acceptodds
Under review as a conference paper at ICLR 2027

Afterimage: A Discrete Adapter for Inspectable Multimodal Fusion

Abstract

Pretrained multimodal large language models (MLLMs) are increasingly used to fuse heterogeneous inputs into a dense, informative latent representation consumed by a downstream model. However, which domain features the fusion carries remains opaque, which limits transparency in high-stakes deployments and leaves unclear whether a prediction reflects the downstream decoder or the quality of the fusion. Prior interpretation approaches usually explain the model after training, such as through attention maps or probes, rather than recording which domain features the fusion itself carries. We present AFTERIMAGE, an adapter that makes part of a frozen MLLM's fusion inspectable on the forward path. AFTERIMAGE discretizes the residual stream at several depths into short code sequences from one shared codebook, giving a discrete inspectable record at each depth. From these records, we extract code motifs and verify them as associations with domain features. With only 0.20–0.71% added parameters, AFTERIMAGE both lets the domain features its own write carries be read from its codes and matches or improves task performance. AFTERIMAGE lowers counterfactual-age FID on 3D brain MRI synthesis from 1.93 to 0.43, raises π0.5 robot success by 26.0 points averaged over shifted LIBERO-Pro suites, and improves video question answering accuracy by 16.0 points over the frozen backbone. On withheld samples, the extracted motifs expose associations with features such as clinical diagnosis, demographic information, objects in the scene, and robot task identity. Substitution tests then support a causal role for these motifs, as overwriting them changes the corresponding features in the output, and swapping two robot task motifs redirects the policy toward the other task. Recorded during fusion, the motifs give a per-input account of which domain features the adapter's write carried, and that account is associated with how the model then behaves.

Primary area:
interpretability and explainable AI
Keywords:
interpretability, generation, language model

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_NbsdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes AFTERIMAGE, a small adapter inserted into a frozen multimodal model that passes its contribution through discrete codes from a shared codebook, so the information it adds can be read and edited. After training, the authors find code patterns linked to features like age, diagnosis, or robot task, and show that swapping these patterns changes the output. They validate it on brain MRI generation, robot manipulation (LIBERO), and video QA (CLEVRER), reporting equal or better performance than baselines with under 1% extra parameters.

Critical strengths

The main strength is the strong performance under distribution shift with very few added parameters nearly 0.2-0.71%. On LIBERO-Pro, AFTERIMAGE beats LoRA and the continuous variant on almost every shifted suite for both robot backbones. This matters because it suggests the adapter is useful even beyond its interpretability goal. The second strength is that the paper tests its codes with interventions, not just correlations. The same-scene task swap redirects the robot to the other task, while a neutral replacement never does, which gives some idea that the codes carry information the downstream model uses.

Critical weaknesses

Major :

  1. Assume I am doing lora fine-tuning of a big model and then I want to know what LoRA changed to improve accuracy. But for that why I need to trained the authors proposed interpretability again which used codebook kind of things. Because in my case I may not want codebook kind of things. Please discuss this limitation.

  2. The main contribution is not clear. Is the paper proposing a better adaptation method or an interpretability method? Both are claimed, but neither is fully developed. The scope is also unclear, is it meant for image/video generation, for VLMs/LLMs, or only for "fusion" pipelines?

  3. If this is an adapter paper, the baselines are not enough. LoRA is missing for MRI and no other PEFT methods are compared. With residual routing, does adding the adapter hurt the backbone’s general abilities, like - baseline model accurecy on medical, trained the model on robotics tasks and then check how much accurecy dropped on the medical tasks.

  4. In Table 2, the OOD gains are very large (e.g., SmolVLA Language-Spatial 43.8 -> 93.8), even though the model never saw these shifts in training and the IID gains are small. I would expect an adapter to mainly help IID. Why does it help OOD this much, especially with the action expert frozen? Please give an explanation or analysis.

  5. If the authors want to position their paper in interpretability domain, then they should plot some visualizations. In Figure 12, I do not clearly see AFTERIMAGE being better than LLM-LDM; in some columns LLM-LDM looks similar or better.

  6. The paper argues that probes and SAEs are weaker, but never compares against them on the same labels and data. Why trigrams (Sec. 2.4)? The paper motivates ordered sequences as carrying composition (Sec. 2.2), but never compares trigrams with single codes or bigrams. If single codes separate the groups equally well, the ordering design (causal mask, consecutive-code motifs) is not needed. Please justify this choice with a simple comparison.


Minor :

  1. what authors mean by 'we extract code motifs ...' in abstract, please present the paper properly what they want to say, else it hard to understand the paper. Things are introduce before define those.
Questions for authors

see the weakness part of the paper

8 comments
AuthorsOpenReview ↗

Response to Reviewer Nbsd (1/3)

We thank the reviewer for the thoughtful review.

  1. Assume I am doing lora fine-tuning of a big model and then I want to know what LoRA changed to improve accuracy. But for that why I need to trained the authors proposed interpretability again which used codebook kind of things. Because in my case I may not want codebook kind of things.

We would like to clarify the role of the codebook and how Afterimage applies to the reviewer's scenario.

The codebook is not the target. It is the material through which the domain features are read. Afterimage finetunes a pretrained model to a domain, as LoRA does, with the added ability to show which domain features take part in each prediction. The codebook turns the model's contribution into discrete units that can be named, counted, and followed across depths, which LoRA leaves implicit, and this transparency is what lets an expert trust a prediction in high-stakes domains such as medical imaging and robot control (Section 1, Appendix B).

Afterimage can replace LoRA for interpretable domain finetuning, or be trained on top of an existing LoRA to read part of the adapted computation.

