acceptodds
Under review as a conference paper at ICLR 2027

The Condescension Trap in Language Models

11 views

Abstract

We introduce the condescension trap, a failure mode in which a language model treats some users as less capable than their request indicates, while appearing considerate. It parallels elderspeak, the simplified and reassuring speech that people direct at older adults and that recipients experience as patronizing. We show that preference optimization amplifies this behavior because the evaluator rewards it and the users who bear its cost are absent from the feedback loop. We test each link of this loop with age as the user attribute, on everyday requests whose answer does not depend on age. First, when the user states an age of 79 rather than 45, six language models change how they address the user: all six refer to the user's age more often, by 8 to 27 percentage points, with smaller increases in referrals to family and in reassurance. Second, LLM judges reward the change: with the answer pair held fixed and only the age told to the judge varied, three judges prefer the answer written for the 79-year-old by 18 to 28 points more when told the user is 79. Third, training on such a judge amplifies the change: three rounds of online DPO raise references to the user's age from 30% to 67% of answers to 79-year-olds, while hiding the age from the judge keeps them near 30%. Finally, adults 65 and over, comparing the same pairs blindly to condition, do not prefer the rewarded answer and rate it as slightly more likely to treat the reader as less capable. Withholding a task-irrelevant attribute from the judge prevents the amplification; grounding the judge in a content rubric reduces it without removing it.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM-as-a-judge, preference optimization, RLAIF, DPO, reward misspecification, bias amplification, ageism, elderspeak, alignment, fairness

Reviews

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

The paper argues that when a model is trained using an AI judge, it can learn to talk down to certain users, because the judge rewards this and the users themselves never get a say. The authors test this using age. They take everyday requests where age doesn't change the right answer, and add either "I'm 45" or "I'm 79" to the request.

The study has four parts. First, six language models give older users roughly the same content, but more often mention their age, suggest asking family for help, or add reassurance. Second, AI judges comparing the same two answers are more likely to pick the one written for the older user when they are told the user is 79. Third, training a model (Qwen3-8B) for three rounds with such a judge raises the share of answers that mention a 79-year-old's age from 30% to 68.5%, while hiding the age from the judge prevents most of this increase. Finally, adults aged 65 and over, comparing the same pairs of answers without knowing which was which, show no preference for the answers written for the older user, and rate them as only slightly more likely to treat the reader as less capable.

Critical strengths
  1. A clean test of the judge. Each pair of answers is judged twice, once with the judge told the user is 45 and once told 79, with the answers unchanged and in both orders. This shows clearly that the judge's choice depends on the stated age. It matters to my assessment because it is the part of the paper that holds up well, and it is a reusable way to check judges for other user attributes.

  2. A control that identifies the cause of the amplification. Training with the same setup but with the age hidden from the judge barely changes how often the model mentions age, while training with the age visible more than doubles it. This shows that the increase comes from the judge seeing the age, and it gives a simple, practical fix. It is the most concrete result in the paper, though as I note below it is also expected.

Critical weaknesses
  1. The findings are largely expected. Language models and AI judges learn the associations people hold about age from their training data, so it is not surprising that they respond to a stated age, and biases of this kind have been documented extensively. That training amplifies whatever an AI judge prefers is also a known pattern, and it would happen in the same way for a preference that is good for users. This matters most to my assessment because it leaves the paper's contribution resting on its claim of harm (point 2). The narrow scope adds to this: one attribute in one direction, markers taken from research on how people talk to older adults (so other biases, for example toward younger users, would not show up), one small trained model, and one judge chosen because it showed the shift.

  2. The harm to older users is assumed rather than shown. The paper calls this a trap because the judge rewards something older users do not want, but its recipient study does not support this. Asked which answer they would rather receive, older raters chose the one written for the older user 49% of the time. The primary result (+4.8 pp on "treats the reader as less capable") is below the authors' own 10 pp threshold, and it and the secondary result (+4.4 pp) barely exclude zero. Under the prompt used for training, the judge's preference does not differ significantly from the raters'. Only older adults were recruited, and in most of the pairs they compared (106 of 144) neither answer mentions age. This is consequential because it is the one finding that would make the paper new, and the title and abstract present it more strongly than the results allow.

