acceptodds
Under review as a conference paper at ICLR 2027

Selective Preference Feedback Changes the Information Order

26 views

Abstract

Can deleting a tie category make preference learning impossible? Under selective response, yes. We study fixed-confidence identification from complete response/nonresponse laws. Binary one-protocol feedback can be nonidentifying; retaining a scored tie yields quartic information and a lower bound, while valid protocols yield quadratic information and a lower bound. After profiling an analytic nuisance manifold, the first normal likelihood jet not reproduced by nuisance or model uncertainty determines the profiled KL order. In a degree- scored model, we derive a sharp support threshold, coefficient, and finite-gap alias radius at scale . Under imperfect exclusion, arbitrary compact joint validity uncertainty yields a product-space coefficient; when the joint set is convex, maximin protocol design becomes a weighted Chebyshev-center problem, while rectangular budgets alone admit a closed-form pair design. A compact deployment-reference family gives positive, negative, and ambiguous decisions, with separate radii for losing a unique sign and attaining an opposite sign. An exchange-symmetric two-protocol product class exhibits a robust quadratic–quartic crossover; correlated response errors with equal marginals can differ. All sample-complexity statements are lower bounds; we do not claim an attaining same-sample algorithm.

Primary area:
learning theory
Keywords:
selective preference feedback, preference learning, fixed-confidence identification, nonregular statistical experiments, information geometry, missing-not-at-random feedback, robust experimental design, ordinal feedback, information-theoretic lower bounds

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_wW3iOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2(−1)
Presentation
1(−1)
Soundness
2
Read review
Paper summary

The paper studies sign identification when preference categories are selectively reported and nonresponse is observed. It relates nuisance-profiled KL information to the first likelihood derivative that survives nuisance cancellation. For an exponential-tilt score model with degree-rr response-logit nuisance, it derives an exact support threshold, an explicit nominal information coefficient, and finite-gap alias radii. It also studies protocol design under joint validity uncertainty, reference-dependent decisions, and a quadratic–quartic crossover for a symmetric two-protocol product class. The evidence consists of analytic proofs and reported numerical certificates; the sample-complexity results are lower bounds.

Critical strengths

S1. An informative exact calculation of nuisance-induced information loss. The interpolation threshold and surviving orthogonal-polynomial coefficient explain why adding a scored category can change the nominal experiment from exact aliasing to higher-order information. The conditioning results also distinguish meaningful additional resolution from a nearly vanishing category or nearly coincident score. These are concrete insights beyond a general claim that richer feedback helps.

S2. Careful treatment of joint validity uncertainty. The distinctions between arbitrary joint sets, convex sets, rectangular budgets, and the symmetric product crossover are consequential. In particular, the examples showing that the same marginal error budgets can yield different information, and that a correlated mechanism can require three protocols, prevent an unjustified generalization of the pair-design formula. The distinction between ambiguity and an opposite unique decision is also valuable.

Critical weaknesses

W1. Nonidentifiable correctness classes limit the fixed-confidence interpretation. Definition 21 requires correctness at every unique-decision model in the class. This can already fail in the one-protocol class K0\mathcal K_0 of equations (80)–(81), without response-error thickening. Take the allowed baseline a=0a=0 and r≥1r\ge1, and set qt(j)=−2tyj+ψ(t)−ψ(−t)q_t(j)=-2ty_j+\psi(t)-\psi(-t) for small t>0t>0. Both (t,qt)(t,q_t) and (−t,−qt)(-t,-q_t) belong to the class because their nuisance coefficients approach zero. The identity σ(x)=exσ(−x)\sigma(x)=e^x\sigma(-x) gives Pt,qt=P−t,−qtP_{t,q_t}=P_{-t,-q_t}, including nonresponse, although the two models require opposite decisions. Thus no uniformly δ\delta-correct procedure exists on this class for δ<1/2\delta<1/2, even above the scored-support threshold. This does not contradict the information calculation at the nominal model (Δ,0)(\Delta,0).

