acceptodds
Under review as a conference paper at ICLR 2027

Towards interpreting any residual stream direction by scaling activation inversion

16 views

Abstract

To interpret a direction in a language model's residual stream, we can look for text that strongly activates it. But when such text is found by scanning a corpus, it is limited to what that corpus contains. We develop *Maximally-Activating Example Metamodels* (MAEMs), language models which map residual stream directions from a subject model to text exemplars that strongly activate those directions. Our objective is to maximise the cosine similarity of a supplied residual-stream direction with the activations of the text generated by the MAEM. This label-free RL objective helps to scale MAEMs on diverse sources of residual stream directions. Empirically, we find the resultant metamodel produces informative exemplars: a Qwen3.6-27B MAEM inverts held-out activations better than corpus search or a natural-language autoencoder (NLA), beats corpus search (at 10M tokens) on a majority of held-out languages/domains, and activates 89% of held-out SAE features. It also identifies concepts behind steering vectors about as well as NLAs, and recovers the payloads of LoRA backdoors more often than corpus search. Unlike explanations produced by other metamodels (e.g. NLAs), MAEM exemplars can be verified with a forward pass on the subject model, making the MAEM objective a grounded alternative for scaling the training of metamodels.

Primary area:
interpretability and explainable AI
Keywords:
activation inversion, meta-model, residual stream, interpretability

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_NK8AOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes Maximally-Activating Example Metamodels (i.e., MAEMs). A MAEM is a language model trained to take a direction in another model's internal activations and write text that strongly activates that direction, a work that normally requires a quite long procedure when, for instance, trying to interpret SAE directions. The authors fine-tune a copy of the model of interest, first with supervised training and then with reinforcement learning. The reward is how closely the generated text's activations match the target direction, so it needs no human or LLM labels, and every output can be checked with a forward pass. Results are strong: a single trained MAEM beats corpus search on activating held-out directions and generalizes to unseen languages and code. It is then applied to four interpretability tasks: identifying the concept behind steering vectors, reading hidden backdoor payloads, reading model "workspace" content, and activating and explaining SAE features.

Critical strengths
  • Well-motivated idea that solves a concrete problem in interpretability: instead of searching a fixed corpus for text that activates a direction, the authors train a model to generate such text. This removes the dependence on what the corpus happens to contain and scales well, since one forward pass replaces scanning millions of tokens.
  • Strong and general results: a single checkpoint outperforms corpus search, GCG and EPO, and generalizes to unseen languages and code. I also appreciated that the authors present thorough results in the context of four different interpretability applications. The same model handles all four without being trained on most of these direction types.
Critical weaknesses

Not critical:

  • Generality and cost: a MAEM is initialized from the model's weights, trained on activations from one layer, and rewarded by that same model. All experiments use a single model and mainly a single layer. Does this mean a new MAEM has to be trained for every model, or even every layer? Training seems expensive. Could the authors report the overall training cost and comment on whether a MAEM might transfer to other layers or models? Not required for acceptance, but for future directions: it'd be great to run a small experiment about transfer to different architectures.
  • Role of RL: the main difference from prior activation inverters, IIUC, seems to be the RL stage with the activation-based reward. However, the authors mention that SFT and RL are not compared at matched compute. How much of the improvement comes from the RL objective itself, as opposed to scale or training data?
Questions for authors

/

2 comments

We thank the reviewer for taking the time to read our work and providing an engaging discussion.

W1. Generality across models and layers, and cost.

Transfer across layers. We run a cross-layer experiment where the layer-42 MAEM reads activations of the same subject model from 10 layers between 6 and 62. We find that

  • activation inversion transfers across the layers tested: on 512 held-out activations, fidelity is 0.70 to 0.80 at all layers, against 0.74 at the training layer
  • target recovery on two WorkspaceBench [1] task families matches the training layer from layer 36 and above, and falls at earlier layers (where J-lens also declines)

Table R1. Zero-shot transfer of a layer-42 MAEM to other layers. Fidelity (held-out): the paper's centred cosine between the target activation and the activation of the MAEM's text read at the target's own layer, mean over 512 held-out activations. Fidelity (control): same score against a different target activation. WorkspaceBench (WB): pass rate under the benchmark's protocol, judged by Claude Sonnet 5. In association (WB.a, n = 91, chance 0.125) the readout must name a concept the prompt implies but never states; in basic readout (WB.b, n = 100, chance 0.20) it must state the one-word answer the model has not yet given.