Questions for authors
  1. In most of the pairs shown to recipients, neither answer mentions age. Could you report the recipient results on the 30 pairs where only the older user's answer mentions age? With only about 50 ratings on these pairs this would be exploratory, but it would show whether the indifference holds where the behavior is actually present. A study designed around this behavior (question 4) would be needed to settle it.
  2. On pairs where neither answer mentions age, which differences between the answers explain the judge's shift (reassurance, family referrals, length, reading level, content)? This would show whether the judge rewards a condescending tone or something harmless.
  3. The reproducibility statement says everything is available at the linked repository, and the paper refers to the code release for the full register patterns. When I checked, the repository only said "Sorry, we are still preparing the code. Please check back soon!". Could you make the code, rubrics, register patterns and recipient data available during the review period?
  4. The paper does not report what recipient study participants were paid, or whether the study received ethics approval or an exemption. Could you state the compensation (and the equivalent hourly rate) and the approval status?
1 comment
AuthorsOpenReview ↗

Response to Reviewer vbcJ

We thank Reviewer vbcJ for the comments and review, and for the questions that led to the correction in our general response #1.


W1 (Expected findings and narrow scope).

  • (Expected findings). We agree that each ingredient has precedent. Models react to a stated age, and training amplifies what a judge prefers, as it would for a preference that is good for users. Our point is that nothing in such a pipeline checks which kind of preference it amplifies, because the users who receive the answers never enter the training signal. Any pipeline that trains on an LLM judge can therefore drift toward an address its users do not prefer without noticing, and we think this needs to be shown rigorously and addressed. We test every link on the same pairs, with training arms that differ only in what the judge sees, and masking the attribute gives a mitigation that any pipeline can apply. The test also produced unexpected results: the largest part of the judges' shift on pairs with age wording is a consistency check (general response #2), and amplification depends on the policy and, for education, on the judge (#3, #4).
  • (Narrow scope). We extended each point the reviewer lists. A disclosed wheelchair moves all three links, education depends on the judge, and gender shows nothing (general response #4). Across seven added ages, references are no higher at 25 or 35 than at 45 (Figure 2b). Three further policies give a third to two fifths of the Qwen3-8B contrast for Llama-3.1-8B and close to zero for Mistral-7B and Yi-1.5-9B (general response #3). The judge was chosen by design, because the experiment asks what training does with a judge that shows the shift (Appendix E), and a second judge, Qwen3-32B, reproduces the contrast in every seed.

W2 ("The harm to older users is assumed rather than shown").

We agree with the facts and with the criticism of the wording. The primary estimate is below our threshold, older adults do not prefer the favored answer (49%), and under the training prompt the judge does not differ from them, so the study shows a level of preference rather than a harm. Please refer to our general responses #5 and #6. We defined the term by the loop of Section 3.1, stated there that perceived condescension is not established, and reworded the abstract and the Introduction to match.


Questions for authors.

Q1 (The pairs in which only the 79-answer mentions age). Please refer to our general response #5 and Table 11 of the revised manuscript. On the 39 such pairs (65 ratings), the estimate is +20.8 points, with no preference for receiving that answer. As the reviewer anticipated, this is exploratory.

Q2 (The judge shift where neither answer mentions age). Part of it came from 12 pairs that our rule misclassified (general response #1). On the remaining 94 pairs, the two open-weight judges are close to zero over five runs, and GPT-5.5 and Claude Sonnet 5 keep a small shift. In an exploratory regression, length, reading grade, list structure, reassurance words, referral and rubric coverage explain almost none of it.

Q3 (Repository). The anonymous repository now contains the analysis code with the full register patterns, the rubrics, the generations and the 144 natural pairs with all judge outputs, and its README links the pseudonymized recipient ratings and the trained adapters. We apologize for the late upload.

Q4 (Compensation and ethics review). Participants were paid £7.87 for a study estimated at 20 minutes (£24.85 per hour at the median completion time of 19 minutes), including the three participants excluded for failing the final attention check. The study was an online survey of adult volunteers on Prolific, with a consent page and no names or contact details collected. We added these points to a new Ethics Statement, with the details in Appendix G.


Please let us know if you have any other comments or questions!

