acceptodds
Under review as a conference paper at ICLR 2027

Over-Personalization Is a Decision Failure: Generation-Induced Apply Bias in LLMs

Abstract

Personalized LLMs must decide, for each stored preference, whether the current context calls for applying or suppressing it, which we call its applicability. They frequently over-personalize, applying preferences the context rules out, yet existing benchmarks score only the final response and cannot tell where this failure arises. We decompose preference handling into three stages and measure each separately: (1) knowing whether a preference applies, (2) deciding on an explicit Apply/Suppress label, and (3) generating a response consistent with that label. Using linear probes, we first show that this applicability signal remains decodable from hidden states during generation. By making the decision explicit, we then find that in most settings wrong decisions faithfully followed outnumber correct decisions lost in generation. We thus locate the failure in the decision, which breaks once the model is also asked to answer. To determine whether this reflects lost sensitivity or a response bias, we propose ABIDE (Apply-Bias Investigation via Decision-scorE), which adapts signal detection theory to Apply-vs-Suppress decision scores read directly from logits. ABIDE reveals a generation-induced Apply bias: merely stating an answer-generation objective shifts the decision score toward Apply while sensitivity is largely preserved, and the shift persists under controls for prompt structure, cascades across preference slots, and prompt wording. Finally, we show that subtracting a single bias scalar, estimated on a held-out split, from the decision score at decoding time reduces leakage while largely preserving fulfillment.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Over Personalization, Large Language Models, Apply Bias

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 5 reviews, above 27% of the others.
Reviewer_PoAUOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies over-personalization: LLMs applying stored preferences selectively in the current context (each attribute in a set of preference attributes could either be applied or suppressed in the generation). The authors first identify three potential sources of over-personalization errors: lack of knowledge, wrong decision, and wrong generation, and then use linear probes to show that the knowledge is retained, and use an explicit Decide+Answer format to show that most errors are caused by wrong decisions, yet the responses could follow decisions correctly. The authors then identified that coupling the answer with the decision shifts the decision logits toward Apply for all preference attributes (compared to only asking the LLM to make the decision), but the AUC among attributes is well preserved, which the authors call a "bias" in the decision. The authors then proposed two fixes: subtracting a scalar for all attributes (which is estimated on the validation set), or making the decision separately before answering.

Critical strengths
  1. The diagnosis is cleanly organized. It narrows the cause step by step: first among the knowledge, decision, and generation stages, then, within the decision stage, between response bias and loss of sensitivity.
  2. The controls are careful: a neutral two-step prompt, gold-prefix scoring, prompt-wording variants, a shuffled-label probe control, and noise margins fixed in advance.
  3. The two fixes are simple and follow directly from the diagnosis. Subtracting a single bias scalar and separating the decision from generation are both cheap and easy to interpret.
Critical weaknesses
  1. Although the diagnosis is cleanly organized and well controlled, the behavioral pattern it uncovers is rather immediate and contributes limited understanding to the community.

  2. The performance of the "Debiasing" intervention is limited and slightly contradicts the phenomenon of "sensitivity preserved." A pure shift could be undone by one constant, yet debiased Decide+Answer is worse than Direct Decision on both AR and SR: Ministral-8B on BenchPreS goes from 0.87/0.840.87/0.84 to 0.73/0.710.73/0.71, and Gemma from 0.88/0.990.88/0.99 to 0.48/0.950.48/0.95. One possible reason is that the AUC evidence uses gold prefixes, while real decoding conditions on the model's own earlier labels, where errors can cascade. In that case, maybe reporting whether "sensitivity preserved" still holds under pred-prefix, and reporting which prefix is used when estimating lambda, would be helpful.

  3. The experiments are limited: there are two small benchmarks; the fixes aren't compared with simple alternatives such as an instruction to suppress, or the mitigations proposed in the benchmark papers. This raises concerns about whether the phenomenon uncovered and the intervention proposed are generalizable

  4. Presentation. The signal-detection-theory framing adds terminology without adding analysis. The actual measurements are a mean logit shift and an AUC compared between prompt conditions, and they can be stated directly. Section 4 lists all conditions, prefix modes, and metrics before saying what each hypothesis predicts. The verdict appears only in Section 4.3, and it is slightly overstated ("proving," "purely").

Questions for authors

Please see weaknesses above

1 comment

W1.

While the pattern may seem intuitive, to our knowledge no prior work has systematically analyzed why it arises. Our contribution is identifying its source. We localize the failure to the decision stage and show that merely stating a generation objective shifts the decision criterion toward Apply while discriminability is largely preserved. This was not obvious in advance: leakage could also have stemmed from applicability no longer being encoded under the generation prompt, from a loss of discriminability, or from insufficient reasoning, and our experiments rule out each of these.

W2.

λ (ABS_GD^held-out) is estimated with pred-prefix on the held-out split and frozen before decoding, matching ordinary decoding.

Discriminability with pred-prefix. We apply the same decision rule (ΔAUC_GD ≥ −ε_AUC; Appendix E), re-estimating ε_AUC with pred-prefix on the held-out split. Discriminability (sensitivity in SDT) is preserved in 9 of 12 cells:

ΔAUC_GD / ε_AUCBenchPreSRPEval-ExRPEval-Im
Ministral-3-8B−0.081/0.079†−0.099/0.107−0.055/0.093
Ministral-3-14B+0.002/0.046−0.029/0.048−0.006/0.087
Qwen3.5-27B−0.050/0.049†−0.019/0.032+0.011/0.043
Gemma-4-31B-it−0.127/0.088†+0.002/0.016−0.024/0.051

The three exceptions (†) include both cells you cite.

Why debiased D+A falls below Direct Decision. The subtraction is an analytical intervention to test whether the Apply bias drives leakage. It generally raises decision accuracy and lowers PLR (Table 3), linking the bias to leakage. Because λ is the mean shift and we do not assume that every preference shifts by this amount, subtracting it moves the decision threshold rather than restoring Direct Decision exactly.

Role of earlier decisions. In the three exceptions, ΔAUC_GD drops further with pred-prefix than with gold-prefix are where ΔABS (pred − gold) is the largest across datasets for each model (Table 2). Across all cells, however, ΔABS (pred − gold) is positive in six and negative in six, so the effect is not consistent, and with gold-prefix ABS_GD stays positive in all 12 cells. Errors in earlier decisions thus coincide with the exceptions but do not cause the shift.