  • As a replacement, it reaches the same or better performance (Tables 1 to 3) while recording the domain features involved in the fusion by design.
  • On top of an existing LoRA, it reads the information passing through the discrete channel and does not by itself explain how the LoRA changed the model.
  • Treating the LoRA-merged model as the frozen backbone and training Afterimage right after the LoRA layers is a direction we see as future work.

  1. The scope is unclear, is it meant for image/video generation, for VLMs/LLMs, or only for "fusion" pipelines? LoRA is missing for MRI and no other PEFT methods are compared. Does adding the adapter hurt the backbone’s general abilities.

We thank the reviewer for these three points and address them in turn.

First, on the scope. Afterimage targets any pipeline in which a frozen pretrained backbone fuses heterogeneous inputs into a representation that a downstream model consumes (Section 2.1). Image generation, VLA policies, and VLM question answering are three instances of this setting, and the method is not limited to them. Afterimage is an interpretability method, and performance (H1) is a constraint. The contributions lie in the causal relation between motifs and features in H2 and H3 (Sections 3.1 and 3.2, Figure 4, Table 12), while Afterimage still matches or improves performance in all three domains (Section 3).

Second, on LoRA for MRI and other PEFT methods. We have added a LoRA arm on brain MRI (Table 1, updated) and a bottleneck adapter arm on CLEVRER, and Afterimage leads both. LoRA is the appropriate reference, since it is the standard choice for adapting a frozen backbone, leaves the pretrained weights untouched as Afterimage does, and is matched in trainable parameters, so the comparison tests how the added capacity is used.

  • On brain MRI, Afterimage leads LoRA on every column of Table 1, and LoRA sits between the continuous variant and the frozen-LLM baseline, consistent with the ordering on CLEVRER and both VLA backbones.
  • On CLEVRER, a Pfeiffer-style bottleneck adapter (width 100, after every feed-forward block, same parameter budget) trails both LoRA and Afterimage on both backbones (table below).
ModelQwen2.5-VL-7BQwen3.5-9B
Frozen backbone71.981.5
+ PEFT Bottleneck adapter (0.24%)82.5 ± 0.488.3 ± 0.7
+ LoRA (0.24%)83.6 ± 0.394.3 ± 0.3
+ Afterimage (0.25%)87.9 ± 0.297.2 ± 0.3

Third, on whether the adapter harms the backbone's general capability. We would like to emphasize that Afterimage is designed for domain-specific adaptation, where it improves over the frozen backbone for a specific task, so some loss on off-domain benchmarks is expected, as with any single-domain finetuning. To quantify it, we evaluated the frozen backbone and both adapted CLEVRER arms on MMLU-Pro, a ten-option multiple-choice benchmark spanning 14 disciplines (1,200 questions, identical prompt, decoding, and training step).

ModelMMLU-Pro acc. (%)Δ vs frozen (pp)Right → wrongWrong → right
Frozen Qwen2.5-VL-7B33.4———
Frozen Qwen2.5-VL-7B + LoRA (finetuned on CLEVRER)25.3−8.1146 (12.2%)49 (4.1%)
Frozen Qwen2.5-VL-7B + Afterimage (finetuned on CLEVRER)31.7−1.851 (4.3%)30 (2.5%)

Afterimage stays within two points of the frozen backbone and changes fewer answers in either direction than the parameter-matched LoRA, which loses four times as much, while reaching higher in-domain accuracy on CLEVRER. With the write disabled, the model is the frozen backbone.

We hope these clarifications and additional results address the reviewer's concerns.

AuthorsOpenReview ↗

Response to Reviewer Nbsd (2/3)

  1. In Table 2, the OOD gains are very large. Why does it help OOD this much, especially with the action expert frozen?

On why the OOD gains are large. The robustness comes from decomposing the fusion into reusable units of domain knowledge, the motifs. Each motif records one feature, such as the task, a scene object, or an action phase, and the shared codebook reuses the same motifs across inputs and depths (Sections 2.2 and 2.4). Under a shift, the surface form of the instruction or layout changes while the features defining the task remain, so the policy receives the write it was trained on. Figure 5 shows the motifs selected on training rollouts stay selective on paraphrased episodes, and Appendix K.4 shows that episodes carrying the instructed task motif succeed far more often than those without it (96.7% against 70.3% on π0.5). With the pretrained backbone plus Afterimage, the burden of multimodal understanding falls on Afterimage, and the frozen action expert keeps the lighter job of mapping a recognized state to actions. Under shift, the task, objects, and phase are still recognized in the codes (Figure 5), so the expert receives a state it already knows how to act on.


  1. They should plot some visualizations. In Figure 12 (It's Figure 13 now), I do not clearly see AFTERIMAGE being better than LLM-LDM; in some columns LLM-LDM looks similar or better.

We thank the reviewer for these comments.

First, on the differences between methods in Figure 12 (Figure 13 in the revised version). Any single column is a plausible scan for every method, and the comparison lies in the progression across columns, where the real row shows peri-hippocampal and ventricular fluid increasing steadily from CN 20 to AD 90. Four points separate the methods.

  1. The age progression. The green regions in the Afterimage maps grow along the row in step with the real row, whereas LLM-LDM shows a similar amount of green in every column and MLP-LDM changes little between CN 60 and AD 90.
  2. The young condition. At CN 20, Afterimage stays close to the real reference, whereas LLM-LDM already shows the fluid pattern of an older scan and MLP-LDM shows a dense red map far from the reference.
  3. The ventricles. Afterimage ventricles enlarge from CN 20 to AD 90 as in the real scans, whereas the LLM-LDM ventricles stay at a similar size across conditions.
  4. The counterfactual edit (boxed CN 80 → AD). Afterimage adds the AD-related fluid to the real CN 80 scan while keeping its anatomy, whereas the LLM-LDM edit differs little from its own CN 80 sample.

