configuration, whether a test answer is correct changes on 14 to 21% of the items for PT but on 9 to 12% for STAVE, in every cell and for every PT mode.

Items whose correctness changes across runs (%)Idefics2 VizWizIdefics2 OK-VQAQwen-VL VizWizQwen-VL OK-VQA
STAVE9.2412.0510.4412.09
Best PT mode of Pre, App, BL14.4216.5317.5718.95

(5) PT makes little use of visual demos. On TrueMICL [1], whose tasks are built to require the demo images, we trained PT with STAVE's data, steps and learning rate, and no demos in its training prompt. Results on Qwen2-VL-7B-Instruct with 4 demos at test (Avg7 is TrueMICL's 7-task average):

Method (+ 4 demos at test)Avg7Task-level Avg.Episode-level Avg.Avg7 change from adding the 4 demos
PT-Pre78.19 ± 2.9390.33 ± 2.9052.19 ± 6.16+7.57
PT-App58.48 ± 9.4574.58 ± 9.3238.96 ± 11.13-7.90
PT-BL67.10 ± 2.1981.67 ± 4.8646.89 ± 1.15-1.14
STAVE86.38 ± 1.1595.58 ± 1.5166.88 ± 0.55+16.57

The 4 demos raise STAVE by 16.57 Avg7 points but PT-Pre by only 7.57, and lower PT-App and PT-BL. STAVE + 4 demos is 8.19 to 27.90 Avg7 points above the three PT modes.

We hope these results resolve both concerns and welcome any further questions.

[1] Chen et al. True Multimodal In-Context Learning Needs Attention to the Visual Context. COLM 2025.

Reviewer_qk6rOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes STAVE, learning only two vectors (one Readout vector and one Context vector) for task adaptation. Both vectors are added directly to the existing input tokens. The authors train these two vectors on pairs of prompts (one with examples and one without) using standard answer loss. Experiments across 6 vision-language models and 5 text LLMs show that STAVE works as well as or better than heavier methods, while using far fewer parameters and keeping zero-shot speed.

Critical strengths
  1. The paper is well organized and written.

  2. The proposed method is easy to follow and efficient.

Critical weaknesses
  1. Proposition 2.1 considers unrestricted per-token perturbations, but STAVE only uses two shared vectors. This raises my concern, since two vectors have much fewer degrees of freedom. Also, theorem 3.1 does not show that two vectors can achieve similar effects as deep-layer interventions. So I think the theory does not fully support the proposed design.

  2. The authors show that the readout and context gradients are nearly orthogonal in Figure 2(c), which supports the two-vector design. However, this analysis is only done for one model and task. It is not clear whether similar gradient patterns also apply to other models and tasks.

  3. Most experiments are conducted with short-form benchmarks (like VQA, classification, or single-token completion). For tasks that require long text generation or multi-step reasoning, it is not clear whether two input vectors are sufficient to guide the entire decoding process.

Questions for authors
  1. How does STAVE perform on tasks that require generating long answers (such as step-by-step math reasoning or code)? Does the effect of the two input vectors fade away as the generated output gets longer?

  2. How many training samples do you really need? If a user only has 8 or 16 labeled examples total, can STAVE still train reliably without overfitting?

  3. How long does the paired training step take on an 8B model compared to standard LoRA fine-tuning?

9 comments
AuthorsOpenReview ↗

Response to Reviewer qk6r (1/4)

We sincerely thank the reviewer for the precise and insightful questions. We ran a new experiment or analysis for each.

W1. Proposition 2.1 considers unrestricted per-token perturbations, while STAVE uses two shared vectors, and Theorem 3.1 does not show that two vectors can match deep-layer interventions. So the theory does not fully support the design.

A1. The theory supports the design, and two shared input vectors match deep-layer interventions on the tasks. The paper's results determine where the update enters, why it is split into a readout and a context vector, and what sharing costs. A direct comparison with a pair of vectors at every layer, which has 32 times more parameters, shows that the added depth fits the training data better but brings no clear gain on new queries.

(1) What each result establishes for the two shared vectors. W1 raises two questions: where the update enters and how few vectors it can use. The paper answers them with different results.

  • Where. Prop 2.1 concerns only the location: an input update acts through every block, while an update after block ℓ reaches only the later blocks (p2, L94 to L103). It is not the argument for the number of vectors.
  • Why two vectors and not one (new identity). Eq. 1 is stated for one vector shared by a set of tokens. It gives an exact identity (derived in A2): at the same update norm, a readout vector and a context vector reach a larger first-order loss decrease than one shared vector whenever their mean gradients differ. Appendix H.5 measures 3.4 to 7.5 times the held-out loss decrease.
  • Why not more. Theorem 3.1 bounds the cost of sharing: tying structural token groups costs at most ρδ_q + κρ² in margin, where δ_q measures the between-group gradient contrasts that tying discards. In practice, one vector per structural token group scores below the two vectors in all four tested cells (Figure 4(a)).