§5.1 now states that λ is estimated with pred-prefix, and we will add the pred-prefix AUC table to the appendix.

W3.

Benchmarks. To our knowledge, BenchPreS and RPEval are the only benchmarks that label whether each stored preference should be applied or suppressed in the given context, so we use both, covering explicit and implicit preferences (3,557 labeled).

Suppress instruction. Our prompts already include one. Decide+Answer tells the model to ignore the preferences it labeled Suppress (Table 11), yet leakage stays high.

RP-Reasoner. We compare against RP-Reasoner, the RPEval's mitigation (Feng et al., 2026). Like Des2Gen, it decides before generating, but in two reasoning calls (query likelihood and intent prior) rather than a single Direct Decision call. We only changed its output format and mapped Support/Dominate to Apply. Its two-call procedure yields no higher SS than a single Direct Decision call on any dataset, and its responses are no better than Des2Gen’s.

DatasetMethodAR ↑SR ↑SS ↑PFR ↑PLR ↓
BenchPreSRP-Reasoner0.660.800.727.262.93
Des2Gen0.980.880.928.822.18
RPEval-ExRP-Reasoner0.880.700.787.714.56
Des2Gen0.850.850.858.213.53
RPEval-ImRP-Reasoner0.820.630.716.864.35
Des2Gen0.840.860.857.533.74

W4.

SDT framing. In Section 4, we removed the SDT background (the analogy to human decision-makers and the 𝑑'/𝑐 discussion) and now state directly that ABIDE measures discriminability and response bias from the decision score. We replaced "sensitivity" with "discriminability," cut the metric definitions from five equations to two, replaced "cascade amplification" with ΔABS (pred − gold), and moved the hit/false-alarm paragraph to Appendix E.1.

Order of Section 4. Both hypotheses are now stated as predictions about the decision score: a loss of discriminability predicts more overlap between Apply and Suppress scores, and a response bias predicts a shift toward Apply. The captions of Figure 4 and Table 2 now state the result directly.

Overstated verdict. We toned down overstated claims throughout the paper. §4.3 now says the results "indicate" the shift is driven "primarily" by the Apply bias, rather than "proving" it is driven "purely" by it. The introduction and conclusion now say discriminability is "largely preserved" rather than "preserved," and §5 says offsetting the bias "reduces leakage" rather than "repairs the behavior."

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

This paper investigates over-personalization in LLMs by decomposing preference handling into knowledge, decision, and generation stages. Linear probing shows that applicability information remains decodable during generation (AUROC 0.8–1.0). The authors find that decision errors dominate generation errors in 11 of 12 model–dataset settings, suggesting that incorrect applicability decisions are a major source of preference leakage. They further introduce ABIDE to characterize generation-induced Apply bias and demonstrate that logit debiasing improves decision accuracy by 6.9 percentage points and reduces the average preference leakage score from 5.28 to 4.63 (on a 1–10 scale), with additional improvements from separating decision and generation stages.

Critical strengths

S1. Explicit Analysis of Decision–Generation Relationships.
The paper explicitly formulates preference applicability as a decision problem and investigates how applicability decisions relate to downstream generation, providing a useful perspective on over-personalization beyond output-level evaluation.

S2. Systematic Decomposition of Failure Stages.

The distinction between knowledge, decision, and generation provides a clear analytical framework. The combination of linear probing and explicit decision analysis helps characterize where personalization failures arise.

S3. Controlled Investigation of Generation-Induced Bias.

The proposed ABIDE framework distinguishes changes in decision sensitivity from response bias. The use of neutral objectives, prefix controls, and prompt ablations, together with evaluations across multiple models and datasets, strengthens the empirical analysis.

S4. Practical Interventions for Reducing Over-Personalization.

The paper explores both logit-level calibration and decision–generation decoupling, demonstrating that modifying explicit applicability decisions can reduce downstream preference leakage.

Critical weaknesses

W1. Ambiguous Distinction Between Knowledge and Decision.
The distinction between knowledge and decision is not sufficiently justified. The paper defines knowledge through the linear decodability of applicability labels from hidden states, while decision is measured through explicit Apply/Suppress predictions. However, decodable information does not necessarily imply that the model can access or utilize such information during decision-making. Thus, the discrepancy between probing accuracy and decision behavior may reflect differences in measurement rather than distinct processing stages. Further justification would strengthen the claim that over-personalization primarily stems from decision failures rather than knowledge failures.

W2. Structural Asymmetry in Error Decomposition.
The error decomposition in Section 3.3 is structurally asymmetric, as generation is explicitly conditioned on the preceding decision. Incorrect decisions propagated into responses are counted as decision errors, whereas generation errors are identified when correct decisions are not followed. Consequently, the dominance of decision errors may partly reflect the experimental design rather than establish decision formation as the primary failure source. This concern is particularly relevant when generalizing the findings to direct generation, where explicit decision labels are absent.

W3. Limited Generalizability and Baseline Comparisons of Logit Calibration.
The proposed debiasing method estimates a model- and dataset-specific scalar from validation data and applies it uniformly across preferences, leaving its effectiveness on unseen contexts and preference distributions unclear. Furthermore, subtracting a constant scalar is equivalent to adjusting the decision threshold, and the observed improvements may reflect a general calibration effect rather than correction of the underlying decision process. Comparisons with simpler alternatives, such as few-shot applicability guidance or explicit contextual decision verification, would help clarify the practical advantages of the proposed intervention.

W4. Limited Analysis of Preference Compositionality.
The paper does not systematically examine how the number and composition of user preferences affect decision and generation failures. In particular, interactions among multiple preferences, conflicting applicability requirements, and increasing preference complexity may introduce different challenges for decision-making and downstream generation. Although Section 4.3 investigates sequential cascade effects, it does not establish whether the dominance of decision errors persists as preference complexity increases. An analysis of these factors would help clarify the scope and robustness of the paper's central claim.

Questions for authors

Q1. Decision–Generation Relationship.
How can the authors ensure that introducing an explicit decision stage does not alter the failure mechanism observed in direct generation?

Q2. Knowledge and Decision.
What evidence shows that the applicability information identified through probing is actually used in decision-making? How might the choice of probing token positions affect this conclusion?

Q3. Source of Decision Failures.
Do decision failures arise from long-term memory retrieval or forgetting, or do they persist even when preferences are explicitly provided in the current input?