The same pattern is measured in Table 1 as age MAE and sex accuracy and in Figure 14 as regional volumes. We will revise the caption to direct the reader to the progression across columns.

Second, the MRI experiment already includes interpretability analyses under H2 and H3 (Section 3.1, Appendix H). Figure 3 and the permutation null show that the motifs selected for diagnosis, age, sex, and regional volume stay selective on withheld inputs beyond what label shuffling produces, and Figure 4 shows that overwriting or removing a motif moves the generated volume in the direction the motif encodes.

Third, the robot experiment visualizes which domain features the fusion carries, and in what order, over a single episode (Section 3.2). Figure 6 tracks keyframe by keyframe which task, object, and phase motifs have fired as the rollout advances, which makes the task and objects behind each predicted action readable and localizes where the behavior departs from the task (SmolVLA in Figure 22). Figure 5 shows these motifs stay selective on withheld and paraphrased episodes.

AuthorsOpenReview ↗

Response to Reviewer Nbsd (3/3)

  1. Why trigrams (Sec. 2.4)? The paper never compares trigrams with single codes or bigrams. If single codes separate the groups equally well, the ordering design (causal mask, consecutive-code motifs) is not needed. Please justify this choice with a simple comparison.

First, we have added a motif-length comparison (Figure 21, new). On π0.5 tasks and action phases, we repeat the motif selection with 1-grams to 4-grams and evaluate on train, held-out, and Language-shift episodes. Single codes and bigrams fire almost everywhere and do not isolate a task or a phase, trigrams concentrate on the diagonal and mark each feature selectively, and 4-grams keep the selectivity but thin the presence, especially on action phases. Trigrams are therefore the shortest run that separates the groups.

Second, the paper already tests the ordering design directly (Section 3.4). Permuting the codes at inference is about as damaging as replacing them at random on brain MRI and robots (FID 3.22 against 3.02 on brain MRI, 22.1% against 24.8% Language success on π0.5), so the write depends on which code occupies which slot. Retraining without the causal mask lowers success on both policies, most under distribution shift (80.3% against 89.2% on LIBERO-Goal under the Language shift for π0.5; Table 11, Appendix L), and its motifs are less selective on held-out and paraphrased episodes (Figure 20). Order therefore carries information the policy uses, which single-code motifs would discard.


  1. The paper argues that probes and SAEs are weaker, but never compares against them on the same labels and data.

We have added Figure 29, which compares a linear probe and a sparse autoencoder (SAE) with the Afterimage record on the same hidden states, task labels, and rollouts, under the LIBERO-Spatial Task shift of π0.5. Both readings use the hidden states entering VLM layer 16, the layer Afterimage reads, and the same 40k demonstration frames. The probe is a 40-way linear head, and the SAE is labelled by assigning each task its most selective feature with the protocol used for the motifs (Appendix H) and decoding each chunk by the most active of these features. On the same Task-shift rollouts, the probe predicts the instructed task on 55% of chunks and the SAE decoder on 30%, both below the motif-task alignment of Figure 18 on the same shift. Both labels are also decoupled from the outcome. The probe keeps the instructed task through a failed rollout and flips once in a successful one, and the SAE label moves among several tasks within a single episode whether or not it succeeds. The Afterimage record marks the task objects and the action phases reached so far, and the failed episode is readable at the frame where the drop replaces the hold after the carry.


Things are introduce before define those. (in abstract)

We have updated the abstract to be more clear.

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

This paper introduces AFTERIMAGE, a discrete adapter that makes part of the fusion process in frozen multimodal models inspectable. It compresses intermediate hidden states into discrete code sequences, from which domain-related motifs can be identified and manipulated. The authors evaluate the method on brain MRI synthesis, robotic manipulation, and video question answering, showing improvements in task performance, associations between code motifs and semantic features, and behavioral changes following motif interventions.

Critical strengths

There are four strengths after carefully reading this paper:

  • An interesting approach to interpretability. Rather than interpreting hidden representations entirely after training, AFTERIMAGE places a discrete code channel directly in the forward computation. This provides a more concrete basis for inspecting and intervening on the adapter's contribution to the model.

  • Strong empirical results across different applications. The improvements are substantial, particularly under distribution shifts in LIBERO-Pro. Evaluating both SmolVLA and π0.5 makes the robotic results more convincing, and the additional MRI and video QA experiments suggest that the approach is not limited to one architecture or task.

  • The intervention experiments are a meaningful addition. The authors go beyond showing correlations between codes and semantic features. The MRI motif substitutions and the robot task-swapping experiments provide useful evidence that some learned motifs have a causal influence on downstream behavior. The held-out signature evaluations also strengthen this part of the paper.

  • Reasonably thorough experimental design. The paper includes continuous-bottleneck and LoRA comparisons, several ablations, multiple random seeds, and fairly detailed implementation information. I particularly appreciate the attempts to separate improvements from simply adding more trainable parameters.

Critical weaknesses

The main concerns during the whole review processing are:

  • The interpretability claim is somewhat stronger than what is actually demonstrated. AFTERIMAGE records the adapter's contribution, not the complete information used by the model. Under residual routing, the original hidden states can bypass the discrete channel, and the robotic action expert can still access representations below the insertion layers. Therefore, while the codes influence predictions, it remains unclear how reliably they can explain the model's overall decisions or identify actual fusion failures.