Eq. 1 and Theorem 3.1 are both stated for shared vectors, so they analyze STAVE with the degrees of freedom it has.

(2) New theoretical analysis: two shared vectors vs deep-layer interventions. Take the deep intervention closest to STAVE, a readout/context pair at the input of every decoder block (2L vectors for L blocks). This class contains STAVE exactly: with STAVE's two vectors in the first pair and zeros in all others, it is STAVE on every prompt. The extra vectors can therefore only lower the best training loss. What they add on new queries depends on the estimation term of the paper's target-risk bound (Section 3 and Appendix D.3), which grows with the number of trained coordinates: 2d for STAVE and 2Ld for the every-layer class at the same radius and loss constants, on the same training samples. Sharing keeps this term small, and the paper's sample-size study agrees: with a tenth of the training samples STAVE loses a few points, while the every-layer methods MimIC and HiFICL lose far more on Idefics2 (Figure 7). (3) tests which of the two effects decides.

(3) New experiment: a pair of vectors at every layer. All other settings are as in STAVE (L = 32, so 32 times more parameters). Full evaluation sets, protocol of Tables 1 and 2:

CellSTAVE (input, 2 vectors)Every layer (2L vectors)
LLaVA VQAv276.07 ± 0.1076.01 ± 0.26
LLaVA COCO130.53 ± 1.17130.26 ± 0.29
Idefics2 VizWiz69.63 ± 0.3769.44 ± 0.71
Idefics2 OK-VQA59.59 ± 0.5059.97 ± 0.91

The every-layer variant is lower in three of the four cells, and every difference is within one standard deviation, so the added depth brings no clear gain.

The losses agree with (2). On the two Idefics2 cells, the every-layer variant ends training with about half of STAVE's training objective (0.55 vs 1.08 on VizWiz, 0.33 vs 0.65 on OK-VQA), yet its answer loss on the validation samples is higher at the selected checkpoint (0.665 vs 0.607 and 0.401 vs 0.329), in all 6 pairs of runs. Both variants select their checkpoint on these validation samples, as in Table 2, so early stopping is already applied. Stopping earlier does not change this: the lowest validation loss that the every-layer variant reaches at any point of training (evaluated every 100 steps) is still above STAVE's (0.644 vs 0.607 and 0.401 vs 0.329), in all 6 pairs of runs. The extra vectors improve the fit to the training samples, not the answers to new queries.

(4) Published deep-layer methods agree. LIVE, MimIC and HiFICL learn a shift in every decoder layer and store 14 to 2,080 times more parameters than STAVE. STAVE still has the best score in every column of Table 1, the best average on both backbones of Table 2, and the best mean in five of the six columns of Table 10. Among single injection depths, the input is best or tied for best on all four curves (Figure 4(b)).

AuthorsOpenReview ↗

Response to Reviewer qk6r (2/4)

W2. The near-orthogonality of the readout and context gradients in Figure 2(c) is shown on one model and task. Does it hold for other models and tasks?

A2. Yes, in every setting we tested. We repeated the analysis on all 11 backbones of the paper: every LMM cell of Tables 1, 2 and 10 (24 cells, with Qwen2.5-VL at two image resolutions) and all TV-benchmark tasks on the five LLMs (90 cells). For each cell we take the frozen model before training and compute the mean gradient of the answer loss with respect to the readout vector and to the context vector.

BackboneCellsReadout/context cosine, range (median)Split-half agreement (median)
LLaVA-Interleave30.004 to 0.189 (0.146)0.999
Idefics280.034 to 0.115 (0.079)0.990
Qwen-VL50.038 to 0.245 (0.105)0.993
Qwen2.5-VL4−0.005 to 0.289 (0.064)0.979
InternVL3.520.003 to 0.014 (0.008)0.970
Idefics32−0.061 to −0.001 (−0.031)0.968
5 LLMs90−0.068 to 0.2090.898 to 0.945 (per LLM)

In all 114 cells the two gradients are nearly orthogonal (cosines within ±0.29, angles of at least 73°), while each gradient is stable across disjoint halves of the prompts (median agreement 0.898 to 0.999 across backbones), so the near-orthogonality is not noise.

Why different gradients support two vectors. Eq. 1 of the paper gives the largest first-order loss decrease that one vector shared by a token set S can reach within norm ρ, Γ(S) = ρ‖Σ_{t∈S} g_t‖/√|S|. Two vectors on disjoint sets A and B with the same total norm reach Γ_split = √(Γ(A)² + Γ(B)²) (Section 2). Subtracting the two gives an exact identity for one prompt at the zero-shot start:

Γ_split² − Γ(A ∪ B)² = ρ²·n_A·n_B/(n_A + n_B)·‖ḡ_A − ḡ_B‖²,

where n_A and n_B count the readout and context tokens and ḡ_A and ḡ_B are their mean per-token gradients. The readout/context split therefore lowers the loss more than one shared vector whenever the two mean gradients differ, and the gain grows with the distance between them. Appendix H.5 measures this gain on LLaVA: at the same update norm, the readout/context split lowers held-out loss 3.4 to 7.5 times as much as one shared vector, while random splits with the same token counts stay within 5% of it. The 114 cells show that the two gradients point in different directions on every backbone, so this gain is not specific to the setting of Figure 2(c).