LayerFidelity (held-out)Fidelity (control)MAEM WB.aJ-lens WB.aMAEM WB.bJ-lens WB.b
60.790.350.010.000.050.01
120.800.320.160.000.260.03
180.730.160.290.020.540.06
240.750.140.590.350.550.18
300.700.150.710.400.590.20
360.750.120.870.570.630.20
42 (train layer)0.740.140.890.590.600.19
480.750.150.870.580.590.17
540.800.140.890.660.570.18
620.720.120.910.290.620.20

Transfer across models. We have not tested this and leave it for future work. It would require addressing the issue of the second model having its own residual stream (different basis). One solution, other than a new MAEM for the second model, would be a learned map between the two activation spaces.

Cost. The released MAEM (SFT and RL) took roughly 120 GPU-hours on B200s.

W2. Role of the RL stage.

Yes, the RL stage is the main difference from prior activation inverters. We now measure and report the effect of the RL objective on the activation inversion task, starting RL from SFT checkpoints at 2M, 4M and 8M examples, and comparing at matched total compute against MAEM SFT-only, InversionView, and InverseScope.

At 1e20 FLOPs, compute-matched SFT reaches 0.54 held-out cosine and SFT+RL reaches 0.77-0.78. SFT-only held-out cosine is highest early in training (0.64 at 1M examples) and declines with more examples. RL improves on SFT on every metric, with the largest gain on SAE features (0.51 to 0.83-0.85 on unseen features). The final result is nearly the same from every SFT starting point.

InversionView and InverseScope trained on the same compute budget reach 0.39-0.42 held-out cosine, which is below the MAEM SFT.

See the new Appendix K for the full training curves. Note in Table R2 we present InversionView on GPT-2 Medium which performed better than InversionView on a 2-layer transformer. In our response to Reviewer 2mwT’s W2b, we report the SFT checkpoint on the Section 5 experiments, where SFT+RL outperforms it on most tasks.

Table R2. All methods at a matched budget of 1e20 FLOPs (including activation collection). MAEM SFT-only is a separate run extended to 14.8M examples to match the budget.

Metric at 1e20 FLOPsMAEM (SFT+RL)MAEM, SFT onlyInversionViewInverseScope
Held-out activations, centred cosine0.77-0.780.540.420.39
Unseen SAE features, activation / corpus peak0.83-0.850.510.220.12
Held-out features of the training SAE, activation / corpus peak0.31-0.320.100.070.07
Mean cosine over all held-out direction families0.600.430.330.30

[1] Camila Blank, Agam Bhatia, Euan Ong, Neel Nanda. WorkspaceBench: Evaluating Interpretability Methods for the Global Workspace. 2026

I thank the authors for their response and maintain my positive score.

Reviewer_54BrOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Maximally-Activating Example Metamodels (MAEMs), which generate text that strongly activates a given residual-stream direction in a target language model. Trained with supervised fine-tuning and reinforcement learning using the target model’s activations as a direct reward, MAEMs outperform corpus search on held-out activation inversion and generalize to unseen languages and domains. The authors demonstrate their use in interpreting steering vectors, reading LoRA backdoor payloads, probing model activations, and activating SAE features.

Critical strengths
  1. This is a helpful topic. MAEM directly optimizes how strongly generated text activates a direction in the target model. Each example can be verified with a forward pass through that model, reducing reliance on human annotations or learned evaluators.

  2. Broad applicability. A single model can handle multiple types of residual-stream directions and generalize to unseen languages and domains, including code and mathematics. The appendix also includes an intriguing analysis of classical Chinese, a topic that could warrant a dedicated paper of its own.

  3. Evaluation across diverse interpretability tasks, which is very solid. The authors evaluate MAEM on activation inversion, steering-vector concept identification, LoRA backdoor recovery, model-workspace interpretation, and SAE feature activation.

  4. MAEM generates highly activating text that may not be present in the corpus, and outperforms corpus search on held-out activation inversion and steering-vector concept identification.