  • The semantic meaning of the discovered motifs is not fully established. The motifs are assigned meanings through post-hoc associations with predefined features, and several of those features are strongly correlated. For example, robot objects are closely tied to task identities in LIBERO, while MRI morphology is associated with age and diagnosis, which are explicitly supplied in the input. The held-out results demonstrate stable associations, but not necessarily that the codes represent independent, reusable semantic concepts. The task-swapping results are interesting, although most swapped episodes still fail to complete either task.

  • The methodological contribution relative to existing discrete bottlenecks needs clearer isolation. Prior work such as [1] has already demonstrated on-path discrete representations and behavioral interventions, while [2] is relevant to cross-layer discrete concepts. AFTERIMAGE has meaningful architectural differences, but the experiments do not fully establish which differences matter. In particular, the continuous baseline also removes the commitment and coverage losses, and there is no direct ablation of the shared codebook against depth-specific codebooks.

  • Some important evaluation limitations weaken the broader claims. The MRI FID is computed using features from a diagnosis classifier built on the same autoencoder as the generator, rather than an independent measure of anatomical realism. Moreover, the diagnostic groups are not equally represented across source studies, leaving possible dataset or scanner confounding. The breadth of multimodal evaluation is also somewhat less than it first appears: MRI conditioning uses a single textual description of clinical attributes, and CLEVRER evaluation is restricted to descriptive questions rather than its more demanding reasoning categories. These limitations make the broader claims about interpretable multimodal fusion less conclusive.

Refs:

  1. Codebook Features: Sparse and Discrete Interpretability for Neural Networks
  2. Cross-Layer Discrete Concept Discovery for Interpreting Language Models
Questions for authors

N/A

4 comments
AuthorsOpenReview ↗

Response to Reviewer Qrbt (1/3)

We thank the reviewer for the thoughtful reviews.

  1. AFTERIMAGE records the adapter's contribution, not the complete information used by the model. Therefore, while the codes influence predictions, it remains unclear how reliably they can explain the model's overall decisions or identify actual fusion failures.

We thank the reviewer for this careful reading and agree on the observations. We would like to clarify and emphasize on some analyses.

Afterimage records the contribution of the trained component, and the claims are scoped to that contribution (Section 2.4, Appendix A). The question the paper tests is whether the part a practitioner trains carries the domain features, and the interventions address exactly this part. Corrupting the codes with the rest of the stream intact degrades performance in all three domains (Section 3.4), and substituting one motif moves the corresponding output feature or redirects the policy toward the other task (Figure 4, Table 12). The bypassing stream and the layers below the insertions stay available in every intervention, so the changes trace to the record.

The record separates a fusion reading from the outcome in the robot experiments. Episodes whose record lacks the instructed task's motif fail at a considerably higher rate on both backbones (success 96.7% with the motif against 70.3% without on π0.5, and 93.2% against 52.5% on SmolVLA), and Figure 28 shows an episode whose record names a task other than the instructed one (Appendix K.3). The record is written on the forward path at the depths the policy reads, so it reports the task the fusion settled on whether or not the episode succeeds, which suggests whether an error arose in fusion or in execution.

We hope this clarifies reviewer's concerns.


  1. The motifs are assigned meanings through post-hoc associations with predefined features, and several of those features are strongly correlated. For example, robot objects are closely tied to task identities in LIBERO, while MRI morphology is associated with age and diagnosis, which are explicitly supplied in the input. The held-out results demonstrate stable associations, but not necessarily that the codes represent independent, reusable semantic concepts.

We thank the reviewer for raising this point and would like to clarify it further.

Regarding correlated features, the motifs are also evaluated on features absent from the input and disentangled from the attributes they correlate with. On brain MRI, the regional volume motifs are associated with morphometry measured on the generated scans, and the substitution grid in Figure 4 (right) overwrites the age and volume motifs in every combination, showing that each contributes separately. On robots, the swap in Table 12 uses two tasks sharing one scene and its objects, so the objects are held fixed while the task motif alone changes the behavior, and the object and phase motifs are evaluated apart from the task motifs (Figure 5, Appendix K). The label-permutation null further confirms that the selectivity exceeds what the selector recovers from shuffled labels (Appendix H).

Regarding the swap result, the informative comparison is against the feature-neutral control, and under that comparison the motif carries a task identity the policy acts upon. Under the task-1 instruction, the instruction and visual tokens below the insertions continue to describe task 1, so the swapped record conflicts with the rest of the stream and completing neither task is the expected outcome. Within this setting, the swap completes task 2 in a sizable fraction of episodes (25.8% in distribution and 21.7% under the Language shift), whereas the feature-neutral replacement completes it in none (Table 12).

AuthorsOpenReview ↗

Response to Reviewer Qrbt (2/3)

  1. The methodological contribution relative to existing discrete bottlenecks needs clearer isolation. Prior work such as [1] has already demonstrated on-path discrete representations and behavioral interventions, while [2] is relevant to cross-layer discrete concepts. AFTERIMAGE has meaningful architectural differences, but the experiments do not fully establish which differences matter. In particular, the continuous baseline also removes the commitment and coverage losses, and there is no direct ablation of the shared codebook against depth-specific codebooks.

We thank the reviewer for this precise framing and agree that the differences from [1] and [2] deserve direct experimental evidence.

