acceptodds
Under review as a conference paper at ICLR 2027

Faithful Activation Verbalization: Reducing Hallucinations in LLM Representation Interpretation

Abstract

Activation verbalization methods such as Activation Oracle and Natural Language Autoencoders decode hidden representations of large language models into human-readable natural language. However, existing methods can produce incomplete or hallucinated descriptions, making their activation verbalizations difficult to trust and use reliably in practice. To this end, we introduce AVPO, a two-stage framework that first reconstructs source text from a hidden activation and then evaluates the resulting text with a separate frozen question-answering model, yielding an explicit and inspectable intermediate readout. We further optimize the inverter with direct preference optimization (DPO), using rewards that capture both semantic recoverability and lexical fidelity. Across six text families, AVPO improves gist- and detail-level information recovery over the strongest baseline by up to 17.1 and 9.3 percentage points, respectively. Crucially, the gains arise from preference optimization rather than fine-tuning on selected reconstructions alone, enabling compact cross-model inverters to surpass donor-matched question-conditioned verbalizers while improving both semantic recoverability and lexical fidelity. Moreover, out-of-distribution case study shows that AVPO better recovers high-level semantics while fabricating fewer details.

Primary area:
interpretability and explainable AI
Keywords:
Representation Verbalization, LLM interpretability

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_XdvTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes AVPO, a two-stage activation verbalization framework that reconstructs text from hidden activations and further improves the inverter with DPO-based rewards. AVPO achieves strong gains over existing baselines across multiple models and evaluation settings.

Critical strengths
  • The problem setting is clear, and the design is intuitive. By separating activation readout from downstream QA, the method makes the intermediate textual representation directly inspectable.
  • The performance gains are substantial and consistent across settings. AVPO consistently outperforms existing activation verbalization baselines across multiple donor and decoder configurations.
  • The experimental analysis is comprehensive. The paper includes analyses of reward design, layer choice, DPO iterations, and several control experiments.
Critical weaknesses
  1. It is unclear whether reconstructing the input text should itself be considered an interpretation of the activation. AVPO aims to provide a human-readable interpretation of hidden representations, but its actual objective is to reconstruct the original source text from the activation. However, the activation may compress or discard parts of the input information, while the inverter may infer missing details from the decoder’s prior. Therefore, strong reconstruction performance does not necessarily imply that the generated text faithfully reflects the information actually represented in the activation.
  2. AVPO may learn reconstructions that are more useful for QA rather than more faithful to the activation. The main DPO reward is based on QA performance over the reconstructed text. Therefore, it is unclear whether the observed gains come from more faithful recovery of the activation’s overall information or from selectively emphasizing information that is useful for downstream QA.
  3. There is limited direct evidence that the generated information actually originates from the activation. The current evaluation mainly measures source information recoverability, but it does not disentangle whether the reconstructed content is determined by the activation itself or supplemented by the inverter’s prior knowledge.
  4. The training reward and the primary evaluation rely heavily on the same QA and judge pipeline. Qwen3-32B is used both for the DPO reward and for the main gist/detail evaluation. Although the judge correlation analysis is useful, it does not directly demonstrate that AVPO’s relative gains over the baselines persist under an independent evaluator.
  5. Direct evaluation of hallucination is limited relative to the paper’s claims of “faithfulness” and “reducing hallucinations.” Anchor F1 only measures overlap in proper nouns and numbers rather than overall factual correctness, while the gist/detail metrics primarily measure information recoverability.
  6. Some closely related baselines are omitted from the main experiments. The paper discusses Rep2Text and NLA as closely related activation verbalization methods, but neither is included in the main comparison.
Questions for authors

Please see the weaknesses.

4 comments
AuthorsOpenReview ↗

Response to reviewer XdvT (1/2)

We thank reviewer XdvT's comprehensive and insightful questions. Please find our response below.

W1. 1) It is unclear whether reconstructing the input text should itself be considered an interpretation of the activation. AVPO aims to provide a human-readable interpretation of hidden representations, but its actual objective is to reconstruct the original source text from the activation. 2). The activation may compress or discard parts of the input information, while the inverter may infer missing details from the decoder's prior. Therefore, strong reconstruction performance does not necessarily imply that the generated text faithfully reflects the information actually represented in the activation.

