acceptodds
Withdrawn from ICLR 2027

EXACT: Explicit Attribute-Guided Decoding-Time Personalization

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Achieving personalized alignment requires adapting large language models to each user’s evolving context. While decoding-time personalization offers a scalable alternative to training-time methods, existing methods largely rely on implicit, less interpretable preference representations and impose a rigid, context-agnostic user representation, failing to account for how preferences shift across prompts. We introduce EXACT, an interpretable decoding-time personalization method that aligns generation with limited pairwise preference feedback using a predefined set of explicit preference attributes. EXACT first identifies user-specific attribute subsets by maximizing the likelihood of preferred responses in the offline stage. Then, for online inference, EXACT retrieves the most semantically relevant attributes for an incoming prompt and injects them into the context to steer generation. We establish theoretical approximation guarantees for the proposed algorithm under mild assumptions, and provably show that our similarity-based retrieval mechanism effectively mitigates contextual preference shifts, adapting to disparate tasks without pooling conflicting preferences. Extensive experiments on human-annotated preference datasets demonstrate that EXACT consistently outperforms strong baselines, including preference modeling accuracy and personalized generation quality, while preserving a transparent and inspectable personalization signal.

Primary area:
generative models
Keywords:
decoding-time personalization, preference alignment

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_D2QPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper aims to adapt a large language model to one user at inference time.

Its key idea is that a user's preferences may shift from prompt to prompt. So to handle this, EXACT, first picks a few attributes, such as Concise or Formal, for each past prompt. And these attributes are chosen greedily. When a new prompt comes in, EXACT finds the most similar past prompt and adds its attributes to the new prompt.

According to the authors, this simple design is competitive with RAG few-shot on PRISM and Summarize from Human Feedback, in both pairwise accuracy and LLM-judged win rates.

Critical strengths

The paper points out that one user may prefer different styles for different prompts - this is a fair motivation for retrieval at the prompt level.

Critical weaknesses

Before assessing scientific novelty, I must raise serious research integrity concerns about this submission. I believe the authors are committed academic misconduct in this work.

It is very likely that large parts of the paper were produced by AI and then submitted without fixing the problems that the AI itself flagged, several key results cannot come from the method and data that the paper describes, and even the paper itself admits that key results are not verified. I listed evidences below.

important!! - I'm here to ask the other reviewers and PC/ACs to help check these points carefully. - I take research integrity seriously and will provide more evidence I have to the AC and other reviewers for assessment, and am willing to discuss these concerns openly.


Evidence 1. The paper itself directly says that many key parts are drafts and are not verified

These notes appear from the Introduction page1 to the Appendix - No human author would leave such content in a submission.

Main text

  • Introduction (line 052): "the draft illustrations in Table 1 await source verification."
  • Table 1 caption (lines 074–077): "Draft illustrations of contextual preference shifts (provenance pending)."
  • Method (lines 130–131): "draft prompt-format illustrations whose source records and vocabulary mapping remain to be verified."
  • Theory (lines 279–281): "Draft theory verification … require correction before this theorem can be used as a guarantee for EXACT."
  • Section 4.3, human study (line 435): "study metadata and item-level records remain to be reconciled."
  • Section 4.4, ablation (line 449): "The retrieval vs. non-retrieval comparison requires metric verification."
  • Section 4.5, case study (lines 458–462): "The draft case record for user thott7Xe … Dataset provenance remains unresolved."
  • Ethics statement (lines 500–503): the study procedures, informed consent, compensation, and IRB review "remain to be documented."
  • Reproducibility statement (line 519): "Draft tables and cases with unresolved metric or provenance conflicts are marked for reconciliation."

Appendix

  • Table 8 (lines 1507–1509): "These draft values require search-log verification."
  • Table 9 caption (lines 1525–1527): "Verification is pending: at k = 3, 0.13 conflicts with 0.64 in Table 2."
  • Table 15 (lines 1669–1681): "Draft result verification … they have not yet been verified" and "they must be confirmed or recomputed."
  • Figures 2–7 (lines 1238–1241): "Figures 2–7 retain draft prompt-format illustrations … remain unverified."