Q4. Alternative Explanations for Decision Failures.
How do the authors distinguish generation-induced Apply bias from other factors, such as instruction-following tendencies or conflicts between stored preferences and the current task? Could these factors contribute to the observed decision errors independently of the proposed bias?

1 comment

W1 & Q2

Knowledge is whether the model encodes a preference's applicability; decision is how it turns this into an Apply or Suppress label. The paper does not infer from decodability that the decision accesses this information; this is tested on the decision score itself (§4.3).

Under Direct Decision, the decision separates the two classes well (SS 0.70–0.96, Table 1). If the generation objective blocked access to the information, the decision score would no longer separate them, but it still does. Its AUC changes from Direct Decision by less than the noise margin in 11 of 12 cells (the other exceeds it by 0.001), while the score shifts toward Apply in all 12 (Table 2). The information thus still reaches the decision; the errors come from an Apply bias, not from failed access.

The probe has a narrower role: it only tests whether applicability is still encoded under Direct Generation, and we do not claim the decision reads the probe direction. The evidence that the information is used comes from the decision score.

Probing position. Each input contains several preferences, so we probe the last token of each preference (each prior conversation in RPEval-Im), where the model has read the query and that preference. The final input token mixes all preferences. The probe only needs to show that the information is present, so decodability at this position is sufficient, and weaker decoding elsewhere would not show its absence.

We will revise §2.2, §3.2, and App. D so that knowledge denotes only the presence of the information.

W2 & Q1

Conditioning the response on the decision in Decide+Answer (D+A) is intended. §2.2 defines the generation stage as carrying the decision into the response, and the prompt instructs the model to respond according to its Step 1 decision. A wrong decision faithfully reflected in the response means the generation stage worked as intended, so it is correctly counted as a decision error. Moreover, on the same inputs, the D+A Step 1 decisions are worse than those under Direct Decision, with SR dropping in 10 of the 12 cells (e.g., Ministral-3-8B on BenchPreS, 0.84 → 0.19). Because Step 1 is generated before the response, this degradation is a failure of the decision itself and is unaffected by the conditioning structure.

Q1. We clarify that the explicit decision stage does change behavior, since Decide+Answer lowers PLR relative to Direct Generation in every cell. We therefore do not claim that the two share an identical mechanism. Nevertheless, by making the decision observable, D+A answers a question that Direct Generation cannot: whether the generation objective changes the decision itself.

W3

The score subtraction in §5.1 is an analytical intervention, not a proposed method. It tests whether the diagnosed Apply bias causally drives leakage, so we do not claim that λ transfers across datasets or prompts. λ is the mean D–G score shift measured on a held-out split disjoint from the evaluation set.

We clarify that subtracting λ is equivalent to shifting the decision threshold. λ is not tuned for leakage or accuracy. Subtracting it raises SR and lowers PLR in 11 of 12 cells, consistent with the Apply bias contributing to leakage.

Explicit contextual decision verification corresponds to Des2Gen (§5.2), reported with the debiased decision in Table 3. Des2Gen makes the decision in a separate call without the generation objective, so it requires no λ.

W4

Our experiments already cover these factors. Every instance contains multiple preferences, with a number and an Apply/Suppress composition that vary across instances (Table 4), so the model decides among preferences with conflicting applicability. Preferences are also given at two levels of complexity: stated explicitly (BenchPreS, RPEval-Ex) or left implicit in prior conversations (RPEval-Im).

Q3

Decision failures persist even when preferences are explicitly provided in the current input. Our setup involves no long-term memory retrieval: all preferences are given directly in the prompt, as explicit statements or in-context prior conversations (§2.1), so retrieval or forgetting cannot account for the failures.

Q4

Preference–task conflict is the definition of a Suppress label (§2.1), not a separate factor. The same conflicts are present in every condition, and models handle them correctly when asked only to decide (Table 1). Errors increase only when the generation objective is added.

To rule out instruction-following, we compare against neutral second steps: a fixed sentence (N) and an arithmetic task (N2). These cause little or no Apply shift, while the generation objective causes a clear shift in every setting (Table 2, Fig. 4, §4.4).

Reviewer_DMQ7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates over-personalization in personalized large language models (LLMs), focusing on why models sometimes apply stored user preferences even when the current context calls for suppressing them. The authors decompose preference handling into three stages: applicability knowledge, explicit decision-making, and response generation. Experiments on BenchPreS and RPEval, covering both explicit and implicit preferences, show that applicability information remains linearly decodable from hidden states during generation, while errors made in Decide+Answer settings are predominantly incorrect decisions that are subsequently followed faithfully.

To investigate the underlying mechanism, the authors introduce ABIDE, a logit-based framework inspired by signal detection theory. They find that adding an answer-generation objective shifts decision scores toward Apply, while AUC-based discriminability remains largely preserved. The effect is examined using neutral multi-step controls, gold-prefix conditions, and prompt ablations. Finally, the authors demonstrate two mitigation strategies: subtracting a validation-calibrated bias scalar from the decision score and decoupling decision-making from response generation. The results suggest that correcting the decision stage can reduce preference leakage, although the extent to which fulfillment is preserved varies across settings.

Critical strengths

(1) The paper goes beyond measuring whether a final response leaks a preference and investigates whether the failure originates in applicability knowledge, decision-making, or generation. The comparison among Direct Decision, Direct Generation, and Decide+Answer conditions is particularly informative. The error decomposition further shows that incorrect decisions faithfully followed by generation are a dominant failure mode in most evaluated settings. This distinction is consequential because it suggests that interventions should target decision formation rather than assume that the model simply lacks the relevant contextual information or cannot generate an appropriate response.

(2) ABIDE is a useful methodological contribution. The neutral two-step condition helps distinguish the effect of an answer-generation objective from the mere presence of a multi-step prompt, while gold-prefix experiments help isolate local effects from sequential error propagation. The authors also examine alternative prompt formulations and test whether score correction improves downstream behavior. These controls make the central finding more convincing than a simple comparison of Apply/Suppress label frequencies. The debiasing and Des2Gen experiments further connect the proposed mechanism to potential practical interventions.