Critical weaknesses
  1. MAEM optimizes the activation similarity between generated text and a direction. This shows that the text does activate a similar direction, but does it still need to be separately demonstrated that the text accurately reveals the direction’s semantics, or does the semantic interpretation still depend on the reader or subsequent evaluation?

  2. In the steering-vector task, GPT-5.6 Sol selects which candidate concept corresponds to the text; in the workspace task, Claude Sonnet 5 judges whether the text mentions the target content. Could introducing other models dilute the contribution’s claim that only the target model is used and no other LLM is introduced? I personally suggest revising the evaluation methods for these tasks.

  3. The result paragraphs could also use some polishing. There is too much data stacked together, and different fonts are mixed (please don’t use italics everywhere; there is no need to change the font for model names either), making it hard to grasp the main points. I still suggest stating the conclusion first. The bulleting in other paragraphs is inconsistent, with (i), (a), and 1) used. These should be standardized. The formatting of cross-references is also inconsistent: some say “see Appendix,” while others just say “Appendix”; capitalization also varies, as do “fig.” and “Fig.”

  4. I suggest revising the contributions to three points. Accessing Limitations, for example, do not need to be listed as a contribution.

Questions for authors

See weakness.

2 comments

We thank the reviewer for taking the time to read our work and providing an engaging discussion.

W1. Does the semantic interpretation still depend on the reader or subsequent evaluation.

Yes, the reviewer is correct that maximum activation and exact semantic interpretation are evaluated separately. The MAEM is trained to output a text that maximally activates a direction, and the objective allows one to directly verify this with a single forward pass. The semantic interpretation is done by the user of the exemplars and their utility depends on the use-case. We highlight this in L489-491.

To evaluate the utility of MAEMs at scale, we design most of our experiments with known ground truths that should become evident either from string matching (word-rule) or having an LLM judge read the MAEM output the way the AI researcher would: identifying the concept behind a steering vector (S5.1), naming a planted backdoor payload (S5.2), naming the target workspace content (S5.3).

In Related work (L110-128), we contrast tools that read residual stream states and return descriptions with MAEMs, which output an exemplar that can be verified to activate the direction, leaving the semantic interpretation to the reader.

W2. External LLM judges in the evaluation.

Our claim concerns training and the reward: the MAEM is trained and its exemplars are verified using only a forward pass from the subject model. In Limitations (L489-491) we state that high cosine alignment does not rule out incorrect semantic account of the direction in some cases.

In practice, the AI researcher using the MAEM would be reading the exemplars directly. In our evaluation we use LLMs to do the task of the AI researcher so we can provide scores at scale.

Regarding the use of two models as judges in our evaluations: we use GPT-5.6 Sol as a judge in the steering vector task (S5.1) because the LLM judge we used throughout the rest of the paper, Claude Sonnet 5 (S5.3), declined to evaluate the text outputs (>50% for some methods) because of safety filters. We chose GPT-5.6 Sol because it is comparable in capability and did not decline the inputs. We highlight this design choice in L1606-1609.

We re-judged results in S5.1, S5.3 and S5.4 with Gemini 4 Argon and obtained high agreement with main conclusions holding:

  • Steering (S5.1): Cohen’s $\kappa$ = 0.86 for AxBench, $\kappa$=0.88 for BiPO.
  • Workspace (S5.3): Cohen’s $\kappa$ = 0.88.
  • Feature description (S5.4): Cohen’s $\kappa$ = 0.73

Gemini was used with an identical pipeline and did not decline more than 0.1% of inputs.

We also believe that our companion judge-free results minimise the risk of relying on LLM judges to highlight the usefulness of our MAEMs, namely: cosine for activation inversion (S4), verbatim payload match (S5.2), word-rule (S5.3), and feature firing (S5.4). Where there are no judge-free results (S5.1) we provide controls as comparators (shuffled, untrained base).

W3. Presentation and formatting of results.

We thank the reviewer for helping us improve the presentation of our paper. In the main text, we have moved some in-line data to relevant appendices, reduced our use of italics and removed unnecessary fonts (model names). We have also standardised bulleting and cross-reference formatting.

We have also added a new Table 4 with qualitative examples of MAEMs, and a discussion at the end of Section 4 on the coherence of MAEM exemplars and how it compares to other methods.