Afterimage differs from [1] in keeping the backbone frozen, in sharing one codebook across depths, and in writing an ordered sequence of a few slots per input, and it differs from [2] in learning the codebook under the downstream task so that the record lies on the forward path of the deployed model. Of these choices, the ordering, the multi-depth insertion, and the frozen backbone are already tested. Permuting the codes at inference is about as damaging as replacing them on brain MRI and robots (FID 3.22 against 3.02 on brain MRI, 22.1% against 24.8% Language success on π0.5), and the Non-causal arm trails on both backbones, most clearly under distribution shift (80.3% against 89.2% on LIBERO-Goal under the Language shift for π0.5; Section 3.4, Table 11). The single-depth arm tests the multi-depth insertion (Appendix L), and LoRA, which adapts the backbone, trails Afterimage under distribution shift (Tables 2 and 11). The codebook and the terms that train it are the subject of the two ablations below.

The paper scopes the discrete-versus-continuous claim to the quantized bottleneck together with the terms maintaining it, and we are running arms that separate the two. Both terms are defined on codebook entries and have no form in a continuous arm, which is why Cont. removes them jointly and why the Cont. + noise arm was added to match the bottleneck strength without its discrete structure (Table 11). To separate the terms from the quantizer, we are training arms that keep the quantizer and remove the terms, starting with the π0.5 Language shift and brain MRI synthesis.

The codebook is shared so that a code denotes the same entry at every depth and a motif stays comparable across depths and inputs, and we are running the depth-specific comparison the reviewer suggests. This choice follows principle P3 in Section 2.2, and the paper makes no performance claim for it. We are training depth-specific codebooks at matched KK and VV and will report task performance together with motif selectivity on held-out episodes. Both ablations are running now, and we will post the results in this thread as soon as they are available.

AuthorsOpenReview ↗

Response to Reviewer Qrbt (3/3)

  1. The MRI FID is computed using features from a diagnosis classifier built on the same autoencoder as the generator. MRI conditioning uses a single textual description of clinical attributes, and CLEVRER evaluation is restricted to descriptive questions rather than its more demanding reasoning categories.

Regarding the FID evaluator, the shared autoencoder applies equally to every arm in Table 1, and we have added an anatomical Fréchet distance independent of the generator, which reproduces the same ranking. The autoencoder is frozen and shared by every arm, and the evaluator is trained separately on real scans, so the latent FID is relative within Table 1 (Appendix F). The new measure segments every generated and real scan with the pretrained FastSurfer pipeline, takes nineteen regional volumes spanning medial temporal and subcortical structures, ventricles, gray matter, and white matter, and computes the Fréchet distance within the same strata as the latent FID, against a real-versus-real split-half floor at matched sample size. It reads the decoded scan and shares nothing with the generator.

Regarding dataset and scanner confounding, the motif–diagnosis associations are free of site information by construction, and the fidelity ranking holds under a single-study reference. The backbone sees only the prompt and the scan enters at the decoder, so study and scanner reach the motifs only through the generator and affect fidelity alone. Every Impaired scan comes from ADNI while the CN reference pools six studies (Appendix E, Table 9), which is the imbalance the reviewer points to. We therefore re-score the same generated scans against a reference restricted to ADNI, within the same strata and the 70 to 99 band, with both measures.

ModelLatent FID, pooled (CN / Imp.)Latent FID, ADNI-only (CN / AD)Anatomical FD, pooled (CN / Imp.)Anatomical FD, ADNI-only (CN / AD)
Real vs. real0.003 / 0.0010.19 ± 0.06 / 0.48 ± 0.1211.8 / 14.116.2 ± 0.5 / 13.6 ± 1.7
LLM-LDM1.21 ± 0.08 / 0.80 ± 0.154.61 ± 0.07 / 5.71 ± 0.4723.5 ± 0.5 / 39.1 ± 1.032.2 ± 2.5 / 53.8 ± 5.5
LoRA1.09 ± 0.09 / 1.39 ± 0.163.12 ± 0.05 / 4.08 ± 0.3222.4 ± 0.7 / 37.8 ± 0.828.8 ± 0.9 / 41.9 ± 1.4
Afterimage0.65 ± 0.05 / 0.58 ± 0.042.40 ± 0.02 / 3.79 ± 0.6016.3 ± 0.1 / 28.1 ± 0.324.2 ± 1.2 / 35.6 ± 5.3

Afterimage leads on both cohorts under both measures and both references and sits closest to the floor in every column, and the ordering of the arms is unchanged, so the pooled CN reference did not favour any arm. All distances rise under the single-study reference, since the reference draw is smaller and every generator was trained on the pooled corpus, and the rise applies to every arm alike. We will add the anatomical measure and the ADNI-only columns alongside Table 1 in the revision.

On brain MRI, the backbone fuses three clinical attributes rendered as text, and fusion across modalities is tested on robots and CLEVRER. The attributes are rendered through 25 template and phrasing combinations, and motifs are scored on withheld combinations, so a motif is tied to the attribute and not to a surface form (Appendix G). The volume motifs are associated with morphometry measured on the generated volumes, a feature the prompt does not state (Figures 12 and 14).

We evaluate the descriptive split for its open-ended answers, each of which names a feature the motif analysis can be checked against (Figure 7). The multiple-choice splits score a per-option judgment that depends on the option text as well as the video, which makes the answer harder to attribute to the fusion. The final revision will report the explanatory, predictive, and counterfactual splits with the same adapter.

We hope these clarifications and additional results address the reviewer's concerns.

AuthorsOpenReview ↗

Continued Response to Reviewer Qrbt (2/3)

We thank the reviewer again for the precise framing of point 3. The two ablations promised above have finished on CLEVRER, and we report them here. Each arm changes one thing relative to the trained Afterimage arm and keeps the data, backbone, insertion depths, KK, VV, adapter width, routing, training protocol, and seeds fixed. The frozen and Afterimage rows are the existing runs of Table 3.

