acceptodds
Under review as a conference paper at ICLR 2027

The Arbitrary-Placement Problem in Entropy-Minimizing Selection, and a Residual-Entropy Formulation

Abstract

Entropy-based selection objectives suffer from a fundamental degeneracy: minimizing Shannon entropy rewards confident selection regardless of whether the selected candidate is informative. We address this limitation with the residual entropy , where is induced directly by candidate trust weights before they are aggregated into scores. We prove the exact identity , where measures whether the score-induced distribution and the underlying trust profile favor the same candidates. The boundary cases establish basic safety: under uniform trust, automatically, so an equal-trust, non-starving state is never penalized, while at any one-hot limit, regardless of the selected candidate. Our main results address the intermediate regime where selection occurs. We prove that when the ordering of candidates by trust agrees pairwise with their ordering by informativeness, giving a checkable pairwise condition for safe informative selection. We further derive a tighter certificate based on the margin of the leading candidate over its competitors, which determines when that candidate can be selected while preserving . These results are independent of the candidate-scoring function and apply to both stationary and dynamically changing information. Experiments with a gradient-based mixture-of-experts router confirm that the ordering conditions can hold during real optimization and show that correct ordering improves downstream performance when candidates are non-interchangeable and selections are used directly rather than averaged. Beyond routing, margin-based reweighting matches or outperforms fixed-strength baselines in a class-imbalance task, while informative selection in a production video-prediction system reduces MSE by approximately 20% and transfers to a related species. Residual entropy, therefore, provides a safety criterion for selection and a usable signal for deciding when that selection is informative.

Primary area:
learning theory
Keywords:
Entropy minimization, Residual entropy, Candidate selection, Information theory, Mixture-of-experts

Reviews

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

The paper examines a limitation of entropy-minimizing selection, i.e., it prioritizes confident selection, but does not consider whether the selected candidate is informative or not. For example, a concentrated distribution can have low entropy regardless of which candidate is selected. This paper aims to address this issue by using residual entropy. The authors proves that the residual entropy D <= 0, when the pairwise ordering of candidates by trust agrees with their ordering by informativeness. The authors conducted empirical validations, real-pipeline validation and downstream applications to evaluate the proposed approach.

Critical strengths
  1. The research question is interesting and this paper points out a concrete limitation of entropy minimization for candidate selection. I am convinced that the informativeness of selected candidates are also important, which, however, is not fully considered in the existing entropy minimization approach. The proposed residual objective is simple, and Proposition 2 gives an exact decomposition. Corollaries 3-4 provide explicit sufficient conditions that can be checked at a given state.

  2. The experiments include useful negative results. Appendix D.5 reports no significant accuracy gain from selecting the initially advantaged expert when participation status is held fixed. Appendix E then examines the condition when there is expert redundancy. These results help clarify when and where the proposed method brings benefits or not.

Critical weaknesses
  1. The theoretical contribution is narrower than the paper’s claim, as it does not always guarantee informative selection. The decomposition (the 2nd equation in the abstract) by itself does not guarantee that minimizing D will always help select an informative candidate or improve downstream performance. Corollary 3 assumes pairwise ordering agreement between beta_t and v_t/beta_t. Corollary 4 assumes that the candidate maximizing beta_t also maximizes v_t/beta_t, and adds a margin condition. Thus, these corollaries show that D< 0 follows when the specified conditions hold; they do not establish that optimizing the residual-entropy objective produces that ordering.

  2. The class-imbalance experiment does not establish that the proposed mechanism adds value beyond a direct count-based baseline. Algorithm 2 computes each class’s sample count from the training labels and uses it to define the class-level scarcity signal. The benchmark, in turn, is constructed by substantially reducing the training examples of one class. In this setup, selecting the least frequent class is a natural baseline for identifying the target class, and the paper does not report how that baseline performs. It makes it unclear what the proposed selection mechanism adds beyond a signal that directly reflects the way the benchmark was constructed. The authors should report a count-based baseline and clarify whether the target remains identifiable from counts under their experimental protocol.

  3. The MoE experiments provide limited evidence that the proposed objective selects the advantaged expert, but do not convincingly connect that ability to the downstream gains. Appendix D.3 reports that the residual objective selects the advantaged expert in 2/25 runs, versus 0/25 for the other objectives. The authors themselves do not treat it as evidence of selection accuracy. Appendix E shows that choosing the better expert can improve performance, but does not test whether the proposed method can identify that expert. Thus, the paper does not convincingly connect the method’s selection ability to downstream gains.