W4. Presentation of contributions.

We agree with the reviewer that the contributions list is too long. We have reduced the number of points in contributions down to 3 in the introduction, and removed the mention of limitations.

Thank you for your response. I will keep the score.

Reviewer_2mwTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper trains a copy of Qwen3.6-27B to take a residual stream direction as input and write text that strongly activates that direction in the original model. The model is first trained with supervised learning on direction and text pairs, then with RL using the cosine between the target direction and the text's activations as the reward. It inverts held out activations better than corpus search and NLA, matches NLA on explaining steering vectors and recovers many LoRA backdoor payloads, but explains SAE features less well than corpus examples.

Critical strengths

Strengths

  1. The paper trains an interpretability model with a reward computed by the subject model itself, so training needs no labels. Because the reward in Equation 1 only needs a direction, SAE directions that have no source text can be used for training, unlike inverters trained to reproduce the original text. In principle each output is text that can be checked with one forward pass, which is a more direct test than a description judged by another model.

  2. The evaluation covers many kinds of directions, including kinds never used in training. The same checkpoint is tested on held out activations, 21 unseen languages and domains, steering vectors, rank 1 LoRA directions, workspace content and SAE features. On steering vectors it identifies 95.8% of AxBench concepts against 61.4% for corpus search, and on theme backdoors its write direction samples name the payload in 85% of cases against 2% for corpus search.

  3. The paper reports its failures and negative results in detail. It shows that about 11% of SAE features cannot be activated, mostly rare and structural ones, that its exemplars are less diverse than corpus windows, that descriptions written from them are worse than those from corpus windows, and that further training on the failed features made results worse (Appendix J.4). Controls such as mismatched targets, random directions and shuffled exemplars are reported for each task.

Critical weaknesses

Weaknesses

  1. (Solved) The method is mainly a combination of existing components, and the closest prior work is not discussed. Activation inverters such as InversionView [1] and InverseScope [2] already train a generator that maps an activation to inputs with similar activations. Investigator agents [3] already train a generator with supervised learning followed by RL, using the target model's response as the reward, to produce inputs that elicit chosen behaviors. The paper cites [3] only for a diversity objective, so it is not clear what is new beyond applying this recipe to activation directions.

  2. (Solved) The effect of the RL objective, which is the main change from prior activation inverters, is not measured. No experiment reports the SFT checkpoint, which is close to an InversionView style inverter [1], and no experiment compares directly with InversionView or InverseScope. Section 6 and Appendix H.7 state that SFT and RL are not compared at matched data or compute. Without this comparison, the gains in Sections 4 and 5 cannot be attributed to the new objective rather than to model scale or the eight million supervised examples.

  3. (Solved) The claim that MAEM outputs are verifiable does not hold in some of the reported experiments. On the write directions of the backdoor adapters, the MAEM samples have a mean cosine of 0.066 with the target, the same as the random direction floor of 0.065 (Table 23), yet 85% of them name the payload. Success there therefore does not come from maximizing the activation, and resembles the describer behavior the paper criticizes in NLAs. In Sections 5.1, 5.3 and 5.4 the final judgment also rests on LLM judges.

references:

[1] Xinting Huang, Madhur Panwar, Navin Goyal, and Michael Hahn. Inversionview: A general-purpose method for reading information from neural activations. arXiv preprint arXiv:2405.17653, 2024. URL https://arxiv.org/abs/2405.17653.

[2] Yifan Luo, Zhennan Zhou, and Bin Dong. Inversescope: Scalable activation inversion for interpreting large language models. arXiv preprint arXiv:2506.07406, 2025. URL https://arxiv.org/abs/2506.07406.

[3] Xiang Lisa Li, Neil Chowdhury, Daniel D. Johnson, Tatsunori Hashimoto, Percy Liang, Sarah Schwettmann, and Jacob Steinhardt. Eliciting language model behaviors with investigator agents. arXiv preprint arXiv:2502.01236, 2025.

Questions for authors

How does MAEM compare with InversionView or InverseScope trained on the same subject model and data?

5 comments
AuthorsOpenReview ↗

Rebuttal Part 1

We thank the reviewer for their time and an engaging discussion.

W1a. Comparison to closely related work (InversionView, InverseScope).