In particular, the continuous baseline also removes the commitment and coverage losses, and there is no direct ablation of the shared codebook against depth-specific codebooks.

ModelAcc. (%) ↑
Frozen Qwen2.5-VL-7B71.9
+ Afterimage, no terms85.7 ± 0.5
+ Afterimage, depth-specific codebooks86.2 ± 0.3
+ Afterimage (shared codebook)87.9 ± 0.2

Ablation A. Keeping the quantizer and removing the commitment and coverage terms separates the discrete bottleneck from the terms that maintain it.

  • The arm reaches 85.7%, against 87.9% for Afterimage.
  • With the quantizer kept, the two terms account for 2.2 points, and the discrete bottleneck carries the rest of the gain over the frozen backbone. The bottleneck is therefore the main source of the gain, and the terms that maintain it add a further, smaller gain, which is the reading the paper scopes the claim to (Section 3.4).

Ablation B. The shared codebook leads the depth-specific codebooks on the task, and it is what lets a motif be followed across depths.

  • The depth-specific arm reaches 86.2%, against 87.9% for the shared codebook. Sharing one codebook across depths therefore keeps accuracy intact, and the paper makes no performance claim for it beyond this.
  • With a shared codebook, a code denotes the same entry at every depth, and a motif found at one depth can be matched at another (Figure 7). With depth-specific codebooks, the same index names unrelated entries at different depths, and the cross-depth reading is lost by construction. This readability is what the shared codebook buys (principle P3, Section 2.2).

We hope these results help isolate which design choices matter, and we will add both ablations to Section 3.4 in the revision.

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

This paper introduces AFTERIMAGE, a discrete adapter that makes it possible to explain which domain features are carried by what the adapter contributes to the residual stream. AFTERIMAGE targets settings where an MLLM serves as the multimodal encoder for a downstream model with a task-specific adapter attached. An adapter with weights and a codebook shared across depths is inserted between several layers of a frozen LLM, and the adapter’s write is produced from, and recorded as, a sequence of discrete codes only. After training, motifs associated with domain features are identified statistically from these code sequences, and substituting one motif with another is shown to shift the output in the corresponding direction. On the downstream tasks of brain MRI synthesis, robot action planning, and video question answering, AFTERIMAGE shows similar or higher performance than baselines such as LoRA, and the gap is especially large under distribution shift.

Critical strengths
  1. The problem is clearly defined. AFTERIMAGE records the adapter’s contribution to the fusion on the deployed forward path through intuitive design principles, unlike existing approaches, which either specialize the model without exposing which knowledge is used or describe the fusion from outside the model’s computation.
  2. Inspectability comes without a performance decrease. Across different downstream tasks, AFTERIMAGE matches or improves task performance compared to the frozen backbone, and it matches or improves on existing adapted baselines such as LoRA, with the margins widening under distribution shift.
  3. Motif analysis supports inspectability. Code motifs are associated with domain features on held-out inputs, and motif substitution supplies intervention-based evidence that they influence the corresponding features in the output.
Critical weaknesses
  1. The persuasiveness of the explanation capability is lacking. For example, the MRI input is a template sentence in which age, sex, and diagnosis are directly filled in. In this case, the association between age and diagnosis and the code motifs may be a trivial result, since it reflects attributes explicitly stated in the input being reflected in the codes.
  2. The problem identified in the Introduction was that, when the desired result is not obtained, existing methods cannot tell whether the trained connecting component (adapter) dropped the information or whether the downstream model (generator) ignored it. However, no formalized method is presented for actually using AFTERIMAGE to analyze, in cases where generation is wrong, which of the two is the problem.
  3. Control groups are lacking in the substitution experiments. The authors conducted experiments replacing the entire codes and experiments replacing the signature, but did not conduct experiments replacing parts other than the signature. The absence of this control group fails to clearly reveal the importance of the signature.
  4. Comparison with interpretation baselines is lacking. There is insufficient quantitative/qualitative analysis of what advantage AFTERIMAGE has in interpreting results over the interpretation methods presented in the Introduction, such as linear probing and SAE.
Questions for authors
  1. To support the explanation capability of AFTERIMAGE, is statistical analysis possible not only on MRI, whose input format is fixed, but also on modalities that are not? (Example: motif analysis of specific objects appearing in video QA such as CLEVRER)
  2. To actually use AFTERIMAGE, is analysis possible on real generation failure cases? Furthermore, can you present guidelines on how such analysis could be performed?
  3. In robot planning, is an experiment replacing trigrams other than the signature possible?
  4. Is it possible to analyze what advantage AFTERIMAGE has over the interpretation methods presented in the Introduction? Alternatively, an example in which AFTERIMAGE reveals a problem that is not exposed by existing interpretation methods would also be sufficient.
3 comments
AuthorsOpenReview ↗

Response to Reviewer sdNx (1/2)

We thank the reviewer for the thoughtful reviews.

  1. The MRI input is a template sentence in which age, sex, and diagnosis. ** The association between age and diagnosis and the code motifs may be a trivial result**.

We would like to clarify that in every domain, the analysis recovers both explicit and implicit features.

  • On brain MRI, the regional volume motifs (hippocampus, ventricles) are associated with morphometry measured on the generated scans, which didn't appear in the input. Substituting these motifs moves the generated volume toward the substituted level, confirming that the generator acts on them (Figure 4).
  • On CLEVRER, we have added Figure 7, where motifs are associated with the question type, the event operators, and the answer token (color, shape, material, count), and the associations reproduce on held-out videos (Section 3.3). The answer-token motifs address the reviewer's concern directly. The answer is absent from the input and is formed by grounding the referenced object and event in the frames, which makes these motifs a record of a feature derived from the video rather than copied from the question.
  • On robots, the object and phase motifs reveal which objects the policy attends to and which action phase it is in at each keyframe (Section 3.2, Figures 5 and 6).