[ANS] (1) We thank the reviewer for raising this fundamental question. Li et al. (2026, §4) explicitly note that "verbalizers could respond based on (implicitly) reconstructed inputs." This possibility motivates examining the reliability of the underlying readout: if an interpretation builds on recovered context, errors in that context can undermine its conclusions. AVPO makes this intermediate step explicit and inspectable, allowing us to assess what was recovered before interpreting its implications.

(2) We agree that the inverter may infer from the decoder's prior, which is a shared challenge. Our goal is to encourage the reconstructed text to remain faithful to the input. With this, we can make sure the information usage during activation readout is inspectable.

W2. AVPO may learn reconstructions that are more useful for QA rather than more faithful to the activation. The main DPO reward is based on QA performance over the reconstructed text. Therefore, it is unclear whether the observed gains come from more faithful recovery of the activation's overall information or from selectively emphasizing information that is useful for downstream QA.

[ANS] We do not claim that the inverted text preserves all information encoded in the activation. Our gist and detail questions are designed to encourage the inverted text to preserve the main meaning and important details of the input.

W3. There is limited direct evidence that the generated information actually originates from the activation. The current evaluation mainly measures source information recoverability, but it does not disentangle whether the reconstructed content is determined by the activation itself or supplemented by the inverter's prior knowledge.

[ANS] Existing work on text inversion has investigated the information source in activation-conditioned decoding. For example, removing activation-conditioned adapter outputs in UAV while retaining the decoder-side LoRA reduces fact-retrieval Token F1 from 0.262 to 0.144 and gist Token F1 from 0.233 to 0.162 (Zhao et al., 2026, §4.5). These results provide evidence that activation information contributes to the generated content.

However, fully disentangling activation-derived information from decoder priors goes beyond the scope of our work. We acknowledge this limitation and will identify direct activation interventions on AVPO as future work in the revised manuscript.

W4. The training reward and the primary evaluation rely heavily on the same QA and judge pipeline. Qwen3-32B is used both for the DPO reward and for the main gist/detail evaluation. Although the judge correlation analysis is useful, it does not directly demonstrate that AVPO's relative gains over the baselines persist under an independent evaluator.

We re-scored the main gist/detail evaluation on 978 sealed-test QA pairs using four additional judges: GPT-4.1-mini, Gemma-4-31B, Llama-3.3-70B, and Qwen2.5-32B, following the same evaluation protocol. The results are shown below. Although the performance gaps between AVPO and other methods vary across judges, they remain consistently positive, and the variation is smaller than the gap magnitudes.

Llama-3.1-8B L31 → Qwen3-4B — Ours = MR-DPO round 4 (n = 978 QAs)

Ours vs.Qwen3-32B (orig.)GPT-4.1Gemma-4-31BLlama-3.3-70BQwen2.5-32BΔ rangeΔ range (%)
Stage-1 (DPO gain)+0.236+0.197+0.214+0.249+0.208[−0.039, +0.013][−17%, +6%]
LatentQA+0.151+0.111+0.135+0.171+0.142[−0.040, +0.020][−27%, +13%]
AO+0.196+0.165+0.178+0.202+0.177[−0.031, +0.006][−16%, +3%]
UAV+0.186+0.142+0.156+0.193+0.166[−0.043, +0.007][−23%, +4%]
SelfIE+0.498+0.409+0.441+0.462+0.410[−0.088, −0.036][−18%, −7%]
Patchscope+0.589+0.515+0.517+0.606+0.515[−0.074, +0.017][−13%, +3%]
AuthorsOpenReview ↗

Response to reviewer XdvT (2/2)

W5. Direct evaluation of hallucination is limited relative to the paper's claims of "faithfulness" and "reducing hallucinations." Anchor F1 only measures overlap in proper nouns and numbers rather than overall factual correctness, while the gist/detail metrics primarily measure information recoverability.

[ANS] We add two complementary evaluations on our curated implication dataset: (1) atomic fact grounding (2) implication QA evaluation that measure how well implied intentions are verbalized. All methods use Llama-3.1-8B from our main settings.

(1) In atomic fact grounding, we decompose reconstrcuted text from pre-dpo and dpo r4 into around 1000 atomic fact for each. We use GPT-4.1-mini to classify each fact as supported, partially supported, contradicted, or not in source. We separately compare reconstrcutions from each setting to the source. The result is shown in the table. We use absolute percentage gap here.