Critical weaknesses

(1) The linear-probing experiments establish that applicability information is recoverable from hidden states, but decodability by an external probe does not necessarily demonstrate that the model's own decision process can reliably access or use that information. This distinction is particularly relevant because Appendix D reports no consistent monotonic relationship between probe confidence and behavioral accuracy.

Similarly, unchanged AUC indicates that rank-based discrimination is largely preserved, but it does not rule out all changes in the class-conditional score distributions, including changes in their shapes or variances. Therefore, the conclusion that the failure is driven purely by response bias, rather than any change in sensitivity-related properties, may be stronger than the current evidence warrants. Additional distribution-level analyses or causal interventions on the relevant representation would help establish the proposed mechanism more firmly.

(2) GPT-5.4 is used for preference decomposition, response-level scoring, reasoning-trace extraction, and translation of RPEval. Although this is a practical evaluation pipeline, the paper does not report a human-annotated validation set, agreement with an independent judge, or another independent reliability assessment. Potential systematic errors in the evaluator could affect the reported PFR, PLR, and reasoning-based error decomposition.

Furthermore, Table 3 primarily reports point estimates for the debiasing results, without corresponding uncertainty intervals for the behavioral changes. The trade-off is not negligible in every setting: for example, on BenchPreS, Qwen3.5-27B's PFR decreases from 8.59 to 7.84 after debiasing, while Gemma-4-31B-it's PFR decreases from 6.66 to 5.93. These results do not invalidate the intervention, but they make the claim that fulfillment is largely preserved less uniformly supported than the overall narrative suggests.

Finally, the correction uses a model- and dataset-specific calibration scalar, and the prompt ablations focus on Ministral-3-14B-Instruct. It remains unclear how well the correction transfers to unseen datasets, prompt distributions, or model versions. This limits the strength of the practical generalization claims.

Questions for authors
  1. Could the authors validate PFR and PLR using a manually annotated subset or an independent judge, and report agreement with GPT-5.4? Could uncertainty intervals or paired significance tests also be provided for the behavioral improvements in Table 3? This would help establish whether the observed reduction in leakage is robust to evaluator choice and whether the fulfillment trade-offs are statistically meaningful.

  2. The paper interprets largely unchanged AUC as evidence that sensitivity is preserved. Could the authors provide additional analyses of the Apply and Suppress score distributions, including their variance and shape, to clarify what properties remain unchanged? Given the probing results in Appendix D, is there evidence beyond linear decodability that the model's own decision process can use the applicability information?

  3. How well does a bias scalar estimated on one dataset or prompt distribution transfer to another? Since the current procedure is model- and dataset-specific, could the authors evaluate cross-dataset or cross-prompt calibration using a fixed scalar estimated on a separate calibration set? Such an experiment would clarify the practical applicability of the proposed correction.

1 comment

W1.1 & Q2.2

The model’s own decisions, rather than the probe, show whether its decision process can use this information. Under Direct Decision, the model judges applicability correctly in most cases (Table 1), which could not happen without access to the information. Under the generation objective, AUC is largely unchanged while scores shift toward Apply (Table 2, Fig. 4; see W1.2): the model still accesses the information, and the errors come from the generation-induced Apply bias.

The inconsistent relationship between probe confidence and accuracy in App. D, with Suppress-gold preferences leaking even at high confidence (Fig. 7), is consistent with the information being present but not reflected in responses (Feng et al., 2025). We therefore use the probe only as evidence that the information is present even under Direct Generation, where no judgment is requested, and have revised §3.2 and App. D accordingly.

W1.2 & Q2.1

We clarify that an unchanged AUC shows only that rank-based discrimination is preserved, not that the variance or shape of the class-conditional distributions is unchanged; since the D and G prompts differ, some change is expected. Our claim is that, under the generation objective, the separation between Apply and Suppress is preserved while the scores of both classes shift toward Apply. Beyond ABS being significantly positive in every cell (Table 2), we compared class-wise shifts and G/D variance ratios under the gold-prefix setting:

RPEval (8 cells)BenchPreS (4 cells)
ΔAUC within εAUC\varepsilon_{\mathrm{AUC}}8/83/4 (other: −0.018 vs. 0.017)
Variance ratio G/DApply 0.95–1.23, Suppress 0.98–1.16Apply 0.83–1.04, Suppress 0.74–1.41
Shift gap (A−S)No consistent direction (−2.12 to +0.54)Suppress larger in all cells (−2.16 to −0.84)

On RPEval, variance ratios are near 1 and the shift gap has no consistent direction: the spread is preserved, so the model still distinguishes applicable preferences, and mainly the decision criterion changes. On BenchPreS, shifts and variance changes are larger for Suppress, yet ΔAUC stays within the margin in three of four models, and the fourth exceeds it only marginally.

This larger Suppress-side shift appears to stem from preference type rather than class itself. BenchPreS labels are fixed by type (style and persona always Suppress; bold formatting, factuality, and conciseness always Apply). Within Apply, bold formatting shifted as much as Suppress style preferences such as emoji and persona (+1.5 to +8.3 nats), while content-oriented ones such as factuality barely moved (−2.5 to +1.1). The class-level gap thus reflects each class’s type composition; RPEval, whose labels depend on relevance to the request, shows none. (Types were assigned post hoc by keyword.)

W2.1 & Q1

The diagnosis in §4 (ABS and ΔAUC) is read from logits and involves no judge. Of the GPT-5.4 uses you list, translation is not an evaluation: RPEval was translated once, manually corrected (App. A), and shared by all conditions. For PFR and PLR, we adopt the checklist-based judge of CUPID (Kim et al., 2025) rather than designing a new one, using both of its prompts verbatim: checklist decomposition and judging.

Paired 95% CIs for Table 3 (debiased vs. plain Decide+Answer):

ΔPFR ↑ (top) / ΔPLR ↓ (bottom)BenchPreSRPEval-ExRPEval-Im
Ministral-3-8B+0.62 [0.46, 0.78]+0.30 [0.05, 0.59]+0.07 [−0.17, 0.35]
−1.90 [−2.19, −1.65]−0.43 [−0.73, −0.14]−0.44 [−0.74, −0.17]
Ministral-3-14B+0.11 [0.003, 0.22]+0.27 [0.00, 0.60]+0.07 [−0.12, 0.33]
−1.67 [−1.98, −1.39]−0.33 [−0.57, −0.11]−0.48 [−0.79, −0.18]
Qwen3.5-27B−0.75 [−1.00, −0.52]0.00 [−0.07, 0.07]−0.05 [−0.20, 0.07]
−1.25 [−1.49, −1.03]−0.34 [−0.58, −0.14]−0.15 [−0.27, −0.05]
Gemma-4-31B-it−0.73 [−0.96, −0.52]0.00 [0.00, 0.00]+0.13 [0.01, 0.28]
−0.69 [−0.94, −0.46]0.00 [0.00, 0.00]−0.10 [−0.21, 0.00]

PLR decreases in every cell where the offset changes decisions (11 of 12), with a CI excluding zero in 10. PFR drops significantly only in two BenchPreS cells, where the offset also lowers AR.