We agree these methods are closely related. The revised Related Work (L131-149) discusses both [1] and [2], stating what MAEMs keep from InversionView and what changes. We added:

InverseScope (Luo et al., 2025) trains a generator conditioned through control layers on (input, noised activation) pairs by maximum likelihood, thus sampling inputs whose activations lie near a target; a hypothesised feature is then scored by how often the samples share it.

W1b. Discussion on Investigator agents.

We agree that Investigator Agents (Li et al., 2025) uses a broadly similar training recipe. We have added it to section 2:

For model outputs, Investigator Agents (Li et al., 2025b) amortise prompt search with a recipe similar to ours: a generator conditioned on a target string (or rubric) is trained with SFT on (prompt, response) pairs obtained by running the target model forwards, then with DPO on the log-probability the target assigns to the desired output (or on an LM verifier’s score).

However, the two solve different problems: IAs invert text output: their generator is conditioned on a target string or rubric and rewarded with the log-probability the target model assigns to it, which requires only output log-probabilities, and their use-case is behaviour elicitation. MAEMs instead invert internal states: our generator is conditioned on a residual-stream vector and rewarded with the cosine between it and the subject’s activations on the generated text (which requires white-box access), and the use-case is interpretability. As a result, a MAEM can take targets that have no text specification (like SAE features, steering vectors, LoRA read/write directions). The MAEM reward doesn’t need an LM judge for any target type, where Li et al. use an LM verifier for open-ended rubric targets. Finally, one MAEM checkpoint covers all of these direction families, where Li et al. train a separate investigator per task.

W2a. The effect of the RL objective.

We now report the effect of RL on the activation inversion task, starting RL from SFT checkpoints at 2M, 4M and 8M examples, and comparing at matched total compute against MAEM SFT-only, InversionView, and InverseScope.

At 1e20 FLOPs, compute-matched SFT reaches 0.54 held-out cosine and SFT+RL reaches 0.77-0.78. SFT-only held-out cosine is highest early in training (0.64 at 1M examples) and declines with more examples. RL improves on SFT on every metric, with the largest gain on SAE features (0.51 to 0.83-0.85 on unseen features). The final result is nearly the same from every SFT starting point.

InversionView and InverseScope trained on the same compute budget reach 0.39-0.42 held-out cosine, which is below the MAEM SFT.

Table R1. Methods at a matched budget of 1e20 FLOPs.

Metric at 1e20 FLOPsMAEM (SFT+RL)MAEM, SFT onlyInversionViewInverseScope
Held-out activations, centred cosine0.77-0.780.540.420.39
Unseen SAE features, activation / corpus peak0.83-0.850.510.220.12
Held-out features of the training SAE, activation / corpus peak0.31-0.320.100.070.07
Mean cosine over all held-out direction families0.600.430.330.30

W2b and Q1. Comparison against InversionView and InverseScope.

We add InversionView and InverseScope (data-matched) as baselines for our Section 5 experiments. MAEM (SFT+RL) outperforms prior inverters across every experiment.

Relative to SFT, RL raises identification of BiPO-trained vectors from 51.4% to 83.7%, backdoor payload identification from 12.5% to 84.6%, the WorkspaceBench mean across 17 task families from 8.5% to 23.8%, and AxBench identification by 3.8 percentage points (p=0.01). In Section 5.3 the two checkpoints perform similarly on association (31.4% vs 33.3%), with RL ahead on multi-hop (51.1% vs 41.3%, though not significantly p=0.23).

For workspace verbalisation we also evaluate all methods on a new benchmark, WorkspaceBench [1], finding that MAEM is at or above prior inverters and SFT-only in each of the 17 task families.

See the new Appendix K in the revision for more details.

Table R2. Pass rate (%) on Section 5 experiments and WorkspaceBench. Prior inverters are trained on the same supervised pairs as MAEM.

ExperimentMAEM (SFT+RL)MAEM (SFT)InverseScopeInversionView
AxBench steering vectors, all concepts (n=500)85.081.276.270.2
BiPO-trained persona vectors (n=119)83.751.417.815.7
Backdoor payload (write, n=32)84.612.50.60.2
Section 5.3 association, judged (n=102)31.433.32.98.8
Section 5.3 multi-hop, judged (n=92)51.141.311.08.8
WorkspaceBench, mean over 17 families (n=1932)23.88.50.70.4
AuthorsOpenReview ↗

Rebuttal Part 2

W3a. The claim that MAEM outputs are verifiable does not hold in some of the reported experiments (backdoors).