In short, the authors doesn't knew that these results were not checked, and still submitted them. This also triggered my concern that the paper was produced by AI and then submitted without any check.


Evidence 2. Incorrect dataset attribution

I checked the examples against the public source records and found incorrect dataset labels:

  • user5 and user9 (Tables 1 and 6) are labeled as Summarize users, but the quoted examples come from PRISM (conversation IDs: c40, c13, c23, and c43).
  • RgH765... (Tables 1 and 5) is labeled as a PRISM user, but the examples match Summarize from Human Feedback, where this ID identifies an annotator.
  • thott7Xe... is assigned to PRISM in Table 8, although it is also an annotator ID in Summarize from Human Feedback.

These errors raise serious concerns about example provenance. The authors should provide exact source-record mappings and clarify whether the errors also affect the experimental pipeline.

Sources: PRISM, Summarize from Human Feedback.


Evidence 3. The exhaustive search in Table 8 cannot have been run

Table 8 and lines 1504–1506 say that the authors compare greedy selection with exhaustive search at k = 3, 5, and 10. At k = 10, exhaustive search must check C(42,10) which about 1.47 billion subsets, and the paper gives this number itself (line 1508). Each subset needs two forward passes of a 7B model: one for the preferred response and one for the dispreferred response.

The paper says that all experiments run on a single H100 GPU (line 1123). Even at an optimistic speed of 500 subsets per second, one preference pair would take more than a month.

Questions for authors

Clarification requests (each answer could change my assessment):

  1. Please provide the per-example scores, per-user test counts, and run logs behind Table 2 and what do the ± values in Table 2 measure?
  2. Help me resolve these conflicts - in particular, please explain how the k=10 search in Table 8 was run on one GPU, which of 0.13 and 0.64 is the correct deviation for Qwen on PRISM, and why Table 15 reports 60.75 for EXACT instead.
  3. Was the train/test split done by pair or by prompt? How did you get 0% overlap?
  4. Who were the human raters, and is there an IRB record for the study?

Suggestions for future work: The related work could also discuss these personalization papers:

  • Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. LaMP: When Large Language Models Meet Personalization. ACL 2024. arXiv:2304.11406
  • Zhaoxuan Tan et al. Can Large Language Models Understand Preferences in Personalized Recommendation? 2025. arXiv:2501.13391
  • Zhaoxuan Tan et al. Instant Personalized Large Language Model Adaptation via Hypernetwork. 2025. arXiv:2510.16282
  • Zhaoxuan Tan et al. Aligning Large Language Models with Implicit Preferences from User-Generated Content. ACL 2025. arXiv:2506.04463
  • Michael J. Ryan, Omar Shaikh, Aditri Bhagirath, Daniel Frees, William Held, and Diyi Yang. SynthesizeMe! Inducing Persona-Guided Prompts for Personalized Reward Models in LLMs. ACL 2025. arXiv:2506.05598 - It also builds a text-based user profile from a few preference pairs on PRISM and adds it to the prompt, which is very close to EXACT.
  • Daiwei Chen, Yi Chen, Aniket Rege, Zhi Wang, and Ramya Korlakai Vinayak. PAL: Sample-Efficient Personalized Reward Modeling for Pluralistic Alignment. ICLR 2025. arXiv:2406.08469 - It studies few-shot personalization on the same Reddit (Summarize from Human Feedback) data.
Reviewer_U3eGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes EXACT, a decoding-time personalization method that represents user preferences as prompt-specific subsets of predefined attributes. Given pairwise preference feedback, EXACT greedily selects attributes that maximize the log-likelihood gap between preferred and dispreferred responses. At inference time, it retrieves the attribute subset associated with the most semantically similar historical prompt and injects it into the generation context without updating model parameters. The paper presents theoretical analyses of greedy attribute selection and contextual retrieval, and evaluates preference modeling accuracy and LLM-judged personalized generation on two datasets and three model backbones against existing personalization methods.

Critical strengths

None

Critical weaknesses

The manuscript is not in a reviewable state.