W2.2 & Q3

The score subtraction in §5.1 is an analytical intervention, not a general-purpose method: it tests whether the diagnosed Apply bias causally drives leakage, so we do not claim cross-dataset or cross-prompt transferability of λ. λ is the mean shift (ABSGD\mathrm{ABS}^{\mathrm{GD}}), estimated on the held-out validation split and frozen before the main analysis; it captures only the direction and average magnitude of the bias. Even this coarse correction, which leaves representations untouched and only moves the decision cutoff, raises SR and reduces PLR in all 11 cells where decisions change, supporting our claim that leakage stems from an Apply-directed response bias rather than a loss of sensitivity.

Reviewer_9XitOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper tackles the problem of over personalization, where a LLM incorporates irrelevant information about a user when responding to a given user query. To better understand over personalization failures, the process of personalized response generation is broken down into three stages (or capabilities): knowledge, decision, and generation. The experimental protocol is then designed to understand at which stage over personalization is occurring. The results indicate that responding to a query changes how a LLM acts on the provided user information by incorporating information it labels as irrelevant in other probing/experimental settings. The paper presents evidence this behavior change can be mitigated.

Critical strengths
  • The paper tackles an important and interesting problem.
  • The experiments seem to be well setup.
Critical weaknesses
  1. Overall the paper writing and structure needs to be cleaned up. The paper seems like it is addressing an important problem, however the way the paper is written leaves the reader to do a lot of work to understand this. The core issue here (more details below), is the overuse of "technical" terms that are used before they are defined and frequently are not necessary. 1.1) The flow of the paper assumes the reader has knowledge of context, framing, or definitions that are introduced either in cited papers or later in the paper. This makes the paper difficult to follow. For example, in the introduction the reader must know the cited papers address problems related to persistent memory for personalization and then must assume this is the problem domain the current paper is tackling. Instead this framing should be explicitly provided to the reader from the start. Beyond identifying persistent memory as the problem domain, the introduction should make clear the types of persistent memory that is the focus (e.g. episodic, rule-based, etc). This lack of framing then further exacerbates understanding what specifically is meant by phrases such as "correct label", "wrong decisions", "correct decisions", "answer-generation objective", "pred-prefixes", and "gold-prefixes". Not providing sufficient context about the problem domain makes it challenging to identify what the paper's contributions are. The abstract and title suffer heavily from this problem. You need to have already read the paper to understand either one. For example, in the introduction, the portion about wrong and correct decisions could be summarized as, "LLMs struggle more to correctly identify which pieces of information are relevant to a given query than to incorporate relevant information into the response", where information relevance does not require knowledge of the experimental set up to understand what counts as a decision. Overloading motivation and findings with details specific to the experimental set up hides the key take away and what the paper is trying to do. 1.2) A number of specific measures for this task are introduced and it is easy to lose track of what each one is measuring and what it tells us about what is happening. Some additional motivation and explanation would be beneficial. Part of the challenge is tracking back to find where variables (e.g. s_i) were defined. References or bolding should be added to make it easier for the reader to track these down. Some of the named measures are probably not necessary. For example, "Cascade Amplification" does not need to be introduced as a named measure. You can just describe what you are measuring. Also in this part of the paper, "Structural Isolation from Sequential Cascade" does not have a lot meaning as a results section title as "Sequential Cascade" is not yet defined.
  2. The distinction between knowledge, decision, and generation is not super clear. The boundaries between one and the other (especially knowledge and decision) is not well articulated. The results seem to suggest the LLM does know consider relevance between generating a response, but has the ability to do so when prompted. The lack of clear boundaries makes it difficult to understand what the linear probe result tells us. We know from Section 3.1 the LLM can generate the correct g_i label, so it is not surprising the linear probe has good AU-ROC. Why should we be surprised by the linear probe results? How is it distinct from the "Direct Decision" result?
Questions for authors
  • How do different forms of the user preferences impact performance?
  • What is the difference between knowledge and decision? Why are these different and distinct in the LLM? What do we learn about model behavior by examining both? Some additional motivation in the paper about why these two should be measured separately would be beneficial.
  • What are some examples of the implicit and explicit preferences? What do they look like after they are decomposed?
  • What is an example of a response's Preference Fulfillment Rate score and Preference Leakage Rate?
  • Why judge responses per apply group and suppress group? Why not score per preference?
  • What is the quality of the decomposer and the judge used to assess preference leakage in responses?
  • Please provide more details about how the linear probe is trained. For example, what is meant by "prefill hidden state"? What are "Direction Decision inputs" and what are "Direct Generation inputs"? These details need to be included in the main body.
  • Please elaborate on how ABS measures the bias towards apply. The components of the ABS metrics seem to rely on apply and suppress recall by taking the mean of the two, however when one goes up the other goes down. So how does this metric specifically show a shift towards apply? Since this is a binary task, why is the δs(k1,k2,p)\delta_s^{(k_1, k_2, p)} sufficient?
2 comments

W1

We thank the reviewer and revised the paper throughout. We have revised the paper throughout. One clarification first: our claim is not that models struggle to identify relevant information. When asked only to decide, models judge applicability well, but when asked to write a response, they tend to apply even preferences that should be suppressed (Table 1; Figure 1).

1.1 Framing and terms. The abstract and introduction now open with the problem setting (§1, para. 1). "LLM assistants keep a persistent memory of user preferences, such as a nickname or a preferred tone, which are given either explicitly or implicitly through past conversations. For each new query, the assistant must decide whether to apply or suppress each preference." Opening this way states the problem domain and the kind of memory we study up front, so readers no longer need to infer them from the cited papers. We then define the three stages of knowledge, decision, and generation in plain language, and redesigned Figure 1 to illustrate each stage.