Sincerely, Authors

Reviewer_3u4YOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper aims to study a behavior similar to elderspeak in people, where models produce more condescending responses conditioned on the age of the user. Specifically, they study cases where evaluators reinforce such behaviors through reported preferences, which they deem the condescension trap. Experiments focus on prompts that explicitly reveal a hypothetical user’s age (either 79 or 45) both in a dialogue setting as well as a LLM-as-a-judge setting, showing that generally, models make more explicit references to age when told the user is 79 and judges prefer such responses when similarly told, even for prompts that should not depend on user age. Additional experiments show that training with such judges lead to increases in the explicit references to age, and that adults’ preferences do not align with that of the judges.

Critical strengths
  1. Social Significance. The paper studies a socially significant and interesting problem, broadly examining how LLMs’ behavior is conditioned on cues of a user’s age grounded in relevant literature on elderspeak.
  2. Human Evaluation. The study includes a human evaluation with adults’ (65+) human judgments to highlight that the judges change in preference is not faithful to human preferences
Critical weaknesses
  1. Age Cue Representativeness and Artifacts. The experiments center around LLM/judge behavior conditioned on users’ age, but age is only indicated through explicit disclosure which may lack ecological validity or representativeness of the harms the experiments attempt to capture. In particular, it appears unnatural or uncommon for a user to explicitly disclose their age when the prompt is intentionally not dependent on age, and in fact, disclosing age in such contexts may prime the audience to think that age is a factor. Additionally, age is strictly provided through these explicit disclosures, while the paper cites other work highlighting that models also act on implicit identity cues as well (lines 123-124), which may be more ecologically valid. Finally, two ages are tested in the main experiments: 45 and 79. While two additional ages appear to be tested in experiments included in the Appendix (25 and 85), the choice of particular ages tested is unmotivated and it is unclear whether the results may in part represent artifacts associated with these particular ages. Examining broad age ranges would ensure that results are robust to particular ages provided, and examining how the frequency of these behaviors correlates with numeric age may provide more insight.

  2. Potential Judge/Result Confound. One of the primary results is that when explicitly told the user is 79, judge models’ preference toward the 79-year-old-conditioned response increases. The main metric summarizing the condescension in models’ responses, however, is explicit references to age (e.g., “As a 79-year-old…”). Therefore, it is unclear whether the judge results are driven by these explicit mentions given that it would be expected that a judge told the user is 45 would avoid the response with incorrect information. A similar effect may impact the human evaluation.

Questions for authors
  1. How is it verified whether answers can be checked against official sources? Is this done manually?
  2. Who are the annotators that judge whether an adult user’s age affects the answer? In this case, the age of the annotators also appears to be important.
  3. Why do you choose the different margins and thresholds used throughout the paper? E.g., why is 10pp the “smallest effect of interest” (Lines 288-289)?
  4. Why is it considered “demonstrated competence” (lines 307-314)? The examples given appear to be explicit statements rather than competence that is demonstrated or shown. Furthermore, where did the sentences stating competence come from?
  5. Can the authors please elaborate on “The effect on the recipients if modest, and its direction is the one that closes the loop”? What is meant by “its direction”?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 3u4Y

We thank Reviewer 3u4Y for the careful reading and the concrete questions. The reviewer's second weakness led to the experiment that changed the paper most.


W1 (Age cue representativeness and artifacts).

  • (Explicit disclosure). We agree that stating an age on a request that does not depend on age is unusual and may itself suggest that age matters. Nevertheless, the paper's contrast compares two stated ages and so holds the disclosure constant. With no age stated, the generators almost never refer to age.
  • (Implicit cues). A profile line that gives only a birth year, with no age stated anywhere, also raises age references (Table 8 of the revised manuscript).
  • (Only 45 and 79). We added seven ages between 25 and 90 (Figure 2b of the revised manuscript). No model refers to age more often at 25 or 35 than at 45, and all rise from about 60 to 90, so the result is not an artifact of the two ages, and the behavior follows numeric age with an onset near 60. Rubric coverage stays within our margin at the added ages, except for Claude Sonnet 5 at 90.

W2 (Judge result driven by explicit age references).

Please refer to our general response #2. We agree with the reviewer on the largest single part, and we changed the paper's wording accordingly.