The codes therefore capture the domain knowledge the model forms during fusion, including features the input does not state.

We hope this clarifies reviewer's concerns.


  1. Control groups are lacking. The authors conducted experiments replacing the entire codes and experiments replacing the signature, but did not conduct experiments replacing parts other than the signature.

We added a brain MRI control (Table 10, new row). The control overwrites a non-motif trigram with a feature-neutral replacement.

ConditionAge MAE (y) ↓Sex acc. (%) ↑
Afterimage (intact CN Syn)11.8 ± 0.776.2 ± 0.8
Feature-neutral motif swap23.4 ± 4.765.2 ± 3.8
Non-motif swap (control, new)12.1 ± 0.875.9 ± 0.7

Editing non-motif codes stays within the intact model's seed noise, whereas editing the motif moves both metrics substantially. The effect in Figure 4 and Table 10 is therefore due to the motif itself, rather than to a code edit of this size.

We hope this clarifies reviewer's concerns.

AuthorsOpenReview ↗

Response to Reviewer sdNx (2/2)

  1. Is statistical analysis possible not only on MRI - fixed template? Example: motif analysis of specific objects appearing in video QA such as CLEVRER.

We would like to clarify that motif analysis on domains without a fixed input format is already in the paper.

  • On robots, episodes vary in length and scene, and the task, object, and phase motifs remain selective on withheld and paraphrased episodes (Section 3.2, Figure 5).
  • On CLEVRER, videos vary in the number and mix of objects, and questions vary in form and target. We have added Figure 7, which reports motifs for three groups: question subtype, scene event operators (collision, partner, moving, stationary), and answer token. The motif selected for a category fires predominantly on that category and remains selective on held-out videos. The scene-event motifs are the component that could help to clarify the reviewer's concern, since whether an object collides or moves is a property of the video rather than of the question, which indicates that the codes record what the model observed in the scene (Section 3.3).

  1. Is analysis possible on real generation failure cases? An example in which AFTERIMAGE reveals a problem that is not exposed by existing interpretation methods.

We would like to clarify that the analysis of real failure cases is already in the submission.

First, Section 3.2 (Figure 6) and Appendix K.4 (Figures 22 and 28) analyze real failed episodes, where the code record localizes the step at which an episode fails and suggests whether the error arose in fusion or in execution. Figure 6 shows a successful and a failed episode of the same task, and the point where the motifs stop advancing localizes the error to the closing step. Reading a failure therefore amounts to checking whether the instructed task motif fires, following the trace to where the motifs stop, and checking which task motif fired instead.

Second, we have added Figure 29, which places a linear task-ID probe and a sparse autoencoder (SAE) on the π0.5 baseline beside the Afterimage record on the same instruction under the LIBERO-Spatial Task shift, as an example where Afterimage exposes a problem that existing readings leave hidden. Both readings use the hidden states entering VLM layer 16, the layer Afterimage reads, and the same 40k demonstration frames. The probe is a 40-way linear head, and the SAE is labelled by assigning each task its most selective feature with the protocol used for the motifs (Appendix H) and decoding each chunk by the most active of these features. On the same Task-shift rollouts, the probe predicts the instructed task on 55% of chunks and the SAE decoder on 30%, both below the motif-task alignment of Figure 18 on the same shift. The probe and SAE rows show the same three baseline episodes, and both labels are decoupled from the outcome. The probe keeps the instructed task through a failed rollout and flips once in a successful one, and the SAE label moves among several tasks within a single episode whether or not it succeeds. The Afterimage record marks the task objects and the action phases reached so far, and the failed episode is readable at the frame where the drop replaces the hold after the carry.

We hope this clarifies reviewer's concerns.

Thank you for the kind and detailed response. I now think that the motifs do not merely copy the input, but are patterns actually learned while the model is fine-tuned, and I can have confidence that changing a key motif can steer the downstream generation in an entirely different direction. On the other hand, I believe I have already given a score that sufficiently reflects the contribution of the paper, so I will keep my score.

Reviewer_kkPqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AFTERIMAGE, an insertable module into multimodal backbone to enable the verbal explanation of multimodal fusion process. Afterimage module is inserted into middle of pretained backbone layers. Given output tokens of certain pretrained layer, Afterimage discrete bottleneck first gather specific number of token positions into several slots. The gathered slots are then feeded into a Quantizer, which utilzes discrete codebook to convert those slots into codebook ids. Those quantized slots are then processed by another slot-attention to obtain corresponding tokens. During the interpretation process, those codebook ids are grouped by features and visualized as graph. Node of graph is asscoiated with verbal expression, contributing to final language description within different domains. This work conduct evaluation across several domains, including brain MRI synthesis, robot action planning, and video question answering, and observe clear improvement over baselines.

Critical strengths
  1. Comprehensive experiment evaluation. This work conduct evaluation of its proposed method across several different domains, and observe clear performance improvement over various baseline methods. Moreover, the ablation and analysis is also quite thorough, providing adequate findings and observations.
  2. This paper is clearly written and easy to follow. The writing of this paper is good, it is convenient to follow the main logic flow and understand the described content.
Critical weaknesses
  1. Doubt of the importance of the paper topic. The main experiment focus of this paper is still improving the model performance, across all 3 evaluated domains. However, if the main goal is to optimize model performance, the most straigtforward way is to improve model architecture, refine and collect training data with higher quality, or devising more effective training algorithm. Studying verbal explanation of multimodal fusion and leverage it to enhance model performance is not a proper manner. Overall, importance of the studied topic is not quite high.
  2. Utilizing codebook id for language explanation is not reaonable. Codebook id is discrete. In contrast, the state to express is quite diverse and uncountable. Taking the robot domain as example, the actions which robot can perform is quite many. Adopting codebook to refer to those action spaces is not appropriate and could limit the final performance.
  3. Model leveraged is out-date for some domain. For the video question answering, the used base model is Qwen-2.5-VL-7B. This model is quite out-of-date. The author is recommended to change to more recent models.
Questions for authors

See Weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer kkPq (1/2)

We thank the reviewer for the thoughtful reviews.