AuthorsOpenReview ↗

Response to Reviewer qk6r (3/4)

W3/Q1. Most evaluations are short-form. How does STAVE perform on long answers such as step-by-step math or code? Does the effect of the two input vectors fade as the output gets longer?

A3. With no demos, STAVE beats the ICL it replaces, as well as MimIC and HiFICL, on both math reasoning and code, and its effect shows no sign of fading: up to 256 generated tokens it is larger than that of the 8 demos, and beyond that both are near zero.

(1) Long-answer benchmarks. GSM8K [1] with chain-of-thought (CoT) answers (1,319 test problems, greedy, up to 512 new tokens) and MBPP [2] code generation (500 test problems, pass@1). ICL uses the standard prompt of each benchmark: the 8 chain-of-thought demos of [3] on GSM8K, and the 3-shot prompt of the MBPP paper [2] (its prompt problems 2 to 4). Trained methods use the backbone's own correct ICL solutions on training questions, use no demos at test, and choose the checkpoint by loss on a held-out validation split. MimIC and HiFICL use our port of their official attention code with the papers' settings. Zero-shot and ICL are deterministic.

GSM8K accuracy (%). Strict match requires the final line "The answer is N", and last-number match extracts the last number, with every output cut at its first blank line.

Method# Params (M)Llama-3.1-8B base, strictLlama-3.1-8B base, last numberLlama-2-7B, strictLlama-2-7B, last number
Zero-shot00.08 ± 0.0045.49 ± 0.000.08 ± 0.004.25 ± 0.00
8-shot CoT ICL054.6657.3214.6315.47
MimIC0.263.035.8411.09 ± 0.5611.52 ± 0.80
HiFICL2.230.000.680.40 ± 0.642.53 ± 1.03
STAVE0.00857.37 ± 0.6957.70 ± 0.8516.76 ± 0.2016.91 ± 0.23

MBPP pass@1 (%), Llama-3.1-8B base:

Method# Params (M)pass@1
Zero-shot040.80 ± 0.00
3-shot ICL048.00
MimIC0.2647.13 ± 2.44
HiFICL2.2328.40 ± 3.30
STAVE0.00849.80 ± 0.35

STAVE is the best method in both tables, with no demos at test and the fewest trainable parameters. MimIC and HiFICL collapse on GSM8K with Llama-3.1-8B base, where nearly all of their outputs run to the 512-token limit (1,315 and 1,305 of 1,319). Their released code computes the attention normalizer without a causal mask, so in teacher-forced training each answer token sees later answer tokens that greedy decoding does not have.

(2) No sign of fading along the answer. We compare STAVE with the 8 demos it replaces at each position of the generated answer (GSM8K, Llama-3.1-8B base, mean over 3 runs). The loss reduction is the per-token loss decrease relative to zero-shot on the same reasoning text (the 8-shot model's own solutions). The attention is the share that each generated token gives to the prompt tokens updated by the two vectors, with and without the vectors on the same text.

Generated tokens1 to 1617 to 6465 to 128129 to 256257 to 512
Loss reduction, STAVE0.5780.3390.4250.2670.005
Loss reduction, 8-shot ICL0.4300.2390.3730.2590.006
Attention to the updated prompt tokens, STAVE (%)24.214.210.77.15.3
Attention to the same tokens, no vectors (%)23.412.48.55.23.6

Up to 256 generated tokens, the two vectors lower the loss more than the 8 demos do in every range. Beyond 256 tokens (under 5% of the tokens), the text is nearly deterministic and both reductions are near zero. The reduction of the 8 demos rises and falls from range to range in the same way, so the vectors' effect lasts as long as that of the demos they replace. The attention that the vectors add does not shrink: after the first 16 tokens it stays at +1.7 to +2.2 points (at least +1.5 in every run), and relative to the model without the vectors it grows from 1.03 times in the first 16 tokens to about 1.5 times beyond 256. Accuracy agrees: with the 1,319 test problems split into quartiles by the length of their reference solution (cut points at 35, 51 and 70 words), STAVE is above 8-shot ICL on the same problems in every quartile and every run (from shortest to longest: 73.23 / 65.89 / 52.86 / 38.02 vs 70.16 / 62.79 / 49.39 / 36.83).

(3) Why. Every answer token after the first reads the cached keys and values of the prompt tokens, which the two vectors change, so it depends on them through both the attended values and their attention weights. The added attention in (2) shows the second path at work, and the full-answer loss trains the vectors at every answer token. By contrast, a shift added after the last block leaves the cache unchanged and has neither path (Prop 2.1(i)).