Questions for authors
  1. How should the ordering-guarantee statements in Section 5.2 and Appendix E.1 be interpreted, given the ordering assumptions in Corollaries 3-4? What evidence supports the stronger claim that optimizing residual entropy produces a useful ordering?

  2. Is there any restriction in Algorithm 2 that prevents selecting the class with the smallest training count? If not, what does residual selection add to this direct rule?

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

The paper discusses the use of differentiable surrogate losses to optimize discrete problems related to choice among TT elements. Examples of such problems notably include Mixtures of Experts. It is argued that using entropy to encourage concentration among few elements cannot distinguish 'good' choices from 'bad' choices. An alternative to entropy is proposed, its theoretical properties are demonstrated, and the algorithm is empirically validated on a variety of use-cases.

Critical strengths
  • The proofs of the properties of the loss L2L_2 are correct and sound. They highlight the merit of the loss compared to entropy.
Critical weaknesses
  • The paper is difficult to follow for someone without prior knowledge on the problem being tackled. For example, the notion of 'the trust weights' is introduced rapidly (abstract and line 68 of the introduction). Are 'trust weights' part of the broader literature or simply a term that is specific to the formulation of algorithm 1 (i.e. the weights βt\beta_t)? Also, when the loss L2L_2 is introduced at line 138, its definition involves two parameters γ\gamma and μ\mu that are not defined. They are actually defined at line 255 of the empirical validation: they are the mean and spread parameters of a Gaussian that parametrizes βt\beta_t. It would help the reader if this definition of βt\beta_t was introduced earlier in the paper, along a citation if it comes from prior work, or a motivation if it is a new contribution. Line 273 of the experiments introduces a parameter KK that does not seem to be defined in the main manuscript. The experiment sections starts discussing settings sequentially (i)-(vi). Then section 6.1 applies an ablation study on the 'production pipeline's upstream uncertainty model', which is suspect is a reference to setting (vii). If it is the case, then why perform an ablation before introducing setting vii? If not, then it is not clear to which setting section 6.1 applies.
  • Algorithm 1 is very general: it can be applied to any problem of choice within KK elements as long was one can define the vtv_t and βt\beta_t functions. However, the definitions and explanations for vtv_t and βt\beta_t for each of the seven experimental settings are left to the appendix. I understand that the main manuscript has limited spaces, but as of right now, it is hard to make connections between the use cases and the theory introduced earlier. It could help to move all proofs to the appendix, and use the extra space for a table that provides some high-level details on the vtv_t and βt\beta_t used in all experiments.
Questions for authors

Corollary 4 claims that βt\beta_t and vt/βtv_t/\beta_t both optimized for the same element t=wt=w whenever the outer loop a settled on a single dominat candidate. I don't understand how this statement generally holds if one divides by βt\beta_t. Won't a very small (near zero) βt\beta_t inflate the term vt/βtv_t/\beta_t?

From my understanding, these ct,kc_{t,k} and ak,ta_{k,t} parameters are designed differently for each experiment setting. Do you have a general rule-of-the-thumb for settings these values in order to solve choice problems?

The usecase (iv) investigate the algorithm to identify the under-represented class in a digit dataset. However, the algorithm assumes access to training labels: the ck,tc_{k,t} parameter directly depends on the count of instances of each class. Why would we need optimization algorithm to identify a minority class when we directly have direct access to the labels?

Is it possible to prove that the loss DD converges to zero while remaining negative (no sign shifts) when a distribution smoothly transition from uniform to a one-hot? Could this be proven for specific cases, e.g., smoothly scaling down βt\beta_t expect for βt0\beta_{t_0}?

Reviewer_C3CXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes residual entropy as a replacement for Shannon entropy. The residual entropy subtracts a second entropy term computed from the trust weights before they are aggregated into candidate scores. The authors claim that the proposed residual-entropy objective removes Shannon entropy's inherent reward for concentration alone. The experimental results in MoE router demonstrate that the residual entropy improves downstream performance.

Critical strengths

This paper attempts to address the inherent shortcomings of Shannon entropy in practical applications and provides a preliminary theoretical justification for the effectiveness of residual entropy.

Critical weaknesses
  1. This paper proposes using candidate trust weights to get the trust profile βt. However, it does not state what βt actually means. In the empirical validation and experiments of MoE router, the specific meaning and calculation method of βt are also unclear.
  2. The formulation of residual entropy is equivalent to the form of KL divergence, to some extent. The authors do not clearly explain the difference between these two methods. Why KL divergence is not used directly as a substitute for residual entropy.
Questions for authors
  1. Besides the MoE router, what advantages does residual entropy have over traditional entropy in other tasks and experimental settings?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.