MethodSame event ↑Supported ↑Partially supportedContradicted ↓Not in source ↓
Pre-DPO30.1%16.8%11.3%12.2%59.7%
Post-DPO51.0%23.3%10.5%11.3%54.9%

DPO increases the same-event rate by 20.9%, showing great improvements at the top level. At the fact level, the supported proportion increases by 6.5%, while contradicted and not-in-source facts decrease by 0.9% and 4.8%, respectively. These results show concrete improvements in both event alignment and factual grounding.

2. Implication QA evaluation.

We additionally evaluate whether the reconstructed information supports correct answers to implied intentions. The dataset contains 300 short passages from SocialIQA (246), INLI (27), and five dialogue sources (27). We retain samples that Llama-3.1-8B answers correctly with the passage but fails on all six question-only attempts, reducing the likelihood of correct answers using model priors.

We use GPT-4.1-mini as readout and judge. GPT-4.1-mini scores all answers against reference answers on a 0–4 scale: 4 denotes fully correct, 2–3 partially correct, and 0–1 incorrect. The average score is normalized to [0, 1].

MethodFully correct ↑Partially correct↑Incorrect ↓Average score ↑
UAV49 (16.3%)115 (38.3%)136 (45.3%)0.403
AO58 (19.3%)99 (33.0%)143 (47.7%)0.407
LatentQA59 (19.7%)116 (38.7%)125 (41.7%)0.441
AVPO SFT54 (18.0%)103 (34.3%)143 (47.7%)0.380
AVPO r162 (20.7%)130 (43.3%)108 (36.0%)0.481
AVPO r469 (23.0%)122 (40.7%)109 (36.3%)0.497

As shown in the above table, DPO r4 increases the fully correct answer rate from 18.0% to 23.0% and reduces the incorrect answer rate from 47.7% to 36.3%.

So both experiments are consistent regarding fact-level results. DPO improves factual grounding in the reconstrcuted text and meanwhile improves downstream activation readout performance.

W6. Some closely related baselines are omitted from the main experiments. The paper discusses Rep2Text and NLA as closely related activation verbalization methods, but neither is included in the main comparison.

[ANS] Thanks for bring up this concern. Our inverter is only different on adapter design from Rep2Text. Our focus are two folds: (1) DPO can improve faithful inversion, which is established in W5. We add another experiment with Rep2Text using 1-round DPO. The result is below in the table and shows that DPO can also improve its inverter (2) Our two-stage approach can make information usage inspectable. Therefore, we focus on comparing with works related to QA-based information readout. Besides our baseline UAV is also a verbalizeation verion of Rep2text.

For NLA, We were hoping to include NLA and tried to replicate it in preliminary experiments. However, even for the smallest released setting (Qwen2.5-7B), a single NLA run is estimated to take about 9.5 days on 16 A100-80GB GPUs based on our measured training throughput, and this cost is incurred separately for every model and layer. Unfortunately, this is far beyond what our limited academic compute budget can afford, so we had to leave NLA out of the main comparison.

QA gistQA detailAnchor F1
Rep2Text0.1960.2150.196
Rep2Text + our DPO0.2590.2310.211
Δ (95% CI)+0.063 [+0.039, +0.088]+0.015 [−0.007, +0.037]+0.015 [+0.003, +0.027]

Reference

  • Zhao, Haiyan, et al. "Rep2Text: Decoding Full Text from a Single LLM Token Representation." arXiv preprint arXiv:2511.06571 (2025).
  • Zhao, Haiyan, et al. "Universal Activation Verbalizer: A Unified Framework for Cross-Model Activation Explanation." arXiv preprint arXiv:2605.25903 (2026).

Dear Reviewer XdvT,

As the author–reviewer discussion period is drawing to a close, we wanted to check whether our response has addressed your concerns, and whether there is anything further we can clarify or provide. We welcome any comments and suggestions you may have at any time, and would be glad to discuss with you how to further improve our paper.

Thank you again for your time and thoughtful review.

Thank you for your response. Regarding W1, I agree with reviewer sz9v. W2 and W3 have not been resolved. Therefore, I maintain my score.