  • (A consistency check). On the 39 pairs in which only the 79-answer refers to age, the training judge told 45 avoids the answer that says "79", but not once the age wording is removed. Over eleven ages told, its choice peaks at 79, the age written in the answers (Appendix E).
  • (What remains). Told 79, the fitting age wording raises the choice, and a smaller shift survives without any age wording. We removed the claim that judges reward the register from Section 4.2.
  • (Human evaluation). We agree that the same applies. The recipients' signal sits on the pairs with an age reference, not on the pairs without one, and we treat this as exploratory (general response #5).

Questions for authors.

Q1 (Checking against official sources). Official pages were located with LLM search agents and approved by the investigator. Each rubric element had to quote a supporting span from those pages, and a script checked every span. Answers are scored against the rubric by Qwen3-32B, validated against the investigator's blind scoring of 200 responses. Appendix B gives the details.

Q2 (Annotators). Three hired annotators in their twenties labeled every request independently, and a request entered the main analysis only if all three judged that an adult user's age does not change the answer. We agree that their age can matter, and the unanimity rule is our protection against requests whose answer depends on age (Appendix A).

Q3 (Margins and thresholds). The content margin of 4.9 points is the coverage lost by omitting one optional rubric element. The other thresholds, 5 points for register markers, 15 for the judge shift and 10 for recipient ratings, were fixed before the confirmatory data were collected and express our judgment of the smallest difference worth acting on. By our own threshold, the primary recipient hypothesis is not supported.

Q4 ("Demonstrated competence"). We agree with this point, and we now call it "stated competence". We wrote one sentence per domain, for example "I'm comfortable with official forms and agency terminology.", fixed before any response was generated, and we list all five in Appendix A.

Q5 ("Its direction"). "Its direction" refers to the effect on the recipients, which runs against the judge's preference rather than with it. The sentence was unclear, and we replaced it with the two estimates of the gap between judge and recipients (Section 4.4).


Please let us know if you have any other comments or questions!

Sincerely, Authors

Thank you for your response and answers to questions.

The addition of more ages satisfies my concern about the experiments focusing on 45 and 79. I would still suggest further motivating the set of ages that were evaluated to clarify this choice in the paper.

My concerns about the representativeness and ecological validity of the age cues however, still remain. I appreciate adding mentions of birth year, although it is not clear to me that birth year is more natural than stating your age, whereas something like discussion of topics that roughly signal age (e.g., college courses/dorms vs. retirement) would seem to better capture harms in natural settings. I also have some concern about the prompt moficiations evaluated given that the number of different domains in one experiment is fairly small and in another case, only a single phrase requesting details appears to be evaluated. There is a question of whether these are reliable modifications to the prompt or artifacts given that they were manually written and the possible artifacts of the phrasing are not evaluated..

I have revised my review accordingly, raising my soundness to 2, but my main concerns prevent me from raising my overall score.

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

This paper defines a phenomenon called "condescension trap", where an AI judge rewards answers that treat the user as less capable. It appears that preference training amplifies this behaviour and the users who bear the cost, never enter the training signal. The authors test this using age as an user attribute on common requests. The six models compared in this work mention the user's age more often when told the user is 79 yrs old, instead of 45, while keeping the content relatively unchanged. The authors have presented extensive experiments to test and document this phenomenon for different LLM judge models. In a study, adults aged 65 and over, found the LLM responses to be slightly more condescending, without any increased preference for the rewarded answer, thus raising a critical mismatch between the training intention and actual effect.

Critical strengths
  1. The end-to-end design is apt. Using the same 144 pairs for the judge and recipient studies allow putting the judge's and the recipient's preferences on the same scale.
  2. The judge's assumption about user is well isolated by holding naturally generated pairs fixed.
  3. The manuscript generally benefits from clear presentation and writing.
Critical weaknesses
  1. It appears that the primary recipient hypothesis fails (+4.8pp, below the 10pp threshold). Also, under the training prompt, the judge's preference exceeds that of the recipients by only +4.9pp [-1.2, 10.9] and this interval includes zero.
  2. Fig 4, and Section 4.2 appears to be a critical experiment, but the details around it is underexplained.
  3. The paper leaves out some closely related works, such as [1] which demonstrates what appears to be a similar mechanism for the reward-gap amplification, along with [2], [3], [4].