[1] Cobbe et al. Training Verifiers to Solve Math Word Problems. arXiv 2021.

[2] Austin et al. Program Synthesis with Large Language Models. arXiv 2021.

[3] Wei et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022.

Reviewer_iY4eOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

STAVE adapts a frozen model with two task-specific vectors added to existing input embeddings: one for the answer-producing positions and one for the context. The vectors are trained with answer supervision on prompts with and without demonstrations. At inference, demonstrations are removed and no extra tokens are added. Experiments on multimodal and text tasks show competitive performance with only 2d trainable parameters. The paper also provides local theoretical analysis and ablations to motivate the split.

Critical strengths
  1. The method is simple and useful. Storing only two vectors per task, without adding prompt tokens or changing backbone weights, makes deployment attractive. The results show that this small update can compete with much larger adaptations on the evaluated tasks.
  2. The readout/context split is well motivated and tested. Comparisons with a single vector, modality-based splits, and more vectors help explain why the grouping matters. The analysis of answer formatting versus query-dependent evidence is also informative.
Critical weaknesses
  1. The model coverage is relatively old, especially for text tasks, which use Pythia, GPT-J, and the original LLaMA. The VQAv2 results rely on LLaVA-Interleave and Idefics2. The additional InternVL3.5, Qwen2.5-VL, and Idefics3 experiments broaden the evaluation, but still leave the gains on current, stronger instruction-tuned models unclear. This matters because the paper attributes part of the improvement to correcting answer format. That problem may be less pronounced in stronger models.
  2. The benchmarks do not establish that STAVE replaces learning from visual demonstrations. Prior work like [1] distinguishes learning from demo images from recognizing a task through demo text. A model can use the query image while largely ignoring the demo images. The current results do not separate these cases, and the paper's own format-correction analysis makes this distinction particularly relevant. The evidence supports efficient supervised task adaptation, but the broader interpretation of replacing multimodal demonstrations needs stronger support.

[1] True Multimodal In-Context Learning Needs Attention to the Visual Context (COLM 2025)

Questions for authors
  1. Do the gains persist on a current, stronger instruction-tuned backbone with a well-matched prompt? In particular, how much improvement remains when the unadapted model already follows the requested answer format?
  2. What happens if STAVE is retrained with demo images removed or shuffled, while keeping demo questions, answers, and the query image unchanged? This control would clarify whether visual information in the demonstrations contributes to the learned vectors. Evaluation on TrueMICL or a similar task requiring demo images would provide stronger evidence, with the per-task training protocol made explicit.

These are requests for evidence that could affect my final assessment. Overall, I like this paper: the writing is clear, the method is compact, practical, and supported by useful ablations with extensive experiments. My main concerns are about the choice of evaluated models and the evaluated MICL tasks

5 comments
AuthorsOpenReview ↗

Response to Reviewer iY4e (1/3)

We sincerely thank the reviewer for the encouraging assessment and the clear, actionable questions. We ran every requested experiment.

W1/Q1. Do the gains persist on a current, stronger instruction-tuned backbone with a well-matched prompt? How much improvement remains when the unadapted model already follows the answer format?

A1. Yes, on both LMMs and LLMs. The paper already includes instruction-tuned LMMs: Table 2 uses "Qwen-VL-7B, used in its chat release" (checkpoint Qwen/Qwen-VL-Chat in Appendix F.1), where STAVE has the highest VizWiz and OK-VQA accuracy, and Table 10 covers 2025 LMMs (Qwen2.5-VL-7B-Instruct, InternVL3.5-8B) and Idefics3-8B, where STAVE exceeds 4-shot ICL by more than 10 points on every column. We add three results, (1) and (2) below and (3) in 2/3.

(1) New LMMs with official and best prompts. Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B-Instruct, Table 2 protocol (checkpoint chosen on validation samples, full test splits). All methods use the official VQA prompt: the model's chat template with the lmms-eval instruction Answer the question using a single word or phrase. (on VizWiz preceded by When the provided information is insufficient, respond with 'Unanswerable'.). For zero-shot and 4-shot ICL, we also fixed four or five prompts per dataset in advance (rewordings of this instruction and the plain template of Table 10), chose the best on 300 validation items, and report the higher of that prompt and the official one. For zero-shot, the best prompt is Answer the question using a single word or phrase. If the question cannot be answered from the image, answer 'unanswerable'. on VizWiz, Please try to answer the question with short words or phrases if possible. on VQAv2, and the official prompt elsewhere. For ICL, it is the official prompt in all four cells.