  1. Doubt of the importance of the paper topic. ...if the main goal is to optimize model performance, the most straigtforward way is to improve model architecture. Studying verbal explanation of multimodal fusion and leverage it to enhance model performance is not a proper manner.

We would like to clarify two aspects of the goal of Afterimage.

First, optimizing task performance is not our main goal. Our aim is to make the fusion computation more transparent, so that the domain features it carries can be read in a human-understandable way, which is essential for expert trust in high-stakes domains such as medicine and robotics.

  • As mentioned in Section 1, the fused representation is a dense hidden state whose involved domain features remain implicit. Adaptation methods such as LoRA improve performance but offer little material to interpret which features the model relied on, and prior interpretation methods observe the model from outside the computation (Section 1).
  • Afterimage makes the fusion transparent within the computation itself and associates its code motifs with named domain features at each depth (Section 2).
  • Task performance (H1) is therefore a constraint rather than the objective. The contributions lie causal relation between motifs and features in H2 and H3 (Sections 3.1 and 3.2, Figure 4, Table 12), while Afterimage still matches or improves performance in all three domains (Section 3).

Second, we would like to note that we do not study verbal explanations. We study a record of the domain features the fusion carries, written as discrete codes on the forward path and therefore faithful to the adapter’s write by construction (Proposition 1).

  • As mentioned in Section 1 and Appendix B, verbal explanations (text sentences) approaches are produced after the fusion, so they may omit the features that determined the output, and they rely on a language head that a diffusion decoder or a policy does not use.
  • The Afterimage record reflects the computation itself, and substituting a motif offers a direct test of whether the model acts on the recorded feature (Sections 3.1 and 3.2).

We hope this clarifies that the paper main contributions.


  1. the used base model is Qwen-2.5-VL-7B is quite out-of-date.

We have added the more recent Qwen3.5-9B backbone to the CLEVRER experiment, keeping the training protocol and settings. The results are consistent across the two backbones. Afterimage improves over the frozen backbone and over the LoRA arm in each case. (We also added PEFT method for comparison)

ModelQwen2.5-VL-7BQwen3.5-9B
Frozen backbone71.981.5
+ PEFT Bottleneck adapter (0.24%)82.5 ± 0.488.3 ± 0.7
+ LoRA (0.24%)83.6 ± 0.394.3 ± 0.3
+ Afterimage (0.25%)87.9 ± 0.297.2 ± 0.3
AuthorsOpenReview ↗

Response to Reviewer kkPq (2/2)

  1. Utilizing codebook id for language explanation is not reaonable. Codebook id is discrete. the state to express is quite diverse and uncountable. The actions which robot can perform is quite many. Adopting codebook to refer to those action spaces is not appropriate and could limit the final performance.

We would like to clarify the expressiveness of discrete codes.

First, the codes occupy a few depths of a continuous computation, so they bound what passes through those depths and do not limit what the model can express. Under residual routing, used on brain MRI and CLEVRER, the original hidden state passes through unchanged and the write adds to it. Under exclusive routing, used on robots, the write at an instrumented depth is decoded from the codes and feeds the next layer, which is what makes the record on-path, while every other layer and the action expert stay continuous and the expert attends to all of them (Section 2.4, Table 4). The continuous capacity of the backbone therefore remains in use in every domain, and the discrete record is the readable part of it.

Second, the codes do not enumerate the task's input or output space. They record the domain features the fusion carries and how these features combine across depths to form a prediction. Taking the robot domain as an example, the action expert still outputs continuous actions. The codes instead mark features such as task identity, phase, and the objects in the scene (Section 3.2, Figures 5 and 6), and the same holds for diagnosis and demographic features on brain MRI and for events on CLEVRER (Sections 3.1 and 3.3). Several of these features are implicit and are still recovered from the codes (Sections 3.1 and 3.2), which shows that the record captures domain knowledge rather than a quantization of the input or output.

Third, the number of code combinations is not small. With a codebook of VV entries and KK ordered slots per depth, each depth admits VKV^{K} ordered sequences, so the combinatorial space grows further with the number of instrumented layers (Appendix D.2, Table 4). The ablations show that both code identity and slot order carry information, since randomizing or permuting the codes collapses performance in all three domains (FID 0.65 to 3.02 and 3.22 on brain MRI, accuracy 87.9% to 9.7% and 41.0% on CLEVRER, π0.5 Language success 92.7% to 24.8% and 22.1%; Section 3.4, Figure 8).

Fourth, the results show that the discrete record does not limit performance, and that it improves robustness under distribution shift.

  • In distribution, Afterimage matches or improves on the continuous and LoRA arms in all three domains (Tables 1 to 3).
  • Under distribution shift, the full-capacity backbone alone degrades sharply (from 98.3% to 41.2% on LIBERO-Goal with π0.5\pi_{0.5} under paraphrased instructions), while Afterimage retains the highest performance (89.2%, Table 2).

We hope this clarifies reviewer's concerns.

Thanks for the author responses, I will maintain my initial score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.