[1] Shapira, Itai, Gerdus Benade, and Ariel D. Procaccia. "How RLHF amplifies sycophancy." arXiv preprint arXiv:2602.01002 (2026). [2] Dewan, Sherwin, Ismail Shaikh, Connie Shaw, Abhilash Sahoo, Akshita Jha, and Alisha Pradhan. "Examining Age-Bias and Stereotypes of Aging in LLMs." In Proceedings of the 27th International ACM SIGACCESS Conference on Computers and Accessibility, pp. 1-9. 2025. [3] Carla Perez Almendros, Luis Espinosa Anke, and Steven Schockaert. 2020. Don’t Patronize Me! An Annotated Dataset with Patronizing and Condescending Language towards Vulnerable Communities. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5891–5902, Barcelona, Spain (Online). International Committee on Computational Linguistics. [4] Xia, Xintao, Zhiqiu Xia, Linjun Zhang, and Zhanrui Cai. "A statistical framework for alignment with biased ai feedback." arXiv preprint arXiv:2602.08259 (2026).

Questions for authors
  1. Can the authors demonstrate clearly how this work delineates itself from the existing works mentioned in weaknesses?
  2. It would be good to have some protocol clarity in experiments, particularly 4.2.
1 comment
AuthorsOpenReview ↗

Response to Reviewer AJjt

We thank Reviewer AJjt for the comments and review, and for the pointers to related work.


W1 (The primary recipient hypothesis; judge against recipients).

We agree with this point, and the paper reports both facts. The primary estimate is below our threshold, and under the training prompt the interval of the difference between judge and recipients includes zero. What the data support is that older adults do not prefer the answer the judge favors. Please refer to our general responses #5 and #6 for the pairs in which only the 79-answer mentions age and for how we narrowed the claims.

W2 (Section 4.2 and Figure 4).

  • (Protocol). Section 3.5 states the protocol, and we added the full judge prompt to Appendix E. In short, each of the 144 pairs holds two unedited answers of one generator to the same request, written for the 45- and the 79-year-old. The judge is told one age, sees both answers in both orders and picks one, and the shift is the share choosing the 79-answer when told 79, minus the same share when told 45.
  • (What changed). The rule behind the submission's Figure 4a missed two forms of age statements (general response #1), and the new experiment shows that the largest part of the shift on pairs with age wording is a consistency check (general response #2). We plotted these results in a new Figure 3 and moved the submission's Figure 4 to Appendix E as Figure 7, with the corrected groups.

W3 (Related work).

We agree with this point, and we now cite all four works in Section 2.

  • [1] Shapira et al. (ICML 2026) show when preference training amplifies a bias such as sycophancy, and we cite this general result for our Equation 1. Their bias is agreement with what the user believes. Ours is a preference that depends on who the user is while the request stays fixed, tested with a training control that masks the attribute and with ratings of the same pairs by the affected group.
  • [2] Dewan et al. (ASSETS 2025) study what models say about aging. We study how a model addresses a user who states an age on a request where age is irrelevant, and what evaluation and training do with it.
  • [3] Pérez-Almendros et al. (COLING 2020) annotate patronizing language about groups in news text. Ours is second-person address conditioned on a stated attribute.
  • [4] Xia et al. (2026) correct AI-judge labels with a small set of human labels. For this bias, those labels would have to come from the affected group, which our recipient study samples.

Questions for authors.

We answered the questions in our responses to the weaknesses above: Q1 – W3; Q2 – W2.


Please let us know if you have any other comments or questions!

Sincerely, Authors

Reviewer_6TLhOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper discusses a failure mode of LLM which may treat elder users differently from the younger users. The contribution of this paper is threefold: First, the LLMs change how they address the user when the user age stated changes to 70 from 45; second, the LLM judges prefer the answer written for the 79-year-old much more when they were told that the user is 79; third, training on such a judge ampliffes the change.

Critical strengths
  1. This work discusses an interesting topic on the condescension trap of LLMs, this work enriches existing LLM bias researches.

  2. This work hypothesizes a feedback-loop mechanism where preference optimization amplifies patronizing behaviour, rather than only documenting static bias outputs.

  3. The experimental design isolates user age as the manipulated variable while holding the other variables, and the experimental results look sound.