CellZero-shotZero-shot, best prompt4-shot ICLLIVEMimICHiFICLSTAVE
Qwen3-VL VizWiz70.53 ± 0.0072.68 ± 0.0070.11 ± 0.0572.51 ± 0.5473.72 ± 0.4872.21 ± 2.8773.85 ± 0.29
Qwen3-VL OK-VQA58.57 ± 0.0058.57 ± 0.0058.58 ± 0.1659.94 ± 0.9061.56 ± 1.0261.84 ± 0.6162.67 ± 0.48
Qwen3-VL VQAv281.12 ± 0.0081.31 ± 0.0081.57 ± 0.0781.95 ± 0.1580.50 ± 1.3981.31 ± 0.5682.39 ± 0.13
Qwen2.5-VL OK-VQA58.92 ± 0.0058.92 ± 0.0061.51 ± 0.1962.17 ± 0.4162.18 ± 1.5558.50 ± 1.0962.81 ± 0.47
# Params (M), Qwen3-VL / Qwen2.5-VL0000.15 / 0.100.30 / 0.202.51 / 1.710.008 / 0.007

STAVE is the best method in all four cells, with the fewest parameters and no demos at test. A better prompt raises zero-shot by up to 2.15 points, and STAVE still leads zero-shot and ICL by 0.82 to 4.10 points. It is also above the task-vector-related methods in every cell.

(2) New LLMs with official chat templates. TV benchmark. The system turn gives a one-line task instruction and Answer with only the answer., so zero-shot is given both the task and format. Baselines keep their Table 3 settings. Strict exact match (%), all tasks / tasks where zero-shot is below 80%:

MethodQwen2.5-7B-InstructLlama-3.1-8B-Instruct
Zero-shot79.75±0.00 / 56.76±0.0078.33±0.00 / 58.35±0.00
1-shot ICL82.03±0.44 / 63.44±0.7079.94±0.48 / 61.23±0.90
5-shot ICL83.77±0.18 / 70.88±0.4587.20±0.56 / 77.26±1.19
15-shot ICL86.59±0.14 / 78.40±0.5390.92±0.32 / 85.52±0.86
LoRA r = 890.41±1.13 / 88.35±0.5993.15±0.44 / 89.75±0.41
Prefix87.47±0.39 / 86.82±0.7791.57±0.45 / 88.49±0.96
PT-Pre84.66±1.33 / 74.91±5.3786.88±1.81 / 77.26±5.29
PT-App85.78±1.25 / 86.73±1.1189.08±3.34 / 88.17±1.97
PT-BL87.07±1.82 / 88.18±1.5991.92±0.66 / 88.33±0.66
FV81.65±0.47 / 61.35±1.3480.16±0.35 / 62.71±0.91
TV81.36±0.39 / 60.88±1.4081.57±0.43 / 65.62±0.61
SITE89.96±0.27 / 86.82±0.8291.02±0.43 / 85.17±0.62
STAVE91.36±0.37 / 88.73±0.1993.18±0.20 / 90.01±0.65

STAVE gains 11.61 to 14.85 points over zero-shot and stays 2.26 to 10.33 points above 15-shot ICL while using no demos. It is above SITE, the closest baseline in Table 3, in every run, and matches or exceeds LoRA with 350 to 420 times fewer parameters. Prefix, the three PT modes, FV and TV are also below STAVE on both backbones and in both columns.

AuthorsOpenReview ↗

Response to Reviewer iY4e (2/3)

A1 (continued).

(3) Gain when the format is already right. We labeled each zero-shot answer by whether it follows the answer format (a short answer for VQA, a single answer for text, no full sentence or template tokens) and measured STAVE's gain only on the queries where it does:

SettingZero-shot answers that follow the formatSTAVE gain on these queries
Qwen3-VL-8B-Instruct, OK-VQA99.45%+3.97
Qwen3-VL-8B-Instruct, VizWiz95.05%+2.03
Qwen2.5-7B-Instruct, TV benchmark99.47%+11.28
Llama-3.1-8B-Instruct, TV benchmark99.13%+14.21

Most of the gain is thus not from format. Two further checks agree. When every method is rescored with a relaxed scorer that also accepts a correct answer inside a longer output, so that format errors cost no method any points, STAVE stays 0.81 to 3.84 points above zero-shot and 0.50 to 2.91 above 4-shot ICL in all four LMM cells. On the two LLMs, STAVE's strict score is 9.64 and 10.32 points above the zero-shot score even under this relaxed scoring. STAVE's gains therefore persist on these stronger models, where answer format is rarely a problem (under 5% of zero-shot answers).

W2/Q2. The benchmarks do not establish that STAVE replaces learning from visual demos. What happens if STAVE is retrained with demo images removed or shuffled, keeping demo questions, answers and the query image unchanged? Evaluation on TrueMICL [1] would provide stronger evidence, with the per-task training protocol made explicit.

A2. Yes, visual information in the demos contributes to the learned vectors: removing or shuffling the demo images in training erases nearly all of STAVE's gain over 4-shot ICL (Avg7 86.38 to 78.52 and 78.05, ICL 77.81). TrueMICL has two kinds of tasks, and STAVE + 4 demos is best on both. Where the rule is fixed, the vectors store it and replace the visual demos: with no demos at test, STAVE is above 4-shot ICL (94.08 vs 90.42). Where the rule changes in every episode, it must be read from the test demos, and STAVE reads it best (see (5)).