Each term the reviewer listed is now written out in plain language at its first use. The generation objective is introduced as the instruction to write a response. Pred-prefix and gold-prefix are described by whether the earlier decisions in the context are the model’s own or the gold labels. We applied the same treatment beyond the listed terms; for example, cascades across preference slots are now described as errors in the model’s earlier decisions (§1, para. 4; §4.3).

1.2 Measures. We removed equations and symbols that are not needed to follow the analysis, so the main text now keeps only two equations: the decision score (Eq. 1) and ABS (Eq. 2). The decision score is also explained in words (Abstract; §1, para. 4) and marked at its definition (§4.1), and later uses refer back to Eq. 1 and Eq. 2, which makes variables easier to trace. Each metric is introduced with what it measures and which direction is better (§2.3, §4.1). In the same spirit, we removed the named measures “Cascade Amplification” and λ and now describe these quantities directly (§4.3, Table 2, §5.1). Finally, we renamed section titles so that they no longer rely on terms defined later. The title the reviewer pointed out is now “Controlling for the Model’s Earlier Decisions,” and §4.3 is now “Ruling Out Loss of Discriminability and Errors in Earlier Decisions.”

W2 & Q2

Knowledge is whether the model’s hidden states encode if a preference applies in the current context. Decision is the Apply/Suppress choice the model makes from that information, which determines whether the preference appears in the response. The two can diverge: a model may encode that a preference should be suppressed and still decide to apply it. A failure at either stage yields the same leaked preference, so the response alone cannot reveal which one occurred, yet the two call for different fixes (intervening on representations vs. correcting the decision).

The reviewer’s reading of §3.1, that the model does not consider applicability when only asked to respond but can when prompted, would be a knowledge failure; the probe tests this. We find that applicability is also decodable from Direct Generation inputs, where no judgment is requested, along the same direction learned under Direct Decision. The model therefore still encodes applicability while responding, so the leak must arise downstream.

We have made this distinction and its motivation explicit in §2.2, §3.2, and Figure 1.

Q1

The form of a preference affects only the difficulty of the task, not the mechanism. Implicit preferences (RPEval-Im), which must first be inferred from prior conversations, are harder to judge (Direct Decision SS 0.71–0.78 vs. 0.80–0.93 for explicit; Table 1), less linearly decodable by the probes, and reflected less in responses (lower PFR and PLR under Direct Generation). Nevertheless, the generation-induced Apply bias appears regardless of the form (Table 2).

Q3 & Q4

Each preference is a single sentence labeled Apply or Suppress, depending on whether the response to the current query should reflect it. The Explicit setting gives the model this sentence verbatim; the Implicit setting gives only the prior conversation that reveals it, and the sentence is used only for evaluation. The decomposer breaks this sentence into yes/no checklist items, and this step is identical in both settings.

Example (RPEval; gold label: Apply).

Request: “I want to take my son to a museum this weekend. Are there any exhibitions suitable for children and any recommended transportation routes?”

Explicit: “The user is full of curiosity about the mysteries of the natural world and likes to actively explore and learn scientific knowledge in everyday life.”

Implicit: “Whenever I come across some curious phenomenon, I like looking up the principle behind it… science is everywhere in life.”

Since RPEval-Im converts RPEval-Ex preferences into conversations, both examples come from the same preference and share the checklist below.

Decomposed: Does the response engage the user’s curiosity about the natural world? / include scientific knowledge? / encourage active exploration rather than passive reception? / connect science to everyday life?

Scoring is per instance: the judge receives all Apply preferences, each with its checklist, and assigns the response a PFR (1–10, higher is better); the Suppress preferences are used the same way for a PLR (1–10, lower is better). A response that satisfies every Apply checklist item and reflects none of the Suppress preferences receives a PFR of 10 and a PLR of 1.

Q5

Our scoring adapts the checklist-based judge of CUPID (Kim et al., 2025). CUPID scores a response against a single target preference, whereas each of our instances contains multiple preferences with opposite expected outcomes (Apply or Suppress). Reflecting an Apply preference counts as success, but reflecting a Suppress preference counts as failure, so a single score would mix the two. We therefore score PFR and PLR separately. We use a graded score rather than a per-preference binary judgment because leakage can appear partially, which a binary judgment cannot capture.

Q6

GPT-5.4 serves as the checklist decomposer and the PFR/PLR judge, following the evaluation protocol of CUPID (COLM 2025), which uses a closed-source LLM (GPT-4o) in the same two roles. We reuse CUPID's prompts, which CUPID adapted from Cook et al. (2024). PLR applies the PFR procedure to the Suppress group; the judge still checks whether each checklist item is reflected in the response, so the PLR judge performs the same task as CUPID's judge.

Q7

By “prefill hidden state,” we meant the hidden state computed when the model encodes the input prompt, before it generates any token. Direct Decision inputs are prompts that ask the model only to label each preference Apply or Suppress. Direct Generation inputs are prompts that ask only for a response to the query. Both prompts contain the same query and preferences. At each layer, we extract the hidden state at the last token of each preference (each prior conversation in RPEval-Im). We then train a [logistic regression] probe to predict its gold label, using [K]-fold cross-validation grouped by instance and preference text.

In the revision, §3.2 now describes this setup in the main text using the wording above instead of “prefill hidden state” and “inputs,” and states the purpose of each view (within-dec, within-gen, cross; also in the Figure 2 caption), the group-held-out splits, and the shuffled-label control against memorization (details in Appendix D).

Q8

ABS is not an average of AR and SR. It is based on the continuous decision score read from the logits, s_i = log p(Apply) − log p(Suppress). ABS averages the shift in the mean score of gold-Apply preferences and that of gold-Suppress preferences, so it measures how far the two distributions move toward Apply together. Under an Apply bias, the two shifts therefore do not cancel but are both positive. Under greedy decoding, the label is Apply when s_i > 0, so when all scores rise together, some preferences previously labeled Suppress cross zero and flip to Apply.

Reviewer_C53MOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2(+1)
Read review
Paper summary

This paper studies why LLMs still include user preferences when they shouldn't be applicable. It explores when this happens, isolating the cause to be a decoding bias introduced by the need to answer a user question. Based on this observation, the authors propose a decoding-time debiasing solution to adjust the decoding scores for LLM responses to no longer misjudge the applicability of user preferences.