Reviewer_6LuZOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a two-stage framework for activation verbalization: it first trains a question-agnostic inverter to reconstruct the source text and then further trains it with DPO using rewards that encourage faithful readouts. Evaluations across multiple text domains and different types of questions show higher faithfulness than baseline methods.

Critical strengths
  1. The paper proposes a two-stage framework with good motivation to disentangle two different problems in the activation verbalization domain.

  2. Detailed ablations are conducted, and the empirical results are strong.

  3. Out-of-distribution results also show the robustness of the method.

Critical weaknesses
  1. The proposed two-stage method highly depends on the quality of the supervised inverter, as the subsequent preference optimization stage requires the inverter to provide reasonable and diverse candidates for sampling and preference-pair construction. However, there is a lack of ablations or analyses on the choice of the inverter. Can we swap an existing inverter into the pipeline, e.g., Rep2Text [1] ? This would affect my evaluation of the contribution of the proposed method. In addition, there is no analysis of the reward distribution of the sampled responses from the inverter: are the negative responses much worse than the positive ones, or are they hard negatives?

  2. The evaluation setup does not completely disentangle "question-agnostic text readout" from question-specific information. The gist QA pairs used for evaluation are constructed such that the answers are almost verbatim copies of the source text, as shown in the two examples in Appendix C.3 (even though the QA generation prompt in Appendix C.2 states that the answers should be concise). Given that, the questions used in the evaluation are basically text readout. This undermines the paper's claim that the main contribution of the method is using a two-stage process to disentangle readout from question answering.

[1] Zhao et al., 2026. Rep2Text: Decoding Full Text from a Single LLM Token Representation.

Questions for authors
  1. The three reward components do not introduce any regularization on response length. Is it possible that the inverter can game the detail reward, and potentially the anchor reward, by generating very long responses?
4 comments
AuthorsOpenReview ↗

Response to reviewer 6LuZ

We thank Reviewer 6LuZ for the thoughtful review and constructive questions, which have helped us strengthen our work.

W1. 1) Lack of ablations on the choice of inverter 2) Lack of analysis of the reward distribution of the sampled response from the inverters: are the negative responses much worse than the positive ones, or are they hard negatives?

[ANS] Thank you for highlighting the issue with choice of inverter and reward distribution of the sampled responses.

(1) We take the released Rep2Text inverter (Llama-3.1-8B, layer 10, with 32 soft tokens) as our pre-DPO setting, then run 1-round DPO on constructed dataset using Rep2Text's Wikipedia data, and evaluate on its official test set with 1,000 samples, with 2,000 questions. The result is shown below in the table.

QA gistQA detailAnchor F1
Rep2Text0.1960.2150.196
Rep2Text + our DPO0.2590.2310.211
Δ (95% CI)+0.063 [+0.039, +0.088]+0.015 [−0.007, +0.037]+0.015 [+0.003, +0.027]

Although the decoder is frozen in Rep2Text's first-stage training, during DPO we jointly optimize the adapter and a LoRA on the decoder (zero-initialized, so the pre-DPO model is identical to the released one). With this, we still see concrete improvements in gist QA and anchor F1. These results show that our preference optimization can also improve an existing Rep2Text inverter.

(2) Following our paper, we use a minimum reward gap thershold of 0.2 when constrcuting pairs. The selected pairs span different degrees of reward separation, as shown in the table below.

Target ModelLayerDecoderRoundNo. of Pairs[0.2, 0.4)[0.4, 0.6)[0.6, 0.8)[0.8, 1.0]
Mistral-Small-24BL39Qwen3-0.6Br121,68630.9%30.8%24.2%14.0%
r222,51238.2%34.2%21.3%6.4%
r322,99641.1%33.5%20.6%4.8%
r4*22,80242.4%34.2%19.2%4.2%
Mistral-Small-24BL39Mistral-24B (self)r123,44022.4%30.1%27.7%19.8%
r221,17138.1%32.1%21.8%8.1%
r321,24640.5%32.5%20.3%6.8%
r4*21,45241.0%32.5%20.4%6.1%

To better illustrate the content differences of preference paris, we also provide one chosen-rejected example per gap bin at the end of the rebuttal.

W2. The evaluation setup does not completely disentangle "question-agnostic text readout" from question-specific information. The gist QA pairs used for evaluation are constructed such that the answers are almost verbatim copies of the source text, as shown in the two examples in Appendix C.3 (even though the QA generation prompt in Appendix C.2 states that the answers should be concise). Given that, the questions used in the evaluation are basically text readout. This undermines the paper's claim that the main contribution of the method is using a two-stage process to disentangle readout from question answering.