We agree that payload naming by MAEM exemplars on the backdoor write directions cannot be verified by their cosine to the direction. In this experiment the MAEM acts as a reader of the direction, and its success is established by comparison with the known payload and by controls rather than by the cosine training objective. The forward pass verifies how strongly an exemplar activates the (whole) direction, and this is correctly reported as low for the backdoor directions used.

(i) The write directions lie mostly outside of the subspace where the model's activations reside. In the top 256 principal directions of layer 42 activations, a real activation has 45.6% of its squared norm, while the write direction has only 5% (about the same as a random direction). As a result, no text scores a high cosine against them: 10M span corpus search reaches 0.088 (against 0.065 on random directions), and the activation of the payload word itself scores only 0.054.

(ii) The write direction does carry a payload component. The payload word's activation scores higher against it than the other 15 adapters' payload words (0.054 vs 0.023 on average), and is the best match of the 16 for 15 of 16 adapters (chance is 1 in 16, p < 0.0001). Removing this component from each write vector (0.6% of its squared norm) lowers payload naming from 81% (in this ablation run) to 38% of samples across 32 adapters, while removing the same share along a random direction leaves it at 79%, and removing another adapter's payload component leaves it at 75%. MAEM's payload naming therefore depends to a substantial extent on the payload component of the direction.

In a new experiment (Table R3), we build directions where a target noun's activation holds a fixed share of the squared norm and the rest is random isotropic noise. At 3% share, the noun naming rate is 16% with cosine 0.042 (close to the random-direction floor of 0.034). This shows that a concept which is a small share of a direction can be named while the cosine stays near the floor, because the cosine measures how well an exemplar activates the whole direction (the training objective).

In the revision, we have reframed the verifiability claim in Section 5.2 to state that "the forward pass verifies how strongly an exemplar activates the direction but does not guarantee payload recovery". We also amend Table 23 (Table 24 in the revision) and the appendix text: 0.065 is corpus search on random directions, and the MAEM cosines are now reported at the same sign as the naming rates (Table 19), where MAEM’s mean per-sample cosine is 0.037, against its own per-sample floor of 0.031. Appendix I.4 gives the full analysis, including the share of each direction family in the top-k principal directions and Table R3.

Table R3. % of MAEM's exemplars naming the target noun, with the mean per-sample cosine to the direction in brackets. MAEM scores 0.034 on a random direction. 60 target nouns, 32 samples each.

Noun share of the direction1%2%3%5%10%20%30%100%
Rest is random noise2 (.035)8 (.038)16 (.042)26 (.050)55 (.077)73 (.124)82 (.161)91 (.319)

W3b. Use of LLM judges in S5.1,5.3, and 5.4.
We re-judged results in S5.1, S5.3 and S5.4 with Gemini 4 Argon and obtained high agreement with main conclusions holding:

  • Steering (S5.1): Cohen’s κ\kappa = 0.86 for AxBench, κ\kappa=0.88 for BiPO.
  • Workspace (S5.3): Cohen’s κ\kappa = 0.88.
  • Feature description (S5.4): Cohen’s κ\kappa = 0.73

Gemini was used with an identical pipeline and did not decline more than 0.1% of inputs.

We also believe our companion judge-free results minimise the risk of relying on LLM judges to highlight the usefulness of our MAEMs, namely: cosine for activation inversion (S4), verbatim payload match (S5.2), word-rule (S5.3), and feature firing (S5.4). Where there are no judge-free results (S5.1) we provide controls as comparators (shuffled, untrained base).

[1] Camila Blank, Agam Bhatia, Euan Ong, Neel Nanda. WorkspaceBench: Evaluating Interpretability Methods for the Global Workspace. 2026

Dear Reviewer 2mwT,

We would greatly appreciate it if you had a chance to review our rebuttal. We are happy to clarify any remaining questions or concerns and would welcome the opportunity to continue the discussion.

We understand this is a busy period and are grateful for your time. Thank you again for your thoughtful and constructive review.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.