(1) Setup and per-task protocol. We use TrueMICL [1], whose tasks are built to require the demo images, with its main backbone Qwen2-VL-7B-Instruct, on its 7 tasks plus CLEVR-hidden (a CLEVR variant we add in which the counted attribute is not named in the query). Task-level tasks (Sudoku, Palindrome, Character, CLEVR) keep one rule for the whole task, and episode-level tasks (Operator, Clock, Outlier, CLEVR-hidden) change it in every episode. STAVE trains one pair of vectors per task on that task's released support set only. Each support item is one training query (on Operator and Clock, each support image is used once per operator). The source prompt holds 4 demos drawn from the other support items by the task's own demo rule, and the target prompt is the same query without demos. All tasks share the same setting: lr 1e-3, 300 steps, 4 pairs per step, final vectors, no checkpoint selection. Following TrueMICL's 4-shot protocol, every method except zero-shot gets the same 4 demos at test.

(2) TrueMICL results. All trained methods train on the same support set:

MethodAvg7Avg6 (no CLEVR)Task-level Avg.Episode-level Avg.
Zero-shot46.14 ± 0.1437.67 ± 0.1758.50 ± 0.2531.36 ± 0.00
Random 4-shot ICL77.81 ± 0.6477.28 ± 1.0890.42 ± 1.0150.66 ± 1.47
LoRA78.43 ± 1.4577.28 ± 0.9291.00 ± 1.6453.49 ± 1.59
DARA (reported in [1])81.83
DARA (our run, released code and settings)69.05 ± 3.7971.78 ± 4.0979.92 ± 1.7747.30 ± 4.79
MimIC + 4 demos75.86 ± 1.9874.56 ± 2.1289.00 ± 2.3855.12 ± 1.74
HiFICL + 4 demos75.67 ± 3.3378.78 ± 1.6483.50 ± 6.6358.96 ± 1.44
PT-Pre + 4 demos78.19 ± 2.9377.28 ± 3.1790.33 ± 2.9052.19 ± 6.16
PT-App + 4 demos58.48 ± 9.4554.56 ± 10.6974.58 ± 9.3238.96 ± 11.13
PT-BL + 4 demos67.10 ± 2.1964.89 ± 0.9681.67 ± 4.8646.89 ± 1.15
STAVE + 4 demos86.38 ± 1.1584.61 ± 1.3595.58 ± 1.5166.88 ± 0.55
STAVE + 4 demos, images removed at test54.90 ± 1.1547.78 ± 1.2767.75 ± 1.5037.83 ± 1.38
STAVE + 4 demos, images shuffled at test72.24 ± 0.5068.11 ± 0.5977.92 ± 0.5259.95 ± 1.47

Avg7 is TrueMICL's 7-task average. The released CLEVR queries name the counted attribute (zero-shot 97.00), so we compare with the DARA value reported in [1] on the other 6 tasks. Our DARA run is below the reported value, likely because the released data include separate training-query and training-support splits for only 3 of the 7 tasks.

  • STAVE + 4 demos is best on every average: 8.57 Avg7 points above 4-shot ICL, 10.52 and 10.71 above MimIC and HiFICL, 8.19 to 27.90 above PT, and above DARA [1], both as reported (84.61 vs 81.83) and in our run (84.61 vs 71.78).
  • STAVE uses the demo images at test. Removing the demo images at test costs it 31.48 Avg7 points and shuffling them costs 14.14.

[1] Chen et al. True Multimodal In-Context Learning Needs Attention to the Visual Context. COLM 2025.

AuthorsOpenReview ↗

Response to Reviewer iY4e (3/3)

A2 (continued).

(3) Control: demo images removed or shuffled in training. We retrained STAVE and changed only the demo images in its training prompts. Demo questions, demo answers, the query image and all settings stay the same. Shuffling pairs each demo image with another demo's question and answer. Every model is tested with 4 normal demos.

Training conditionAvg7Task-level Avg.Episode-level Avg.
Normal demo images86.38 ± 1.1595.58 ± 1.5166.88 ± 0.55
Demo images removed78.52 ± 2.3287.58 ± 1.9456.84 ± 1.89
Demo images shuffled among the demos78.05 ± 1.8082.75 ± 4.8765.36 ± 1.84

Both controls lower STAVE in every run, by 6.57 to 9.00 Avg7 points when the demo images are removed and by 6.00 to 11.57 when they are shuffled. Trained without the correct demo images, STAVE + 4 demos scores about the same as the frozen model given the same 4 demos (Avg7 78.52 and 78.05 vs 77.81 for 4-shot ICL). Removing the demo images adds no wrong pairing and changes nothing else, and it alone takes away 7.86 of STAVE's 8.57-point gain over ICL (92%). Shuffling takes away 8.33. The demo images are therefore the source of STAVE's gain over ICL, and visual information in the demos does contribute to the learned vectors.