[ANS] Our paper focuses on activation readout, and we indeed hope the inverted text stays true to the input. Therefore, truthful verbatim is preferred.

For separation, we mean the process of information recoverage and information usage. The point is that, without explicit inverted text, we can't decide if the model is answering with activation-grounded information or using prior knowledge. However, explicit inverted text can make the information usage process inspectable.

W3. The three reward components do not introduce any regularization on response length. Is it possible that the inverter can game the detail reward, and potentially the anchor reward, by generating very long responses?

[ANS] Thank you for raising this important question. In our preliminary experiments, we tested length regularization following R-DPO (Park et al., 2024). We retrained the first DPO round on the same preference pairs with all other hyperparameters fixed (Llama-3.1-8B L31 → Qwen3-4B). Results on the 2,172 validation inputs are as below:

MethodGistDetailOverallAvg. length (words)
SFT inverter0.3500.4590.40528.6
DPO (α=0\alpha=0)0.4390.5290.48441.6
DPO + length reg. (α=0.02\alpha=0.02)0.4430.5270.48536.2
DPO + length reg. (α=0.05\alpha=0.05)0.4390.5280.48437.6

Length regularization shortened outputs while preserving the QA gains, with overall scores changing by at most 0.001. So the improvement is unlikely to come primarily from gaming response length. Given its negligible effect on performance, we did not emphasize this analysis in the main paper. We will include the ablation in the revised appendix.

Reference

  • Park, Ryan, et al. "Disentangling length from quality in direct preference optimization." Findings of the Association for Computational Linguistics: ACL 2024. 2024.
AuthorsOpenReview ↗

Examples of preference pairs

GapOriginal textChosenRejected
[0.2, 0.4)<br>gap = 0.35Develop a Powershell script<br>that list all ppt file in a directory<br>and it's sub directory,<br>then call powerpoint<br>to convert it to pdf file.I want to convert a pptx file to pdf format.<br>Write a powershell script to do it,<br>only using the native modules.<br><br>gist 1.00 / detail 0.75 / anchor F1 0.00<br>R = 0.61Write a powershell script,<br>which takes an Excel file from a location<br>and saves it as pdf in an other location.<br><br>gist 0.75 / detail 0.00 / anchor F1 0.00<br>R = 0.26
[0.4, 0.6)<br>gap = 0.50On a distant planet,<br>the inhabitants experienced weightlessness<br>once every orbit<br>when their twin moons aligned perfectly,<br>nullifying gravity for a single day.On a distant planet,<br>the inhabitants experienced weightlessness<br>once every orbit<br>when their twin moons aligned perfectly,<br>nullifying gravity for a single day.<br><br>gist 1.00 / detail 1.00 / anchor F1 –<br>R = 1.00On the planet of Novia,<br>the inhabitants experienced a perpetual motion day<br>every 100 years<br>when the twin moons aligned perfectly,<br>causing gravity to fluctuate unpredictably<br>for 24 hours.<br><br>gist 0.50 / detail 0.50 / anchor F1 –<br>R = 0.50
[0.6, 0.8)<br>gap = 0.70The Vietnamese National Football First League<br>(V.League 2)<br>is the second highest league in Vietnamese football<br>after the V.League 1<br>with 14 teams competing in it.V.League 2 (V.League 2)<br>is the second highest level of professional football in Vietnam.<br>The league consists of 14 teams<br>who will play each other team once at home and once away<br>for a total of 26 games a year.<br><br>gist 1.00 / detail 1.00 / anchor F1 0.43<br>R = 0.83The Vietnamese League Cup,<br>known as V.League Cup for sponsorship reasons,<br>is the men's national football cup competition<br>for clubs in Vietnam<br>and is contested in a tournament format.<br><br>gist 0.00 / detail 0.00 / anchor F1 0.43<br>R = 0.13
[0.8, 1.0]<br>gap = 0.90Bordeaux's producers target Denmark<br>to fight wine hangover (AFP)<br>AFP - A delegation of producers<br>from the prestigious French wine-growing region of Bordeaux<br>launched a major campaign<br>try to convince Danes<br>to start drinking their wines again.French look for new markets for Bordeaux in Denmark (AFP)<br>AFP - French wine-makers are on a mission<br>to remind the Danes<br>how much they once liked their Bordeaux,<br>despite a drop in consumption in recent years.<br><br>gist 1.00 / detail 1.00 / anchor F1 1.00<br>R = 1.00Danish winemakers on charm offensive in France (AFP)<br>AFP - Danish wine producers<br>are mounting a charm offensive in France<br>to convince the French<br>to drink less of their own produce<br>and more of their own.<br><br>gist 0.00 / detail 0.00 / anchor F1 0.33<br>R = 0.10