Critical strengths

The main strength is the study on the root cause of user preference applicability in the LLM inference pipeline. Additionally, the finding that there exists a logit-level bias introduced by the response could plausibly lead to a significant scientific insight. Although the current level of diagnosis is sound, it is regrettably incomplete, stopping at the observation level instead of digging into the mechanistic explanation. A careful major revision should focus genuinely and primarily on studying the why behind the numerical bias itself. Instead of staying at this level and proposing a suboptimal solution based on surface-level insight, this work could become a significant diagnostic study that may or may not need to include a more fundamental model-level solution inspired by deeper mechanistic finding.

Critical weaknesses
  1. This paper's main issue is its writing. In its current state, this work is far from ready for publication and requires a careful pen-and-paper rewrite.

    1. The most stark weakness is the repeated, unnecessary definition of new terms, concepts, metrics, phrases, equations, methodologies, technical phenomena, without sufficient explanation, making it difficult to comprehend this work which could otherwise be simply and clearly explained. Littered across the text are undefined and under-discussed terms/concepts like "selectivity," "sensitivity," "applicability," "linear probes," "leakage," "Knowledge," "Decision," "Generation," "Apply Recall," "Suppress Recall," "Selectivity Score," "decomposer," "Preference Fulfillment Rate," "Preference Leakage Rate, "contrast," "cascade," "simplified generation objective," "arithmetic neural objective," among countless others. Out of dozens, at most five terms should remain in the revised manuscript. Of the five equations, only one (either Eq. 4 or the unlabeled Eq. 5) shall be kept. Similarly, do not retain the lettering such as variants of g and r whenever they can be simply replaced by "label" and "score", etc. When an author feels encouraged to come up with a new term, they must severely question its necessity and potential harm to readability. For example, if you could unearth a mechanistic explanation finding mentioned above, the work would in turn require complete mathematical proofs.

    b. In contrast to the multitude of unfamiliar terms and phrases, the manuscript is devoid of detailed analysis or unfolding of an argument or complex ideas. It resembles an LLM-facing memory document rather than a scientific study.

    c. Similarly, the figures (except for Figure 1) and tables do not independently convey findings or conclusions.

    d. Many paragraphs retain LLM-response-like phrasing and cadence that make them unnatural to follow and understand.

  2. There is also substantial concern as to the soundness of the solution, pending author clarification. First, the proposed solution of decoding-time score subtraction seems practically intractable. It requires a model-specific, pre-computed bias score that must be calculated on a particular set of user preferences and queries. At inference time, a user using the same model and having a similar distribution of the user preferences and queries, must have the decoding-time scalar subtraction solution applied in their inference pipeline solely for particular preferences. This seems impractical in typical LLM generation settings. Furthermore, although I cannot understand the broader role of the LLM judge/decomposer (GPT-5.4) based on the current manuscript, I highly doubt that an under-probed, closed-source model should serve as the sole judge of applicability.

Questions for authors
  1. How would a solution based on decoding-time scalar subtraction help in one-step generation? When a user just wants a one-step answer to one question, how would your proposed solution identify each of the user's stored preferences and apply the bias subtraction?
  2. Why might the need to respond introduce the bias in applicability? What are some mechanistic reasons causing this phenomenon?
  3. Which scores in the experiment results are derived from logits? Which are from natural language answers? Which are from the LLM judge? There are a lot of metrics and results, but it's unclear where they ultimately stem from.
3 comments

W1

We thank the reviewer for the detailed feedback and have revised the manuscript as follows.

Terms. Every term the reviewer listed is now removed, merged, or defined in plain language at first use. Beyond that list, we merged every group of near-synonyms in the manuscript into one term and replaced coined phrases with plain descriptions:

SubmittedRevised
selectivitypreference handling
sensitivity, discriminative ability, discrimination capabilitydiscriminability
decision variable, decision criterion, cutoffdecision score
simplified generation objective, arithmetic neutral objectivereworded Decide+Answer (G2), arithmetic Neutral 2-step (N2)
cascade, error propagation across preference slots, Cascade Amplificationerrors in the model’s earlier decisions, ΔABS (pred − gold)
contrast, generation/structural contrast, Step 2 presence effect“compare a condition X with Direct Decision,” “multi-step structure”
decomposer“GPT-5.4 decomposes each preference into checklist items”
base floor, rerun floor, paired precision floor, default margin, equivalence marginnoise margin

The listed terms that remain are a standard method (linear probes), ordinary words used in their usual sense (applicability, knowledge, decision, generation, leakage), and evaluation metrics (AR, SR, SS, PFR, PLR). Each is defined in plain language at first use, as are terms the reviewer did not list, such as generation objective, response bias, within-dec/within-gen/cross, and pred-/gold-prefix.

Symbols. We removed the condition and prefix indices (k, k₁k₂, p) and all quantities indexed by them, as well as λ, s_G, s̃_G, CascadeAmp, the selectivity axis w, and the noise-floor symbols η_m. Conditions now appear only as subscripts (e.g., ABS_GD), and prefix modes are stated in words.

Equations. Displayed equations went from five to two: the decision score (Eq. 1), from which every ABIDE quantity is computed, and a single ABS equation (Eq. 2) replacing submitted Eqs. 1–2. AUC (submitted Eq. 3) now uses the standard definition with a citation, ΔAUC (Eq. 4) and the debiased score are stated inline, and the appendix noise-floor equation is replaced by a noise margin defined in words (Appendix E).

Unfolding of the argument. Each section now opens by stating which explanation it tests and what it rules out, and each result is followed by a sentence stating what it implies for the argument. For example, §3.2 opens by asking whether applicability is still encoded once the model is asked to respond, concludes from the probes that it is, ruling out the knowledge stage, and passes the remaining two stages to §3.3.

Figures and tables. We revised the captions so that each states the setup, and the key result is pointed to in the text.

W2.1 & Q1

As stated in §5.1, our decoding-time score subtraction is not a practical, general-purpose solution but an analytical intervention to test whether the diagnosed Apply bias causally drives the observed leakage. That said, its requirements are lighter than the review describes. λ (= ABS_GD) is a single scalar per model and dataset, estimated once on the held-out split and frozen; it is not computed for each user’s preferences. It is subtracted uniformly from the decision score at every decision slot, so no particular preference needs to be identified. It does require an explicit decision step, so it does not directly apply to plain one-step generation. Subtracting λ alone raises SR and decision accuracy in nearly all model-dataset pairs and lowers PLR in the final responses (Table 3), consistent with a causal role of the generation-induced Apply bias in over-personalization.