(4) The learned vectors also change. Cosine similarity to the vectors trained on normal demo images (mean over 8 tasks and 3 runs, same training episodes unless noted):

Trained withReadout vectorContext vector
Normal demo images, retrained (reference)0.940.97
Normal demo images, different training episodes (reference)0.740.86
Demo images shuffled0.660.79
Demo images removed0.520.64

In all 8 tasks, changing the demo images moves both vectors farther than retraining does, and removing the images moves them farther than using different training episodes does. The shift is consistent across runs: vectors trained without images agree with each other as much as the normal ones do (0.85 vs 0.86, context vector).

(5) On replacing visual demos (W2). The two kinds of tasks answer this separately.

  • Task-level tasks: the vectors replace the visual demos. One rule holds for the whole task, and on average the frozen model finds it only from the demo images: zero-shot scores 58.50 and 4 demos without their images score 59.67, while 4 full demos score 90.42. So ICL must read the support examples and their images at test. STAVE reads the same support examples once, in training, and then scores 94.08 ± 0.76 with no demos at test, above 4-shot ICL in every run. For example, on Palindrome the frozen model scores 28.00 with no demos and 96.67 with 4 demos, while STAVE scores 99.67 with no demos. Adding the 4 demos back raises it by only 1.50 points. The frozen model and the demo-free prompt are the same as in zero-shot, so the two vectors are the only source of this gain: the fixed visual rule is stored in the two vectors, and the demos can be removed at test. This is the setting of Tables 1 and 2, where each task also has one fixed rule.
  • Episode-level tasks: the vectors make the model use the visual demos better. The rule changes in every episode, so the 4 demos at test carry it. There, the 4 demos raise STAVE by 27.66 points, and STAVE + 4 demos is 16.22 points above 4-shot ICL (66.88 vs 50.66) and above every other method.

We hope these results address both concerns, and we are happy to discuss any further questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their careful and constructive reviews. We are encouraged that reviewers find STAVE simple, efficient and practical to deploy, with two vectors per task at zero-shot prompt length and cost (Yu4N, iY4e, qk6r), its gains consistent across backbones and competitive with much larger adaptations (Yu4N, iY4e), the readout/context split well motivated and supported by ablations (Yu4N, iY4e), and the paper clear and well organized (Yu4N, iY4e, qk6r).

To address the reviewers' concerns, we ran every requested experiment. The main results are summarized below as means over 3 runs, and each reviewer's thread gives the full tables (mean ± std) and protocols.

1. Current instruction-tuned backbones (iY4e). STAVE is the best method in all four Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B-Instruct cells, even when zero-shot and ICL use the best of up to five prompts:

CellZero-shot, best prompt4-shot ICL, best promptLIVEMimICHiFICLSTAVE
Qwen3-VL VizWiz72.6870.1172.5173.7272.2173.85
Qwen3-VL OK-VQA58.5758.5859.9461.5661.8462.67
Qwen3-VL VQAv281.3181.5781.9580.5081.3182.39
Qwen2.5-VL OK-VQA58.9261.5162.1762.1858.5062.81

On the TV benchmark with Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, STAVE reaches 91.36 and 93.18, above 15-shot ICL, FV, TV, SITE and LoRA. On queries that the unadapted model already answers in the required format, it still gains +2.03 to +14.21 points, so the gain is not an answer-format correction.

2. Visual demos (iY4e). On TrueMICL [1], whose tasks are built to require the demo images, STAVE + 4 demos is the best method:

TrueMICL, 4 demos at test4-shot ICLLoRAMimICHiFICLPT-PreSTAVE
7-task average77.8178.4375.8675.6778.1986.38

Retrained with only the demo images removed or shuffled (demo text and query image unchanged), STAVE falls in every run, from 86.38 to 78.52 and 78.05, close to 4-shot ICL with no training (77.81). Only the demo images changed, so visual information in the demos does contribute to the learned vectors.

3. Theory and design (qk6r). The theory explains where the update enters and why it uses two shared vectors, and a variant with a pair of vectors at every decoder layer (32 times more parameters) brings no clear gain: it is lower in three of four cells, and every difference is within one standard deviation. The near-orthogonal readout and context gradients of Figure 2(c) hold in all 114 cells on all 11 backbones (cosines within ±0.29).

4. Long outputs, reasoning and code (qk6r). With no demos, STAVE beats 8-shot chain-of-thought (CoT) ICL on GSM8K [2] and standard 3-shot ICL on MBPP [3] (49.80 vs 48.00 pass@1, zero-shot 40.80, Llama-3.1-8B base), and is also above MimIC and HiFICL on both benchmarks. GSM8K strict-match accuracy (%):

MethodLlama-3.1-8B baseLlama-2-7B
Zero-shot0.080.08
8-shot CoT ICL54.6614.63
STAVE57.3716.76

The effect shows no sign of fading in long outputs: up to 256 tokens, the vectors lower the per-token loss more than 8 demos do, and STAVE beats 8-shot ICL in every length quartile.