Dear Reviewer 6LuZ,

As the author–reviewer discussion period is drawing to a close, we wanted to check whether our response has addressed your concerns, and whether there is anything further we can clarify or provide. We welcome any comments and suggestions you may have at any time, and would be glad to discuss with you how to further improve our paper.

Thank you again for your time and thoughtful review.

I appreciate the responses from the authors and I think they strengthen the paper. I will keep my already positive scores.

Reviewer_sz9vOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors present a pipeline that given a text and the activation at its last token reconstructs the text from the activation. They first train the model to try to reconstruct the exact text with SFT, then use DPO to optimize for text reconstructions that retains key facts about the text. They show that this generalizes to other datasets but not really to other objectives. In contrast to activation verbalization techniques, their reconstruction focuses on reproducing what is already in the text and by design seems they can't beat a simple "give the context to an LLM" baseline which is the whole point of activation verbalizer.

Critical strengths

None. To be clear this doesn't mean all of the paper is bad, there is some effort in ablations etc that are appreciated, but I think the method in the first place is not well motivated. I encourage the authors to carefully think about their research questions before going deep into experiments for their future projects!

Critical weaknesses
  • Weakness 1: The point of activation verbalization is to go beyond text inversion and to give insight in stuff that the model did not verbalize (Li et al., 2026). This paper is an activation inversion paper framed as activation verbalization which doesn't make sense. The abstract of Bauer et al. (2026) literally warns that we need to go beyond that! The authors seem to have not a clear understanding of the problem they are trying to tackle. Like why would I ever use that to understand model behavior while the "ceiling" is literally just give the context to an LLM. Activation verbalization is supposed to go beyond the ceiling but the task chosen makes this impossible!!!

  • Weakness 2: I don't think the authors took sufficient care in the baselines. For example the activation oracle they train is on a single Q/A task and give a single layer while previous work emphasized that to get the best out of activation oracles you need a lot of different tasks (Karvonen et al., 2026), and multi-token/multi-layer inputs, on-policy rollouts, and an improved injection formula (Bauer et al., 2026).

  • Bauer et al. (2026): Jan Bauer, Celeste De Schamphelaere, Adam Karvonen, Niclas Luick, and Neel Nanda. Building better activation oracles. arXiv preprint arXiv:2606.02609, 2026.

  • Karvonen et al. (2026): Adam Karvonen, James Chua, Clément Dumas, Kit Fraser-Taliente, Subhash Kantamneni, Julian Minder, Euan Ong, Arnab Sen Sharma, Daniel Wen, Owain Evans, and Samuel Marks. Activation oracles: Training and evaluating LLMs as general-purpose activation explainers. In Forty-third International Conference on Machine Learning (ICML), 2026. (arXiv:2512.15674 / OpenReview)

  • Li et al. (2026): Millicent Li, Alberto Mario Ceballos Arroyo, Giordano Rogers, Naomi Saphra, and Byron C. Wallace. Do activation verbalization methods convey privileged information? In Proceedings of the 43rd International Conference on Machine Learning (ICML), 2026. (OpenReview)

Questions for authors

None

typos and remarks

As ICLR don't have a section for typos / other slightly confusing stuff, I include some notes here in case they are useful to authors for future work. No need to respond to this in your rebuttal as those are not influencing my score: Some of the writing was quite hard to understand and makes sense only after reading the paper, I'd encourage the authors to avoid such dense and jargony text in the abstract.

both semantic recoverability and lexical fidelity

to surpass donor-matched question-conditioned verbalizers while improving both semantic recoverability and lexical fidelity.

yielding an explicit and inspectable intermediate readout

Also even after reading the paper, figure 1 is not easily understandable.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.