The same class-wide issue affects the two-protocol uncertainty classes. Fix any row of equation (267) with γ>0\gamma>0. A model with u=t>0u=t>0 and zero nuisance and deviations is exactly matched under both protocols by a model with u=−tu=-t, q=θ=0q=\theta=0, and

\bigl(2ty_j+\psi(-t)-\psi(t)\bigr).$$ Equations (272)–(273) give equality of the complete laws, including nonresponse. These deviations approach zero with $t$, so both models belong to the same class for sufficiently small $t$, despite requiring opposite decisions. No uniformly $\delta$-correct procedure therefore exists on that row for $\delta<1/2$, even when the selected nominal instance has positive information. Shrinking $\gamma$ across rows does not remove this problem within each positive-radius row. Corollary 80's conditional lower bound consequently has no class-correct procedure to which it applies in these cases. The paper explicitly claims lower bounds, and these observations leave its nominal KL calculations and conditional inequalities intact. The concern is the fixed-confidence interpretation of equations (10), (20), and (38), including Corollary 80, on the stated correctness classes. Positive information at the selected nominal instance does not make class-wide identification feasible on those classes. **Update.** I agree with reviewer JrQD and 7JEX about the presentation issues. I lower my rating from 3 to 2.
Questions for authors

Clarification concerning W1: Is “higher-order identification” intended only at the nominal response baseline? Which correctness class is intended for the fixed-confidence conclusions, and how does it exclude the opposite-decision aliases in W1, including the example without response-error thickening?

Reviewer_7JEXOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The paper studies preference learning and claims that deleting a tie-category can change the lower bound from Δ4\Delta^{4} to Δ2\Delta^{2}.

Critical strengths
  • I cannot understand the paper but the premise of how choices like tie-categories and admissible user input affect learning is, in my opinion, quiet nice. I'm sure this could be a nice paper (AI written or not). However, as is, I feel like it is not in a shape for me to understand!
  • Paper seems to acknowledge limitations and discuss them carefully.
Critical weaknesses
  • I think paper needs to be rewritten. Abstract in particular is, in my opinion, not readable. See comments below.
Questions for authors
  • What is "analytic nuisance manifold"?
  • Please say distribution instead of "laws" if that's what you mean.
  • Figure 1. What is I(Δ)I(\Delta)? You shouldn't just put a figure and assume reader knows the problem you are working on with all terminology.
  • Are you only comparing lower bounds for the two settings? It seems more instructive to compare upper bound for setting claimed to be good and lower bound for setting claimed to be bad. Comparing two lower bounds can be misleading unless they are both known to be tight somehow. People do sometimes use "improved upper bound" to conclude "probably easier setting" even without known lower bounds but even then I think readers would correctly interpret the nuance; with lower bounds I'm more skeptical as obtaining lower bounds can be very difficult because of limitation of techniques not because the problem is "easy".
Reviewer_JrQDOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper present two lower bounds for a problem of preference learning.

Critical strengths

The paper may have genuinely interesting results, but it is almost impossible to assess, given the presentation. See weakness.

Critical weaknesses

The paper is written in a way that it is almost impossible for a human to understand, follow and verify.

Some examples of the lack of care in the exposition:

  • Section 2, which follows the introduction and should therefore introduce the notation, the model and the setting, does not do that at all but it introduces instead concepts without making the effort to explain them or contextualizing them. The reader should always at least be able to understand the model and notation. It is not the case for this paper.

  • The figures, which could be a great tool to help a reader understand, are confusing. The text in the figures is sometimes hard to read, sometimes hidden below other parts of the figure, the captions do not really explain what is happening in the figures.

  • The abstract, which should be the most understandable part of a paper, is written in a very confusing manner. The authors prefer to use very technical terms and sentence construction, failing to give an high level idea of what the paper is about. The final sentence of the abstract is emblematic of this approach "All sample-complexity statements are lower bounds; we do not claim an attaining same-sample algorithm" . This sentence shows how the presentation is not well curated. The fact that the main contribution of the paper is a lower bound and not an algorithmic approach should be clear from the rest of the abstract, not something that you feel the need to remark in the last sentence of around 17 lines hyper-technical abstract.

  • The font for theorems is the same as the rest of the text, meaning that a reader has difficulties even understand where the theorems end.

  • The appendix is something that I have honestly some difficulties imagine an human carefully reading and submitting as it is. See table 3.

To conclude I am in no way comfortable with this paper being accepted as it is.

Questions for authors

_

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.