5. Few training examples and training cost (qk6r). About 100 training samples suffice (Figure 7). With only 8 examples, STAVE stays 9 points above zero-shot on held-out Idefics2 queries, while MimIC and HiFICL overfit and fall below it. It is above both for n = 8 to 64, and above ICL and SITE on the TV benchmark. A paired STAVE step is 1.5 times faster than a LoRA step, and full Table 1 training takes 19 to 33 minutes, against 70 to 104 for LIVE.

6. Stricter evaluation protocols (Yu4N). With STAVE's checkpoints chosen on a new validation split, its Table 1 scores change by at most 0.34 points, and it still outperforms or matches the strongest methods in every column, although those methods keep their evaluation-set checkpoints. 16/32-shot ICL scored on the full evaluation sets is far below STAVE in every column (32-shot shown).

Table 1 columnLLaVA VQAv2LLaVA OK-VQALLaVA COCOIdefics2 VQAv2Idefics2 OK-VQAIdefics2 COCO
STAVE, validation-chosen75.8155.47130.4473.3960.91134.04
Strongest other method in Table 1HiFICL 75.36MimIC 55.74HiFICL 130.24PT-Pre 72.75HiFICL 60.29LoRA 132.84
32-shot ICL, full set62.9141.81116.2853.6753.42114.72

Even with the full-answer loss and 4,000 training steps, 4 times as many as STAVE uses, all three PT modes stay below STAVE in every Table 2 VQA cell.

We hope these results resolve the concerns, and we welcome any further questions.

[1] Chen et al. True Multimodal In-Context Learning Needs Attention to the Visual Context. COLM 2025.

[2] Cobbe et al. Training Verifiers to Solve Math Word Problems. arXiv 2021.

[3] Austin et al. Program Synthesis with Large Language Models. arXiv 2021.

AuthorsOpenReview ↗

General Response: Summary of Revisions

Dear Reviewers,

We sincerely thank you for the time and effort you have devoted to our submission, and for your thoughtful and constructive feedback, which has helped us improve the paper. We have uploaded a revised manuscript that includes the results from our responses. Revised content is in blue. For your convenience, we list below where each concern is addressed.

Reviewer iY4e

  • W1/Q1 (instruction-tuned backbones). Table 3 adds Qwen2.5-7B-Instruct and Llama-3.1-8B-Instruct, and Table 8 adds Qwen3-VL-8B-Instruct and Qwen2.5-VL-7B-Instruct (Section 4.5). The prompts are in Appendices F.4 and G.1, and Table 20 gives the gain on queries whose zero-shot answer already follows the answer format.
  • W2/Q2 (visual demos). Table 10 and Section 4.5 report TrueMICL, including STAVE retrained with the demo images removed or shuffled. Appendix G.3 gives the per-task training protocol and the full comparison (Tables 21 and 22).

Reviewer qk6r

  • W1 (theory and design). Section 2 states the role of Proposition 2.1 and the exact gain of two vectors over one (Appendix C.3). Table 4 compares STAVE with a pair of vectors at every decoder layer (Section 4.4, Appendix H.7).
  • W2 (gradient analysis). Section 2 and Appendix C.4 (Table 11) report the readout and context gradients in 114 cells on 11 backbones.
  • W3/Q1 (long-form generation). Table 9 and Section 4.5 report GSM8K and MBPP, and Appendix G.4 adds the effect of the vectors along the generated answer (Tables 23 and 24).
  • Q2 (few training samples). Appendix G.5 (Tables 25 and 26).
  • Q3 (training time). Appendix E.4 (Table 14).

Reviewer Yu4N

  • W1/Q1 (evaluation protocol). Table 1 now scores 16/32-shot ICL on the full evaluation sets. Appendix F.2 (Table 16) reports STAVE with its epoch chosen on a held-out validation split, which changes its scores by at most 0.34 points.
  • W2/Q2 (PT budget). The PT rows of Table 2 now use the full-answer loss and up to 4,000 training steps (Appendix F.3, Tables 17 and 18). STAVE keeps the best average on both backbones.

Notes. Table 4 and Tables 8 to 10 are new, so the former Tables 4 to 6 are now Tables 5 to 7 and the appendix tables are renumbered.

We understand that you have a demanding schedule and a multitude of responsibilities, and we truly appreciate your attention. As the discussion deadline is approaching, we would be very grateful if you could kindly take a moment to look at our responses and the revised manuscript, and let us know whether any concern remains. We would be glad to provide any further clarifications.

Thank you once again for your time and consideration.

Best regards,

The Authors of Submission 2397

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.