W2.2

We clarify that GPT-5.4 does not judge applicability. The gold Apply/Suppress labels come from the BenchPreS and RPEval annotations, and the probing and ABIDE analyses (§3.2, §4) use these labels without any LLM judge. GPT-5.4 serves as the checklist decomposer and the PFR/PLR judge, following the evaluation protocol of CUPID (Kim et al., 2025; COLM 2025), which uses a closed-source LLM (GPT-4o) in the same two roles. We reuse CUPID’s prompts, which CUPID adapted from Cook et al. (2024), and apply the same template to Suppress preferences for PLR.

S & Q2

We conjecture that this bias is a byproduct of preference-based post-training. Human preference data favor responses that agree with the user’s stated preferences and beliefs (Sharma et al., 2024), and optimizing against such rewards amplifies user-endorsing behavior (Shapira et al., 2026). A stored preference is a strong user signal, and applying it is the endorsing move. An answer objective may therefore invoke this reward-shaped response mode and pull the criterion toward Apply without affecting discriminability. We added this discussion to Related Work (§6).

Q3

The source of each reported score is as follows:

  • Logits. The decision score s_i = log p(Apply) − log p(Suppress) is read from raw logits at the Step 1 decision position, with Step 1 teacher-forced and Step 2 never generated. ABS and ΔAUC are computed from this score, as is the difference in ABS_GD between pred- and gold-prefix (formerly CascadeAmp). The debiased Step 1 decisions in Table 3 use the same score during ordinary decoding, minus ABS_GD from the held-out split (formerly λ).
  • Hidden states. Linear probes on the hidden state at the last token of each preference yield the probe AUROC.
  • Natural-language answers. For each preference, we parse the Apply/Suppress label the model outputs in Direct Decision and in Step 1 of Decide+Answer. AR, SR, and SS are computed from these labels against the gold labels. In Fig. 3 (Decide+Answer), whether each Step 1 label matches the gold label determines decision correctness (d✓/d✗).
  • LLM judge (GPT-5.4). Checklist-based 1–10 scoring yields PFR and PLR. Label extraction from reasoning traces and responses yields the generation side (g) of Fig. 3 and the decision side (d) for Latent Decide+Answer.

Thank you for the detailed response and manuscript revisions. I appreciate the authors' efforts to clarify the experimental methodology, simplify the presentation, and address my questions. The revised manuscript provides a clearer account of the empirical findings. However, several of my central scientific concerns remain.

(Q1) Practicality of the proposed solution. The authors now clarify that decoding-time bias subtraction is intended as an analytical intervention rather than a practical, general-purpose solution. This clarification addresses the ambiguity in the original presentation. However, the method still requires explicit preference-level decisions and model- and dataset-specific calibration, and does not apply directly to ordinary one-step generation. I encourage the authors to ensure that the manuscript consistently reflects this narrower scope.

(Q2) Mechanistic explanation of the observed bias. This remains my primary concern. The authors conjecture that preference-based post-training produces the generation-induced Apply bias. While plausible, this explanation is not experimentally investigated. The additional controls in §§4.2–4.4 help establish that the generation objective induces a systematic shift in decision scores, rather than an artifact of multi-step prompting, earlier decisions, or particular instruction wording. However, identifying the experimental condition that induces a bias is not equivalent to explaining the underlying mechanism that produces it.

Similarly, the intervention in §5.1 demonstrates that adjusting the decision threshold can reduce preference leakage, but it does not explain why the model develops the original bias. The paper therefore provides stronger evidence for where and under what conditions the failure occurs, without establishing the mechanistic why that motivated my original question. This distinction is central to my assessment of the paper's scientific contribution.

(Q3) Provenance of evaluation metrics. The authors' clarification distinguishing raw logits, hidden-state probes, generated decision labels, and GPT-5.4 evaluations substantially addresses my question. I also appreciate the explanation that gold applicability labels originate from benchmark annotations rather than GPT-5.4. Nevertheless, conclusions about response-level leakage and the decision-versus-generation decomposition still depend partly on LLM-based evaluation, whose reliability is not independently validated in the revised manuscript.

Additional concern regarding the conclusions. Some of the conclusions in §5 appear stronger than the reported results support. For example, Table 3 shows that Des2Gen increases preference leakage for Ministral-3-14B on RPEval-Im (5.16 to 6.22) and Ministral-3-8B on RPEval-Im (5.84 to 5.98), despite the broader claim of improved generation performance. These exceptions should be explicitly acknowledged and investigated.

Overall assessment. The revisions improve the paper's clarity and strengthen its empirical diagnosis of generation-induced Apply bias. I recognize the value of systematically separating knowledge, decision, and generation failures. However, the fundamental distinction between observing a numerical bias, demonstrating its influence on downstream behavior, and identifying its underlying mechanism remains unresolved.

I encourage the authors either to provide direct evidence for the proposed mechanistic explanation or to narrow their scientific claims to the experimentally established phenomenon.

In recognition of the authors' substantial revision efforts, improved presentation, and stronger empirical evidence, I will revise my score from 1 to 2. This increase acknowledges the meaningful progress made during the rebuttal, although my primary concern regarding the depth of scientific explanation and the strength of the resulting conclusions remains unresolved.

Comments on the paper

2 comments

We thank all reviewers for their constructive feedback. All changes made during the rebuttal are marked in blue in the revised manuscript. The main revisions are:

  • Writing and presentation : We rewrote the paper for readability, stating the problem setting up front and keeping only the terms and equations the argument needs, each defined at first use.
  • Knowledge vs. decision : To clarify the distinction between the two stages, we rewrote their definitions (§2.2) and the probing section (§3.2, App. D) to make clear that the probe tests whether applicability is still encoded when the model is asked to respond.
  • Scope of claims : We toned down overstated claims and added confidence intervals to Table 2.
  • Possible source of the bias : §6 now discusses a possible source: preference-based post-training that rewards responses endorsing the user.
AuthorsOpenReview ↗

Gentle reminder

Thank you for your thorough review of our work. As the discussion period is coming to a close, we would be grateful if you could take a moment to assess whether our responses have sufficiently addressed your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.