Critical weaknesses
  1. The current evaluation only uses age as a single user attribute, and it is unclear whether the condescension trap generalizes to other vulnerable attributes such as disability or low income.

  2. It does not disentangle whether the observed shift stems from the judge’s preference reward alone, or from pre-existing pretraining biases that predispose models toward elderspeak-style language.

  3. Only a small set of six LLMs are tested, the selection of LLms sholuld be clearly given. Moreover, model family, model size, and alignment recipe differences are not systematically controlled, limiting the robustness of the claim.

  4. The work does not propose or evaluate mitigation strategies; it only diagnoses the phenomenon, the possible methods which can reduce the condescension trap are not discussed.

Questions for authors
  1. The current evaluation only uses age as a single user attribute, and it is unclear whether the condescension trap generalizes to other vulnerable attributes such as disability or low income. Should this be discussed?

  2. A small set of six LLMs are tested, the reason of choosing this set of LLMs should be discussed. That is, whether this set of LLMs are representative?

  3. Would the condescension trap disappear or become stronger when the task genuinely requires age-related context or domain adaptation?

  4. What specific design guidance would you give to alignment pipelines to avoid inadvertently rewarding this seemingly polite but patronizing behaviour?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 6TLh

We thank Reviewer 6TLh for the comments and review, and for recognizing the strengths of the submission.


W1 (Age as the only user attribute).

Please refer to our general response #4. We agree that generality beyond age was a hypothesis in the submission. We now test it with exploratory pilots and two training runs (Section 4.5).

W2 (Judge reward against pretraining biases).

  • (Design). The arms share the policy, prompts, starting candidates and settings, and differ only in what the judge sees, so the difference between arms cannot come from the policy's prior. The prior is the first link: before training, Qwen3-8B already refers to age far more often for 79-year-olds than for 45-year-olds.
  • (Dependence on the policy). As the reviewer suggested, how far training moves depends on the starting policy (general response #3). With a second judge, the Llama-3.1-8B contrast stays small, so this dependence lies in the policy rather than in the judge.

W3 (Choice of the six LLMs).

The list was fixed before any data were collected. It covers the three major closed providers and open-weight models from two families, one at two sizes, so that a shift found in all of them would not belong to one provider or one size. To add breadth beyond this set, we screened five further open-weight models from five families on the 30 held-out requests. Four of them show the gap between the 79 and the 45 cue at lower levels than Qwen3-8B, and Mistral-7B almost never refers to age (Figure 9a of the revised manuscript).

W4 (Mitigation and design guidance).

  • (Evaluated mitigations). Masking the age from the judge and grounding the judge in a content rubric are arms of the training experiment (Section 4.3, Table 2). Masking leaves age references at 37.8% of answers to 79-year-olds, against 68.5% when the judge sees the age, and a rubric halves the rise. Over nineteen runs, masking prevents most of the rise on average, but not in every run.
  • (Design guidance). (1) Audit the judge on fixed pairs before training, with and without the attribute and its wording, to separate a consistency check from a preference. (2) Check that the judge repeats its own verdicts, and report training over several runs. (3) Withhold attributes that the task does not need, keeping in mind that a masked judge also votes down references it cannot check. (4) Calibrate the judge with ratings from the affected group, since rubric grounding reduced the rise without removing it. We added these points to Section 5.

Questions for authors.

We answered Q1, Q2 and Q4 in our responses to the weaknesses above: Q1 – W1; Q2 – W3; Q4 – W4.

Q3 (Tasks that require age-related context). On the six held-out requests whose answer depends on age, the untrained policy already refers to age in almost all answers at the 79 cue, because age rules belong to the correct answer. Training changes this rate little under either judge, while referrals to family or friends rise more under the visible judge. Appendix F (Table 10) gives the details.


Please let us know if you have any other comments or questions!

Sincerely, Authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response to all reviewers (1/2)

We sincerely thank all reviewers for the careful review and constructive suggestions. The following are our responses to some central concerns.