The manuscript contains numerous results explicitly marked by the authors as requiring verification or correction. The Reproducibility Statement also acknowledges that draft tables and cases remain in the submission with unresolved metric or prevenance conflicts. Results that the authors themselves identify as unverified cannot serve as a reliable basis for evaluating them.

The following issues appear in the main text alone:

  1. Table 1 (section 1): These examples appear to form part of the paper's central motivation. However, the caption explicitly states "provenance pending" and that it "require source-record verification before these cases can serve as empirical evidence."
  2. Theorem 3.2 (section 3.2): The authors explicitly label this section "Draft theory verification" and acknowledge that "The variance decomposition and query-noise argument require correction before this theorem can be used as a guarantee for EXACT."
  3. Retrieval vs. non-retrieval comparison (section 4.4): The authors explicitly acknowledge that this comparison "requires metric verification." This is particularly consequential because the comparison directly concerns the benefit of the proposed context-dependent attribute retrieval, a central component of EXACT.
  4. Case study (section 4.5): The authors explicitly state that "Dataset provenance remains unresolved" and that "This illustrates variation in the model's retrieved representation, not independent evidence of a true preference shift." The case study therefore cannot serve as verified evidence for the paper's motivating phenomenon.

Numerous additional results, tables, figures, and case studies in the appendices are also explicitly marked as unverified or requiring reconciliation. I do not enumerate them here.

W2. Unclear advantage over its simplest baseline (RAG few-shot)

In Table 2, EXACT achieves the best accuracy in only 3/6 settings, with performance generally comparable to RAG few-shot (differences mostly within 1-2%p).

Questions for authors

Have all results, theoretical claims, and examples explicitly marked as requiring verification, correction, or reconciliation been fully verified and resolved? If so, please provide the verified results and explain how exactly the identified issues were resolved.

Reviewer_YXvtOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

EXACT learns compact attribute subsets from pairwise feedback and retrieves them to personalize generation without parameter updates. Its contribution is an explicit, context-dependent preference representation. Experiments on two datasets and three model families show competitive preference-ranking and LLM-judged generation results against RAG few-shot.

Critical strengths
  1. I think the paper addresses a meaningful limitation of static user profiles. A user may prefer different response styles across tasks, so retrieving prompt-specific attribute subsets is a reasonable way to account for contextual preference shifts.

  2. The method is straightforward and uses a compact preference representation. EXACT converts historical pairwise feedback into a small set of explicit attributes, then retrieves and injects those attributes at inference time without updating the generation model’s parameters. This makes the personalization signal easy to inspect and potentially convenient to edit.

Critical weaknesses
  1. The attribute-selection objective does not distinguish user preferences from other factors that affect the model’s likelihood scores. For example, adding Formal may increase the log-likelihood gap between the preferred and dispreferred responses, but this does not establish that formality explains the user’s choice. The gain could instead reflect response length, factual quality, or the base model’s own biases. I therefore find the interpretation of selected attributes as faithful descriptions of user preferences insufficiently supported.

  2. The evidence for prompt-conditioned retrieval remains incomplete. The retrieval versus non-retrieval comparison in Table 15 is explicitly marked as requiring metric verification, since the previously reported values may measure improvement in the baseline-anchored likelihood gap rather than strict pairwise ranking accuracy. This is a central ablation: without a verified comparison against a fixed user-level attribute set, it is difficult to assess how much contextual retrieval contributes.

  3. The efficiency analysis is incomplete. Short attribute prompts should reduce online context length relative to RAG few-shot, but offline selection requires many likelihood evaluations. Offline indexing cost, online latency, and token usage should be reported together to establish when this tradeoff pays off.

Questions for authors
  1. How much user-specific information do the default three attributes capture? I would be interested in comparisons using the current user’s memory, another user’s memory, and attributes selected from the test prompt alone, as well as a control with shuffled preference labels. If performance changes little under these controls, the gains may reflect task adaptation rather than personalization.

  2. How stable and faithful are the attributes? Does removing one change generation as expected, and do synonymous names, different attribute orders, or a different generation backbone preserve its effect? Do users or independent annotators agree that the attributes describe preferences supported by the historical comparisons?

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

Related papers

Discussion (0)

Sign in to comment.