1. Corrections (3u4Y, vbcJ, AJjt).

Questions from Reviewers 3u4Y and vbcJ led us to an error, thanks for pointing out. Our automatic rule for age references missed "at 79," and "you’re 79" with a typographic apostrophe, so 12 of the 106 pairs in which neither answer was thought to refer to age do contain an age statement. On the remaining 94 pairs, the judges' shift changed from '11 to 18' to '5 to 12' points. We also re-estimated the starting rate from fifteen answers per request (37.1%, not 30.0%), which the contrast between arms does not use. Under the age rule fixed in advance, no confirmatory test changes.

2. Judges and explicit age references (3u4Y, vbcJ, AJjt).

We judged all 144 pairs again with every age-revealing expression removed, added a fourth judge (Qwen3-32B), and ran both open-weight judges five times. We report the results in Figure 3 and Table 13 of the revised manuscript. The four judges shift by 16 to 25 points on the original answers and by 2 to 12 points without the age wording. On pairs with age wording, a judge told 45 turns away from an answer that says the user is 79, which accounts for 46% to 70% of the shift. We agree with Reviewer 3u4Y that this largest part is a consistency check, not a view about older users, and we rewrote Section 4.2 accordingly.

3. Training (vbcJ, 6TLh).

Table 2 and Figures 4 and 9 of the revised manuscript report the training results. Over nineteen runs of the paper's configuration, the visible judge exceeds the masked one in every run, by 23.3 points on average, below the 30.7 of the paper's three seeds. A second judge, Qwen3-32B, reproduces the contrast in every seed. A judge that knows only that an age was stated ends close to the masked one, so most of the rise needs the judge to know the age. The contrast depends on the policy: a third to two fifths of the Qwen3-8B value for Llama-3.1-8B, and close to zero for Mistral-7B and Yi-1.5-9B.

4. Other user attributes (6TLh, vbcJ).

We ran exploratory pilots that add a sentence disclosing one of three further attributes to each of the 94 requests, and report them in Section 4.5, Figure 5 and Appendix H. A disclosed wheelchair moves the generators, the training judge and training, and training amplifies it most on the requests where it should not change the answer. Education depends on the judge, and gender moves neither the generators nor the judge. Therefore, we could say, the amplification is not specific to age though it does not appear for every attribute.

5. Recipients and the term "condescension trap" (vbcJ, AJjt, 3u4Y).

We agree that the recipient link is the weakest. The primary estimate is +4.8 points, below our threshold of 10, and older adults do not prefer the answer the judge favors (49% of ratings). We added Table 11 for the pairs Reviewer vbcJ asked about. On the 39 pairs in which only the 79-answer refers to age, raters find it more likely to treat the reader as less capable by 20.8 points, and on the 94 pairs in which neither does, they do not. This analysis is exploratory. We do not use "condescension trap" to claim that recipients feel talked down to, but to name the loop of Section 3.1. In this loop, the evaluator's preference moves with the stated age, a policy trained on it moves the same way, and the recipients do not prefer the result.

AuthorsOpenReview ↗

General response to all reviewers (2/2)

6. How the claims of the paper changed (all reviewers).

The revised manuscript is uploaded. We revised each claim to state only what the results support, and the table compares the submitted abstract with the revised one.

SubmissionRevision
Thesis"We show that preference optimization amplifies this behavior because the evaluator rewards it""We test it as a feedback loop with three parts"
Judges"LLM judges reward the change"The largest part of the judges' shift is a check of the answer's age wording against the age told, and a smaller shift remains without age wording
TrainingReferences to age rise from 30% to 68.5% of answersThe visible judge exceeds the masked one by 23.3 points over nineteen runs, and the size depends on the policy
Masking"prevents the amplification""prevents most of the amplification on average"
Recipients"slightly more likely to treat the reader as less capable"This rating stays below our 10-point threshold, and Section 3.1 states that perceived condescension is not established

The main changes by section are as follows.

  • Section 4.1 now covers the generators across eleven stated ages and a birth-year cue.
  • We rewrote Section 4.2 around the results of #2, shown in the new Figure 3, and moved the submission's Figure 4 to Appendix E as Figure 7, with the corrected groups.
  • Section 4.3 now reports the corrected baseline, the second judge, the hidden-age arm, the further policies and run-level estimates.
  • Section 4.4 adds the subset analysis of #5, marked as exploratory.
  • The new Section 4.5 reports the other attributes (#4), with details in Appendix H.
  • Section 2 now cites the four related works suggested by Reviewer AJjt.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.